Product carbon footprint accounting method and device

Abstract

The invention provides a product carbon footprint accounting method and device. The method comprises the following steps: firstly, acquiring full life cycle information of a product; then extracting local features, global features and dynamic features of the full life cycle information according to a feature extraction model; and finally, predicting the carbon footprint of the product according to the local features, the global features, the dynamic features and a deep learning model. According to the method, the accuracy of carbon footprint accounting is improved by extracting multiple types of features, and meanwhile, the relationship between the carbon footprint and the product data is learned according to deep learning, so that the carbon footprint is predicted, and inaccurate carbon footprint accounting caused by incomplete data acquisition is avoided.

Description

Technical Field

[0001] The present invention belongs to the technical field of carbon footprint calculation, and in particular relates to a product carbon footprint calculation method and device. Background Art

[0002] In order to meet the needs of reducing carbon emissions in the power system, the key is to accurately process and analyze carbon footprint data, that is, to scientifically calculate the factors and their extent that affect the carbon footprint to obtain accurate and reliable data. At a time when global climate change is attracting much attention, product carbon footprint, as a key indicator to measure the greenhouse gas emissions of products throughout their life cycle from raw material acquisition to waste disposal, is of great significance for companies to respond to carbon reduction policies, achieve green and sustainable development, and meet consumers' environmental protection needs.

[0003] At present, the more commonly used product carbon footprint accounting technologies include life cycle assessment method and input-output analysis method. However, the existing technology has many shortcomings. For example, there are many difficulties in data acquisition and limitations in feature extraction. Therefore, the accuracy of product carbon footprint accounting in the existing technology is not high. Summary of the invention

[0004] In view of this, the present invention provides a product carbon footprint accounting method and device, aiming to solve the problem of low accuracy of product carbon footprint accounting in the prior art.

[0005] A first aspect of the present invention provides a product carbon footprint calculation method, comprising:

[0006] Obtain information about the entire product life cycle;

[0007] According to the feature extraction model, local features, global features and dynamic features of the whole life cycle information are extracted;

[0008] Predict the carbon footprint of products based on local features, global features, dynamic features and deep learning models.

[0009] In a possible implementation, the feature extraction model includes an input layer, a local feature extraction layer, a global feature extraction layer, a dynamic feature extraction layer, a feature fusion layer, and an output layer; according to the feature extraction model, local features, global features, and dynamic features of the full life cycle information are extracted, including:

[0010] The whole life cycle information is input into the local feature extraction layer through the input layer to obtain the initial local features. At the same time, the feature fusion layer fuses the initial local features and outputs the local features through the output layer.

[0011] The whole life cycle information is input into the global feature extraction layer through the input layer to obtain the initial global features. At the same time, the feature fusion layer fuses the initial global features and outputs the global features through the output layer.

[0012] The whole life cycle information is input into the dynamic feature extraction layer through the input layer to obtain the initial dynamic features. At the same time, the feature fusion layer fuses the initial dynamic features and outputs the dynamic features through the output layer.

[0013] In a possible implementation, the whole life cycle information is input into the local feature extraction layer through the input layer to obtain the initial local features. At the same time, the feature fusion layer fuses the initial local features and outputs the local features through the output layer, including:

[0014] The whole life cycle information is input into the local feature extraction layer through the input layer to obtain the initial time local features and initial space local features;

[0015] The initial temporal local features and the initial spatial local features are fused in the feature fusion layer, and the local features are output through the output layer.

[0016] In a possible implementation, the full life cycle information is input into the global feature extraction layer through the input layer to obtain the initial global features. At the same time, the feature fusion layer fuses the initial global features and outputs the global features through the output layer, including:

[0017] The whole life cycle information is input into the global feature extraction layer through the input layer to obtain the initial time global feature and the initial space global feature;

[0018] The initial temporal global features and the initial spatial global features are fused in the feature fusion layer, and the global features are output through the output layer.

[0019] In a possible implementation, the feature extraction model further includes an auxiliary extraction layer; and the method further includes:

[0020] Inputting the initial local feature into the auxiliary extraction layer to obtain the first auxiliary feature;

[0021] The initial temporal global features and the initial spatial global features are fused in the feature fusion layer, and the global features are output through the output layer, including:

[0022] The first auxiliary feature, the initial temporal global feature and the initial spatial global feature are input into the feature fusion layer for fusion, and the global feature is output through the output layer.

[0023] In a possible implementation, the full life cycle information is input into the dynamic feature extraction layer through the input layer to obtain the initial dynamic features. At the same time, the feature fusion layer fuses the initial dynamic features and outputs the dynamic features through the output layer, including:

[0024] The whole life cycle information is input into the dynamic feature extraction layer through the input layer to obtain the initial time dynamic features and initial space dynamic features;

[0025] The initial temporal dynamic features and the initial spatial dynamic features are fused in the feature fusion layer, and the dynamic features are output through the output layer.

[0026] In a possible implementation, the feature extraction model further includes an auxiliary extraction layer; and the method further includes:

[0027] Inputting the initial local features and the initial global features into the auxiliary extraction layer to obtain the second auxiliary features;

[0028] The initial temporal dynamic features and the initial spatial dynamic features are fused in the feature fusion layer, and the dynamic features are output through the output layer, including:

[0029] The second auxiliary feature, the initial temporal dynamic feature and the initial spatial dynamic feature are fused in the feature fusion layer, and the dynamic feature is output through the output layer.

[0030] In one possible implementation, the carbon footprint of a product is predicted based on local features, global features, dynamic features, and a deep learning model, including:

[0031] Local features, global features, and dynamic features are input into the deep learning model to obtain the prediction results of the product's carbon footprint.

[0032] In one possible implementation, the full life cycle information of the product is obtained, including:

[0033] Obtain initial full life cycle information of the product;

[0034] The initial full life cycle information is processed for missing values, outliers and normalization to obtain the full life cycle information.

[0035] A second aspect of the present invention provides a product carbon footprint calculation device, comprising:

[0036] Acquisition module, used to obtain the full life cycle information of the product;

[0037] An extraction module is used to extract local features, global features and dynamic features of the whole life cycle information according to the feature extraction model;

[0038] The prediction module is used to predict the carbon footprint of products based on local features, global features, dynamic features and deep learning models.

[0039] The product carbon footprint calculation method and device provided by the embodiment of the present invention first obtains the full life cycle information of the product; then extracts the local features, global features and dynamic features of the full life cycle information according to the feature extraction model; finally, the carbon footprint of the product is predicted based on the local features, global features, dynamic features and deep learning model. The present invention improves the accuracy of carbon footprint calculation by extracting multiple types of features, and at the same time learns the relationship between carbon footprint and product data based on deep learning, thereby predicting the carbon footprint and avoiding inaccurate carbon footprint calculation caused by incomplete data acquisition. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0041] Figure 1 is a flow chart for implementing a product carbon footprint calculation method provided by an embodiment of the present invention;

[0042] Figure 2 It is a schematic diagram of the structure of a product carbon footprint accounting device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0044] Figure 1 FIG. 1 is a flowchart of the method for calculating the carbon footprint of a product provided by an embodiment of the present invention. Figure 1 As shown in the figure, the product carbon footprint accounting method includes:

[0045] S110, obtain the full life cycle information of the product;

[0046] S120, extracting local features, global features, and dynamic features of the full life cycle information according to the feature extraction model;

[0047] S130 predicts the carbon footprint of the product based on local features, global features, dynamic features and a deep learning model.

[0048] In the embodiment of the present invention, obtaining the full life cycle information of the product is to comprehensively and systematically collect all data and situations related to each stage of the product from raw material acquisition, manufacturing, transportation, use to final disposal. This includes but is not limited to information on the origin and mining methods of raw materials, the types and quantities of energy consumed in the production process, the means of transportation used in the transportation link and the transportation distance, the energy consumption and usage time of the product during use, and the disposal methods after disposal. Through such detailed collection, a solid foundation can be laid for the subsequent accurate calculation of the carbon footprint of the product, because the product may generate carbon emissions at every stage of its entire life cycle. Only by fully mastering this information can its carbon impact on the environment be fully evaluated.

[0049] In some embodiments, the feature extraction model includes an input layer, a local feature extraction layer, a global feature extraction layer, a dynamic feature extraction layer, a feature fusion layer, and an output layer; according to the feature extraction model, the local features, global features, and dynamic features of the full life cycle information are extracted, including: inputting the full life cycle information into the local feature extraction layer via the input layer to obtain initial local features, and after the feature fusion layer fuses the initial local features, the local features are output through the output layer; inputting the full life cycle information into the global feature extraction layer via the input layer to obtain initial global features, and after the feature fusion layer fuses the initial global features, the global features are output through the output layer; inputting the full life cycle information into the dynamic feature extraction layer via the input layer to obtain initial dynamic features, and after the feature fusion layer fuses the initial dynamic features, the dynamic features are output through the output layer.

[0050] In an embodiment of the present invention, the main function of the input layer is to receive the full life cycle information as the input data of the model. Its structure is relatively simple, mainly defining the input format and dimension of the data. For example, if the full life cycle information contains image data, the size of the input image (such as 224x224 pixels), the number of color channels (such as RGB is 3 channels), etc. may be specified; if it is text data, the maximum length of the text and the encoding method (such as word vector encoding, character encoding, etc.) may be defined. It usually does not contain complex operation structures, mainly to ensure that the subsequent layers can correctly receive and process the input data. The input layer is directly connected to the local feature extraction layer, the global feature extraction layer, and the dynamic feature extraction layer, and the received full life cycle information is passed to these three feature extraction layers at the same time, so that they can each perform feature extraction operations.

[0051] In an embodiment of the present invention, for processing data with spatial structures such as images, multiple convolution layers may be included. For example, the first convolution layer may use a smaller convolution kernel (such as 3x3), and the number of convolution kernels may be set to 32. Each convolution kernel slides on the image to perform a convolution operation to extract features of the local area of ​​the image. The size of the obtained feature map will change according to factors such as the size of the convolution kernel and the stride (for example, if the input image size is 224x224, the convolution kernel is 3x3, and the stride is 1, then the output feature map size is 222x222, assuming the filling method is VALID). There may be other convolution layers in the future, and the size and number of convolution kernels may gradually change. For example, the size of the convolution kernel of the second convolution layer becomes 5x5, and the number of convolution kernels increases to 64, so as to further extract local features.

[0052] Between convolutional layers, activation functions are usually used to introduce nonlinear factors. Common activation functions include ReLU (Rectified Linear Unit). The ReLU function is defined as: f(x) = max(0, x), which can keep the output of the neuron linear when it is greater than 0, and output 0 when it is less than 0, effectively avoiding the gradient vanishing problem and speeding up model training. In order to reduce the data dimension, pooling layers may be added after some convolutional layers. For example, the maximum pooling layer, whose window size can be set to 2x2 and stride is 2, will select the maximum value in each pooling window as the output, thereby achieving downsampling of the feature map, reducing the amount of data while retaining important features.

[0053] The global feature extraction layer may also perform preliminary feature extraction through a series of convolutional layers for data such as images. Its structure is similar to that of the convolutional layer of the local feature extraction layer, but it may differ in the size and number of convolution kernels. The purpose is to capture some macro features of the image as a whole. For example, the initial convolutional layer may use a larger convolution kernel (such as 7x7) and a relatively small number of convolution kernels (such as 16) to quickly obtain features of larger areas of the image. Activation functions such as ReLU are also used between convolutional layers to introduce nonlinearity. Different pooling strategies may be used to distinguish it from the local feature extraction layer. For example, an average pooling layer can be used with a window size of 3x3 and a stride of 3. By calculating the average value within the pooling window as the output, the image features can be aggregated as a whole to obtain some attribute features about the image as a whole, such as color distribution, overall texture, etc. For data such as text, some statistical-based methods may be combined with neural network structures. For example, the text is first converted into a vector representation through a word vector model, and then the global features of the text, such as the topic distribution of the text, are calculated through a fully connected layer. Activation functions (such as ReLU) are also used between fully connected layers to increase nonlinearity.

[0054] Dynamic Feature Extraction Layer When processing full life cycle information with time series characteristics (such as changes in sensor data over time, video frame sequences, etc.), recurrent neural networks (RNNs) and their variants may be used. For example, long short-term memory networks (LSTMs) or gated recurrent units (GRUs). Taking LSTM as an example, its internal structure includes input gates, forget gates, output gates, etc. The input gate determines how new information enters the cell state, the forget gate determines which information is to be forgotten from the cell state, and the output gate determines how the information in the cell state is output. Each gate has its corresponding weight and bias. Through the synergy of these gates, LSTM can effectively capture long-term dependencies in time series data. At each time step, the LSTM unit will calculate based on the input data (the value of the full life cycle information at this time step) and the cell state and output of the previous time step to obtain the output of the current time step and the new cell state.

[0055] Similarly, activation functions are used between LSTM units, such as the tanh function used to calculate the update of the cell state, and the sigmoid function used to calculate the output of each gate. The tanh function is defined as: f(x) = (e^xe^-x) / (e^x+e^-x), which can map the input value to the (-1,1) interval; the sigmoid function is defined as: f(x) = 1 / (1+e^-x), which can map the input value to the (0,1) interval. If other types of dynamic data are processed, other methods based on time series analysis may be used, such as the autoregressive moving average model (ARMA) combined with a neural network structure to extract dynamic features by analyzing the historical values ​​of the data.

[0056] The main task of the feature fusion layer is to fuse the initial features from the local feature extraction layer, the global feature extraction layer, and the dynamic feature extraction layer. Its structure may be adopted in many ways. A common way is weighted summation. For example, the initial feature vectors from the three layers can be set with weight vectors respectively, and then fused by weighted summation. The weight vector here can be automatically learned based on experiments, experience, or through the training process. Another way is splicing. That is, the three initial feature vectors are spliced ​​together in a certain order to form a new fused feature vector.

[0057] In some embodiments, the full life cycle information is input into the local feature extraction layer via the input layer to obtain initial local features, and after the feature fusion layer fuses the initial local features, the local features are output through the output layer, including: inputting the full life cycle information into the local feature extraction layer via the input layer to obtain initial time local features and initial space local features; fusing the initial time local features and the initial space local features in the feature fusion layer, and outputting the local features through the output layer.

[0058] In an embodiment of the present invention, local time feature extraction: extract the time-related characteristics of each specific link in the entire life cycle of the product. For example, in the raw material procurement stage, the time interval between each purchase is recorded, which can reflect the potential impact of the stability and timeliness of raw material supply on the carbon footprint. In the production and manufacturing stage, the time cycle of each batch of product production is counted, including processing time, waiting time, etc. If the processing time is too long, it may mean that the equipment is inefficient, consuming more energy and thus increasing the carbon footprint; long waiting time may involve equipment idling or additional storage requirements, which are also related to the carbon footprint. For the product use stage, monitor the distribution of product usage time in different time periods, such as the difference in usage time of an electrical appliance during the day and at night. Because the power supply source and carbon emission factors may be different in different time periods, the time distribution characteristics of the usage time will affect the total carbon footprint calculation.

[0059] In an embodiment of the present invention, local spatial feature extraction: consider the spatial attributes of specific links in the entire life cycle of the product. In the raw material mining stage, determine the geographical coordinates of the mining site and the geological characteristics of the area (such as ore burial depth, terrain complexity, etc.), because these spatial factors will affect the use of mining equipment, transportation routes and energy consumption, which are related to carbon footprint. In the production and manufacturing stage, analyze the spatial layout characteristics inside the factory, such as the distance between different production workshops, the length of the spatial path for material transportation, etc. Shorter transportation paths may reduce the energy consumption of material handling equipment and reduce carbon footprint. For product sales and distribution, record the geographical location of the sales point or distribution center and the spatial distance from the production site. Different spatial distances involve different transportation mode selections and transportation costs (including carbon costs).

[0060] In some embodiments, the full life cycle information is input into the global feature extraction layer via the input layer to obtain initial global features, and after the feature fusion layer fuses the initial global features, the global features are output through the output layer, including: inputting the full life cycle information into the global feature extraction layer via the input layer to obtain initial time global features and initial space global features; fusing the initial time global features and the initial space global features in the feature fusion layer, and outputting the global features through the output layer.

[0061] In an embodiment of the present invention, global time feature extraction: consider the features from the time span of the entire product life cycle. Calculate the total time length of the entire process from raw material acquisition to final waste disposal of the product. A longer life cycle may mean more maintenance, storage and other links, which increase energy consumption and carbon footprint. Analyze the comprehensive impact of differences in the time distribution of product production and use in different seasons or years on carbon footprint. For example, some products are used frequently in the summer and in special environments (such as air conditioning in high temperature seasons), which may lead to a sharp increase in energy consumption during this time period. From a global time perspective, this seasonal fluctuation feature has an important impact on the overall carbon footprint.

[0062] Using time series decomposition technology, the time series data of the entire life cycle is decomposed into trend, seasonality and residual components. The trend component reflects the long-term change trend of product carbon footprint related factors over time. For example, with the advancement of technology, the energy efficiency of the product throughout its life cycle may show a trend of gradual improvement; the seasonal component clarifies the impact of periodic time fluctuations on carbon footprint, such as the seasonal characteristics of product use mentioned above; the residual part can be used to analyze the impact of special events or abnormal situations on carbon footprint, such as additional carbon footprint fluctuations caused by production interruptions due to natural disasters.

[0063] In an embodiment of the present invention, global spatial feature extraction: focuses on the impact of the spatial layout of the entire product supply chain on the carbon footprint. Determine the distribution range of all geographical locations involved in the entire life cycle of the product. For example, the raw materials of a multinational company's products come from multiple countries, the production bases are distributed in different regions, and the sales market is spread all over the world. This wide spatial distribution range will affect the complexity and energy consumption of the transportation network. Calculate the logistics flow and flow characteristics between different regions, such as the main transportation direction and cargo transportation volume from the raw material production site to the production site, and the production site to the sales site, which helps to analyze the transportation energy consumption hotspots in the global supply chain.

[0064] Using methods such as spatial cluster analysis, regions with similar carbon footprint-related spatial characteristics are clustered. For example, regions with similar energy supply structures and transportation infrastructure conditions are classified into one category, and the carbon footprint contribution characteristics of each category of regions in the entire product life cycle are analyzed. At the same time, considering the differences in environmental policies within the global spatial scope, such as carbon emission standards and renewable energy policies in different countries or regions, these spatial difference characteristics will affect the production and operation decisions of enterprises in different regions, thereby affecting the overall carbon footprint of the product.

[0065] In some embodiments, the feature extraction model also includes an auxiliary extraction layer; the method also includes: inputting the initial local feature into the auxiliary extraction layer to obtain the first auxiliary feature; fusing the initial time global feature and the initial space global feature in the feature fusion layer, and outputting the global feature through the output layer, including: inputting the first auxiliary feature, the initial time global feature and the initial space global feature into the feature fusion layer for fusing, and outputting the global feature through the output layer.

[0066] In an embodiment of the present invention, the production link time cycle data in the local time feature can be integrated into the calculation of the total time length of the whole life cycle by the global time feature extraction module, and the seasonal fluctuation information in the local time feature can assist the global time feature in decomposing a more accurate seasonal component. For example, the high failure rate of a certain equipment in the local production link in the summer leads to an extension of the production time. This local time feature can be taken into account in the global time feature to analyze its impact on the overall product life cycle time and carbon footprint.

[0067] The spatial information of the production workshop layout in the local spatial features can provide the global spatial feature extraction module with details about the internal space utilization efficiency of the production base, which helps to consider the optimization potential within the production base when analyzing the global supply chain spatial layout. For example, if the local spatial features show that the material transportation path of a workshop is too long, the layout of the production base can be adjusted in the global spatial feature analysis, or the logistics arrangement of the base can be optimized in the supply chain planning to reduce the carbon footprint.

[0068] In some embodiments, the full life cycle information is input into the dynamic feature extraction layer via the input layer to obtain the initial dynamic features, and after the feature fusion layer fuses the initial dynamic features, the dynamic features are output through the output layer, including: inputting the full life cycle information into the dynamic feature extraction layer via the input layer to obtain the initial time dynamic features and the initial space dynamic features; fusing the initial time dynamic features and the initial space dynamic features in the feature fusion layer, and outputting the dynamic features through the output layer.

[0069] In an embodiment of the present invention, time dynamic change feature extraction: tracking the dynamic evolution of carbon footprint related factors over time during the entire life cycle of the product. In the raw material procurement link, analyze how the fluctuation of raw material prices over time affects the company's procurement strategy and transportation plan, and thus affects the carbon footprint. For example, when raw material prices rise, companies may reduce the frequency of purchases but increase the amount of each purchase, which may change the mode of transportation and inventory management strategy, leading to dynamic changes in carbon footprint. In the production link, monitor the downward trend of energy consumption over time during the production process improvement process, and the dynamic impact of this trend on the carbon footprint of the product. By establishing a time series model, such as the differential autoregressive moving average model (ARIMA), the future dynamic changes of carbon footprint related factors are predicted, providing a basis for companies to adjust their strategies in advance.

[0070] Analyze the time transition characteristics of products between different life cycle stages. For example, from the product use stage to the waste stage, observe how the product scrap rate changes over time, and how this change triggers changes in waste treatment methods (such as from a higher recycling ratio to an increased landfill ratio), thereby leading to dynamic changes in carbon footprint during the product life cycle transition process. At the same time, consider the impact of external time factors, such as the driving effect of the adjustment of policies and regulations over time on the dynamic changes in product carbon footprints. For example, the gradual tightening of carbon tax policies will prompt companies to take different emission reduction measures at different time points, which is reflected in the dynamic reduction trend of product carbon footprints.

[0071] In an embodiment of the present invention, spatial dynamic change feature extraction: focus on the dynamic changes of spatial related factors throughout the product life cycle. In the process of adjusting the global supply chain layout of an enterprise, such as moving the production base from a high-cost, high-carbon emission area to a low-carbon area, analyze the dynamic impact of this spatial migration on the carbon footprint of the product. This includes changes in the spatial relationship between the new production base and the origin of raw materials and the sales market, as well as the impact of dynamic changes in spatial factors such as local infrastructure and energy supply on the carbon footprint. In the process of product transportation, with the development of transportation networks (such as new highways and railway lines), the spatial dynamic optimization of transportation routes will change the energy consumption and carbon footprint during transportation.

[0072] Use spatial dynamic simulation technology, such as agent-based models, to simulate the flow and carbon footprint changes of products in different spatial scenarios. For example, when urban traffic congestion is alleviated (dynamic changes in spatial traffic conditions), the driving speed of product delivery vehicles is increased and the driving routes are optimized, thereby reducing the carbon footprint of the distribution link. At the same time, consider the impact of spatial dynamic factors such as urbanization and land use changes in different regions on the carbon footprint of the product throughout its life cycle. For example, urban expansion leads to a reduction in production land, and companies may relocate to the suburbs, changing the spatial dynamic processes such as raw material transportation and product distribution, which in turn affects the carbon footprint.

[0073] In some embodiments, the feature extraction model also includes an auxiliary extraction layer; the method also includes: inputting the initial local features and the initial global features into the auxiliary extraction layer to obtain a second auxiliary feature; fusing the initial temporal dynamic features and the initial spatial dynamic features in the feature fusion layer, and outputting the dynamic features through the output layer, including: fusing the second auxiliary features, the initial temporal dynamic features and the initial spatial dynamic features in the feature fusion layer, and outputting the dynamic features through the output layer.

[0074] In an embodiment of the present invention, the combination of local and global features can help the dynamic feature extraction module determine which factors' dynamic changes are more critical to the carbon footprint. For example, if the global time feature shows that the sales and usage of a certain type of product in a specific season increase dramatically, and the local time feature shows that the energy supply stability of the production link in that season is poor, then the dynamic feature extraction module will focus on the time dynamic change characteristics of the energy supply-related factors in the production link in that season, such as the frequency of energy supply interruptions, the start-up time of emergency energy, etc., and the impact of these dynamic changes on the carbon footprint.

[0075] From a spatial perspective, if the global spatial characteristics show that a certain region is the main sales market for a product and the environmental protection policies in the region are becoming stricter, and the local spatial characteristics indicate that the local logistics distribution center is far away from the sales point, then the dynamic feature extraction module will focus on analyzing the spatial dynamic change characteristics of transportation methods, transportation efficiency, etc. caused by policy pressure and spatial layout restrictions in the logistics and distribution links in the region, as well as the impact on carbon footprint, to see whether it will prompt companies to adopt more environmentally friendly but more costly local distribution methods, thereby changing the spatial distribution dynamics of carbon footprint.

[0076] Similarly, the extracted initial spatial dynamic features can also assist in the extraction of local features and global features.

[0077] The time-dynamic change feature can prompt the local time feature extraction module to update and optimize the extracted features. For example, during the product usage phase, the user usage frequency increases significantly in a short period of time due to marketing activities. After the time-dynamic change feature of the usage frequency is extracted, it will be fed back to the local time feature extraction module. This module may re-examine the previously extracted local time features of the product usage time distribution, and further subdivide the time features under different usage scenarios (such as normal use, promotional period use, etc.) to more accurately analyze the changes in carbon footprint under different usage scenarios.

[0078] The dynamic spatial change characteristics also have a similar effect on local spatial characteristics. When a company builds a new sales outlet in a certain area due to changes in market demand, the dynamic spatial change feature extraction module captures this change and feeds the information back to the local spatial feature extraction module. The local spatial feature extraction module may re-evaluate the spatial layout characteristics of the sales network in the area, including the spatial relationship between the new outlets and the original outlets and distribution centers, as well as the impact of this change on the local spatial characteristics of product storage, transportation and other links, and then more accurately analyze the impact of these local spatial feature changes on carbon footprint.

[0079] The dynamic change characteristics of time affect the extraction of global time features. For example, with technological innovation, the production cycle of products has been greatly shortened. Once the dynamic change characteristics of time in this production link are discovered, it will prompt the global time feature extraction module to re-evaluate the total time length of the product life cycle and the time allocation ratio of different stages. It may be found that the part that originally took a long time in the production stage now has a significantly reduced time, while the time in the R&D or after-sales stage has increased relatively. This requires the adjustment of the global time feature to more accurately reflect the changes in the product's carbon footprint in the time dimension of the entire life cycle.

[0080] The dynamic spatial change characteristics are also critical to the adjustment of global spatial characteristics. For example, due to geopolitical factors, the supply stability of a raw material supply source has changed, and the company is forced to adjust the supply chain and purchase raw materials from other regions. After the spatial dynamic change feature extraction module identifies this change, the global spatial feature extraction module will re-analyze the global spatial characteristics such as the geographical distribution range and logistics flow characteristics of the product supply chain. It may be found that the supply of raw materials that was originally concentrated in a certain area has become dispersed and the logistics routes have become more complicated. This requires updating the global spatial features in order to more comprehensively consider the impact of these spatial dynamic changes on the carbon footprint of the product, such as changes in raw material transportation costs and carbon emissions in different regions.

[0081] In some embodiments, the carbon footprint of a product is predicted based on local features, global features, dynamic features and a deep learning model, including: inputting local features, global features, and dynamic features into the deep learning model to obtain a prediction result of the carbon footprint of the product.

[0082] In an embodiment of the present invention, before the three features are input into the deep learning model, they need to be preprocessed and integrated. Local features contain detailed information about each specific link in the entire life cycle of the product, such as local time features such as the processing time and energy consumption of a certain component in the production process, as well as local spatial features such as the spatial layout of a workshop in the factory and the location of equipment. Global features describe the entire product life cycle from a macro perspective, such as the total time span from the acquisition of raw materials to the final disposal of the product, the spatial distribution of all regions involved in the global supply chain, etc. Dynamic features reflect characteristics that change over time or space, such as the change curve of the energy efficiency of products at different years of use (time dynamic change characteristics), or the spatial changes in transportation paths in the supply chain due to the opening of new transportation routes (spatial dynamic change characteristics).

[0083] These features may have different data types, value ranges, and magnitudes. Therefore, data cleaning is required to handle missing values, outliers, and so on. For example, for some missing time data in local features, mean filling or model-based filling methods can be used. Then feature scaling is performed to unify the numerical features to a similar magnitude range, such as using normalization or standardization methods. Normalization maps the data to the [0,1] interval, and standardization makes the data have zero mean and unit variance.

[0084] Next, the processed local features, global features, and dynamic features are integrated in a certain order or manner to form a unified feature vector or feature matrix for input into the deep learning model. For example, different types of features can be concatenated into a long vector, or a two-dimensional matrix can be constructed, where rows represent different feature categories (local, global, dynamic), and columns represent specific feature values.

[0085] Deep learning model selection and architecture design:

[0086] Recurrent Neural Networks (RNN) and their variants (such as LSTM, GRU):

[0087] Since the product life cycle information has a certain order, RNN and its variants are very suitable for processing this type of sequence data. The input layer receives the integrated feature data, and each time step corresponds to the feature information of a stage (the local, global, and dynamic features can be reasonably divided in the time dimension). For example, the first time step inputs the local, global, and dynamic feature combinations related to the raw material acquisition stage, and the second time step inputs the feature combinations of the production and manufacturing stage, etc.

[0088] The hidden units inside RNN learn the forward and backward dependencies between features through cyclic connections. The forget gate, input gate, and output gate in LSTM can effectively handle long-term dependency problems in long sequences and avoid gradient vanishing or exploding. GRU improves computational efficiency through a simplified gating structure. The number of hidden layers and the number of hidden units can be adjusted according to the complexity of the data and the performance of the model. For example, for complex product life cycle data, 2-3 hidden layers may be set, each with 100-200 hidden units.

[0089] The output layer usually has only one neuron, which is used to output the predicted carbon footprint value of the product. The activation function can be selected according to the characteristics of the carbon footprint value. For example, for non-negative carbon footprint values ​​with a wide range, a linear activation function can be used to directly output the predicted value; if you want to map the predicted value to a specific interval (such as [0,1]), you can use the Sigmoid function, etc.

[0090] Convolutional Neural Network (CNN): CNN can be used if the feature data is properly structured, such as combining features at different stages into a two-dimensional matrix similar to an image (rows represent different feature categories, and columns represent different stages or attributes). The input layer receives this feature matrix, and the convolution layer performs convolution operations by sliding multiple convolution kernels on the matrix. Convolution kernels of different sizes can extract features of different scales. Small convolution kernels focus on the extraction of local features, while large convolution kernels can capture more global feature patterns, similar to the extraction process of local and global features.

[0091] The pooling layer is used to reduce the data dimension, reduce the amount of calculation and retain important feature information. For example, the maximum pooling or average pooling operation is used. The fully connected layer integrates and maps the features after convolution and pooling, learns the complex nonlinear relationship between the features, and finally the output layer outputs the predicted value of the product carbon footprint.

[0092] Deep Belief Network (DBN):

[0093] DBN is composed of multiple stacked restricted Boltzmann machines (RBMs). First, the input data consisting of local features, global features, and dynamic features is input into the first RBM, which learns the feature representation of the data through unsupervised learning and extracts some preliminary feature patterns. Then the output of the first RBM is used as the input of the second RBM, and so on. The entire DBN model is constructed through layer-by-layer pre-training.

[0094] After pre-training is completed, the back-propagation algorithm is used to fine-tune the entire network to adapt to the task of carbon footprint prediction. The output layer of the last layer outputs the predicted results of the product's carbon footprint.

[0095] During the training process, the training set data is input into the deep learning model according to the set batch size (such as 32, 64, etc.), the loss function value is calculated, and then the error is back-propagated according to the selected optimization algorithm to update the model parameters. After each training round (epoch), the performance of the model is evaluated using the validation set data, such as calculating the loss value on the validation set, the accuracy (if the carbon footprint prediction results are classified) or other evaluation indicators. According to the evaluation results of the validation set, the hyperparameters of the model are adjusted, such as the learning rate, the number of hidden layers, the number of hidden units, etc. This process is repeated until the performance of the model on the validation set no longer improves or reaches the preset upper limit of the training round.

[0096] In some embodiments, obtaining full life cycle information of a product includes: obtaining initial full life cycle information of the product; performing missing value processing, outlier processing and normalization on the initial full life cycle information to obtain full life cycle information.

[0097] In the embodiment of the present invention, internal system data extraction is first performed:

[0098] Enterprise Resource Planning System: Obtain information such as product production plans, raw material inventory, purchase orders, production costs, etc. from the system. For example, in the manufacturing industry, the system records the raw material usage and cost of each production batch of the product. By querying the relevant modules, the detailed resource input of the product in the production stage can be obtained.

[0099] Customer relationship management system: The system contains customer information, sales records, customer feedback, etc. Taking the sales of electronic products as an example, the system can provide data such as the purchase frequency, purchase time, customer complaints and suggestions of different customer groups for products. This information is very critical for understanding the situation of products in the sales and use stages.

[0100] Product data management system: The system mainly stores product design documents, such as design drawings, technical specifications, product version change records, etc. For the development of complex products such as automobiles, the system can track the changes in design files throughout the entire process from product concept design to finalization, providing a rich resource for obtaining information at the product design stage.

[0101] Second, external data collection is required:

[0102] Supplier data: Establish a data sharing mechanism with raw material suppliers and component suppliers to obtain information such as raw material characteristics, component quality inspection reports, supply cycles, etc. For example, in electronic product manufacturing, chip suppliers can provide chip performance parameters, reliability test data, and production process improvements. This information helps to gain a deeper understanding of the quality and supply stability of key product components.

[0103] Market research agency data: Purchase industry reports, product evaluation reports, etc. published by market research agencies. These reports cover market size, competitive product comparison, consumer preferences, etc. Taking smartphones as an example, market research agency reports can provide information such as the market share of different brands of mobile phones and user satisfaction rankings of the functions of each brand of mobile phones, helping companies understand the position of their products in the market from a macro perspective.

[0104] Data from government departments and industry associations: Pay attention to information such as industrial policies, product standards, and environmental regulations issued by relevant government departments, as well as industry statistics and technology trend reports provided by industry associations. For example, product hazardous substance restriction regulations issued by environmental protection departments will affect the selection of raw materials and recycling methods of products; industry average production efficiency data compiled by industry associations can be used as a reference for companies to evaluate their own production levels.

[0105] Social media and online platform data: Monitor user-generated content on social media platforms, product review websites, e-commerce platforms, etc. The information such as product experience, evaluation, and fault feedback shared by users on these platforms is an important source for understanding the true situation of the product during the usage phase. For example, user reviews on e-commerce platforms can reflect the actual quality of the product, packaging satisfaction, delivery timeliness and other issues.

[0106] Missing value processing: There are many ways to handle missing data. If there are few missing values, you can use the mean, median or mode to fill them. For example, in sales data, if the sales volume of a certain region in a certain month is missing, you can fill it based on the mean sales volume of other regions during the same period. If there are many missing values ​​and they are related to other variables, you can consider using regression analysis or machine learning algorithms to predict and fill them.

[0107] Outlier processing: Identify and process outliers through statistical methods (such as box plot method). For data that obviously deviates from the normal range, it is necessary to further analyze the reasons for its occurrence. If it is a data entry error, it can be corrected directly; if it is an outlier caused by special circumstances, such as a sales peak caused by promotional activities, it can be marked separately and its particularity can be considered in subsequent analysis. For example, in product price data, if a price is found to be much lower than the normal price, it is necessary to make a correction if it is found that the price is wrong.

[0108] Variable normalization: When data from different sources have different dimensions and value ranges, they need to be normalized. Common methods include minimum-maximum normalization and Z-score normalization. For example, in product quality assessment, there are different quality indicators, some of which have a value range of 0-100 and some of which are 0-1. Through minimum-maximum normalization, these indicators can be unified into the range of 0-1, which is convenient for subsequent data comparison and analysis.

[0109] Data format unification: Ensure that data from different sources are consistent in format. For example, the date format may be expressed differently in different systems (such as "YYYY-MM-DD" and "DD / MM / YYYY"), and they need to be unified into one format; text data may have inconsistent uppercase and lowercase letters, which also needs to be unified.

[0110] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0111] Figure 2 Schematic diagram of the structure of the product carbon footprint calculation device provided by the embodiment of the present invention. Figure 2 As shown, the product carbon footprint accounting device, the types of nodes in the power system include power supply nodes, transmission nodes, load nodes and energy storage nodes; the device includes:

[0112] The acquisition module 210 is used to acquire the full life cycle information of the product;

[0113] An extraction module 220 is used to extract local features, global features and dynamic features of the whole life cycle information according to a feature extraction model;

[0114] The prediction module 230 is used to predict the carbon footprint of a product based on local features, global features, dynamic features and a deep learning model.

[0115] Optionally, the feature extraction model includes an input layer, a local feature extraction layer, a global feature extraction layer, a dynamic feature extraction layer, a feature fusion layer, and an output layer; the extraction module 220 is used to: input the full life cycle information into the local feature extraction layer via the input layer to obtain initial local features, and after the feature fusion layer fuses the initial local features, the local features are output via the output layer; input the full life cycle information into the global feature extraction layer via the input layer to obtain initial global features, and after the feature fusion layer fuses the initial global features, the global features are output via the output layer; input the full life cycle information into the dynamic feature extraction layer via the input layer to obtain initial dynamic features, and after the feature fusion layer fuses the initial dynamic features, the dynamic features are output via the output layer.

[0116] Optionally, the extraction module 220 is used to input the full life cycle information into the local feature extraction layer via the input layer to obtain initial time local features and initial space local features; fuse the initial time local features and initial space local features in the feature fusion layer, and output the local features via the output layer.

[0117] Optionally, the extraction module 220 is used to input the full life cycle information into the global feature extraction layer through the input layer to obtain initial time global features and initial space global features; fuse the initial time global features and initial space global features in the feature fusion layer, and output the global features through the output layer.

[0118] Optionally, the feature extraction model also includes an auxiliary extraction layer; an extraction module 220, used to input the initial local feature into the auxiliary extraction layer to obtain a first auxiliary feature; the first auxiliary feature, the initial temporal global feature and the initial spatial global feature are input into the feature fusion layer for fusion, and the global feature is output through the output layer.

[0119] Optionally, the extraction module 220 is used to input the full life cycle information into the dynamic feature extraction layer through the input layer to obtain initial time dynamic features and initial space dynamic features; fuse the initial time dynamic features and initial space dynamic features in the feature fusion layer, and output the dynamic features through the output layer.

[0120] Optionally, the feature extraction model also includes an auxiliary extraction layer; an extraction module 220, used to input the initial local features and the initial global features into the auxiliary extraction layer to obtain a second auxiliary feature; the second auxiliary feature, the initial time dynamic feature and the initial space dynamic feature are fused in the feature fusion layer, and the dynamic feature is output through the output layer.

[0121] Optionally, the prediction module 230 is used to input local features, global features, and dynamic features into the deep learning model to obtain a prediction result of the carbon footprint of the product.

[0122] Optionally, the acquisition module 210 is used to obtain the initial full life cycle information of the product; perform missing value processing, outlier processing and normalization on the initial full life cycle information to obtain the full life cycle information.

[0123] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0124] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A product carbon footprint calculation method, characterized in that: include: Obtain information about the entire product life cycle; Extracting local features, global features and dynamic features of the full life cycle information according to the feature extraction model; The carbon footprint of a product is predicted based on the local features, the global features, the dynamic features and the deep learning model.

2. The product carbon footprint calculation method according to claim 1, characterized in that: The feature extraction model includes an input layer, a local feature extraction layer, a global feature extraction layer, a dynamic feature extraction layer, a feature fusion layer, and an output layer; According to the feature extraction model, local features, global features and dynamic features of the full life cycle information are extracted, including: The full life cycle information is input into the local feature extraction layer through the input layer to obtain initial local features, and the feature fusion layer fuses the initial local features and then outputs the local features through the output layer; The full life cycle information is input into the global feature extraction layer through the input layer to obtain initial global features, and the feature fusion layer fuses the initial global features and then outputs the global features through the output layer; The full life cycle information is input into the dynamic feature extraction layer through the input layer to obtain the initial dynamic features. Meanwhile, after the feature fusion layer fuses the initial dynamic features, the dynamic features are output through the output layer.

3. The product carbon footprint calculation method according to claim 2, characterized in that: The whole life cycle information is input into the local feature extraction layer through the input layer to obtain the initial local features, and the feature fusion layer fuses the initial local features and then outputs the local features through the output layer, including: Inputting the full life cycle information into the local feature extraction layer via the input layer to obtain initial time local features and initial space local features; The initial temporal local features and the initial spatial local features are fused in the feature fusion layer, and the local features are output through the output layer.

4. The product carbon footprint calculation method according to claim 2, characterized in that: The whole life cycle information is input into the global feature extraction layer through the input layer to obtain the initial global feature, and after the feature fusion layer fuses the initial global feature, the global feature is output through the output layer, including: Inputting the full life cycle information into the global feature extraction layer via the input layer to obtain initial time global features and initial space global features; The initial temporal global feature and the initial spatial global feature are fused in the feature fusion layer, and the global feature is output through the output layer.

5. The product carbon footprint calculation method according to claim 4, characterized in that: The feature extraction model also includes an auxiliary extraction layer; the method also includes: Inputting the initial local feature into an auxiliary extraction layer to obtain a first auxiliary feature; The initial temporal global feature and the initial spatial global feature are fused in the feature fusion layer, and the global feature is output through the output layer, including: The first auxiliary feature, the initial temporal global feature and the initial spatial global feature are input into the feature fusion layer for fusion, and the global feature is output through the output layer.

6. The product carbon footprint calculation method according to claim 2, characterized in that: The whole life cycle information is input into the dynamic feature extraction layer through the input layer to obtain the initial dynamic features, and the feature fusion layer fuses the initial dynamic features and then outputs the dynamic features through the output layer, including: Inputting the full life cycle information into the dynamic feature extraction layer via the input layer to obtain initial time dynamic features and initial space dynamic features; The initial temporal dynamic features and the initial spatial dynamic features are fused in the feature fusion layer, and the dynamic features are output through the output layer.

7. The product carbon footprint calculation method according to claim 6, characterized in that: The feature extraction model also includes an auxiliary extraction layer; the method also includes: Inputting the initial local feature and the initial global feature into an auxiliary extraction layer to obtain a second auxiliary feature; The initial temporal dynamic features and the initial spatial dynamic features are fused in the feature fusion layer, and the dynamic features are output through the output layer, including: The second auxiliary feature, the initial time dynamic feature and the initial space dynamic feature are fused in the feature fusion layer, and the dynamic feature is output through the output layer.

8. The product carbon footprint calculation method according to claim 1, characterized in that: Predicting the carbon footprint of a product according to the local features, the global features, the dynamic features and the deep learning model, including: The local features, the global features, and the dynamic features are input into a deep learning model to obtain a prediction result of the carbon footprint of the product.

9. The product carbon footprint calculation method according to claim 1, characterized in that: Get full life cycle information of products, including: Obtain initial full life cycle information of the product; The initial full life cycle information is subjected to missing value processing, outlier processing and normalization to obtain the full life cycle information.

10. A product carbon footprint calculation device, characterized in that: include: Acquisition module, used to obtain the full life cycle information of the product; An extraction module, used to extract local features, global features and dynamic features of the full life cycle information according to a feature extraction model; A prediction module is used to predict the carbon footprint of a product based on the local features, the global features, the dynamic features and the deep learning model.

Images