Carbon footprint intelligent evaluation method based on process modularization algorithm mechanism

Abstract

The application provides a kind of carbon footprint intelligent evaluation method based on process modularization algorithm mechanism, comprising the following steps: S1. Multi-granularity deconstruction and standardization coding of production system process element characteristics;S2. Process carbon footprint element fusion based on evidence reasoning rules;S3. Process modularization carbon footprint transmission network topology construction;S4. Abnormal carbon emission process identification of Bayesian hierarchical variable point detection;S5. Carbon footprint tracing of monte carlo sharple value decomposition graph attention network;S6. Multi-objective carbon footprint optimization method of inter-module coordination.The application comprehensively solves the core technical problems existing in the prior art, such as low accounting accuracy, lack of uncertainty processing, repeated calculation, unreliable anomaly detection, uninterpretable tracing, unfeasible optimization and lack of self-learning ability, and realizes fine, intelligent and sustainable evaluation and optimization of industrial production whole process carbon footprint.

Description

Technical Field

[0001] This invention relates to the field of intelligent evaluation technology of carbon footprint in industrial production, and in particular to an intelligent evaluation method of carbon footprint based on a modular algorithm mechanism for process. Background Technology

[0002] As a major source of global carbon emissions, industrial production requires precise accounting, scientific evaluation, and systematic optimization of its carbon footprint. Existing carbon footprint assessment technologies mainly include life cycle assessment, process analysis, and input-output methods, but they still have many fundamental shortcomings in practical industrial applications. First, the imbalance in process granularity makes it impossible to balance accounting accuracy and computational efficiency. Existing technologies either employ coarse-grained accounting at the workshop or production line level, leading to unclear carbon emission responsibilities and omissions of key emission links; or they blindly pursue fine-grained accounting at the operational level, causing an explosion in computational complexity that cannot be adapted to large-scale complex production systems.

[0003] Secondly, the ability to fuse multi-source heterogeneous data is lacking, and the handling of uncertainty is severely inadequate. Industrial carbon emission data comes from a wide range of sources, covering various data sources such as real-time energy consumption of equipment, fuel combustion, process data, and waste treatment. These sources generally suffer from measurement errors, data conflicts, and system uncertainties. Most existing technologies directly use a single data source or simple weighted averaging without quantifying uncertainty, resulting in poor reliability of the calculation results. Some methods that introduce evidence theory also fail to clarify the rules for generating basic probability allocations, and the evidence combination suffers from normalization defects, making them completely impractical.

[0004] Third, the carbon footprint transfer network is not rigorously constructed, posing a risk of systematic double counting. Existing technologies do not strictly define the system boundaries of accounting units, leading to the overlapping of carbon emissions from upstream raw materials with the unit's own emissions. Furthermore, the method for calculating carbon footprint output is overly simplistic, failing to consider the conversion and distribution relationships of materials and energy during the process, resulting in significant deviations in the calculation of transferred carbon emissions.

[0005] Fourth, abnormal carbon emission detection relies on manual thresholds, resulting in poor robustness. Most existing technologies use fixed threshold methods to identify anomalies, which cannot adapt to dynamic changes such as production load fluctuations and process parameter adjustments, leading to persistently high false alarm and false negative rates. The few methods employing Bayesian variable point detection also fail to properly set prior parameters, resulting in systematic biases in the detection results.

[0006] Fifth, carbon footprint tracing methods suffer from poor interpretability and high computational complexity. Existing technologies often employ black-box models for tracing, failing to quantify the contribution of each upstream link to abnormal carbon emissions and hindering the identification of core issues. Some methods incorporating Shapley values ​​also fail to address the exponential complexity of accurate calculations, making them unsuitable for large-scale production networks.

[0007] Sixth, carbon footprint optimization models suffer from a lack of substance and practical operability. Most existing technologies only define optimization objectives in a general way, without specifying adjustable decision variables or quantifiable constraints. This results in optimization algorithms lacking actual optimization targets and failing to generate feasible emission reduction solutions. Furthermore, modular optimization does not consider the strong coupling relationships between processes, and individual optimization inevitably leads to global suboptimal results.

[0008] Seventh, the lack of a closed-loop learning mechanism leads to a decline in model accuracy over time. Existing technologies all employ static model architectures, which cannot dynamically update model parameters and data reliability weights based on actual operating data during the production process. As production conditions change, such as equipment aging and process adjustments, model accuracy will continue to decrease. Summary of the Invention

[0009] This invention provides a carbon footprint intelligent evaluation method based on a modular process algorithm mechanism. By constructing a multi-granularity process deconstruction system, a standardized evidence reasoning fusion model, an accurate carbon footprint transfer network, adaptive Bayesian variable point detection, interpretable Monte Carlo Shapley source tracing, inter-module coordinated multi-objective optimization, and a dynamic closed-loop learning mechanism, it comprehensively solves the core technical problems existing in the prior art, such as low accounting accuracy, lack of uncertainty handling, redundant calculation, unreliable anomaly detection, uninterpretable source tracing, infeasible optimization, and lack of self-learning ability. It realizes refined, intelligent, and sustainable evaluation and optimization of the carbon footprint of the entire industrial production process.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent carbon footprint evaluation based on a modular process algorithm mechanism includes the following steps: S1. Deconstruct the target production system layer by layer at multiple granularities, standardize and encode the smallest unit obtained from the deconstruction, extract the core meta-features of each encoded unit, and construct the process meta-feature encoding matrix. S2. Collect corresponding multi-source heterogeneous carbon emission data based on the process element feature encoding matrix, construct an evidence reasoning rule fusion model to fuse the multi-source heterogeneous data, and generate a process carbon footprint element library; S3. Construct a carbon footprint transfer network topology based on the process element feature coding matrix, with coding units as nodes. Combined with the process carbon footprint primitive library, calculate the carbon footprint input, carbon footprint output, and carbon transfer coupling degree between each pair of upstream and downstream nodes, and identify strongly coupled node pairs. S4. Extract the time series of carbon footprint input for each node based on the topology of the carbon footprint transmission network, construct a Bayesian hierarchical change point detection model, set the model prior parameters based on the process carbon footprint primitive library, perform change point detection on the carbon footprint input time series, and identify abnormal carbon emission processes. S5. Starting with the abnormal carbon emission process, reverse source tracing reasoning is performed along the carbon footprint transmission network topology to construct a Monte Carlo Shapley value decomposition graph attention network model, quantify the contribution of upstream nodes to abnormal carbon emissions, and identify key influencing nodes. S6. Based on key influencing nodes and strongly coupled node pairs, the topology of the carbon footprint transfer network is divided into several optimization modules. Decision variables and optimization constraints are clearly defined. An improved multi-objective particle swarm optimization algorithm is used to solve the problem. The optimal solution is selected through fuzzy decision-making, and a set of carbon footprint optimization schemes is generated.

[0011] In this specification, S1 uses a multi-granularity, layer-by-layer deconstruction approach to deconstruct the entire process from three progressive granularities: process layer, step layer, and operation layer. The process layer is divided according to the key nodes where the physical form of the product undergoes irreversible changes. The step layer is divided according to the switching nodes where a single piece of equipment completes an independent action cycle. The operation layer is divided according to the individual worker completing an indivisible action unit.

[0012] In this specification, the multi-source heterogeneous carbon emission data in S2 includes three categories: direct carbon emission data, on-site indirect carbon emission data, and baseline carbon emission data. The construction of the evidence reasoning rule fusion model is as follows: a recognition framework is defined that includes true carbon emission values, higher carbon emission values, and lower carbon emission values. For each data source and each coding unit, a basic probability allocation function is generated using the relative deviation method. Initial evidence weights are set according to the data source type. The weighted basic probability allocation function and the unallocated probability are calculated. The normalized evidence reasoning rule combination operator is used to combine all weighted basic probability allocation functions one after another. The carbon emission fusion value and carbon emission uncertainty variance of each coding unit are calculated.

[0013] In this specification, the construction of the carbon footprint transfer network topology in S3 is as follows: Based on the process element feature coding matrix, the material flow, energy flow and information flow transfer relationships between each coding unit are determined, and a directed weighted graph is constructed as the carbon footprint transfer network topology. In the graph, the nodes correspond to each coding unit, the directed edges correspond to the material flow transfer relationships between coding units, and the weight of the edge corresponds to the implicit carbon emissions transferred from the upstream coding unit to the downstream coding unit.

[0014] In this specification, the carbon footprint input of each node in S3 is calculated as the sum of the implicit carbon emissions transmitted by all upstream nodes and the combined carbon emissions of the node itself. The carbon footprint output of each node is calculated using the mass allocation method, based on the total material output, total material input, and auxiliary material added during the production process of the coding unit. The carbon transmission coupling degree between each pair of upstream and downstream nodes is calculated as follows: the implicit carbon emissions transmitted from the upstream node to the downstream node are divided by the total upstream transmitted carbon emissions of the downstream node. When the downstream node has no upstream transmitted carbon emissions, the carbon transmission coupling degree is 0. Node pairs with a carbon transmission coupling degree greater than 0.7 are classified as strongly coupled node pairs, and node pairs with a carbon transmission coupling degree less than 0.3 are classified as weakly coupled node pairs. A strongly coupled node association table is established to record all strongly coupled node pairs.

[0015] In this specification, the Bayesian hierarchical change point detection model constructed in S4 is divided into three layers: observation layer, change point layer, and parameter layer. The prior parameters of the model are set based on the process carbon footprint primitive library as follows: the parameter layer adopts the normal inverse gamma distribution as the prior distribution, and the scale parameter of the prior distribution adopts the weighted average method to integrate the historical variance of the carbon footprint input time series and the carbon emission uncertainty variance in the process carbon footprint primitive library. When the number of historical production cycles is greater than or equal to 20, the historical variance is assigned a weight of 0.7 and the carbon emission uncertainty variance is assigned a weight of 0.3. When the number of historical production cycles is less than 20, the carbon emission uncertainty variance is directly used as the prior scale parameter.

[0016] In this specification, the specific method for detecting change points in the carbon footprint input time series in S4 is as follows: the Markov chain Monte Carlo method is used to estimate the parameters of the model to obtain the posterior probability of each production cycle as a change point. Production cycles with a posterior probability greater than 0.95 are marked as change point cycles. The carbon emission anomaly degree of each node corresponding to each change point cycle is calculated. Nodes with a carbon emission anomaly degree greater than 3 are marked as abnormal carbon emission processes.

[0017] In this specification, the Monte Carlo Shapley value decomposition graph attention network model constructed in S5 includes a graph attention layer and a Monte Carlo Shapley value decomposition layer. The graph attention layer introduces a node dropout strategy during training. In each iteration, a preset proportion of node features are randomly set to zero to simulate the case of subset input, enabling the model to learn the predictive ability when some nodes are missing.

[0018] In this specification, the specific steps for quantifying the contribution of upstream nodes to abnormal carbon emissions in S5 are as follows: Set the number of Monte Carlo samplings to a preset number of samplings, randomly generate an upstream node arrangement for each sampling, calculate the marginal contribution of each node under that arrangement, take the average of all the marginal contributions obtained from the sampling to obtain the Monte Carlo Shapley value of that node, and mark upstream nodes with Monte Carlo Shapley values ​​greater than the preset contribution threshold as key impact nodes.

[0019] In this specification, the optimization modules in S6 are divided according to the following rules: each optimization module contains a key influencing node and all its direct upstream and downstream nodes; all nodes that have a strong coupling relationship with the key influencing node, regardless of whether they are direct upstream or downstream, are included in the same optimization module; if there is a strong coupling relationship between two key influencing nodes, then these two key influencing nodes and their related nodes are merged into one optimization module. In summary, the present invention has at least the following beneficial effects: This invention realizes refined carbon footprint accounting across the entire process at multiple granularities. Through a three-layer progressive granularity deconstruction system, it effectively controls computational complexity while ensuring accounting accuracy, resolves the contradiction between omissions and computational explosion in traditional methods, and clarifies the carbon emission responsibilities of each production unit.

[0020] This invention establishes a standardized framework for the fusion of uncertainties in multi-source heterogeneous data, clarifies the generation rules for the basic probability allocation of various data sources, and adopts a normalized evidence reasoning combination operator to quantify the system uncertainty of the fusion process, thereby significantly improving the accuracy and reliability of carbon footprint basic data.

[0021] This invention constructs a carbon footprint transfer network with clear boundaries and accurate calculations, fundamentally avoiding the duplication of calculations of upstream hidden carbon and unit emissions. It uses a mass allocation method to accurately reflect the carbon transfer and distribution relationship in the process, laying a solid foundation for subsequent anomaly detection and source tracing.

[0022] This invention realizes adaptive thresholdless abnormal carbon emission detection. The Bayesian hierarchical change point detection model integrates prior information on historical data fluctuations and uncertainties in the fusion system, which can automatically adapt to dynamic changes in the production process and significantly reduce false alarm rate and false negative rate.

[0023] This invention provides a fair and interpretable carbon footprint tracing method. The Monte Carlo Shapley value decomposition graph attention network model combines the nonlinear feature extraction capability of deep learning with the fair interpretability of Shapley values. While reducing the computational complexity from exponential to linear, it accurately quantifies the contribution of each upstream node.

[0024] This invention constructs a feasible multi-objective carbon footprint optimization model, clearly defines adjustable decision variables and quantifiable constraints, and introduces an inter-module coordination mechanism to solve the global suboptimal problem caused by individual optimization. It can generate feasible emission reduction schemes that take into account carbon emission reduction, production costs and production efficiency.

[0025] This invention establishes a complete closed-loop self-learning mechanism, which dynamically updates the evidence weights through the source tracing results, continuously improves the accuracy of multi-source data fusion, and enables the model to adapt to changes in production conditions such as equipment aging and process adjustments, maintaining long-term high-precision operation. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating the intelligent carbon footprint evaluation method based on a modular process algorithm mechanism involved in this invention.

[0028] Figure 2 This is a schematic diagram of the process of multi-source carbon emission data fusion and carbon footprint primitive generation involved in this invention.

[0029] Figure 3 This is a flowchart illustrating the process of identifying and locating abnormal carbon footprints involved in this invention.

[0030] Figure 4 This is a schematic diagram illustrating the process of generating the modular optimization and emission reduction scheme for carbon footprint involved in this invention. Detailed Implementation

[0031] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0032] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0033] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] like Figure 1 As shown, this embodiment provides a carbon footprint intelligent evaluation method based on a modular process algorithm mechanism. First, the target production system is deconstructed layer by layer from three progressive granularities: process layer, step layer, and operation layer. The ISO14067 (International Standard for Product Carbon Footprint) extended coding system is used to perform globally unique standardized coding on each smallest production unit, and ten core meta-features are extracted to construct a process meta-feature coding matrix.

[0035] For each coding unit, three types of multi-source heterogeneous data are collected: direct emissions, on-site indirect emissions, and baseline emissions. An evidence reasoning rule fusion model is constructed, the generation method of basic probability allocation of various data sources is clarified, and a normalized combination operator is used to fuse the data to generate a process carbon footprint primitive library containing carbon emission fusion values ​​and uncertainty variance.

[0036] A directed weighted carbon footprint transfer network topology is constructed based on the transfer relationships of material flow, energy flow, and information flow. The system boundary is strictly defined to avoid redundant calculations. The carbon footprint input and output of each node are calculated using the mass allocation method. The carbon transfer coupling degree between processes is quantified and strong / weak coupling node pairs are divided.

[0037] A Bayesian hierarchical change point detection model is constructed that integrates historical data random fluctuations and system uncertainties to automatically identify change point cycles where carbon emissions undergo statistically significant changes, calculate carbon emission anomalies, and mark abnormal carbon emission processes.

[0038] A Monte Carlo Shapley value decomposition graph attention network model is constructed to trace the source of abnormal carbon emissions backwards. A node dropout training strategy ensures the accuracy of predictions based on subset inputs, while Monte Carlo sampling reduces the computational complexity of the Shapley value. The model quantifies the contribution of each upstream node to abnormal carbon emissions and identifies key influencing nodes. Simultaneously, the source tracing results are used as feedback signals, and an upper bound truncation method is employed to update the evidence weights, forming a closed-loop learning mechanism.

[0039] With key influencing nodes as the core, optimization modules are divided according to the principle of strong coupling nodes being in the same module. Four types of decision variables, namely process parameters, equipment selection, material routes, and personnel scheduling, and three types of constraints, namely capacity, quality, and equipment load, are clearly defined. An improved multi-objective particle swarm optimization algorithm with an inter-module coordination mechanism is adopted to solve the global Pareto optimal frontier. The optimal solution is selected through fuzzy decision and a feasible carbon footprint optimization scheme is generated.

[0040] Finally, a dynamic update mechanism combining scheduled updates and event triggers is set up to achieve automatic updates and continuous optimization of carbon footprint assessment results.

[0041] The specific steps are as follows: S1. Multi-granularity deconstruction and standardized coding of process element characteristics in the production system; The target production system is deconstructed layer by layer from three progressive granularities: process layer, step layer, and operation layer. This three-layer deconstruction system can cover all carbon emission-related links from macro-production process to micro-manual actions, avoiding carbon emission omissions due to overly coarse granularity or computational complexity explosion due to overly fine granularity. The process layer is strictly divided according to the key nodes where the physical form of the product undergoes irreversible changes, including five core processes: raw material pretreatment, forming, assembly, inspection, and packaging. The step layer is divided according to the switching nodes where a single piece of equipment completes an independent action cycle, including six standard steps: clamping, cutting, tool retraction, unloading, cleaning, and drying. The operation layer is divided according to the individual worker completing an indivisible action unit, including five basic operations: handling, button operation, parameter adjustment, quality inspection, and tool replacement.

[0042] The ISO 14067 (International Standard for Product Carbon Footprint) extended coding system is adopted to assign a globally unique standardized code to each deconstructed smallest unit. The code uses a continuous character concatenation format, consisting of a system code, process code, step code, and operation code. The system code is a six-digit number used to distinguish different production systems or production lines; the process code is a four-digit number used to distinguish different processes within the same production system; the step code is a three-digit number used to distinguish different steps within the same process; and the operation code is a two-digit number used to distinguish different operations within the same step. This coding system enables rapid location and traceability of any production unit, providing a unified identifier for the association and matching of carbon footprint data.

[0043] The core meta-features of each coding unit are extracted. Unit duration is extracted from production cycle records, accurate to the second, and used to calculate energy consumption and labor costs per unit time. Participating equipment models are matched from equipment ledgers to correlate rated energy consumption parameters and carbon emission factors. Material input and output quantities are obtained from the production material balance sheet to calculate implicit carbon emissions from material transfer and material conversion rates. Energy consumption types include electricity, coal, natural gas, and steam, used to match the corresponding carbon emission factors. Energy consumption power is obtained from equipment nameplate parameters to calculate theoretical energy consumption. Waste generation types include solid waste, liquid waste, and gaseous waste, used to match the carbon emission factors of corresponding waste treatment methods. The theoretical waste generation per unit product is extracted from production process documents to calculate theoretical waste treatment emissions. Equipment load rate is obtained from the equipment operation monitoring system to calculate actual energy consumption and equipment operating costs. Auxiliary material addition amounts are selected to supplement the process material balance, improve carbon footprint allocation, and support output carbon calculation.

[0044] Construct a process meta-feature coding matrix. The rows of the matrix correspond to each uniquely coded production unit, the columns of the matrix correspond to ten meta-features, and the matrix elements are the quantized values ​​or classification codes of the corresponding meta-features.

[0045] S2. Process carbon footprint primitive fusion based on evidence reasoning rules; multi-source carbon emission data fusion and carbon footprint primitive generation process as follows: Figure 2 As shown, for each coding unit in the process element feature coding matrix, three types of multi-source heterogeneous carbon emission related data are collected simultaneously. The data covers three dimensions: direct emissions, on-site indirect emissions, and baseline emissions. The system boundary is strictly defined as the boundary of the coding unit itself, excluding carbon emissions from upstream raw material production and downstream product use.

[0046] Direct carbon emission data includes real-time equipment energy consumption data, fuel combustion emission data, and process emission data. Real-time equipment energy consumption data is collected every 15 minutes through the enterprise's energy management system; fuel combustion emission data is calculated using fuel consumption and corresponding emission factors; and process emission data is calculated using chemical reaction equations and material inputs. Indirect carbon emission data includes on-site waste treatment carbon emission data and on-site transportation carbon emission data. On-site waste treatment carbon emission data is calculated using waste generation and corresponding treatment emission factors; and on-site transportation carbon emission data is calculated using in-plant transportation distance, transportation mode, and corresponding emission factors. Benchmark carbon emission data includes industry average carbon emission data, benchmark enterprise carbon emission data, and data from the national carbon emission factor database, obtained from annual reports published by industry associations, publicly available benchmark enterprise carbon disclosure reports, and the carbon emission factor database published by the Ministry of Ecology and Environment, respectively.

[0047] A fusion model of evidence reasoning rules is constructed to address the conflicts and uncertainties inherent in multi-source data. A recognition framework is defined. ,in Indicates the true carbon emissions value. This indicates a relatively high carbon emission value. This represents a relatively low carbon emission value. Define the basic probability assignment function. ,in Indicates the first One data source, Indicates the first One coding unit, , Indicates the first The data source for the first Each coding unit proposition The level of support.

[0048] For each data source and each encoding unit, the basic probability allocation function is generated according to the following method, with all data sources using the relative bias method: 1. Real-time energy consumption data of the equipment: The theoretical energy consumption value equals the unit duration multiplied by the energy consumption power multiplied by the equipment load rate. Calculate the relative deviation between the measured energy consumption value and the theoretical energy consumption value. .when hour, , , ;when hour, , , ;when hour, , , .

[0049] 2. Fuel Combustion Emission Data: Theoretical fuel consumption equals the fuel consumption per unit of product specified in the production plan multiplied by the product output of the coded unit. Theoretical fuel combustion emissions equal theoretical fuel consumption multiplied by the carbon emission factor of the corresponding fuel. Calculate the relative deviation between measured fuel combustion emissions and theoretical fuel combustion emissions. The deviation threshold is the same as the equipment's real-time energy consumption data.

[0050] 3. Process Emission Data: The theoretical process emission is equal to the theoretical emission per unit product calculated based on the chemical reaction equation multiplied by the product output of the coded unit. The relative deviation between the measured process emission and the theoretical process emission is calculated. The deviation threshold is the same as the equipment's real-time energy consumption data.

[0051] 4. On-site waste treatment carbon emission data: Theoretical waste generation equals the theoretical waste generation per unit of product multiplied by the product output of that coded unit. Theoretical waste treatment emissions equal the theoretical waste generation multiplied by the carbon emission factor of the corresponding waste treatment method. Calculate the relative deviation between the measured waste treatment emissions and the theoretical waste treatment emissions. The deviation threshold is adjusted as follows: when hour, , , ;when hour, , , ;when hour, , , .

[0052] 5. On-site transportation carbon emission data: The theoretical transportation distance equals the unit product transportation distance specified in the standard operating procedure multiplied by the product output of that coded unit. The theoretical transportation emissions equal the theoretical transportation distance multiplied by the carbon emission factor of the corresponding transportation mode. Calculate the relative deviation between the measured transportation emissions and the theoretical transportation emissions. The deviation threshold is the same as the carbon emission data for on-site waste treatment.

[0053] 6. Baseline carbon emission data: The theoretical value equals the average carbon emission per unit of product for that type of coded unit, as specified in the national carbon emission factor database or industry report, multiplied by the product output of that coded unit. Calculate the relative deviation between the measured carbon emissions and the theoretical value. The deviation threshold is adjusted as follows: when hour, , , ;when hour, , , ;when hour, , , .

[0054] Define the weight of evidence , indicating the first The data source for the first The reliability level of each coding unit, with a value range of [value range missing]. The initial evidence weights are set according to the type of data source: real-time energy consumption data from equipment has an initial weight of 0.9, fuel combustion emission data has an initial weight of 0.8, process emission data has an initial weight of 0.7, on-site waste treatment carbon emission data has an initial weight of 0.6, on-site transportation carbon emission data has an initial weight of 0.5, industry average carbon emission data has an initial weight of 0.3, benchmark enterprise carbon emission data has an initial weight of 0.4, and national carbon emission factor database data has an initial weight of 0.5. It is clearly defined that each data source and coding unit has a one-to-one binding relationship; that is, each coding unit is equipped with an independent metering device, and there is no situation where a single data source provides data to multiple coding units simultaneously.

[0055] Calculate the weighted basic probability assignment function This step incorporates evidence weights into the basic probability allocation, reflecting the differences in reliability among different data sources. The calculation formula is as follows: ; Calculate the unassigned probability , indicating the first The data source cannot explicitly support the first one. The probability of any proposition in a coding unit is calculated using the following formula: ; An evidence reasoning rule combination operator is employed to successively combine all weighted basic probability assignment functions. This operator effectively handles evidence conflict problems and avoids counterintuitive results that occur with the traditional Dempster combination rule in high-conflict situations. The resulting basic probability assignments are shown below. The calculation formula is: ; in The denominator represents the total number of data sources, and the sum of the basic probability distributions of all propositions is used for normalization to ensure... .

[0056] Calculate the measured carbon emissions for each coding unit. This value is the sum of the measured values ​​of all direct emissions and indirect emissions from the coding unit itself, calculated using the following formula: ; in This represents the total number of emission sources, where s is the emission source index. For example, s=1 represents equipment energy consumption, s=2 represents fuel combustion, s=3 represents the process, s=4 represents waste treatment, and s=5 represents in-plant transportation, etc. Indicates the first The coding unit of the first Activity level data for emission sources of this type Indicates the first Carbon emission factors corresponding to each type of emission source.

[0057] Calculate the carbon emission fusion value for each coding unit. This value integrates information from all data sources and is the optimal estimate of the actual carbon emissions of the coding unit itself. The calculation formula is as follows: ; in This indicates a relatively high carbon emission value. This indicates a relatively low carbon emission value; based on the typical error range (±20%) for industrial carbon emission monitoring, let... , This coefficient can be adjusted in practical applications.

[0058] Calculate the carbon emission uncertainty variance for each coding unit. This value reflects the dispersion of the carbon emission fusion value, quantifies the system uncertainty generated during the multi-source data fusion process, and is used to set the prior parameters for the subsequent Bayesian hierarchical change point detection model. The calculation formula is as follows: ; Construct a process carbon footprint primitive library, in which each entry contains a coding unit code and a fused carbon emission value. Carbon emission uncertainty variance Evidence weights for each data source The database consists of six fields: data source, update time, etc.

[0059] S3. Construction of the modular carbon footprint transfer network topology for each process: Based on the process meta-feature coding matrix, the material flow, energy flow, and information flow transfer relationships between each coding unit are determined one by one. The material flow transfer relationships are extracted from the material inbound and outbound records of the production logistics management system, clarifying the source of input materials and the destination of output materials for each coding unit; the energy flow transfer relationships are extracted from the equipment energy flow analysis report, clarifying the energy transfer and sharing relationships between equipment; the information flow transfer relationships are extracted from the instruction records of the production scheduling system, clarifying the transmission order of production instructions between each coding unit.

[0060] A directed weighted graph is constructed as the topology of the carbon footprint transfer network. This topology can intuitively display the flow path and transfer relationship of the carbon footprint in the production system. The nodes in the graph correspond to each coding unit in S1, and the directed edges in the graph correspond to the material flow transfer relationship between coding units. The direction of the edges is consistent with the direction of the material flow.

[0061] Calculate the weight of a directed edge. ,in Indicates the upstream coding unit, The upstream coding unit represents the downstream coding unit, and the edge weight represents the implicit carbon emissions passed from the upstream coding unit to the downstream coding unit. The calculation formula is as follows: ; in Indicates upstream coding unit Downstream coding unit The output material quantity is obtained from the production material balance sheet; This indicates the implicit carbon emission factor of the material, obtained from the IPCC National Greenhouse Gas Inventory Guidelines and localized according to industry characteristics. This factor includes the carbon emissions of the raw materials throughout their entire life cycle, from mining to transportation to the plant.

[0062] Calculate the carbon footprint input for each node. This represents the total carbon emissions entering the coding unit, including implicit carbon emissions from upstream sources and carbon emissions generated by the unit itself. The calculation formula is as follows: ; in Represents all directional coding units Input the upstream coding unit set of the material. For coding unit Its own carbon emission combined value, due to The emissions are strictly limited to those within the unit's own boundary and do not include carbon contained in upstream raw materials, thus avoiding the problem of double counting.

[0063] Calculate the carbon footprint output for each node. This represents the total carbon emissions transferred from the coding unit to the downstream unit. It is calculated using the mass allocation method, taking into account the material and energy conversion relationships during the process. The calculation formula is: ; in Represents coding unit Total material output Represents coding unit Total material input Represents coding unit The amount of auxiliary materials added during the production process. This method is more accurate than simple material conversion rate and can rationally allocate the carbon emissions generated during the process to the output product.

[0064] Calculate the carbon transport coupling degree between each pair of upstream and downstream nodes. This indicator reflects the degree to which upstream carbon emissions affect downstream carbon emissions, and the calculation formula is as follows: ; in Indicates downstream coding unit The total upstream carbon emissions. When When the node is defined as the starting process of the production line, there is no upstream carbon emission transmission. Therefore, its coupling degree with all upstream nodes is defined as 0 to avoid division by zero error.

[0065] Node pairs with a carbon transport coupling degree greater than 0.7 are marked as strongly coupled node pairs, indicating that changes in carbon emissions from upstream units have a significant impact on downstream units; node pairs with a carbon transport coupling degree less than 0.3 are marked as weakly coupled node pairs, indicating that changes in carbon emissions from upstream units have a negligible impact on downstream units. A strongly coupled node association table is established to record all strongly coupled node pairs.

[0066] S4. Identification of anomalous carbon emission processes using Bayesian hierarchical change point detection; the process for identifying and locating anomalous carbon footprint processes is as follows: Figure 3 As shown. From the carbon footprint transfer network topology of S3, the past performance of each node is extracted. Time series of carbon footprint inputs within a production cycle ,in Indicates the first 1 node Indicates the first One production cycle, Represents a node In the Carbon footprint input per production cycle The production cycle was set to 100 cycles to ensure that the time series has sufficient statistical significance.

[0067] Calculate the historical mean of the carbon footprint input time series for each node. and historical variance The calculation formula is: , ; A Bayesian hierarchical change point detection model was constructed. This model can automatically identify time points in time series where carbon emissions change significantly without the need for pre-setting thresholds, and it exhibits strong robustness and adaptability. The model consists of a three-layer structure, from bottom to top: the observation layer, the change point layer, and the parameter layer.

[0068] The observation layer describes the probability distribution of the carbon footprint input time series. It is assumed that the carbon footprint input within each segment follows a normal distribution, and the likelihood function is: ; in Represents a node In the Average carbon emissions for each segment Represents a node In the The carbon emission variance of each segment This represents the probability density function of the normal distribution.

[0069] The change point layer describes the probability of a change occurring and defines change point indicator variables. ,when Time indicates the first The production cycle is the variable point, when Time indicates the first The production cycle is not a turning point. Assume the occurrence of turning points follows a Bernoulli distribution, with the prior distribution as follows: ; in Represents a node The prior probability of a change point occurring is obtained from historical production data statistics, with an initial value set at 0.01. It can be adaptively adjusted according to the actual frequency of change points occurring.

[0070] The parameter layer describes the prior distribution of the mean and variance of carbon emissions within each segment, using the inverse normal gamma distribution as the conjugate prior to simplify the posterior calculation. The prior distribution simultaneously incorporates the random fluctuation uncertainty of historical data and the systematic uncertainty of multi-source data fusion; the calculation formula is as follows: ; in For nodes Historical mean of carbon footprint inputs over time series Set to 1. Set to 2. The scale parameter of the inverse gamma distribution is used, and a weighted average method is adopted to integrate historical variance and carbon emission uncertainty variance. The calculation formula is as follows: ; in For nodes The corresponding encoding unit, Let Variance be the carbon emission uncertainty of this coding unit. When the historical production cycle number... When the number of historical production cycles is greater than or equal to 20, the historical data has sufficient statistical significance, therefore a weight of 0.7 is assigned to the historical variance and a weight of 0.3 is assigned to the fusion uncertainty variance. When the number of historical production cycles is less than 20, historical data is insufficient, and the carbon emission fusion uncertainty variance is directly used as the prior scale parameter. This design can simultaneously consider the random fluctuations of the production process and the systematic errors of data measurement fusion, making the prior distribution more consistent with the actual situation and improving the accuracy of change point detection.

[0071] The Markov chain Monte Carlo method was used to estimate the parameters of the model. A No-U-Turn sampler was used for sampling, and the total number of iterations was set to 10,000. The first 5,000 iterations were the combustion period to eliminate the influence of initial values, and the last 5,000 iterations were used to calculate the posterior probability distribution of the parameters. The posterior probability of each production cycle as a variable point was obtained. .

[0072] Production cycles with a posterior probability greater than 0.95 are designated as change-point cycles, indicating that carbon emissions have undergone statistically significant changes within that cycle. For each change-point cycle... Calculate the carbon emission anomaly degree of the corresponding node. This indicator reflects the magnitude of the change in carbon emissions before and after the turning point, and is calculated using the following formula: ; in The average carbon emissions before the change point. The average carbon emissions after the change point. This represents the standard deviation of carbon emissions before the change point.

[0073] Nodes with a carbon emission anomaly score greater than 3 are marked as anomalous carbon emission processes. This threshold is based on the 3σ principle, indicating that the average carbon emission after the change point exceeds three times the normal fluctuation range before the change point, which is considered a significant anomaly. A set of anomalous carbon emission processes is then formed.

[0074] S5. Carbon Footprint Tracing Using Monte Carlo Shapley Value Decomposition Graph Attention Network: Starting with anomalous carbon emission processes, deep tracing reasoning is performed along the reverse edges of the carbon footprint transmission network topology, covering all upstream nodes that may lead to anomalous carbon emissions. A Monte Carlo Shapley value decomposition graph attention network model is constructed. This model integrates the nonlinear feature extraction capability of graph attention networks with the interpretability and computational efficiency advantages of Monte Carlo Shapley values, accurately quantifying the contribution of each upstream node to anomalous carbon emissions.

[0075] The model's input consists of the adjacency matrix and node feature matrix of the carbon footprint transfer network topology. The adjacency matrix is ​​constructed based on the carbon footprint transfer network topology; if there are slave nodes... To the node If there is a directed edge, the element at the corresponding position in the adjacency matrix is ​​1; otherwise, it is 0. The rows of the node feature matrix correspond to each coding unit, and the columns correspond to four static features: the unit's own carbon emissions, upstream carbon emissions, material input, and material output. The feature values ​​are all obtained from the calculation results of S2 and S3, avoiding the use of dynamically changing carbon transfer coupling degree as static input features.

[0076] The input to the graph attention network is the feature matrix. and adjacency matrix , where N is the total number of nodes, d=4 (four static characteristics: carbon emissions of the unit itself, carbon emissions transmitted upstream, material input, and material output).

[0077] The network output is a scalar , representing the predicted carbon footprint input value of node i in the abnormal carbon emission process. During training, for each abnormal process node, the features of all its upstream nodes and the full graph adjacency matrix are used as input, and the actual carbon footprint input value of the abnormal process is used as the supervision signal.

[0078] When calculating the marginal contribution of the Shapley value, the node features corresponding to the upstream node subset S to be evaluated are retained, while the features of the remaining nodes are set to zero. The input to the network yields the predicted value f(S), which is the carbon footprint input of the abnormal process under subset S.

[0079] The first layer of the model is a graph attention layer, used to capture non-linear dependencies between nodes. The model training process is as follows: historical production data and corresponding carbon footprint data from the past 12 months are collected to construct the training dataset; a node dropout strategy is introduced during training, randomly setting 10% to 30% of node features to zero in each iteration to simulate subset input, enabling the model to learn predictive capabilities when some nodes are missing; the loss function is defined as the mean squared error loss function, calculated using the following formula: ; in Indicates the number of training samples. This represents the predicted carbon footprint input value for process node i with abnormal carbon emissions. This represents the actual carbon footprint input; the Adam optimizer is used to optimize the model parameters, with the learning rate set to 0.001, the batch size set to 32, and the training epochs set to 100; when the validation set loss no longer decreases for 10 consecutive epochs, training is stopped early, and the optimal model parameters are saved.

[0080] After training is completed, the computing nodes Its neighboring nodes Attention coefficient This coefficient represents the node Features of nodes The degree of influence of a feature is calculated using the following formula: ; in Represents a node The initial feature vector, This represents the linear transformation weight matrix obtained during training. This represents the attention vector obtained during training. This represents a vector concatenation operation. Represents a node The set of all neighboring nodes, This represents the leakage correction linear unit activation function, with its negative slope set to 0.2.

[0081] By introducing the posterior probability of change points to weight and correct the attention coefficient, a deep fusion of the Bayesian hierarchical change point detection model and the graph attention network is achieved. This allows the model to pay more attention to nodes where carbon emissions change significantly, improving the accuracy of source tracing. (Corrected attention coefficient) The calculation formula is: ; in The nodes obtained in S4 The posterior probability is the current production cycle as the point of change.

[0082] compute nodes Update feature vector This vector combines the node's own features with the weighted features of its neighboring nodes, and the calculation formula is as follows: ;in This represents the sigmoid activation function.

[0083] The second layer of the model is a Monte Carlo Shapley value decomposition layer, used to quantify the marginal contribution of each upstream node to anomalous carbon emissions, addressing the problem of high computational complexity in accurate Shapley value calculation. The set of all upstream nodes is defined as... Set the number of Monte Carlo sampling times to , The sampling time is set to 1000. For each sampling, a node permutation is randomly generated. Determine the node Position in the arrangement The node in the generated permutation Previous node set .

[0084] In calculation At that time, the same method of zeroing nodes as in the training process is used: nodes not in the subset are zeroed out. All feature vectors of all nodes are set to zero, while the adjacency matrix of the graph remains unchanged. The processed feature matrix and adjacency matrix are then input into a trained graph attention network. After passing through the graph attention layer and fully connected layers, the predicted carbon footprint input value of abnormal carbon emission processes is calculated. This method ensures the consistency of the distribution between the prediction and training processes, avoiding prediction inaccuracies caused by out-of-distribution inputs.

[0085] In this model, the graph attention network The output is fixed as the predicted carbon footprint input value of node t in the abnormal carbon emission process. Therefore, for different upstream node subsets S, the network output is the predicted value of the same target node t. The Shapley value for each upstream node i... This represents the average marginal contribution of node i to the predicted carbon emissions of target node t, regardless of whether node i appears in the subset.

[0086] compute nodes marginal contribution The calculation formula is: ; compute nodes Monte Carlo Shapley value This value represents the node. The average marginal contribution to anomalous carbon emissions is a fair quantification of the degree of contribution of each node, and the calculation formula is as follows: ; in Indicates the first The node arrangement obtained from the second sampling. This method reduces the computational complexity from Reduce to This meets the needs of practical applications. Shapley value The value can be positive or negative. A positive value indicates that the presence of node i increases abnormal carbon emissions (positive contribution), while a negative value indicates that the node actually inhibits carbon emissions (negative contribution). In this embodiment, only nodes with a Shapley value greater than 0.1 that contribute positively are marked as key impact nodes.

[0087] A key impact factor quantification matrix is ​​constructed, where rows correspond to anomalous carbon emission processes, columns correspond to upstream nodes, and matrix elements are the Monte Carlo Shapley values ​​of the corresponding upstream nodes for anomalous carbon emission processes. .

[0088] Upstream nodes with a Shapley value greater than 0.1 are marked as critical impact nodes, indicating that these nodes contribute more than 10% to anomalous carbon emissions and are the main cause of the carbon emission anomalies. A set of critical impact nodes is thus formed.

[0089] The calculated Shapley value is used as a feedback signal to update the evidence weights of all data sources related to key influencing nodes in S2. This achieves reverse optimization of the basic data fusion process based on the source tracing results, forming a closed-loop learning mechanism. Updated evidence weights The calculation formula is: ; in The original evidence weights in S2, For nodes The Shapley value. This formula uses an upper limit truncation method, which can both increase the weight of data sources related to key influencing nodes, enhancing their role in the subsequent fusion process, and ensure that the weight always remains within the range of... Within a reasonable range, the weights are prevented from growing indefinitely, while maintaining the physical meaning of the evidence weights representing absolute reliability.

[0090] S6. A multi-objective carbon footprint optimization method with inter-module coordination; the process of modular carbon footprint optimization and emission reduction scheme generation is as follows: Figure 4 As shown, the carbon footprint transfer network topology is divided into several optimization modules, with key influencing nodes at the core. The module division follows these rules: each optimization module contains one key influencing node and all its direct upstream and downstream nodes; all nodes with strong coupling relationships to the key influencing node, regardless of whether they are direct upstream or downstream, are included in the same optimization module; if two key influencing nodes have a strong coupling relationship, these two key influencing nodes and their related nodes are merged into one optimization module. This division method ensures that strongly coupled nodes are always in the same optimization module, avoiding the global suboptimal problem caused by individual optimization.

[0091] Four categories of decision variables are clearly defined, all of which are adjustable production parameters. The first category is process parameter variables, including equipment operating temperature, equipment operating pressure, cutting speed, feed rate, etc.; the second category is equipment selection variables, which determine which equipment to use for production when multiple machines with the same function exist; the third category is material routing variables, which determine which route to use for material transportation when multiple material transportation routes exist; and the fourth category is personnel scheduling variables, which determine the arrangement of different numbers of workers for production at different times.

[0092] Establish a quantitative relationship model between decision variables and carbon footprint inputs; for continuous process parameters such as temperature, pressure, and cutting speed, use multiple linear regression or random forest regression to establish the mapping. ,in For the first One process parameter, The regression coefficients are obtained by training with the least squares method based on historical production data (at least 100 sets).

[0093] For discrete equipment, a table lookup method is used to select variables: the carbon emission per unit product of each piece of equipment under standard operating conditions is pre-determined and denoted as... Then the increase in carbon emissions , This indicates the number of products processed by the equipment per unit time (pieces / hour) or the batch size of products processed in a single run.

[0094] For material route variables, carbon emissions are calculated by multiplying the transport distance by the emission factor of the transport vehicle: ,in The distance is the route distance (km). Emission factor (kgCO2 / t·km). The weight of the material is in tons (t).

[0095] For personnel scheduling variables, the impact of carbon emissions is mainly reflected indirectly through changes in production efficiency: increasing personnel may shorten the time per unit, but it will increase auxiliary energy consumption such as lighting and air conditioning. This can be assessed by fitting curves using historical data. Make an estimate.

[0096] Three conflicting multi-objective optimization functions are defined to comprehensively consider the objectives of carbon reduction, cost, and efficiency. The first objective function minimizes the total carbon emissions of the optimization module, and its calculation formula is as follows: ; in This represents the set of coding units contained in the optimization module. Represents coding unit The carbon footprint input is a decision variable vector. The function.

[0097] The second objective function is to minimize the production cost of the optimization module. The production cost includes four parts: energy cost, raw material cost, labor cost, and equipment maintenance cost. The calculation formula is as follows: ; in Represents coding unit Energy costs, Represents coding unit The cost of raw materials, Represents coding unit Labor costs, Represents coding unit The equipment maintenance costs are all decision variable vectors. The function.

[0098] The third objective function is to maximize the production efficiency of the optimization module, and the calculation formula is: ; in This represents the average time (in hours / piece) required for the optimization module to produce each qualified product; production efficiency is inversely proportional to the production time per unit, and minimizing this time is key. Equivalent to maximizing .

[0099] It should be noted that, (Minimize carbon emissions) (Minimize production costs) and There are typical conflicts between maximizing production efficiency (i.e., minimizing the time per unit). For example, increasing production efficiency (reducing the time per unit) may lead to high-load operation of equipment, increasing carbon emissions and costs; while reducing carbon emissions may require slowing down the production pace and extending the time per unit. The multi-objective optimization algorithm of this invention aims to obtain the Pareto front, allowing decision-makers to choose an equilibrium solution based on their actual preferences.

[0100] Establish mathematical expressions for three core optimization constraints. Capacity constraint: The hourly capacity of the optimization module must not be lower than the hourly output requirement specified in the production plan. The expression is: ; in To optimize the module's output of qualified products (pieces / hour) per unit time (hour). It is calculated by dividing the daily output of the production plan by the daily planned production hours.

[0101] The quality constraint is that the product qualification rate of the optimization module shall not be lower than the company's stipulated quality standard of 95%, expressed as: ; in The product pass rate of the optimization module is represented by a vector of decision variables. The function is obtained by building a quality prediction model using historical data.

[0102] The equipment constraint is that the average load rate of all equipment within the optimization module does not exceed 85% of the equipment's rated load rate, expressed as: ; in Represents coding unit The average load rate of the corresponding equipment This indicates the rated load rate of the equipment.

[0103] An improved multi-objective particle swarm optimization (PSO) algorithm is employed to solve the aforementioned multi-objective optimization problem. This algorithm introduces an inter-module coordination mechanism based on the traditional PSO algorithm. First, each optimization module is optimized independently, yielding its local Pareto optimal front. Then, an inter-module coordination model is established, using the output of the upstream module as the input of the downstream module for iterative optimization. Iteration stops when the change in input / output parameters between adjacent modules is less than 1%, yielding the global Pareto optimal front. This method leverages the advantage of modular optimization in reducing computational complexity while addressing the global suboptimal problem caused by independent module optimization.

[0104] The algorithm iterations are set to 200, and the population size is set to 50. After the iterations, each solution on the global Pareto optimal front is a feasible solution that achieves optimal balance among the three objectives.

[0105] A fuzzy decision-making method is used to select the final optimal solution from the Pareto optimal frontier. A linear membership function is defined for each objective function, the comprehensive membership degree of each non-dominated solution is calculated, and the non-dominated solution with the largest comprehensive membership degree is selected as the optimal solution.

[0106] A set of carbon footprint optimization schemes is generated based on the optimal solution. Each scheme includes five parts: optimization module number, specific optimization measures, expected carbon emission reduction, expected cost change, and expected efficiency change. Possible optimization measures include five types: replacing equipment with high-efficiency equipment, adjusting process parameters, optimizing material transportation routes, improving waste disposal methods, and optimizing personnel scheduling.

[0107] The detailed process of the improved multi-objective particle swarm optimization algorithm (with inter-module coordination): Input: Set of optimization modules (Arranged in process order), decision variable boundaries and objective function for each module. Constraints.

[0108] Output: Global Pareto optimal frontier .

[0109] step: 1. For each module Independently run standard multi-objective particle swarm optimization (MOPSO) to obtain the local Pareto front. .

[0110] 2. Initialize global variables This represents a feasible solution for each module.

[0111] 3. Repetition: For k=1 to K (sequential iteration): fix the output of the upstream module (material flow rate, quality, etc.) as the input boundary of this module; in Select the solution that minimizes the upstream and downstream connection error, update the decision variables of module k; calculate the output of this module. It is passed on to the downstream.

[0112] If the rate of change of input and output parameters between all adjacent modules is less than Then stop.

[0113] 4. Combine the final solutions of each module into a global solution to obtain a set of global Pareto fronts.

[0114] Particle velocity and position update formulas (Standard MOPSO): ; ; Where w=0.4, , It is a random number in the range [0,1].

[0115] In the coordination process between the above modules, the exemplary definition of upstream and downstream connection error is as follows: Let the upstream module be... The output material quantity (or energy) is Downstream modules The input requirement is Then the connection error The calculation is as follows: ; When multiple materials or energy flows exist, take the root mean square of the error for each component: Where L is the number of material / energy flow types. A smaller value indicates a better match between upstream and downstream modules. In iterative optimization, when all adjacent modules... The iteration stops when all values ​​are less than 1% of the preset threshold.

[0116] The number of iterations is set to 200, and the population size is 50.

[0117] S7. Dynamic updating of carbon footprint assessment results; A dynamic update mechanism combining scheduled updates and event triggers is established to ensure the real-time nature and accuracy of the carbon footprint assessment results. The time interval for scheduled updates... It is set to automatically collect real-time production and energy consumption data from the previous day every day at midnight, 24 hours a day. Event-triggered updates are triggered immediately when events that may affect carbon emissions occur, such as equipment maintenance, raw material replacement, process adjustments, or personnel changes.

[0118] The complete dynamic update process is as follows: First, re-execute the evidence reasoning rule fusion in S2 using the newly acquired data to update the carbon emission fusion value, uncertainty variance, and evidence weights in the process carbon footprint primitive library. Second, recalculate all parameters of the carbon footprint transfer network topology in S3 using the updated process carbon footprint primitive library. Third, re-execute the Bayesian hierarchical change point detection in S4 using the updated carbon footprint transfer network and new historical data to update the prior probability of change point occurrence. If a new abnormal carbon emission process is discovered, the Monte Carlo Shapley value decomposition graph attention network source tracing in step S5 is automatically executed, and the parameters of the graph attention network are updated using an incremental learning method without retraining the entire model; the multi-objective optimization in step S6 is automatically executed to generate a new carbon footprint optimization scheme.

[0119] The incremental learning strategy employs the Elastic Weight Consolidation (EWC) method to update the graph attention network parameters. EWC calculates the importance of parameters to the old task (diagonal elements of the Fisher information matrix) and applies a secondary penalty to important parameters during updates to avoid catastrophic forgetting. Specifically, the loss function is: ,in For the prediction loss on new data, For the old model parameters, For Fisher information, diagonal element is the regularization coefficient. Each time an event update is triggered, parameter adaptation can be achieved by training on the new dataset for only 5-10 epochs, without requiring a complete retraining.

[0120] The system generates standardized carbon footprint assessment reports, which include five parts: total carbon emissions of the production system, the carbon emission percentage of each process, analysis of processes with abnormal carbon emissions, analysis of key influencing factors, and recommendations for carbon footprint optimization. The reports can be exported in both PDF and Excel formats, providing data support for enterprises' carbon management decisions, carbon disclosure, and carbon trading.

[0121] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0122] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0123] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0124] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0125] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0126] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

Claims

1. A method for intelligent evaluation of carbon footprint based on a modular process algorithm mechanism, characterized in that, Includes the following steps: S1. Deconstruct the target production system layer by layer at multiple granularities, standardize and encode the smallest unit obtained from the deconstruction, extract the core meta-features of each encoded unit, and construct the process meta-feature encoding matrix. S2. Collect corresponding multi-source heterogeneous carbon emission data based on the process element feature encoding matrix, construct an evidence reasoning rule fusion model to fuse the multi-source heterogeneous data, and generate a process carbon footprint element library; S3. Construct a carbon footprint transfer network topology based on the process element feature coding matrix, with coding units as nodes. Combined with the process carbon footprint primitive library, calculate the carbon footprint input, carbon footprint output, and carbon transfer coupling degree between each pair of upstream and downstream nodes, and identify strongly coupled node pairs. S4. Extract the time series of carbon footprint input for each node based on the topology of the carbon footprint transmission network, construct a Bayesian hierarchical change point detection model, set the model prior parameters based on the process carbon footprint primitive library, perform change point detection on the carbon footprint input time series, and identify abnormal carbon emission processes. S5. Starting with the abnormal carbon emission process, reverse source tracing reasoning is performed along the carbon footprint transmission network topology to construct a Monte Carlo Shapley value decomposition graph attention network model, quantify the contribution of upstream nodes to abnormal carbon emissions, and identify key influencing nodes. S6. Based on key influencing nodes and strongly coupled node pairs, the topology of the carbon footprint transfer network is divided into several optimization modules. Decision variables and optimization constraints are clearly defined. An improved multi-objective particle swarm optimization algorithm is used to solve the problem. The optimal solution is selected through fuzzy decision-making, and a set of carbon footprint optimization schemes is generated.

2. The intelligent carbon footprint evaluation method based on process modularization algorithm mechanism according to claim 1, characterized in that, S1 uses a multi-granularity, layer-by-layer deconstruction approach to deconstruct the entire process from three progressive granularities: process layer, step layer, and operation layer. The process layer is divided according to the key nodes where the physical form of the product undergoes irreversible changes. The step layer is divided according to the switching nodes where a single piece of equipment completes an independent action cycle. The operation layer is divided according to the individual worker completing an indivisible action unit.

3. The intelligent carbon footprint evaluation method based on process modularization algorithm mechanism according to claim 1, characterized in that, The multi-source heterogeneous carbon emission data in S2 includes three categories: direct carbon emission data, on-site indirect carbon emission data, and baseline carbon emission data. The construction of the evidence reasoning rule fusion model is as follows: a recognition framework is defined that includes true carbon emission values, higher carbon emission values, and lower carbon emission values. For each data source and each coding unit, a basic probability allocation function is generated using the relative deviation method. Initial evidence weights are set according to the data source type. The weighted basic probability allocation function and the unallocated probability are calculated. The normalized evidence reasoning rule combination operator is used to combine all weighted basic probability allocation functions one after another. The carbon emission fusion value and carbon emission uncertainty variance of each coding unit are calculated.

4. The intelligent carbon footprint evaluation method based on process modularization algorithm mechanism according to claim 1, characterized in that, The construction of the carbon footprint transfer network topology in S3 is as follows: Based on the process element feature coding matrix, the material flow, energy flow and information flow transfer relationships between each coding unit are determined, and a directed weighted graph is constructed as the carbon footprint transfer network topology. In the graph, the nodes correspond to each coding unit, the directed edges correspond to the material flow transfer relationships between coding units, and the weight of the edge corresponds to the implicit carbon emissions transferred from the upstream coding unit to the downstream coding unit.

5. The intelligent carbon footprint evaluation method based on process modularization algorithm mechanism according to claim 1, characterized in that, In S3, the carbon footprint input of each node is calculated as the sum of the implicit carbon emissions transmitted by all upstream nodes and the combined carbon emissions of that node itself. The carbon footprint output of each node is calculated using the mass allocation method, based on the total material output, total material input, and auxiliary material added during the production process of the coding unit. The carbon transfer coupling degree between each pair of upstream and downstream nodes is calculated as follows: the implicit carbon emissions transferred from the upstream node to the downstream node are divided by the total upstream carbon emissions transferred to the downstream node. When the downstream node has no upstream carbon emissions, the carbon transfer coupling degree is 0. Node pairs with a carbon transfer coupling degree greater than 0.7 are classified as strongly coupled node pairs, and node pairs with a carbon transfer coupling degree less than 0.3 are classified as weakly coupled node pairs. A strongly coupled node association table is established to record all strongly coupled node pairs.

6. The intelligent carbon footprint evaluation method based on process modularization algorithm mechanism according to claim 1, characterized in that, The Bayesian hierarchical change point detection model constructed in S4 consists of three layers: observation layer, change point layer, and parameter layer. The prior parameters of the model are set based on the process carbon footprint primitive library as follows: the parameter layer adopts the normal inverse gamma distribution as the prior distribution, and the scale parameter of the prior distribution adopts the weighted average method to integrate the historical variance of the carbon footprint input time series and the carbon emission uncertainty variance in the process carbon footprint primitive library. When the number of historical production cycles is greater than or equal to 20, the historical variance is assigned a weight of 0.7 and the carbon emission uncertainty variance is assigned a weight of 0.

3. When the number of historical production cycles is less than 20, the carbon emission uncertainty variance is directly used as the prior scale parameter.

7. The intelligent carbon footprint evaluation method based on process modularization algorithm mechanism according to claim 1, characterized in that, The specific method for detecting change points in the carbon footprint input time series in S4 is as follows: the Markov chain Monte Carlo method is used to estimate the parameters of the model to obtain the posterior probability of each production cycle as a change point. Production cycles with a posterior probability greater than 0.95 are marked as change point cycles. The carbon emission anomaly degree of the corresponding node for each change point cycle is calculated. Nodes with a carbon emission anomaly degree greater than 3 are marked as abnormal carbon emission processes.

8. The intelligent carbon footprint evaluation method based on process modularization algorithm mechanism according to claim 1, characterized in that, The Monte Carlo Shapley value decomposition graph attention network model built in S5 includes a graph attention layer and a Monte Carlo Shapley value decomposition layer. The graph attention layer introduces a node dropout strategy during training. In each iteration, a preset proportion of node features are randomly set to zero to simulate the case of subset input, enabling the model to learn the predictive ability when some nodes are missing.

9. The intelligent carbon footprint evaluation method based on process modularization algorithm mechanism according to claim 1, characterized in that, In S5, the specific steps for quantifying the contribution of upstream nodes to abnormal carbon emissions are as follows: Set the number of Monte Carlo samplings to a preset number of samplings, randomly generate an upstream node arrangement for each sampling, calculate the marginal contribution of each node under that arrangement, take the average of all the marginal contributions obtained from the sampling to obtain the Monte Carlo Shapley value of the node, and mark the upstream nodes whose Monte Carlo Shapley values ​​are greater than the preset contribution threshold as key impact nodes.

10. The intelligent carbon footprint evaluation method based on process modularization algorithm mechanism according to claim 1, characterized in that, In S6, the optimization modules are divided according to the following rules: each optimization module contains a key influencing node and all its direct upstream and downstream nodes; all nodes that have a strong coupling relationship with the key influencing node, regardless of whether they are direct upstream or downstream, are included in the same optimization module; if there is a strong coupling relationship between two key influencing nodes, then these two key influencing nodes and their related nodes are merged into one optimization module.

Images