Carbon ripple dynamic heterogeneous graph printing quality regulation method and system

Abstract

The application provides a carbon emission ripple dynamic heterogeneous graph printing quality regulation method and system, and belongs to the technical field of printing quality control and carbon emission management. The method comprises the following steps: collecting multi-source heterogeneous data in real time, time aligning and causally marking, forming quality carbon event data flow; constructing a dynamic heterogeneous graph containing quality nodes, process nodes and carbon emission nodes, configuring regulation edges, energy consumption dependent edges and confrontation edges in the graph; generating a candidate regulation path when detecting quality defects, accumulating and calculating carbon emission ripple cost along the path; combining the quality carbon emission constraint relationship carried by the confrontation edge to prune the candidate path, and outputting the recommended quality carbon emission collaborative regulation scheme. The application integrates quality repair decision and carbon emission increment prediction into the same graph model for collaborative quantification, realizes the dynamic balance of quality target and carbon emission target, continuously optimizes the decision through online feedback learning, and effectively reduces the additional carbon emission caused by quality correction.

Description

Technical Field

[0001] This invention relates to the fields of printing quality control and carbon emission management, and more specifically, to a method and system for controlling printing quality using a dynamic heterogeneous graph of carbon emission ripples. Background Technology

[0002] In current green printing practices, online correction of quality defects and carbon emission control in the production process are typically handled as two separate systems. When operators or automated control systems detect quality defects such as color difference or uneven density, they often adjust parameters solely from the perspective of process optimization, neglecting the cascading energy consumption and carbon emission impacts of these adjustments on subsequent drying, power, and other processes. This control model, lacking quantitative linkage and collaborative decision-making mechanisms, easily leads to excessive environmental costs in ensuring print quality or sacrificing product quality to meet carbon emission targets, making it difficult to achieve a dynamic balance and global optimization between the two. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, the present invention provides a method for controlling the printing quality of carbon emission ripple dynamic heterogeneous maps, the method comprising: Real-time acquisition of multi-source heterogeneous data during the printing production process, including at least visual quality data of printed materials, equipment energy consumption data, and process parameter data, and time-series alignment and causal labeling of the multi-source heterogeneous data to form a correlated carbon event data stream; Based on the carbon emission event data stream, a dynamic heterogeneous graph is constructed. The nodes in the dynamic heterogeneous graph include quality nodes representing the quality status of printed materials, process nodes representing adjustable process variables, and carbon emission nodes representing carbon emission intensity. The edges in the dynamic heterogeneous graph include at least: control edges connecting the process nodes and the quality nodes, energy consumption dependency edges connecting the process nodes and the carbon emission nodes, and antagonistic edges connecting the quality nodes and the carbon emission nodes to quantify the contradiction between quality compliance requirements and carbon emissions. In response to a detected quality defect event, multiple candidate control paths capable of correcting the defect are determined in the dynamic heterogeneous graph; along the energy consumption attachment edge along each candidate control path, the energy consumption increment and carbon emission factor of each process node on the path are cumulatively calculated to obtain the carbon emission ripple cost of the candidate control path, wherein the carbon emission ripple cost represents the increase in carbon emissions caused by performing control along the path. Based on the carbon emission ripple cost and the quality-carbon emission constraints carried by the adversarial edges connecting quality nodes and carbon emission nodes in the dynamic heterogeneous graph, the control edges on the candidate control paths are pruned, including: obtaining the expected quality repair amount for each candidate control path; determining a preset tolerance threshold based on the quality-carbon emission constraints; using the ratio of the carbon emission ripple cost to the expected quality repair amount as the matching degree; removing control edges whose matching degree exceeds the preset tolerance threshold; and determining and outputting a recommended quality-carbon emission collaborative control scheme from the control schemes corresponding to the remaining control edges.

[0004] Furthermore, the present invention also provides a carbon emission ripple dynamic heterogeneous pattern printing quality control system, which is used to store and execute the above-described method.

[0005] Based on the above aspects, compared with the prior art, the present invention has the following beneficial technical effects: First, this invention fundamentally changes the traditional disconnect between printing quality control and carbon emission management. By innovatively introducing adversarial edges into a dynamic heterogeneous graph, the quality acceptance standard is quantified into a specific carbon emission constraint upper limit. Based on this, the matching degree of carbon emission ripple cost for each candidate control path is evaluated and pruned. This ensures that every quality repair decision naturally carries quantifiable carbon cost considerations, achieving real-time collaboration and game balance between quality goals and carbon emission goals within the same framework. This effectively avoids the dilemma of blindly sacrificing carbon emissions in pursuit of quality or reducing product qualification rate due to carbon emission restrictions.

[0006] Secondly, this invention transforms the attribution of quality defects and the prediction of control effects, which were originally based solely on operator experience, into a traceable and quantifiable data-driven model by constructing a quality-carbon causal rule base and a causal confidence weighting calculation mechanism. This provides a solid data foundation and reasoning basis for subsequent path search, cost calculation, and solution recommendation, significantly reducing over-adjustment and repeated trial and error caused by human misjudgment or reliance on experience, and curbing unnecessary carbon emission increases at the source.

[0007] Furthermore, this invention possesses powerful online learning and self-evolution capabilities. By monitoring the actual intervention behaviors of operators and the actual effects of adjustments, the system can continuously perform reverse calibration of the dynamic heterogeneous graph, adaptive correction of tolerance thresholds, and growth and verification of virtual process nodes. This closed-loop feedback mechanism allows the efficient control experience implicit in the operator group to be systematically captured, verified, and solidified into part of the graph model, enabling the system's decision-making capabilities to continuously evolve during the production process, and its quality and carbon emission synergy efficiency to continuously improve under long-term operation.

[0008] Finally, this invention also introduces forward-looking mechanisms such as avalanche critical state early warning and cross-production line knowledge transfer, which respectively solve the risk avoidance problem of the production system during the period of accelerated efficiency deterioration and the model cold start problem in the early stage of new production line commissioning, ensuring the robustness and applicability of the entire technical solution at different production line life cycle stages. By comprehensively applying the above technical means, while maintaining the print quality pass rate, an average of 15% to 25% of additional carbon emissions caused by quality defect correction can be reduced, significantly improving the greening and intelligence level of the printing manufacturing process. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the virtual process node credibility verification and progressive activation process in this invention. Figure 3 A schematic diagram illustrating the system provided in this embodiment. Detailed Implementation

[0010] like Figures 1-3 As shown.

[0011] This technical solution provides a method and system for controlling the printing quality of carbon emission ripple dynamic heterogeneous images. The technical details are explained in detail below using a specific application scenario.

[0012] This embodiment applies the method of the present invention to a typical book and magazine printing production line. The production line, in sequence according to the printing process, includes: a four-color offset printing press for color printing on paper; a hot air drying unit for accelerating ink drying and curing; and an automatic optical inspection unit installed at the dryer's outlet for real-time acquisition of visual quality data of the printed materials. A slitting and binding unit is also provided at the end of the production line. Smart meters and sensors are installed on key equipment to collect energy consumption and process parameter data. Taking the production of a batch of high-end brochures as an example, this batch requires extremely high color fidelity. The core indicator in the customer's quality acceptance standard is the color difference tolerance threshold, i.e., the ΔE value in the CIELAB color space must not exceed 2.5. Any quality defect will lead to the downgrading or scrapping of the entire batch of products, while any downtime or parameter adjustments to correct defects will further increase energy consumption and carbon emissions. The goal of this embodiment is to achieve optimal synergistic control of quality and carbon emissions in this dilemma.

[0013] Based on the above scenario, a complete control cycle involving this method consists of four stages: data acquisition and preprocessing, dynamic heterogeneous graph construction, candidate path generation and carbon emission ripple cost calculation, and pruning and scheme recommendation.

[0014] During the data acquisition and preprocessing stage, the system collects multi-source heterogeneous data in real time during the printing production process. Multi-source heterogeneous data is a general concept; in this embodiment, it specifically refers to data from three different types of sensors, with different data structures and sampling frequencies. The first type is visual quality data of the printed matter, acquired by a high-resolution linear CCD camera on an automated optical inspection device. The data format is an uncompressed BMP image sequence, acquired at 10 frames per second, with each frame corresponding to a printed sample. The second type is equipment energy consumption data, acquired by three-phase smart meters installed at the power input of the offset printing press and drying equipment. The data format is a structured JSON message, acquired once per second, and includes fields such as voltage, current, instantaneous power, and cumulative power consumption. The third type is process parameter data, provided by the equipment's own programmable logic controller. The data format is a semi-structured key-value pair log, sampled at a frequency of 50 times per second, including parameters such as the offset printing press's ink key opening, dampening solution supply, printing pressure, and the drying equipment's operating temperature and hot air velocity.

[0015] These three types of data are not synchronized in time, and their sources and formats are quite different. Therefore, time-series alignment and causal labeling are necessary to form a correlated carbon quality event data stream. In this invention, time-series alignment specifically refers to mapping the timestamps of the three types of data uniformly to the main clock system of the production line, and using the timestamp of the equipment energy consumption data with the slowest sampling frequency as a reference point to downsample or interpolate the high-sampling-frequency visual quality data and process parameter data. For example, window mean or center value extraction is performed on the image sequence and process parameters at one time point per second, so that the three form a neat time-series matrix.

[0016] Causal labeling is a core step in constructing a carbon emission event data stream. It's not simply data annotation, but rather a process of discovering and constructing causal event pairs in a time series based on a pre-built carbon-quality causal rule base. The specific construction and causal labeling process of this rule base is as follows: First, from the aligned visual quality data, image processing algorithms extract the trigger timestamps, defect type identifiers, and affected quality node identifiers of quality defect events. For example, at timestamp T1, the image processing algorithm detects that the cyan component density of the image is lower than the standard value, determining a density deficiency defect and triggering a quality defect event; the affected quality node identifier is Cyan_Density_Zone3. Second, from the equipment energy consumption data, the peak timestamps of carbon emission events and the corresponding carbon emission node identifiers are extracted. Carbon emission events are not simply energy consumption pulses, but high-emission periods bound to specific process states. For example, if the drying equipment experiences a power spike lasting several seconds at timestamp T2, forming a carbon emission event, the corresponding carbon emission node identifier is Dryer_Peak.

[0017] Finally, the execution timestamp, adjustment parameter type, adjustment amount, and corresponding process node identifier are extracted from the process parameter data for the process adjustment event. For example, at timestamp T3, the cyan ink key opening of the offset printing press is manually or automatically adjusted from 0.35 mm to 0.40 mm, forming a process adjustment event with the process node identifier Cyan_Ink_Key_Adjustment.

[0018] After extracting the above three types of events, the system uses the process adjustment event as a causal bridge to establish causal relationships. Centering on the timestamp T3 of the process adjustment event Cyan_Ink_Key_Adjustment, it traces backwards within a preset causal time window, such as 5 seconds, to check for the existence of a quality defect event; it then traces forwards for 5 seconds to check for the existence of a carbon emission event. If, within the time window, a quality defect event Cyan_Density_Zone3 is found forwards, and a carbon emission event Dryer_Peak is found backwards, these three events are packaged into a triplet-structured quality-carbon causal event pair, with the structure: quality defect - process adjustment - carbon emission peak. To quantify the strength of this causal relationship, the system assigns a causal confidence weight to each quality-carbon causal event pair. The comprehensive calculation formula for this weight is: W_causal = exp(-α * ΔT_quality) * exp(-β * ΔT_carbon) * S_rule; Where ΔT_quality is the time difference between a quality defect event and a process adjustment event, and ΔT_carbon is the time difference between a carbon emission event and a process adjustment event, in seconds. α and β are attenuation coefficients, both taken as empirical values ​​of 0.2 in this embodiment. S_rule is the strength coefficient of the mapping relationship retrieved from the quality-carbon causal rule base. This coefficient is a value between 0 and 1, pre-determined by expert experience and historical statistics. For example, an increase in ink bond opening leads to an increase in density, which in turn leads to an increase in drying power. This mapping relationship has high determinism, and its strength coefficient S_rule is set to 0.95. The calculated causal confidence weights are encapsulated together with the triplet to form a basic data unit of the quality-carbon event data stream. This set of weighted event pairs constitutes the sole input basis for all subsequent analysis and decisions.

[0019] Based on this data stream, the system constructs a dynamic heterogeneous graph. The dynamic heterogeneous graph is a graph data structure whose dynamism is reflected in the fact that whenever new carbon event data flows in, the attributes of the corresponding nodes and edges in the graph are updated, and even nodes and edges are automatically added when a new pattern is detected.

[0020] In this embodiment, the nodes in the graph are finely divided into three categories. Quality nodes represent the digital quality status of the printed material. For example, a specific quality node is QE_201_Cyan_Density, whose attributes include the current measured density value, the target density value range of 1.45 to 1.55, and the customer-required color difference tolerance threshold ΔE≤2.5. Process nodes represent directly adjustable physical or chemical process variables. For example, a specific process node is PE_301_InkKey_Cyan_Zone3, whose attributes include the current opening value, adjustable range, and energy efficiency conversion factor. The energy efficiency conversion factor refers to the change in energy consumption caused by each unit adjustment of this process node, expressed in kilowatt-hours per unit adjustment. Carbon emission nodes represent the intensity of carbon emissions from different sources or types. For example, a specific carbon emission node is CE_501_Dryer_Energy, whose attributes include the real-time carbon emission factor obtained from the power grid, expressed in kilograms of carbon dioxide per kilowatt-hour.

[0021] The edges in the graph define the interaction relationships between nodes, including at least three types. The edge connecting the process node and the quality node is called the control edge. It is a directed edge, with the direction from the process node to the quality node, indicating that adjustments to the process will affect the quality index.

[0022] For example, a control edge pointing from PE_301_InkKey_Cyan_Zone3 to QE_201_Cyan_Density has an edge weight representing the influence coefficient of that process parameter on quality. An edge connecting a process node and a carbon emission node is called an energy-dependent edge, which is also a directed edge, pointing from the process node to the carbon emission node. This indicates that the execution of the process consumes energy and is dependent on the corresponding carbon emission source.

[0023] For example, an energy consumption-dependent edge pointing from PE_301_InkKey_Cyan_Zone3 to CE_501_Dryer_Energy indicates that the adjustment of the inkkey changes the ink layer thickness, thus affecting the energy consumption and carbon emissions required for subsequent drying. The edge connecting the quality node and the carbon emission node is called an adversarial edge, a special type of undirected constraint edge. Adversarial edges are used to quantify the contradiction between quality compliance requirements and carbon emissions, carrying a key quality-carbon emission constraint relationship. In this embodiment, the specific form of this constraint relationship is: starting from the color difference tolerance threshold ΔE≤2.5 of the quality node QE_201_Cyan_Density, it is mapped to a carbon emission constraint upper limit, that is, the upper limit of the carbon emission increase brought about by the control measures acceptable for this quality defect point, for example, 1.5 kg of carbon dioxide. The adversarial edge internally encapsulates this mapping logic and the upper limit value.

[0024] When the system detects a new quality defect event, such as the measured density value of the QE_201_Cyan_Density node deviating from the target range, it triggers the generation process of candidate control paths. The system uses this quality node as the endpoint and traces back along the control edge to all process nodes that can affect it, forming several control paths. Each path starts at one or more process nodes, connects to a carbon emission node via an energy consumption dependency edge, and ends at the defective quality node. A complete pathway of "process node - control edge - quality node" plus "process node - energy consumption dependency edge - carbon emission node" is defined as a candidate control path. Assume the system generates path A: increasing the ink bond opening, and path B: increasing the drying temperature, two candidate control paths.

[0025] Next, the system calculates the carbon ripple cost for each path. The carbon ripple cost is a core indicator defined in this invention, referring to the predicted increase in carbon emissions across the entire chain caused by the control operation performed along the candidate control path. Taking path A as an example, its cumulative calculation method is as follows: First, obtain the energy efficiency conversion coefficient of process node PE_301 on path A, for example, 0.05 kWh per unit adjustment, with a required adjustment of 0.05 mm, then the predicted energy consumption increase is 0.0025 kWh. Then, obtain the real-time carbon emission factor of the carbon emission node CE_501 connected to it, for example, 0.6 kg CO2 per kWh. Then, the single-node carbon emission increase caused by this process node is 0.0025 * 0.6 = 0.0015 kg CO2. If path A involves only one process node, then the carbon ripple cost is 0.0015 kg CO2. For complex paths involving multiple process nodes, the product of each process node and its corresponding carbon emission node on the path needs to be summed node by node to obtain the total cost. Simultaneously, the system also needs to calculate the expected mass restoration amount for each path. The expected mass restoration amount is a predicted value, derived based on the historical influence coefficient of the control edge; for example, the expected density restoration amount for path A is 0.05.

[0026] After obtaining the carbon emission ripple cost and the expected mass restoration amount, the system enters the pruning and scheme recommendation stage. The system prunes using the mass carbon emission constraints carried by the adversarial edges. First, a matching degree is defined as the ratio of the carbon emission ripple cost to the expected mass restoration amount, essentially representing the carbon emission cost per unit of mass restoration. For path A, the matching degree is 0.0015 kg of carbon dioxide per 0.01 density increment, and for path B, it is 0.005 kg of carbon dioxide per 0.01 density increment. The system obtains a preset tolerance threshold, which is the maximum acceptable carbon emission cost per unit of restoration. In this embodiment, it is dynamically calculated based on the upper limit of the carbon emission constraints of the adversarial edges; assuming the currently calculated tolerance threshold is 0.003 kg of carbon dioxide per 0.01 density increment. The system traverses all control edges on candidate control paths, removing control edges with a matching degree exceeding the preset tolerance threshold. Path B's matching degree of 0.005 exceeds the threshold of 0.003; therefore, the control edge corresponding to path B—increasing the drying temperature—is pruned and removed.

[0027] Finally, from the remaining control schemes corresponding to the control edges, a recommended quality carbon emission coordinated control scheme is determined and output. In this embodiment, path A becomes the only acceptable scheme. The system will generate a control instruction containing the process node identifier PE_301, the adjustment type as ink key opening, and the adjustment amount as +0.05 mm. This instruction is pushed to the operator through the human-machine interface or directly issued to the programmable logic controller of the offset printing machine for execution.

[0028] A better alternative to the above pruning process is to introduce Pareto optimality and user preference. Specifically, a two-dimensional evaluation space is constructed with carbon emission ripple cost as the first coordinate axis and expected quality restoration amount as the second coordinate axis, mapping all candidate paths that pass the initial screening through the tolerance threshold into this space. Within this space, a set of candidate control paths constituting the Pareto front is identified. Points constituting the Pareto front have the following characteristic: for any path in the set, there is no other path that simultaneously offers a lower carbon emission ripple cost and a higher expected quality restoration amount. The system then obtains an externally input carbon emission tolerance parameter, representing the maximum carbon emission increment the user is willing to tolerate for each unit of color difference restoration, for example, set by the user to 0.002 kg of carbon dioxide per 0.01 density increment. Based on this parameter, the system determines the optimal equilibrium point on the Pareto front. The optimal equilibrium point is the point on the front curve where its marginal carbon emission cost equals the user-input tolerance parameter. If the user's tolerance parameter is extremely high, the optimal balance point may be the point with the largest amount of repair on the frontier; if the tolerance parameter is extremely low, the optimal balance point may be the point with the smallest carbon emission cost. The control scheme corresponding to the optimal balance point is output as the final recommended scheme, achieving the best match between user preferences and the characteristics of the objective physical system.

[0029] Based on the above embodiments, the construction process of the dynamic heterogeneous graph and the calculation of the carbon emission ripple cost of the candidate control path are described in more detail.

[0030] The construction of a dynamic heterogeneous graph is not a one-time event, but a continuous evolutionary process. During the initialization of the dynamic heterogeneous graph, the system configures differentiated initial weight parameters and response functions for each type of edge according to the domain rules of the printing process. The response function defines how an edge calculates and outputs a response value when it receives an input signal. For example, for a control edge, its response function can be configured as a linear mapping function or a sigmoid function with saturation characteristics.

[0031] When using a linear mapping function, assuming a control edge connecting the inking bond aperture process node and the ink density quality node has an initial weight of 0.8, it means that for every 0.01 mm adjustment of the inking bond aperture, the ink density value is expected to change by 0.008. If a sigmoid function is used, the response will tend to flatten near the limit adjustment range of the process parameters, preventing unexpected nonlinear distortions caused by over-adjustment. This is more in line with the actual law that the adjustment effect decreases when the printing process approaches its physical limits. For energy consumption dependent edges, their response function is usually configured as a product operation, that is, the input is the energy efficiency conversion coefficient of the process node, and the response output is the product of this coefficient and the real-time carbon emission factor of the connected carbon emission node.

[0032] The construction process of the adversarial edge is the most unique, as it embodies the quantitative contradiction between quality compliance requirements and carbon emission constraints. This process unfolds in two aspects. First, the system extracts the current print quality acceptance standard from the carbon event data stream. In this application scenario, this standard is a color difference tolerance threshold ΔE of no more than 2.5 for the brochure. The system then maps this color difference tolerance threshold to the upper limit of the carbon emission constraint embodied in the adversarial edge.

[0033] The mapping rules are defined by a preset rule base. For example, for every ΔE unit increase in color difference requirement, the allowable increase in carbon emissions is reduced by a certain percentage. In this embodiment, a color difference tolerance of ΔE of 2.5 maps to a carbon emission constraint limit of 1.5 kg of carbon dioxide per printed sheet. This means that any control operation performed to repair quality defects, if resulting in an increase in carbon emissions exceeding 1.5 kg of carbon dioxide, will be considered unacceptable. On the other hand, the system acquires data from historical printing batches and constructs a historical mapping matrix of control effectiveness and carbon emission costs. This matrix is ​​a two-dimensional query structure, where the row index is a discretized range of expected quality repair amounts, and the column index is the corresponding statistical value of the actual increase in carbon emissions.

[0034] For example, a historical record shows that increasing the ink bond opening corrected a density deviation of 0.05, ultimately resulting in an actual carbon emission increase of 0.0015 kg of CO2. This record is stored in a matrix with row keys representing density corrections of 0.04 to 0.06 and column keys representing carbon emission increases of 0.001 to 0.002, with values ​​representing the frequency of the record and the average carbon emission increase. The system uses this carbon emission constraint upper limit and the historical mapping matrix of the carbon emission cost of regulation effectiveness as static attributes of the adversarial edge, injecting it into the dynamic heterogeneous graph during initialization. After the adversarial edge is constructed, a complete quality carbon emission constraint relationship is established. This relationship is not a simple threshold, but a joint evaluation system composed of the carbon emission constraint upper limit as a hard constraint and the historical mapping matrix as a soft reference.

[0035] In one run of this embodiment, the system detected a quality defect event of insufficient cyan density. The dynamic heterogeneity graph was configured to generate two candidate control paths in response to this event: path A is to increase the ink bond opening, and path B is to increase the drying temperature. The system then determines the carbon emission ripple cost and expected quality restoration amount for each candidate control path. For path A, the expected quality restoration amount is a density increase of 0.05, and the carbon emission ripple cost is 0.0015 kg of carbon dioxide; for path B, the expected quality restoration amount is an equivalent density increase of 0.04, but the carbon emission ripple cost is 0.005 kg of carbon dioxide.

[0036] To determine whether the carbon emission costs of these paths are acceptable, the system performs a joint comparison and calculates the tension value. The tension value is a quantitative indicator that measures the degree of tension in the quality-carbon emission conflict of the candidate control paths, and its calculation process is as follows: First, based on the expected quality restoration amount of 0.05 for path A, the system looks up the corresponding reference carbon emission increment from the historical mapping matrix of control effectiveness carbon emission costs, and the lookup result is 0.0018 kg of carbon dioxide. Next, the system jointly compares the carbon emission ripple cost of path A (0.0015) with the reference carbon emission increment (0.0018) and the upper limit of the carbon emission constraint 1.5 carried by the counter-edge.

[0037] The comparison logic is as follows: when the carbon emission ripple cost is simultaneously less than both the upper limit of carbon emission constraints and a preset upward floating multiple of the reference carbon emission increment, such as 1.2 times (i.e., 0.00216), it is considered acceptable, and the tension value is 0; otherwise, it is considered unacceptable, and the tension value is 1. The carbon emission ripple cost of path A, 0.0015, is less than 0.00216 and much less than 1.5, therefore its tension value is 0. The carbon emission ripple cost of path B, 0.005, although still less than the upper limit of carbon emission constraints (1.5), is greater than 1.2 times its reference carbon emission increment, and is therefore considered too high, with a tension value of 1. When the tension value indicates that the carbon emission ripple cost is unacceptable (i.e., a value of 1), the control edge on this candidate control path is considered a control edge where the carbon emission ripple cost exceeds a preset tolerance threshold. In subsequent pruning operations, the control edge of path B that increases the drying temperature is removed. The control edge of path A that increases the ink bond opening is retained and enters the subsequent scheme recommendation stage.

[0038] In calculating the carbon emission ripple cost of candidate control paths, this invention also introduces a forward-looking risk assessment mechanism to identify and address the control-carbon emission avalanche critical state. The control-carbon emission avalanche critical state refers to a systemic risk state where, when multiple stages of a production system simultaneously enter a period of accelerated efficiency decline, even minor quality improvement controls may trigger a nonlinear surge in carbon emissions.

[0039] Specifically, while accumulating energy consumption and carbon emission factors along candidate control paths, the system continuously monitors the changing trends of the energy efficiency conversion coefficients at each process node along the path. The energy efficiency conversion coefficient is an efficiency indicator of how a process node converts energy consumption into effective process output; its unit is the amount of adjustment achievable per kilowatt-hour of process adjustment. This coefficient is not constant and will slowly drift with equipment aging and changes in operating conditions. The system maintains a time series for each process node with a preset time window length of 120 minutes, storing the average energy efficiency conversion coefficient per minute.

[0040] The system calculates the second derivative of the rate of decline for the time series in real time using a sliding window. The first derivative of the rate of decline represents the speed at which the coefficient decreases, while the second derivative indicates whether the rate of decline is accelerating or decelerating. When it detects that multiple process nodes on the same candidate control path have continuously decreasing energy efficiency conversion coefficients within a preset time window, and the second derivative of their rate of decline is positive, the system determines that the candidate control path has entered the control-carbon emission avalanche critical state. A positive second derivative means that the rate of coefficient decline is accelerating, which is a typical signal of accelerated system degradation. For example, in a certain operation of this production line, the energy efficiency conversion coefficients of the offset printing machine ink key motor and the hot air circulating fan of the drying equipment on path A both showed an accelerated decline within 120 minutes due to aging of the motor windings and dust accumulation on the fan impeller, respectively. The second derivative detection showed a positive value, and the system immediately determined that path A had entered the avalanche critical state.

[0041] For candidate control paths at the critical state of regulation-carbon emission avalanche, the system generates a critical state warning flag independent of the carbon emission ripple cost. This warning flag is a flag appended to the path data structure, which sends a clear signal to the decision-making system: if this path is adopted, even if the currently calculated carbon emission ripple cost is within an acceptable range, there is an extremely high risk of triggering a nonlinear surge in actual carbon emissions, and the actual carbon emission cost may far exceed the predicted value. In subsequent pruning operations, this warning flag has a veto power.

[0042] When performing matching degree comparison and pruning, the system directly treats control edges on candidate control paths carrying critical state warning markers as control edges whose carbon emission ripple cost exceeds a preset tolerance threshold and prioritizes their removal, eliminating the need for detailed calculations such as matching degree. This mechanism ensures a safe boundary for control decisions, preventing the worst-case scenario of both quality and carbon emissions spiraling out of control due to improper decision-making when the production system is in a vulnerable period. For example, in the scenario above, although the carbon emission ripple cost and matching degree of path A were originally within an acceptable range, this path was prioritized for removal because it carried a critical state warning marker. The system then prompts operators to perform preventative maintenance, such as checking the motor windings or cleaning the fan impeller, fundamentally eliminating the risk of avalanche before proceeding with quality defect repairs.

[0043] After pruning, there may be more than one control scheme corresponding to the remaining control edges. The core of subsequent decision-making lies in determining the final recommended control scheme that best balances the needs for quality restoration and carbon emission constraints among multiple feasible options.

[0044] In this embodiment, the expected quality repair amount is specified as the expected color difference repair amount, which is the predicted reduction in CIELAB color difference value ΔE brought about by the control operation. This predicted value is jointly determined by the historical influence coefficient of the control edge and the proposed adjustment amount of the process parameters. For example, for a candidate control path that increases the ink key opening by 0.05 mm, the system calculates, based on the historical influence coefficient of its control edge, that the color difference value of the target area is expected to be repaired from the current 3.8 to 1.9, that is, the expected color difference repair amount is a reduction of ΔE by 1.9.

[0045] The system acquires two core indicators for each remaining control edge: the carbon ripple cost and the expected color difference repair amount for its candidate control path. The carbon ripple cost is the predicted increase in carbon emissions along the entire chain caused by the control operation along the path, expressed in kilograms of carbon dioxide. The expected color difference repair amount is the quantified value of the quality improvement that the path can bring, expressed in ΔE.

[0046] To intuitively evaluate and compare the comprehensive effectiveness of various candidate control pathways, the system constructs a two-dimensional evaluation space based on carbon emission and carbon emissions. This space uses the carbon emission ripple cost as the first coordinate axis (horizontal axis), with units of kilograms of carbon dioxide; and the expected color difference repair amount as the second coordinate axis (vertical axis), with units of ΔE. Each remaining candidate control pathway is mapped to a data point in this two-dimensional space based on its calculated carbon emission ripple cost and expected color difference repair amount. For example, the coordinate of pathway A is 0.0015 kilograms of carbon dioxide, ΔE 1.9; the coordinate of pathway C is 0.003 kilograms of carbon dioxide, ΔE 2.3; and the coordinate of pathway D is 0.002 kilograms of carbon dioxide, ΔE 2.1.

[0047] Within this evaluation space, the system identifies a set of candidate regulatory paths that constitute the Pareto front. The Pareto front is defined as follows: for any path in this set, there is no other path that simultaneously offers a lower carbon emission ripple cost and a higher expected color difference repair. In other words, points on the Pareto front cannot be surpassed by other points in either the cost or benefit dimensions. Taking the three paths mentioned above as examples, path A has the lowest carbon emission cost but also the smallest repair amount; path C has the largest repair amount but the highest cost; and path D is inferior to path A in cost but inferior to path C in repair amount. If there is no point that is smaller than path D on the x-axis while larger than path D on the y-axis, then path D is also on the front. The system filters out all such non-dominated paths through a comparison-by-comparison method, forming the Pareto front path set.

[0048] Pareto fronts alone cannot automatically select a single solution from multiple non-dominated options because the points on the front represent different trade-offs. Path A favors low carbon emissions, while path C favors high-quality remediation. To make the final decision, the system introduces a user preference mechanism.

[0049] The system acquires externally input carbon emission tolerance parameters. These parameters are user-defined values ​​representing the maximum carbon emission increase the user is willing to tolerate for each unit of color difference repair, expressed in kilograms of carbon dioxide per ΔE. In this application scenario, this parameter is input by the workshop director or process engineer through the human-machine interface of the production management system. For example, for this batch of high-end brochure orders, the user has set a relatively strict carbon emission tolerance parameter of 0.0012 kilograms of carbon dioxide per ΔE, meaning that for every ΔE of color difference repaired, the maximum allowed carbon emission increase is 0.0012 kilograms of carbon dioxide.

[0050] It is important to clarify that the carbon emission tolerance parameter and the preset tolerance threshold used in the aforementioned pruning operation belong to the same conceptual category in this invention, but operate at different decision-making stages. The preset tolerance threshold is a rigid safety boundary automatically generated by the system internally based on the upper limit of carbon emission constraints carried by the adversarial edge and the historical mapping matrix, for rapid coarse screening. Its unit is also kilograms of carbon dioxide per unit of repair. The carbon emission tolerance parameter is a business preference boundary directly input by the user. The two can be identical in value or differ. In a preferred embodiment, the preset tolerance threshold is pre-filled as a system suggestion value on the human-machine interface. The user can manually adjust it to form the final carbon emission tolerance parameter based on the profit margin of the actual order or the customer's rigid requirements for carbon footprint. The system uses the user-confirmed carbon emission tolerance parameter simultaneously for determining the Pareto optimal equilibrium point and updating the preset tolerance threshold for the next round of coarse screening, thus unifying the standards for coarse screening and fine-tuning.

[0051] The system determines the optimal equilibrium point on the Pareto front based on carbon emission tolerance parameters. The logic for determining the optimal equilibrium point is as follows: on the Pareto front, find candidate control paths that maximize the expected color difference repair amount while satisfying the carbon emission tolerance constraint. Specifically, for each point on the Pareto front, calculate its matching degree, which equals the carbon emission ripple cost divided by the expected color difference repair amount, i.e., the horizontal axis divided by the vertical axis. This matching degree is essentially the carbon emission cost per unit mass of repair.

[0052] The calculated matching degree is compared with the carbon emission tolerance parameter input by the user. If the matching degree at a point is less than or equal to the carbon emission tolerance parameter, it means that the unit repair cost of that path is within the user's acceptable range. From all points that meet the conditions, the system selects the one with the largest expected color difference repair amount as the optimal balance point. If path A has the largest repair amount among the constraints, it is determined as the optimal balance point. The corresponding control scheme for this point, increasing the ink key opening by 0.05 mm, will be the recommended quality-carbon emission coordinated control scheme, output to the operator through the human-machine interface or directly issued to the offset printing machine controller for execution.

[0053] After outputting the recommended solution, the system does not end its task but enters a cycle of continuous monitoring and closed-loop feedback. During subsequent production, the system continues to monitor changes in process parameters. These changes are identified by comparing the real-time collected process parameter data stream with the expected values ​​of the recommended solution. When the system detects that the actual control operations performed by the operator are inconsistent with the recommended quality-carbon emission coordinated control solution, it initiates a special recording and learning process.

[0054] The inconsistency is determined when the deviation between the type and amount of adjustment parameters in the actual process adjustment event and those in the recommended scheme exceeds a preset dead zone. For example, if the recommended scheme requires increasing the indenter key opening by 0.05 mm, but the operator manually inputs an increase of 0.08 mm, or the operator increases the dampening solution supply instead of adjusting the indenter key, both are considered inconsistencies. The system records the control edge corresponding to this actual control operation as a manually reused control edge. The term "manually reused" means that the control edge originally existed in the system, but this time the operator manually selected and executed parameters different from the recommended scheme. This is equivalent to the operator reusing a control path that may not be the system's preferred choice based on their own experience and judgment.

[0055] The system then obtains the actual color difference correction amount and the actual carbon emission increase generated after the manually reused control edge is actually executed. The actual color difference correction amount is the change in color difference value detected and calculated by automatic optical inspection equipment in the next printing cycle after the operation is executed. The actual carbon emission increase is the change in carbon emissions accumulated and calculated by smart meters within the same time window.

[0056] The system compares these actual values ​​with historical records to determine whether the manual operation has potential value superior to the system's recommendation. Specifically, the system generates a reactivation trigger signal when two conditions are met simultaneously. Condition one: The actual color difference repair amount reaches the preset quality repair standard. The preset quality repair standard is a configurable benchmark, defined in this embodiment as reaching more than 90% of the color difference repair amount expected by the recommended scheme, meaning that although the operation deviates from the recommended value, the repair effect is still acceptable. For example, if the recommended scheme expects to repair ΔE 1.9, and the manual operation actually repairs ΔE 1.8, reaching 94.7%, the condition is met. Condition two: The actual carbon emission increment is lower than the average carbon emission ripple cost of the manually reused control edge in the historical record. The average carbon emission ripple cost in the historical record refers to the exponentially weighted moving average of the carbon emission ripple costs recorded when the control edge was adopted or evaluated by the system in the past, used to smooth historical fluctuations and emphasize recent performance. If the actual carbon emission increment this time is 0.0012 kg of carbon dioxide, and the historical average is 0.0015 kg of carbon dioxide, then condition two is also met.

[0057] When both conditions are met, a resurrection trigger signal is generated. In response to this signal, the system removes the pruning status of the manually reused control edge in the dynamic heterogeneous graph, restoring it to a usable state. This means that if the edge was previously marked for removal during pruning, this mark will be cleared, and in subsequent defect events, candidate control paths containing that edge will be re-evaluated.

[0058] Simultaneously, the system calculates a tolerance threshold correction based on the difference between the actual increase in carbon emissions and the historical average ripple cost of carbon emissions. The formula for calculating the tolerance threshold correction is as follows: ΔThreshold = learning_rate * (avg_cost - actual_cost) / avg_cost *base_threshold; Where avg_cost is the historical average carbon emission ripple cost of this edge, actual_cost is the actual carbon emission increment in this instance, base_threshold is the local tolerance threshold used by the current set of control edges of this type, and learning_rate is a learning rate coefficient, which is 0.1 in this embodiment. The calculated ΔThreshold is a positive value, indicating that the threshold should be adjusted in the direction of relaxing constraints. The system applies this tolerance threshold adjustment only to the pruning determination of the control edge set that belongs to the same process node type and the same quality defect type as the manually reused control edge. For example, if the manual reuse involves a process node of ink bond opening type and a quality defect of insufficient density type, then the adjustment amount only applies to all control edges connecting ink bond type process nodes and density type quality nodes. By subtracting this adjustment amount from the local tolerance threshold of these edges, the local tolerance threshold corresponding to the control edge set is shifted in the direction of relaxing pruning constraints, that is, allowing a slightly higher unit repair carbon emission cost to be tolerated in subsequent decisions. This reflects the system's adaptive learning from the efficient experience of operators.

[0059] Based on the closed-loop feedback learning mechanism, the system further introduces the ability of graph structure evolution and knowledge transfer to address the challenges of long-term operation and rapid deployment in actual production.

[0060] First, the autonomous growth mechanism of dynamic heterogeneous graphs is explained. When a manually reused control edge is revived, the system does not simply restore its original state, but initiates a long-term statistical and pattern recognition process to capture new control knowledge contained in the collective wisdom of the operators that has not yet been explicitly represented by the model.

[0061] The system accumulates records of the quality defect type, process node type, and actual control operation parameters corresponding to each manually reused control edge. These records are persistently stored in a dedicated operation event log database. Each record includes a timestamp, batch number, quality defect type identifier (e.g., Density_Low_Cyan), process node type identifier (e.g., InkKey_Cyan_Adjustment), and the specific adjustment amount of the operation parameter (e.g., ±0.08 mm).

[0062] Based on these accumulated records, the system continuously monitors the frequency of occurrence of specific patterns. When a manually reused control edge of the same quality defect type and the same process node type is detected, and its frequency exceeds a frequency threshold within a preset statistical period, the system determines that the combination of the quality defect type and the process node type constitutes a frequent manual correction pattern. The preset statistical period and frequency threshold are two configurable parameters. In this embodiment, the statistical period is set to the most recent 100 printing batches, and the frequency threshold is set to 5 times. This means that if an operator manually adjusts the cyan inking bond opening to repair a quality defect of low cyan density, and this action occurs 6 times in the most recent 100 batches, then this action is determined to be a frequent manual correction pattern. The settings of 100 batches and 5 times represent a trade-off between statistical significance and response sensitivity. Experiments show that when the statistical period is too short or the frequency threshold is too low, occasional operational fluctuations may be misjudged as patterns, causing the system to frequently generate meaningless virtual nodes. Conversely, if the period is too long or the threshold is too high, the system will be too slow to respond to new patterns and may miss the valuable window for solidifying experience.

[0063] In response to the determination result of the frequent manual correction mode, the system generates a node growth trigger signal. This signal triggers the structural evolution of the dynamic heterogeneity graph. In response to this node growth trigger signal, the system creates a virtual process node in the dynamic heterogeneity graph. The virtual process node does not correspond one-to-one with a specific physical actuator on the production line, but rather is an abstraction and solidification of a class of high-frequency manual control behaviors. The type of this virtual process node is a new node type added to the existing process node types in the dynamic heterogeneity graph. For example, the system can create a process node identified as PE_Virtual_Macro_CyanFix, which does not simply represent the ink bond aperture or dampening liquid volume, but rather represents a combined adjustment strategy or an operating tool for cyan density defects.

[0064] While creating the virtual process node, the system establishes connections between it and other nodes in the graph. A new control edge is created between this virtual process node and the quality node corresponding to the quality defect type involved in the frequent manual correction mode. For example, a new control edge pointing from PE_Virtual_Macro_CyanFix to the quality node QE_201_Cyan_Density. Simultaneously, a new energy consumption dependency edge is created between this virtual process node and the carbon emission node corresponding to the process node type involved in the frequent manual correction mode. For example, a new energy consumption dependency edge pointing from PE_Virtual_Macro_CyanFix to the carbon emission node CE_501_Dryer_Energy. In this way, the virtual process node and its associated edges are fully integrated into the structure of the dynamic heterogeneous graph.

[0065] From its creation, this virtual process node serves as a new starting point for candidate control paths, participating in the generation of all subsequent quality-carbon emission coordinated control schemes. Its energy efficiency conversion coefficient needs initialization. The initialization method is as follows: Calculate the weighted average of the energy efficiency conversion coefficients of the original process nodes involved in all historical control operations during the frequent manual correction mode that triggered the creation of this virtual process node. Here, the weight of the weighted average is set as the proportion of each historical control operation's frequency to the total frequency; the higher the frequency of the operation, the greater the proportion of its original process node's energy efficiency conversion coefficient in the initialization calculation. For example, if this frequent mode includes 6 operations, of which 4 primarily adjust the ink keys with an energy efficiency conversion coefficient of 0.05 kWh per unit adjustment, and 2 primarily adjust the dampening solution with an energy efficiency conversion coefficient of 0.01 kWh per unit adjustment, then the initial energy efficiency conversion coefficient of the virtual node is calculated as follows: (4*0.05+2*0.01) / 6=0.0367 kWh per unit adjustment. This virtual process node and its energy efficiency conversion coefficient are seamlessly integrated into the previously mentioned monitoring range of the control-carbon emission avalanche critical state after generation. The system will also maintain a 120-minute sliding window to monitor the changing trend of its energy efficiency conversion coefficient in real time, ensuring that the control path it participates in is also constrained by avalanche risk.

[0066] The following section details how, during the data preprocessing stage, multi-source heterogeneous data are aligned temporally and labeled causally to form a correlated carbon event data stream. This process forms the data foundation for all subsequent intelligent decisions.

[0067] The processing of multi-source heterogeneous data streams begins with the extraction of event elements. The system extracts the trigger timestamp, defect type identifier, and affected quality node identifier from the visual quality data of printed materials. For example, when the image processing algorithm identifies a deviation of the cyan density value from the standard value in the 3rd frame of a high-resolution image sequence at 14:30:15.200 on October 26, 2023, a quality defect event is triggered. Its defect type identifier is Density_Low_Cyan, and the affected quality node identifier is QE_201_Cyan_Density. Simultaneously, the system extracts the peak timestamp, peak power, and corresponding carbon emission node identifier from the equipment energy consumption data. For example, at 14:30:18, the smart meter records a 3-second power spike in the drying equipment, with a peak value of 15.2 kilowatts. This information is packaged as a carbon emission event, with the corresponding carbon emission node identifier being CE_501_Dryer_Energy. Meanwhile, the system extracts the execution timestamp, adjustment parameter type, adjustment amount, and the corresponding process node identifier from the process parameter data for process adjustment events. For example, at 14:30:17, the programmable logic controller log displays an instruction to adjust the cyan ink key opening from 0.35 mm to 0.40 mm. This record is extracted as a process adjustment event, with the process node identifier PE_301_InkKey_Cyan_Zone3.

[0068] After extracting these discrete events, the system uses process adjustment events as a causal bridge and, based on a pre-defined quality-carbon causal rule base, establishes a causal relationship between quality defect events and carbon emission events. The quality-carbon causal rule base is a core knowledge base of this invention. This rule base stores two types of key mapping relationships. The first type is the quality-effect mapping relationship, which maps various process adjustment events to the types of quality defects they can cause. For example, a positive adjustment of the ink bond aperture can lead to increased density, and an increase in dampening solution supply can lead to dot gain. The second type is the carbon-effect mapping relationship, which maps various process adjustment events to the types of carbon emission peaks they can cause. For example, an increase in the ink bond aperture leads to a thicker ink layer, which in turn triggers a power peak in the subsequent drying process; an increase in drying temperature directly leads to a power peak in the drying process itself. Each mapping relationship is accompanied by an intensity coefficient, a real value between 0 and 1, determined by domain experts based on the degree of certainty of the physical mechanism and the conditional probability calculated from historical data, used to quantify the reliability of the causal relationship.

[0069] The specific process of causal association involves a combination of pattern matching and time window constraints. The system uses the extracted execution timestamp of the process adjustment event, i.e., 14:30:17, as the center, and traces back a preset causal time window (set to 10 seconds in this embodiment) to find time-proximity quality defect events. The logic of forward tracing is based on the fact that operators or automatic control systems typically observe quality defects before performing adjustment operations. Within the 10-second window, the system successfully traced back to the Density_Low_Cyan quality defect event occurring at 14:30:15:200. Next, the system traces backward to a preset causal time window of the same 10 seconds to find carbon emission events caused by the process adjustment. The logic of backward tracing is based on the fact that the impact of process adjustments on energy consumption must occur after the adjustment action is performed. Within the backward 10-second window, the system successfully traced back to the drying power spike event CE_501_Dryer_Energy occurring at 14:30:18. Thus, these three events are linked temporally and logically, forming a quality-carbon causal event pair with a "quality defect - process adjustment - carbon emission peak" ternary structure. This event pair structure can be represented as: Quality defect: Density_Low_Cyan at T1; Process adjustment: InkKey_Cyan_Zone3 +0.05mm at T3; Carbon emission peak: Dryer_Energy_Peak at T2; Finally, the system assigns causal confidence weights to each formed mass-carbon causal event pair. This quantifies the credibility of the causal relationship. The weight is calculated by comprehensively considering the time interval and the rule base strength coefficient. The specific calculation formula is as follows:

[0070] in, This is the strength coefficient of the corresponding mapping relationship retrieved from the mass-carbon causal rule base. In this example, the strength coefficient of the entire causal chain where an increase in ink bond opening leads to an increase in density, which in turn leads to an increase in drying power. Set as .

[0071] Let the time of occurrence of the quality defect event be... Seconds, the time when the process adjustment event occurs is Seconds, the time when the carbon emission event occurred was Seconds. The mass time difference and carbon emission time difference are respectively:

[0072]

[0073] Time decay coefficient is taken (Its value makes the product in the exponential term...) (Dimensionless, here we directly substitute the experimentally given values). Substitute into the formula for calculating the causal confidence weight:

[0074] This causal confidence weight of 0.718, as metadata, is encapsulated along with the event pair itself within the triplet structure, forming a data record in the associated carbon-mass event data stream. This data stream serves as the information source for subsequent dynamic heterogeneous graph construction and online incremental updates, continuously driving the operation and evolution of the entire system.

[0075] After the system outputs the recommended quality-carbon emission coordinated control scheme and it is executed by the operator or control system, its workflow does not end. Instead, it enters a continuous online learning and model calibration phase. The core purpose of this phase is to use the actual execution results of each control operation to provide feedback correction to the key prediction parameters in the dynamic heterogeneity diagram, so as to continuously improve the accuracy of the model's prediction of future carbon emission ripple costs.

[0076] Once the entire printing production process is completed, in this embodiment, after the last sheet of the high-end brochure has dried and been removed from the production line, the system initiates a complete model calibration process. First, the system obtains the actual total carbon emissions generated during the production of this batch, based on the recommended quality-carbon emission coordinated control scheme. The actual total carbon emissions are obtained by tracing back the cumulative carbon emission readings of the smart meters within a complete production cycle after the execution of the control command. For example, from the moment the command to increase the ink key opening is issued until the entire process of drying and cutting the adjusted sheet is completed, if the smart meter records a cumulative carbon emission of 0.0018 kg of carbon dioxide from the drying equipment and 0.0003 kg of carbon dioxide from the offset printing machine, then the actual total carbon emissions from this control operation are the sum of the two, i.e., 0.0021 kg of carbon dioxide.

[0077] Simultaneously, the system retrieves the predicted value of the carbon ripple cost recorded during the decision-making process for the recommended route. In this embodiment, the predicted value is the carbon ripple cost of path A, which is 0.0015 kg of carbon dioxide, calculated previously.

[0078] Next, the system calculates the prediction deviation between the actual total carbon emissions and the predicted value of the carbon ripple cost. The prediction deviation is calculated using a simple difference: the actual value minus the predicted value. In this example, it's 0.0021 minus 0.0015, which equals 0.0006 kg of carbon dioxide. This positive deviation indicates that the model overestimated the carbon emission efficiency of the control path, meaning the actual carbon emissions are higher than predicted. The system compares the absolute value of this prediction deviation with a preset calibration trigger threshold. The calibration trigger threshold is a sensitivity control parameter used to prevent frequent model oscillations due to small random fluctuations. In this embodiment, the threshold is set to 20% of the predicted value, i.e., 0.0015 multiplied by 20% equals 0.0003 kg of carbon dioxide. Since the absolute value of the deviation, 0.0006, is greater than the trigger threshold of 0.0003, the system determines that the model's prediction accuracy has significantly drifted and immediately generates a model calibration trigger signal.

[0079] In response to the model calibration trigger signal, the system performs a reverse allocation and parameter correction. The goal of the reverse allocation is to reasonably decompose and attribute the global prediction deviation to each process node on the candidate control path that constitutes the recommended scheme, and to accurately correct the core prediction parameters of each node.

[0080] The reverse allocation process strictly adheres to the contribution ratio of each process node in this control. The contribution ratio is calculated as follows: for each process node on the candidate control path, the product of the energy efficiency conversion coefficient of that process node and the corresponding process parameter adjustment amount in that candidate control path is calculated; then, the ratio of this product to the sum of the similar products of all process nodes on the candidate control path is calculated. In path A of this embodiment, the control path is relatively simple, involving only one process node, namely PE_301_InkKey_Cyan_Zone3. The energy efficiency conversion coefficient of this node is 0.05 kWh per unit adjustment, and the adjustment amount is 0.05 mm, so the product is 0.05 * 0.05 = 0.0025. Since there is only one process node on the path, its contribution ratio is 100%. If a path involves multiple process nodes, such as both ink key adjustment and drying temperature adjustment, then the respective products need to be calculated separately and summed, and then the percentage of each node's product needs to be calculated.

[0081] Based on the contribution ratio, the system decomposes the global prediction deviation into node-level calibration errors for each process node. In this embodiment, the prediction deviation of 0.0006 kg of carbon dioxide multiplied by the contribution ratio of 100% yields a node-level calibration error of 0.0006 kg of carbon dioxide for node PE_301. This means that the energy efficiency conversion coefficient of this node is overestimated, and its true value should be higher than the current value, i.e., lower efficiency.

[0082] Based on the node-level calibration error, the system corrects the energy efficiency conversion coefficient for that process node. The goal of the correction is to make the carbon ripple cost recalculated using the corrected energy efficiency conversion coefficient approach the actual total carbon emissions. Specifically, "approaching" in this invention means that the deviation between the recalculated predicted value and the actual value is reduced to below the calibration trigger threshold. The specific correction process is as follows: the system first establishes a conversion relationship between the node-level calibration error and the correction amount of the energy efficiency conversion coefficient.

[0083] The aforementioned model calibration mechanism ensures the system's continuous evolution capability on mature production lines. However, when a printing production line is put into operation for the first time, and sufficient local historical control data has not yet been accumulated in the dynamic heterogeneous graph, all the mechanisms that rely on historical data, such as historical mapping matrices, control edge weights, and energy efficiency conversion coefficients, will face a cold start problem. Sufficient local historical control data is a clearly defined criterion in this invention. This criterion is quantified as follows: the estimated variance of the control edge weights in the local dynamic heterogeneous graph is lower than a preset convergence threshold, such as 0.01. When the variance is higher than this threshold, the system determines that the local data is insufficient to support independent decision-making and instead initiates a cross-production line knowledge transfer mechanism.

[0084] First, the system acquires the source dynamic heterogeneity diagram of at least one source printing production line. The source printing production line is a mature production line that has been running and has undergone long-term feedback calibration; the parameters in its dynamic heterogeneity diagram have become stable and accurate. In this embodiment, source production line B is an old production line that produces similar brochures; the dynamic heterogeneity diagram running on it has been calibrated through thousands of batches and is a highly reliable source model.

[0085] Next, the system extracts a general control knowledge subgraph from the source dynamic heterogeneous graph, which is independent of equipment specifications. This general control knowledge subgraph is an abstract knowledge graph stripped of specific equipment physical parameters; its extraction is a decoupling operation. Specifically, the system retains quality node types independent of equipment specifications in the source dynamic heterogeneous graph, such as density defects and color difference defects; process node types, such as ink bond opening adjustment and dampening solution adjustment (abstract types, not specific motor models); and carbon emission node types, such as drying energy consumption and power consumption. Simultaneously, the system retains the topological structure of control edge types, energy consumption dependent edge types, and antagonistic edge types between node types, as well as their interconnections. For example, the subgraph of source production line B retains abstract relationships such as "ink bond opening adjustment type nodes affect density defect type nodes" and "ink bond opening adjustment type nodes depend on drying energy consumption type nodes," as well as constraint relationships such as "there are antagonistic edges between density defect type nodes and drying energy consumption type nodes." However, all specific numerical parameters, such as the initial weight of a certain control edge being 0.8, and the energy efficiency conversion coefficient of a certain energy consumption dependent edge being 0.05, are stripped away, leaving only the parameterized positions.

[0086] Then, the system acquires the target equipment characteristic parameters of the printing production line in this printing production process. The target equipment characteristic parameters are quantitative indicators describing the characteristics of each physical piece of equipment in the local production line, including at least: the rated power of each printing unit, for example, the offset printing machine in the local production line is 15 kW, while the source production line B is 12 kW; the energy efficiency level of the drying unit, for example, the local production line is level 1 energy efficiency, while the source production line B is level 2 energy efficiency; the carbon emission factor of the ink type, for example, the local production line uses plant-based ink with a carbon emission factor of 0.3 kg CO2 per kilogram of ink, while the source production line B uses mineral-based ink with a carbon emission factor of 0.5 kg CO2 per kilogram of ink.

[0087] Based on these target equipment feature parameters, the system constructs a domain adaptation layer. The domain adaptation layer is a computational module configured to complete cross-production line parameter mapping. Its core function is to map the node types and edge types in the general control knowledge subgraph to the corresponding node types and edge types in the target dynamic heterogeneous graph that adapts to the feature parameters of the target equipment in the local production line, and to perform inter-domain scaling on the weight parameters of each edge type.

[0088] The specific method for inter-domain scaling is to calculate a scaling factor based on the ratio of the specifications of the source and target production lines on similar equipment, and then apply it to the corresponding edge weights. For example, for the initial weight of the energy consumption-dependent edge, i.e., the initial value of the energy efficiency conversion coefficient, the scaling factor is defined as the ratio of the rated power of the drying unit of the target production line to the rated power of the drying unit of the source production line. In this embodiment, the rated power of the target offset printing machine is 15 kW, and the rated power of the source offset printing machine is 12 kW, so the ratio is 15 / 12 = 1.25. If the parameterized position of the ink key adjustment energy consumption-dependent edge in the general knowledge subgraph extracted from source production line B needs to be filled with a base value, and this base value originates from the calibration value of 0.05 of source production line B, then the initial energy efficiency conversion coefficient after inter-domain scaling is 0.05 multiplied by 1.25, which equals 0.0625 kWh per unit adjustment. For example, the carbon emission factor scaling factor is defined as the ratio of the target ink carbon emission factor to the source ink carbon emission factor. 0.3 divided by 0.5 equals 0.6, so the basic values ​​for calculating carbon emissions related to ink are all scaled down by this ratio.

[0089] The general control knowledge subgraph, mapped and scaled by the domain adaptation layer, is used as the initial skeleton of the local production line dynamic heterogeneity graph. Only after the initial skeleton is built does the system start real-time data acquisition and incrementally update the dynamic heterogeneity graph with the incoming carbon event data stream.

[0090] During incremental updates, the system implements a gradual migration strategy rather than immediately discarding source knowledge. The core of this strategy is dynamically adjusting the weighted contribution ratio of migrated edges to local edges in regulatory decisions. This gradual transition mechanism ensures a smooth and seamless transition from relying on source production line experience to relying on local production line data, guaranteeing the robustness and accuracy of the model throughout its lifecycle.

[0091] Building upon the aforementioned dynamic heterogeneous graph self-growth mechanism, this embodiment further introduces a credibility verification and progressive weight activation process for newly generated virtual process nodes. The core purpose of this process is to prevent occasional, non-universally applicable manual operation patterns from being prematurely solidified into graph structures, thereby contaminating the decision-making model and leading to a decline in the quality and carbon emission efficiency of subsequent automatically recommended solutions.

[0092] When the system triggers node growth due to the detection of frequent manual correction modes and creates a virtual process node in the dynamic heterogeneous graph, it does not immediately grant it the same decision-making participation rights as the original normal nodes. Instead, the system assigns an initial reliability score to the virtual process node. The initial reliability score is a comprehensive rating that quantifies whether the control mode represented by the new node is reliable and effective, with a score range of 0 to 100. This score is calculated based on the historical control operation records in the set of frequent manual correction modes that triggered the creation of the virtual process node.

[0093] The specific calculation rule is as follows: the initial credibility score equals the weighted value of the actual color difference repair success rate score, plus the weighted value of the carbon emission increment deviation score. The actual color difference repair success rate refers to the proportion of times, after manual adjustment by the operator, the subsequent online detection equipment determines that the color difference repair meets the preset quality repair standard out of the total number of operations. This proportion is multiplied by 100 to obtain the base value of the success rate score. The carbon emission increment deviation refers to the average percentage deviation between the actual carbon emission increment generated by each manual adjustment in this mode and the historical average carbon emission ripple cost of that adjustment edge. The lower the deviation, the better the predictability and stability of carbon emissions, and the higher the corresponding score. After mapping the deviation to a score from 0 to 100, it is weighted and summed with the success rate score.

[0094] In this embodiment, the success rate is weighted at 0.7, and the carbon emission deviation is weighted at 0.3. For example, a virtual node has a set of 6 historical operations with a success rate of 5 divided by 6, which is approximately 83.3%, resulting in a success rate score of 83.3. Its average carbon emission increment deviation is 15%, corresponding to a deviation score of 85. Therefore, the initial credibility score of this virtual node is calculated as 83.3*0.7 + 85*0.3 = 58.31, plus 25.5, resulting in a score of 83.81.

[0095] The virtual process node and its newly added control edges and energy consumption dependent edges are immediately put into use after creation, but their state is marked as a candidate state. A candidate state is a restricted transient state, distinct from the normal state of the original node. The impact of the candidate state on the decision-making process is reflected on two levels. During the generation of candidate control paths, candidate state control edges can be normally included in the path search, which means that control schemes originating from this virtual node can be explored and evaluated by the system.

[0096] However, in subsequent pruning operations, the system applies a stricter tolerance threshold bias to candidate control paths that pass through candidate state control edges than to normal control edges. The tolerance threshold bias is an offset added to a preset tolerance threshold, making the pruning criteria more stringent for candidate paths. Specifically, after calculating the matching degree of a path containing a candidate state control edge, its matching degree is compared with a preset tolerance threshold plus a positive bias value, rather than with the original threshold. In this embodiment, this bias value is set to 20% of the original tolerance threshold. For example, if the original tolerance threshold is 0.003 kg of carbon dioxide per unit of remediation, then the effective threshold after applying the bias is 0.003 multiplied by 1.2, which equals 0.0036. This means that the unit carbon emission cost of a candidate state path must be about 20% lower than that of a normal path to pass pruning, thus making it more likely to be removed in the pruning decision, thereby suppressing the impact of unverified emerging patterns on stable decisions.

[0097] The system sets a preset verification period for candidate nodes. This verification period is not simply a time length, but rather a data accumulation window driven by the actual number of adjustments. Within the verification period, the system continuously monitors the actual adjustments performed by the candidate node's adjustment edge. These actual adjustments include two scenarios: first, the system ultimately recommends the candidate path after evaluation and it is adopted by the operator; second, the operator manually selects an operation consistent with the adjustment mode represented by the candidate node. After each adjustment, the system records the actual color difference repair amount and the actual carbon emission increment.

[0098] When the cumulative number of actual adjustments within the verification period reaches the preset minimum verification sample size, the system triggers a credibility update and final state determination for the virtual process node. The minimum verification sample size is a preset statistic, set to 10 times in this embodiment. This value is set because experiments show that when the sample size is less than 10 times, the calculated success rate and deviation indicators fluctuate too much, and the credibility interval is too wide, which is insufficient to make a reliable activation or abandonment decision.

[0099] Once the cumulative execution count reaches 10, the system calculates an updated credibility score based on the data from these 10 actual executions, using the same formula as the initial credibility score: a success rate score weighted at 0.7, plus a carbon emission deviation score weighted at 0.3. The system then compares this updated credibility score with a preset activation threshold. The activation threshold determines whether a candidate node can transition to a normal state; in this embodiment, it is set to 75 points. This value is based on extensive offline simulation experiments, which show that a threshold of 75 points achieves the best balance between capturing truly effective new patterns and filtering out random interference. If the threshold is too low, some unstable patterns may be mistakenly activated; if the threshold is too high, some effective improved patterns may fail to pass verification for a long time.

[0100] If the updated confidence score reaches or exceeds 75, the system will switch the virtual process node and its connected control edges and energy consumption dependent edges from the candidate state to the normal state. After the state switch, the system immediately removes the 20% tolerance threshold bias applied to the control edge during the pruning operation. The node and its associated edges will then enjoy the same decision weight as the original normal node, marking the formal acceptance and solidification of this new knowledge emerging from operational experience.

[0101] If the updated credibility score is below 75, the system determines that the virtual process node has failed verification, possibly representing only occasional operator preferences or short-term emergency behavior in a specific batch, and lacks universality. In this case, the system completely removes the virtual process node and its connected control edges and energy consumption dependent edges from the dynamic heterogeneous graph. Simultaneously, the system generates and sends a node abandonment notification to the operator's terminal, i.e., the human-machine interface or workshop management terminal. This notification is not a simple deletion prompt, but carries reason analysis data for the removal of the virtual process node. The reason analysis data includes: the actual color difference repair success rate within the verification period, the average carbon emission increment deviation, the final calculated credibility score, and a comparison of each score with the activation threshold.

[0102] For example, a notification might indicate that the virtual node PE_Virtual_Macro_CyanFix has been abandoned due to a color difference repair success rate of only 60%, a carbon emission deviation as high as 35%, and a final credibility score of 62, which is below the threshold of 75. Operators can use this information to understand why the system refuses to solidify its operating mode and proactively check and optimize their own control habits when necessary. This mechanism builds a closed-loop feedback channel between the system's autonomous learning and human-machine interaction.

[0103] In the model calibration process described above, the system continuously optimizes the model's prediction accuracy by reversing the prediction bias and correcting the energy efficiency conversion coefficients of each process node. However, in actual production environments, the physical characteristics and stability of different process nodes vary significantly. Some nodes, such as the energy consumption characteristics of large-scale drying equipment, are typically stable, with fluctuations in their energy efficiency conversion coefficients caused only by slow equipment aging. Other nodes, such as the response characteristics of ink key motors, may exhibit more frequent and larger fluctuations due to the interplay of complex factors such as ink batches and ambient temperature and humidity. Applying the same correction logic to all nodes could lead to overly slow calibration of stable nodes and overly aggressive calibration of highly volatile nodes, thus introducing prediction oscillations.

[0104] To address this discrepancy, this embodiment introduces an adaptive calibration mechanism based on the historical fluctuation characteristics of nodes, which is used to finely control the actual correction magnitude of each model calibration.

[0105] After calculating the node-level calibration error for each process node, the system does not directly use this error for coefficient updates. Instead, it first performs a step to calculate the correction step size constraint coefficient. The correction step size constraint coefficient is an adjustment factor between 0 and 1, which adaptively determines the magnitude of the correction based on the node's historical calibration performance.

[0106] The generation process of this coefficient is as follows. First, the system obtains the energy efficiency conversion coefficient correction sequence for each process node in the historical calibration record. The historical calibration record is a list stored in chronological order of the correction amount and direction generated by each valid calibration event. For a specific process node, its correction sequence may contain the following values: positive 0.01, negative 0.005, positive 0.02, positive 0.001, etc., all in kilowatt-hours per unit adjustment. The system statistically analyzes this correction sequence and calculates its fluctuation characteristic value. The fluctuation characteristic value is an indicator used to quantify the stability of the historical correction behavior of the node. In this embodiment, the standard deviation of the correction sequence is used as the fluctuation characteristic value. The larger the standard deviation, the more drastic the magnitude and direction of the historical correction of the energy efficiency conversion coefficient of the node, and the more unstable its intrinsic physical characteristics or external operating conditions. For example, the standard deviation of the historical correction sequence of the offset printing press ink key node PE_301 is 0.008, while the standard deviation of the drying equipment hot air fan node PE_502 is 0.001.

[0107] Based on this fluctuation characteristic value, the system generates correction step size constraint coefficients for each process node in this calibration. The generation function of the correction step size constraint coefficients is configured to have a negative correlation with the fluctuation characteristic value: the larger the fluctuation characteristic value, the smaller the correction step size constraint coefficient. The design logic of this inverse relationship is to apply a more conservative correction magnitude to process nodes with large historical fluctuations, avoiding the destruction of the overall stability of the model by chasing random fluctuations.

[0108] After completing this coefficient update, the calibration process does not end immediately but enters a preset calibration verification window. The calibration verification window is a well-defined verification period used to verify whether the correction truly improves the model's predictive ability. This window is not defined by absolute time but by the number of events. In this embodiment, the window is set to three consecutive printing batches in which the process node is reactivated and participates in control during subsequent production. Setting it to three batches is a balance between verification effectiveness and timely feedback. One batch might be too random, while more than five batches would result in an excessively long verification period, potentially delaying subsequent calibration decisions.

[0109] Within the calibration verification window, the system continuously tracks the performance of the node. Whenever the node is included in the control path and actually executed due to a new quality defect event, the system acquires the predicted new carbon ripple cost and the actual total carbon emissions for that production activity, and calculates the new prediction deviation. When enough new prediction deviation data is collected within the verification window, or when the window ends, the system makes a final judgment. If the absolute value of any new prediction deviation is greater than or equal to a preset calibration trigger threshold, such as 0.0003 kg of carbon dioxide, the calibration is deemed invalid, failing to achieve the expected accuracy improvement, and may even have introduced larger errors. In this case, the system performs a rollback operation, restoring the energy efficiency conversion coefficient of the process node to its original value before the correction, i.e., 0.05, completely revoking the calibration. Simultaneously, the system generates a calibration instability flag for the process node. This flag is attached to the node as a persistent attribute, serving as an additional input factor in any subsequent calibration process for that node, further reducing its correction step constraint coefficient. For example, when an instability flag exists, the calculated constraint coefficient will be multiplied by a decay factor of 0.5, making the correction of the node more cautious.

[0110] Conversely, if the absolute values ​​of all new prediction deviations within the verification window are less than the calibration trigger threshold, the calibration is deemed valid. The system permanently retains the updated energy efficiency conversion coefficient of 0.069872 and records the actual correction amount of this calibration (positive 0.019872) and its correction direction in the historical calibration record of this node. This record also updates the fluctuation characteristic value of its correction sequence, providing the latest historical fluctuation reference for the next calibration cycle. Through this closed-loop calibration mechanism with verification rollback and adaptive fluctuation suppression, the system's underlying prediction model can maintain long-term robustness and accuracy during continuous learning.

[0111] This application also provides a carbon emission ripple dynamic heterogeneous graph printing quality control system, which stores and executes the methods described in any of the foregoing embodiments. This system, as a hardware and software integrated system, is deployed in the control room of the printing production line and interconnected with various printing equipment, testing equipment, and sensing devices via industrial Ethernet, forming a complete online closed-loop control system.

[0112] The overall architecture of the system can be divided into the perception and execution layer, the data fusion layer, the graph computing engine layer, and the decision interaction layer from bottom to top. Its hardware foundation and logical functional modules together constitute the following four core components.

[0113] The first part is the data acquisition and preprocessing module. The physical foundation of this module includes sensor arrays deployed at key workstations throughout the printing production line, a data acquisition gateway, and an edge computing server. The sensor array specifically includes: a linear CCD camera and its associated image acquisition card installed inside the automated optical inspection equipment, used for high-speed capture of visual quality data of printed materials; a three-phase smart meter connected in series at the power input of major energy-consuming equipment such as the offset printing press and drying equipment, used to collect equipment energy consumption data such as voltage, current, instantaneous power, and cumulative power consumption at a frequency of 1 Hz; and a data interface module that communicates with the programmable logic controllers of each device via the OPC UA protocol, used to read process parameter data such as ink key opening, dampening solution supply, printing pressure, drying temperature, and hot air velocity in real time at a frequency of 50 Hz. The data acquisition gateway is responsible for aggregating the above multi-source heterogeneous data and transmitting it to the edge computing server via gigabit Ethernet.

[0114] The data preprocessing software running on the edge computing server is the core of the module. This software includes a time-series alignment engine and a causal labeling engine. The function of the time-series alignment engine is to map three types of data with different sampling frequencies and timestamp systems to the main clock system of the production line. Specifically, it operates by using the timestamp of the smart meter data with the slowest sampling frequency as a reference point, generating an alignment moment every second, extracting frames from high-sampling-frequency image data or averaging multiple frames within that second, and extracting the center value of the process parameter data within that second, thus forming a neat multi-dimensional time series matrix. The causal labeling engine, based on a pre-defined mass-carbon causal rule base, performs event detection and correlation on this time series matrix.

[0115] The quality-carbon causal rule base is stored as a database table on the solid-state drive of the edge server. The table stores the quality-effect mapping relationships between various process adjustment events and the types of quality defects they can cause, as well as the carbon-effect mapping relationships between various process adjustment events and the types of carbon emission peaks they can cause. Each mapping record is accompanied by an intensity coefficient. The causal tagging engine's operation logic is as follows: It scans the aligned data matrix to identify quality defect events, carbon emission peak events, and process adjustment events; using the timestamp of the process adjustment event as the center, it traces backwards through a preset causal time window (e.g., 10 seconds) to find quality defect events, and then traces backwards through the same window to find carbon emission events; it encapsulates the three events that satisfy the time correlation into a triplet structured quality-carbon causal event pair, and calculates the causal confidence weight based on the time interval between each pair of events and the rule base intensity coefficient. Finally, the output of this module is a data stream of associated quality-carbon events, which is sent to subsequent modules via the ZeroMQ message middleware in the form of a structured message queue.

[0116] The second part is the dynamic heterogeneous graph construction module. This module is implemented in the graph computing engine layer, and its physical carrier is a graph computing server equipped with a general-purpose graphics processor. The graph computing server internally maintains a dynamic heterogeneous graph instance based on a graph database management system, such as Neo4j or TigerGraph. This module is responsible for receiving carbon event data streams and using them to initialize and continuously incrementally update the dynamic heterogeneous graph.

[0117] The nodes in the dynamic heterogeneous graph are divided into three categories: quality nodes, whose attributes include node identifier, associated sheet area, target value range, current measured value, and color difference tolerance threshold; process nodes, whose attributes include node identifier, associated equipment, current setpoint, adjustable range, energy efficiency conversion coefficient, historical correction sequence of energy efficiency conversion coefficient, and calibration instability flag; and carbon emission nodes, whose attributes include node identifier, associated energy source, and real-time carbon emission factor. Edges are also divided into at least three categories: control edges, which are directed edges from process nodes to quality nodes, and whose attributes include influence coefficient and response function type; energy consumption dependent edges, which are directed edges from process nodes to carbon emission nodes, and whose attributes include energy efficiency conversion coefficient; and adversarial edges, which are undirected edges connecting quality nodes and carbon emission nodes, and whose attributes include upper limit of carbon emission constraints and a reference to the historical mapping matrix of control effectiveness and carbon emission cost.

[0118] The dynamic heterogeneous graph construction module also has a dedicated embedded domain adaptation submodule for enabling cross-production line knowledge transfer. This submodule maintains a configurable adaptation rule base. Based on the target equipment feature parameters input from the outside, it performs inter-domain scaling calculations on the edge weights of the general control knowledge subgraph extracted from the source production line dynamic heterogeneous graph, and uses the result as the initial skeleton of the local production line dynamic heterogeneous graph.

[0119] The third part is the candidate path generation and evaluation module. This module, also running on the graph computing server, is the core computing unit connecting problem discovery and solution decision-making. Responding to quality defect event signals from the data acquisition and preprocessing module, this module initiates a graph traversal and path search algorithm. It retrieves all control edges and their upstream process nodes that can affect the defective quality node from the graph database, and recursively explores carbon emission nodes connected to these process nodes along energy consumption dependency edges, thereby generating several complete candidate control paths. For each candidate control path, this module calls the energy efficiency and carbon emission calculation submodule to perform the cumulative calculation of carbon emission ripple costs, i.e., multiplying and summing the energy efficiency conversion coefficients of each process node on the path, the proposed process parameter adjustments, and the real-time carbon emission factors of the corresponding carbon emission nodes node by node. Simultaneously, this module calls the quality prediction submodule to calculate the expected quality repair amount based on the historical influence coefficients of the control edges. After calculation, this module associates and stores the carbon emission ripple costs of all candidate control paths with the corresponding control edge identifiers, forming a historical record of carbon emission ripple costs. This historical record is persisted as a time-series database for subsequent model calibration, tension value calculation, and dynamic updates of pruning tolerance thresholds. Furthermore, this module integrates avalanche critical state monitoring functionality, generating critical state warning markers for risky paths through second-order derivative analysis of the time series of energy efficiency conversion coefficients at each process node.

[0120] The fourth part is the pruning and scheme recommendation module. This module is deployed in the decision interaction layer, and its hardware foundation consists of the system's central processing server and human-machine interface terminal. This module receives all candidate control paths, their costs, repair amounts, and various labeling information output by the candidate path generation and evaluation module, and executes a two-level screening logic. The first level is hard pruning based on tension values: it reads the upper limit of carbon emission constraints and historical mapping matrix carried by the adversarial edges on each path, calculates the tension value by combining the carbon emission ripple cost of the path and the expected quality repair amount, and directly removes the paths and their control edges that indicate unacceptable tension values; at the same time, it performs priority pruning on paths carrying critical state warning labels. The second level is fine screening based on Pareto optimality and user preferences: it maps the remaining paths to the quality-carbon emission two-dimensional evaluation space, identifies the Pareto front, and determines the optimal balance point on the Pareto front based on the carbon emission tolerance parameters input by the user through the human-machine interface, and outputs the control scheme corresponding to this point as the recommended quality-carbon emission coordinated control scheme.

[0121] This module is also responsible for continuous online learning and feedback. It continuously monitors the consistency between the actual control operations and the recommended schemes. When it detects manual control edge reuse and the actual effect is better than the historical average, it triggers the revival mechanism and tolerance threshold correction. After each batch of production, it also triggers the model calibration process, which, based on the deviation between the actual total carbon emissions and the predicted values, reversely corrects the energy efficiency conversion coefficients of each process node and performs calibration verification and rollback judgment. At the same time, it counts the frequency of manual edge reuse and, when conditions are met, triggers the creation of virtual process nodes and their candidate state verification process, realizing the autonomous evolution of the dynamic heterogeneous graph structure.

[0122] The four modules mentioned above work together under the scheduling of a unified system clock and message middleware to form a closed-loop system that spans the entire chain from data perception, graph modeling, path evaluation to decision feedback, thus fully realizing the joint management and control method for carbon emissions of green printed materials as described in any of the aforementioned method embodiments.

[0123] In summary, the carbon emission ripple dynamic heterogeneous graph printing quality control method and system provided by the present invention is based on the construction of a data-driven, causally related and continuously evolving dynamic heterogeneous graph model, and on this basis, realizes closed-loop collaborative decision-making for quality repair and carbon emission control.

[0124] In practical applications, when automated optical inspection equipment detects quality defects such as color difference or density inhomogeneity, the system does not seek correction solutions in isolation. Instead, it examines each potential regulatory operation within a multi-dimensional evaluation framework. Specifically, the system traces back the potential causes of defects through causal labeling, explores forward along the regulatory and energy-dependent edges of the dynamic heterogeneity graph, calculates the carbon emission ripple cost of each feasible regulatory path, and jointly compares it with the quality carbon emission constraints carried by the antagonistic edges.

[0125] This process concretizes the abstract concept of carbon neutrality into quantifiable carbon emission increment predictions and tolerance thresholds, ensuring that every action in quality control has a clear and calculable carbon cost awareness. Subsequently, through a Pareto frontier-based decision-making mechanism, the system automatically identifies the optimal control scheme for quality restoration while meeting users' carbon emission tolerance, thus avoiding the situation at its source where carbon emission targets are blindly sacrificed in pursuit of quality, or quality standards are relaxed due to excessive restrictions on carbon emissions.

[0126] More importantly, this method and system are not a static rule engine, but an intelligent agent with online learning and autonomous evolution capabilities. The system continuously monitors the actual intervention behaviors of operators and the resulting quality and carbon emission effects, performs closed-loop feedback calibration on the model, and structurally extracts and verifies the credibility of high-performing operational experiences, ultimately solidifying them into the graph model in the form of virtual process nodes. This mechanism enables the systematic capture, quantification, and reuse of tacit knowledge from domain experts.

[0127] Meanwhile, the system's adaptive calibration of the energy efficiency conversion coefficient and monitoring of its avalanche critical state endow the production control system with forward-looking early warning and adaptability capabilities to long-term changes such as equipment aging and operating condition drift. Experiments and production practice show that, by adopting this method and system, printing production lines can reduce additional carbon emissions caused by quality corrections by an average of approximately 15% to 25% while maintaining a product qualification rate no lower than the original level, significantly improving the greening and intelligentization level of the printing manufacturing process. The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for controlling the printing quality of carbon emission ripple dynamic heterogeneous diagrams, characterized in that, The method includes: Real-time acquisition of multi-source heterogeneous data during the printing production process, including at least visual quality data of printed materials, equipment energy consumption data, and process parameter data, and time-series alignment and causal labeling of the multi-source heterogeneous data to form a correlated carbon event data stream; Based on the carbon emission event data stream, a dynamic heterogeneous graph is constructed. The nodes in the dynamic heterogeneous graph include quality nodes representing the quality status of printed materials, process nodes representing adjustable process variables, and carbon emission nodes representing carbon emission intensity. The edges in the dynamic heterogeneous graph include at least: control edges connecting the process nodes and the quality nodes, energy consumption dependency edges connecting the process nodes and the carbon emission nodes, and antagonistic edges connecting the quality nodes and the carbon emission nodes to quantify the contradiction between quality compliance requirements and carbon emissions. In response to a detected quality defect event, multiple candidate control paths capable of correcting the defect are determined in the dynamic heterogeneous graph; along the energy consumption attachment edge along each candidate control path, the energy consumption increment and carbon emission factor of each process node on the path are cumulatively calculated to obtain the carbon emission ripple cost of the candidate control path, wherein the carbon emission ripple cost represents the increase in carbon emissions caused by performing control along the path. Based on the carbon emission ripple cost and the quality-carbon emission constraints carried by the adversarial edges connecting quality nodes and carbon emission nodes in the dynamic heterogeneous graph, the control edges on the candidate control paths are pruned, including: obtaining the expected quality repair amount for each candidate control path; determining a preset tolerance threshold based on the quality-carbon emission constraints; using the ratio of the carbon emission ripple cost to the expected quality repair amount as the matching degree; removing control edges whose matching degree exceeds the preset tolerance threshold; and determining and outputting a recommended quality-carbon emission collaborative control scheme from the control schemes corresponding to the remaining control edges.

2. The method according to claim 1, characterized in that, The construction of the dynamic heterogeneous graph specifically includes: When initializing the dynamic heterogeneous graph, differentiated initial weight parameters and response functions are configured for each type of edge according to the domain rules of the printing process; The process of constructing the adversarial edge includes: Extract the current print quality acceptance standard from the carbon event data stream, and map the color difference tolerance threshold in the quality acceptance standard to the carbon emission constraint upper limit carried by the anti-edge edge. Obtain the actual color difference repair amount and actual carbon emission increment corresponding to different control paths adopted for the same quality defect event in historical printing batches, and construct a historical mapping matrix of control effectiveness and carbon emission cost. The carbon emission constraint upper limit and the historical mapping matrix of the carbon emission cost of the regulation effectiveness are used as static attributes of the adversarial edge and injected into the dynamic heterogeneous graph. The dynamic heterogeneous graph is configured as follows: Each time a quality defect event is detected, candidate control paths are generated, and the carbon emission ripple cost and expected quality repair amount of each candidate control path are determined. Based on the estimated quality restoration amount, the corresponding reference carbon emission increment is queried from the historical mapping matrix of the carbon emission cost of the regulation effectiveness; The carbon emission ripple cost is jointly compared with the reference carbon emission increment and the carbon emission constraint upper limit to determine whether the carbon emission ripple cost is acceptable, and the determination result is used as the tension value of the candidate regulation path. When the tension value indicates that the carbon emission ripple cost is unacceptable, the control edge on the candidate control path is regarded as a control edge whose carbon emission ripple cost exceeds a preset tolerance threshold and is removed in the pruning operation.

3. The method according to claim 1, characterized in that, The energy consumption dependency edges along each candidate control path are used to accumulate the energy consumption increment and carbon emission factor of each process node on the path to obtain the carbon emission ripple cost of the candidate control path, specifically including: For each candidate control path, obtain the real-time carbon emission factor of all carbon emission nodes connected to the energy consumption dependent edges on the path, and the energy efficiency conversion coefficient of the process nodes corresponding to all control edges on the path. The real-time carbon emission factor and the energy efficiency conversion coefficient are accumulated side-by-side along the candidate control path to obtain the carbon emission ripple cost of the candidate control path. The carbon emission ripple cost represents the predicted value of the whole-link carbon emission increment caused by the control operation along the path. While calculating the carbon emission ripple cost, the changing trend of the energy efficiency conversion coefficient at each of the process nodes is monitored; When it is detected that the energy efficiency conversion coefficient of multiple process nodes on the same candidate control path continuously decreases within a preset time window, and the second derivative of the rate of decrease is positive, it is determined that the candidate control path has entered the control-carbon emission avalanche critical state. For candidate control paths that are in the control-carbon emission avalanche critical state, a critical state warning flag independent of the carbon emission ripple cost is generated. The critical state warning flag is used to indicate that when the candidate control path corrects quality defects, there is a risk of triggering a nonlinear surge in carbon emissions. The step of pruning the control edges in the dynamic heterogeneous graph according to the carbon emission ripple cost further includes: treating the control edges on the candidate control paths carrying the critical state warning marker as control edges whose carbon emission ripple cost exceeds a preset tolerance threshold, and pruning them preferentially.

4. The method according to claim 3, characterized in that, The expected quality restoration amount is the expected color difference restoration amount; the step of determining and outputting a recommended quality carbon emission coordinated control scheme from the control schemes corresponding to the remaining control edges specifically includes: Obtain the carbon emission ripple cost of the candidate control path corresponding to each remaining control edge, and the expected color difference repair amount of the candidate control path; Using the carbon emission ripple cost as the first coordinate axis and the expected color difference repair amount as the second coordinate axis, a two-dimensional quality-carbon emission evaluation space is constructed, and each candidate regulation path is mapped to the two-dimensional quality-carbon emission evaluation space. In the two-dimensional quality-carbon emission assessment space, a set of candidate regulatory paths constituting the Pareto front is identified. The Pareto front is defined as: for any path in the set, there is no other path that results in a lower carbon emission ripple cost while simultaneously achieving a higher expected color difference repair amount. Obtain the carbon emission tolerance parameter input from the outside, which represents the upper limit of carbon emission increment that the user is willing to accept for a unit of color difference repair. Based on the carbon emission tolerance parameter, an optimal balance point is determined on the Pareto front. The optimal balance point is the candidate control path that maximizes the expected color difference repair amount under the carbon emission tolerance constraint on the Pareto front. The control scheme corresponding to the optimal balance point is output as the recommended quality-carbon emission coordinated control scheme.

5. The method according to claim 4, characterized in that, Following the recommended quality-carbon emission synergistic control scheme, the following is also included: Monitor changes in process parameters during the printing production process. When it is detected that the actual control operation performed by the operator is inconsistent with the recommended quality-carbon emission coordinated control scheme, the control edge corresponding to the actual control operation is recorded as a control edge that is manually reused. The actual color difference repair amount and actual carbon emission increment generated after the manually reused control edge is actually executed; When the actual color difference repair amount reaches the preset quality repair standard, and the actual carbon emission increment is lower than the average carbon emission ripple cost of the manually reused control edge in the historical record, a resurrection trigger signal is generated. In response to the resurrection trigger signal, the pruning state of the manually reused control edge in the dynamic heterogeneous graph is released, restoring it to a usable state; Simultaneously, based on the difference between the actual carbon emission increment and the average carbon emission ripple cost of the manually reused control edge in historical records, the tolerance threshold correction amount is calculated. The tolerance threshold correction is applied only to the pruning determination of the set of control edges that belong to the same process node type and the same quality defect type as the manually reused control edge, so that the local tolerance threshold corresponding to the control edge set shifts in the direction of relaxing the pruning constraints.

6. The method according to claim 5, characterized in that, After the manually reused control edge is depruned from its pruning state in the dynamic heterogeneity graph, the method further includes: The quality defect type, process node type, and actual control operation parameters corresponding to each manually reused control edge are recorded cumulatively. When the frequency of manually reused control edges of the same quality defect type and the same process node type exceeds the frequency threshold within a preset statistical period, it is determined that the combination of the quality defect type and the process node type constitutes a frequent manual correction mode. In response to the determination result of the frequent manual correction mode, a node growth trigger signal is generated; In response to the node growth trigger signal, a virtual process node is created in the dynamic heterogeneous graph. The virtual process node is a new node type in addition to the original process node types in the dynamic heterogeneous graph. A new control edge is established between the virtual process node and the quality node corresponding to the quality defect type involved in the frequent manual correction mode, and a new energy consumption dependency edge is established between the virtual process node and the carbon emission node corresponding to the process node type involved in the frequent manual correction mode. The virtual process node, the newly added control edge, and the newly added energy consumption dependent edge are used as new starting points for candidate control paths and incorporated into the subsequent generation process of quality-carbon emission coordinated control schemes. The energy efficiency conversion coefficient of the virtual process node is initialized based on the weighted average of the energy efficiency conversion coefficients of the original process nodes involved in all historical control operations during the frequent manual correction mode that triggered the creation of the virtual process node. The virtual process nodes and their energy efficiency conversion coefficients are included in the monitoring range of the carbon emission avalanche critical state after candidate control paths are generated.

7. The method according to claim 1, characterized in that, The step of performing time-series alignment and causal labeling on the multi-source heterogeneous data to form a correlated carbon event data stream specifically includes: From the visual quality data of the printed matter, extract the trigger timestamp, defect type identifier, and affected quality node identifier of the quality defect event; From the energy consumption data of the equipment, extract the peak timestamp, peak power and corresponding carbon emission node identifier of the carbon emission event; From the process parameter data, extract the execution timestamp, adjustment parameter type, and adjustment amount of the process adjustment event, as well as the process node identifier corresponding to the process adjustment event; Using the process adjustment event as a causal bridge, and based on a preset quality-carbon causal rule library, the quality defect event and the carbon emission event are causally correlated. The quality-carbon causal rule base stores: quality-effect mapping relationships between various process adjustment events and the types of quality defects they can cause, and carbon-effect mapping relationships between various process adjustment events and the types of carbon emission peaks they can cause. The causal association process includes: taking the execution timestamp of the process adjustment event as the center, tracing back the quality defect event within a preset causal time window, and tracing back the carbon emission event within a preset causal time window to form a quality-carbon causal event pair with a "quality defect-process adjustment-carbon emission peak" triplet structure; A causal confidence weight is assigned to each of the quality-carbon causal event pairs. The causal confidence weight is calculated comprehensively based on the time interval between the quality defect event and the process adjustment event, the time interval between the carbon emission event and the process adjustment event, and the strength coefficient of the corresponding mapping relationship in the quality-carbon causal rule base. The set of quality-carbon causal event pairs carrying the causal confidence weights is used as the associated quality-carbon event data stream.

8. The method according to claim 3, characterized in that, Following the recommended quality-carbon emission synergistic control scheme, the following is also included: After the printing production process is completed, the actual total carbon emissions generated by the recommended quality-carbon emission coordinated control scheme are obtained, as well as the predicted value of the carbon emission ripple cost corresponding to the recommended quality-carbon emission coordinated control scheme. Calculate the prediction deviation between the actual total carbon emissions and the predicted value of the carbon ripple cost; When the absolute value of the prediction deviation exceeds a preset calibration trigger threshold, a model calibration trigger signal is generated; In response to the model calibration trigger signal, the prediction bias is back-allocated to each of the process nodes on the candidate control path constituting the recommended mass-carbon emission synergistic control scheme, along the structure of the dynamic heterogeneity graph. The reverse allocation process includes: calculating the contribution ratio of each process node to the carbon emission ripple cost in the candidate control path. The contribution ratio is determined by the ratio of the product of the energy efficiency conversion coefficient of the process node and the corresponding process parameter adjustment amount of the process node in the candidate control path to the sum of the similar products of all process nodes in the candidate control path. The prediction deviation is decomposed into node-level calibration error of each process node according to the contribution ratio of each process node. Based on the node-level calibration error, the energy efficiency conversion coefficient of each process node is corrected so that the carbon emission ripple cost recalculated by the corrected energy efficiency conversion coefficient approaches the actual total carbon emissions.

9. The method according to claim 2, characterized in that, When the printing production line corresponding to the printing production process is in its first production run, and sufficient local historical control data has not yet been accumulated in the dynamic heterogeneity graph, the construction of the dynamic heterogeneity graph, before the real-time collection of multi-source heterogeneous data in the printing production process, further includes: Obtain a source dynamic heterogeneity map of at least one source printing production line, wherein the source dynamic heterogeneity map is a dynamic heterogeneity map that has been running on the source printing production line and has been calibrated by feedback. From the source dynamic heterogeneous graph, extract a general control knowledge subgraph that is independent of equipment specifications. The general control knowledge subgraph includes: the quality node type, process node type, carbon emission node type that is independent of equipment specifications in the source dynamic heterogeneous graph, as well as the control edge type, energy consumption dependent edge type, and antagonistic edge type between each node type. The target equipment characteristic parameters of the printing production line during the printing production process are obtained. The target equipment characteristic parameters include at least: the rated power of each printing unit, the energy efficiency level of the drying unit, and the carbon emission factor of the ink type. Based on the target device feature parameters, a domain adaptation layer is constructed; the domain adaptation layer is configured to: map each node type and edge type in the general regulation knowledge subgraph to the corresponding node type and edge type in the target dynamic heterogeneous graph that adapts to the target device feature parameters, and perform inter-domain scaling on the weight parameters of each edge type; The general regulation knowledge subgraph mapped by the domain adaptation layer is used as the initial skeleton of the dynamic heterogeneous graph; After the initial skeleton is constructed, the real-time acquisition is initiated to incrementally update the dynamic heterogeneous graph with the carbon event data stream. During the incremental update process, the contribution ratio of the migration edge weights in the general regulation knowledge subgraph to the regulation decision is gradually reduced, while the contribution ratio of the weights of the edges generated by the local incremental update is increased.

10. A carbon emission ripple dynamic heterogeneous image printing quality control system, the system being used to store and execute the method described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data in the printing production process in real time, including at least printed visual quality data, equipment energy consumption data and process parameter data, and to perform time-series alignment and causal labeling on the multi-source heterogeneous data to form a related carbon event data stream. A dynamic heterogeneous graph construction module is used to construct a dynamic heterogeneous graph based on the carbon event data stream; The nodes in the dynamic heterogeneous graph include quality nodes representing the quality status of printed materials, process nodes representing adjustable process variables, and carbon emission nodes representing carbon emission intensity. The edges in the dynamic heterogeneous graph include at least: control edges connecting the process node and the quality node, energy consumption dependence edges connecting the process node and the carbon emission node, and antagonistic edges connecting the quality node and the carbon emission node to quantify the contradiction between quality compliance requirements and carbon emissions. The candidate path generation and evaluation module is used to determine multiple candidate control paths that can correct the defect in the dynamic heterogeneous graph in response to the detected quality defect event; to accumulate and calculate the energy consumption increment and carbon emission factor of each process node along the energy consumption attachment edge of each candidate control path, so as to obtain the carbon emission ripple cost of the candidate control path, wherein the carbon emission ripple cost represents the increase in carbon emissions caused by the control along the path; and to associate and store the carbon emission ripple cost of the control edge on each candidate control path with the corresponding control edge identifier to form a carbon emission ripple cost history record. The pruning and scheme recommendation module is used to prune the control edges on the candidate control paths based on the carbon emission ripple cost and the quality-carbon emission constraint relationship carried by the adversarial edges connecting the quality nodes and carbon emission nodes in the dynamic heterogeneous graph. This includes: obtaining the expected quality repair amount for each candidate control path; determining a preset tolerance threshold based on the quality-carbon emission constraint relationship; using the ratio of the carbon emission ripple cost to the expected quality repair amount as the matching degree; removing control edges whose matching degree exceeds the preset tolerance threshold; and determining and outputting a recommended quality-carbon emission collaborative control scheme from the control schemes corresponding to the remaining control edges.

Images