Method, apparatus and electronic equipment for controlling molten pool carbon content based on digital twin

Abstract

This application provides a method, apparatus, and electronic device for controlling the carbon content of a molten pool based on digital twins, belonging to the field of intelligent control technology in steelmaking. The method includes mapping real-time sensing data of the target molten pool to a pre-constructed digital twin model, and evaluating the current furnace condition of the target molten pool based on the digital twin model; dynamically adjusting the fusion weights of the mechanistic model and the data-driven model according to the current furnace condition; weighted fusion of the first carbon content prediction value output by the mechanistic model and the second carbon content prediction value output by the data-driven model based on the fusion weights to obtain the target carbon content prediction value; and sending a lance-lifting control command to the PLC layer in response to the target carbon content prediction value meeting preset control conditions to execute the corresponding lance-lifting operation. This application uses a digital twin model to sense the furnace condition in real time and dynamically adjusts the fusion weights of the mechanistic model and the data-driven model accordingly, ensuring the stability and accuracy of the carbon content prediction results across the entire operating range.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology in steelmaking, and more specifically, to a method, apparatus, and electronic device for controlling the carbon content of molten pool based on digital twins. Background Technology

[0002] Converter steelmaking, as the core process of modern iron and steel smelting, has a blowing process divided into the main blowing stage and the post-blowing stage. The post-blowing stage is a supplementary blowing operation carried out after the main blowing is completed because the carbon content of the molten steel is not up to standard or the temperature is insufficient. It is a key link to accurately control the final carbon content and temperature of the molten steel, which directly determines the quality of the molten steel, smelting energy consumption and furnace lining life.

[0003] The carbon composition of molten steel changes rapidly and exhibits significant nonlinear characteristics during the post-blowing stage. Traditional single dynamic detection methods (such as secondary lance endpoint detection) cannot achieve continuous real-time prediction of carbon composition, have strong detection lag, and are difficult to match the rapid response process characteristics of the post-blowing stage, resulting in low detection accuracy of carbon content in molten steel. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method, apparatus and electronic device for controlling the carbon content of molten steel based on digital twins, so as to improve the detection accuracy of carbon content in molten steel.

[0005] Firstly, a method for controlling the carbon content of a molten pool based on digital twins is provided, including: The real-time sensing data of the target molten pool is mapped to a pre-built digital twin model, and the current furnace condition of the target molten pool is evaluated based on the digital twin model. The fusion weights of the mechanistic model and the data-driven model are dynamically adjusted based on the current furnace condition; the mechanistic model is an empirical model of exponential decay based on oxygen consumption time series; the data-driven model is a gradient boosting regression tree model. Based on the fusion weight, the first carbon content prediction value output by the mechanism model and the second carbon content prediction value output by the data-driven model are weighted and fused to obtain the target carbon content prediction value. In response to the target carbon content prediction value meeting the preset control conditions, a gun-raising control command is sent to the PLC layer to execute the corresponding gun-raising operation.

[0006] Optionally, the fusion weights of the mechanistic model and the data-driven model can be dynamically adjusted based on the current furnace condition, including: The current feature vector is constructed based on real-time sensing data, and its local density estimate and information entropy in the historical feature space are calculated. The confidence index of the data-driven model is determined based on the local density index and information entropy. In response to a confidence index exceeding a preset confidence threshold, the fusion weight of the data-driven model is increased while the fusion weight of the mechanistic model is decreased. In response to a confidence index falling below or equal to a preset confidence threshold, the fusion weight of the mechanistic model is increased while the fusion weight of the data-driven model is decreased.

[0007] Optionally, the process of using a mechanistic model to predict the first carbon content prediction includes: Obtain the actual cumulative oxygen consumption during the simulated main blowing phase and the real-time total oxygen consumption during the subsequent blowing phase, and add the two together to obtain the total oxygen consumption. Based on the total oxygen consumption, the preset first amplitude coefficient and the first characteristic oxygen consumption constant, the decarbonization response component in the initial stage of the post-blowing phase is calculated using the exponential decay function. Based on the total oxygen consumption, the preset second amplitude coefficient and the second characteristic oxygen consumption constant, the decarbonization response component in the middle stage of the post-blowing phase is calculated using the exponential decay function. Based on the total oxygen consumption, the preset third amplitude coefficient, and the third characteristic oxygen consumption constant, the decarbonization response component at the end of the post-blowing stage is calculated using an exponential decay function. Among them, the first amplitude coefficient, the second amplitude coefficient, and the third amplitude coefficient represent the intensity of carbon-oxygen reaction in the early, middle, and late stages of the post-blowing stage, respectively; the first characteristic oxygen consumption constant, the second characteristic oxygen consumption constant, and the third characteristic oxygen consumption constant represent the response sensitivity to oxygen consumption in each stage. The initial first carbon content prediction value is obtained by weighting and fusing the various decarbonization response components.

[0008] Optionally, after obtaining the initial predicted carbon content, the method further includes: Identify the type of steel being smelted in the current process; Retrieve historical furnace data belonging to the same steel grade from the historical database; Calculate the residual between the actual detected carbon content of each historical furnace and the theoretical prediction value of the mechanism model under the same operating conditions for that historical furnace. The residuals of all historical furnaces were arithmetically averaged, and the resulting mean was used as the baseline correction term for carbon content. The initial predicted value of the first carbon content is corrected based on the baseline correction term to obtain the final predicted value of the first carbon content.

[0009] Optionally, the process of using a data-driven model to predict the second carbon content forecast includes: The instantaneous time series values ​​of carbon monoxide concentration, carbon dioxide concentration, and oxygen concentration, as well as their corresponding cumulative values, are obtained in real time during the carbon-oxygen reaction in the target molten pool. An input feature vector is constructed based on the collected data. By using a pre-trained gradient boosting regression tree model, regression prediction is performed on the input feature vector to obtain the predicted value of the second carbon content.

[0010] Optionally, real-time sensing data of the target molten pool can be mapped to a digital twin model, including: A virtual entity of the target molten pool is constructed based on 3D modeling technology; Acquire multi-dimensional sensing data collected by the sensing layer. The multi-dimensional sensing data includes furnace batch planning information, flue gas data, material data, operating parameters, and process test data. The multidimensional perception data is preprocessed, including: converting the original data column names into standardized column names within the model, effectively converting non-numerical, infinite, and missing values, and processing key boundary data. The preprocessed multidimensional sensing data is mapped to virtual entities to build the global working condition sensing foundation for the digital twin model.

[0011] Optionally, after sending the gun-raising control command to the PLC layer, the method further includes: Receive verification and confirmation signals from the PLC layer in response to control commands; Perform a validity check on the verification confirmation signal; Upon successful validity verification, a start execution signal is sent to the PLC layer.

[0012] Optionally, after the gun-raising operation is completed, the method also includes: Obtain actual production data, which includes actual carbon composition data, actual oxygen consumption data, and actual flue gas data; Calculate the prediction error between actual production data and the predicted target carbon content; Based on the prediction error, the arithmetic mean learning algorithm is used to dynamically adjust the model parameters of the mechanistic model and update the model parameters of the data-driven model.

[0013] Secondly, a digital twin-based device for controlling the carbon content of a molten pool is provided, comprising: The evaluation unit is used to map real-time sensing data of the target molten pool to a pre-built digital twin model and evaluate the current furnace condition of the target molten pool based on the digital twin model. The adjustment unit is used to dynamically adjust the fusion weights of the mechanism model and the data-driven model according to the current furnace condition; the mechanism model is an empirical model of exponential decay based on oxygen consumption time series; the data-driven model is a gradient boosting regression tree model. The fusion unit is used to perform weighted fusion of the first carbon content prediction value output by the mechanism model and the second carbon content prediction value output by the data-driven model based on the fusion weight, so as to obtain the target carbon content prediction value. The execution unit is used to send a gun-lifting control command to the PLC layer in response to the target carbon content prediction value meeting the preset control conditions, so as to execute the corresponding gun-lifting operation.

[0014] Thirdly, an electronic device is provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements any of the methods of the first aspect.

[0015] This invention provides a digital twin-based method, apparatus, and electronic device for controlling the carbon content of a molten pool. It maps real-time sensing data of the target molten pool to a pre-constructed digital twin model and evaluates the current furnace condition based on the digital twin model. The fusion weights of the mechanistic model and the data-driven model are dynamically adjusted according to the current furnace condition. Based on the fusion weights, a weighted fusion of the first carbon content prediction value output by the mechanistic model and the second carbon content prediction value output by the data-driven model is performed to obtain the target carbon content prediction value. In response to the target carbon content prediction value meeting preset control conditions, a lance-lifting control command is sent to the PLC layer to execute the corresponding lance-lifting operation. This application uses a digital twin model to sense the furnace condition in real time and dynamically adjusts the fusion weights of the mechanistic model and the data-driven model accordingly. When the furnace condition is stable and the historical data matching degree is high, the data-driven model's fitting advantage for nonlinear characteristics is fully utilized. When the furnace condition is abnormal or fluctuates drastically, the weight of the mechanistic model is automatically increased, utilizing its physical interpretability and boundary constraint capabilities as a fallback. This adaptive fusion mechanism ensures the stability and accuracy of the carbon content prediction results across the entire operating range.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of a method for controlling the carbon content of a molten pool based on digital twins, provided by an embodiment of the present invention, is shown. Figure 2A flowchart of another method for controlling the carbon content of a molten pool based on digital twins provided by an embodiment of the present invention is shown; Figure 3 A schematic diagram of a molten pool carbon content control device based on digital twin provided in an embodiment of the present invention is shown. Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] This invention provides a method for controlling the carbon content of a molten pool based on digital twins, such as... Figure 1 As shown, it includes the following steps: Step S101: Map the real-time sensing data of the target molten pool to a pre-built digital twin model, and evaluate the current furnace condition of the target molten pool based on the digital twin model.

[0021] By collecting multidimensional heterogeneous data from the converter smelting process, including but not limited to furnace schedule information, flue gas composition, material weight, operating parameters, and process test data, these real-time perceived data from the physical world are cleaned and standardized, and then mapped in real time to the corresponding virtual entities in the virtual space.

[0022] This mapping relationship enables the digital twin model to synchronously reflect the dynamic changes of the physical molten pool, thereby providing a unified, clean, and spatiotemporally consistent data view for assessing the current furnace condition.

[0023] For example, when the CO concentration collected by the flue gas analyzer changes abruptly, the digital twin model can immediately reflect the change in the virtual flue and, combined with the current oxygen lance position and feeding records, comprehensively determine whether the current situation is in a normal decarbonization reaction period or an abnormal splashing or back-drying phenomenon has occurred.

[0024] This state assessment method based on full data fusion overcomes the limitations of traditional methods that rely on a single sensor or local data for judgment, providing reliable state input for subsequent prediction and decision-making.

[0025] Step S102: Dynamically adjust the fusion weights of the mechanism model and the data-driven model according to the current furnace condition.

[0026] The mechanistic model is an empirical model of exponential decay based on oxygen consumption time series; the data-driven model is a gradient boosting regression tree model.

[0027] Specifically, due to the highly nonlinear and time-varying nature of the steelmaking process, a single model struggles to maintain optimal performance under all operating conditions. Mechanistic models, built upon the principles of metallurgical thermodynamics and kinetics, possess clear physical meaning and boundary constraints, but may exhibit biases when dealing with complex nonlinear disturbances. In contrast, data-driven models excel at uncovering implicit nonlinear patterns from historical data and achieve high fitting accuracy for subtle changes under normal operating conditions, but are prone to failure when encountering unseen abnormal conditions.

[0028] Therefore, this embodiment dynamically adjusts the fusion weights of the two by evaluating the furnace condition in real time. For example, when the furnace condition is stable and the production conditions match the historical typical operating conditions well, the weight of the data-driven model is automatically increased to make full use of its ability to capture nonlinear features and improve prediction accuracy; conversely, when the furnace condition fluctuates violently or an abnormal pattern is detected, the weight of the mechanism model is automatically increased to make a fallback prediction by utilizing the stability of its physical laws.

[0029] This adaptive adjustment mechanism essentially seeks the optimal balance between "data fitting" and "physical constraints" under the current operating conditions, thereby significantly improving the robustness of the prediction model in complex and ever-changing environments.

[0030] Step S103: Based on the fusion weight, the first carbon content prediction value output by the mechanism model and the second carbon content prediction value output by the data-driven model are weighted and fused to obtain the target carbon content prediction value.

[0031] For example, the formula for calculating the predicted target carbon content is as follows:

[0032] in, This represents the predicted target carbon content. This represents the first predicted carbon content. This indicates the predicted value for the second carbon content; This indicates the fusion weight.

[0033] The fused target carbon content prediction retains the physical trends implied by the mechanistic model while incorporating the data-driven model's refined correction of the current real-time operating conditions.

[0034] For example, in the early stage of the post-blowing phase, if the furnace conditions are stable, the target prediction value may be more dominated by the data-driven model to quickly respond to small changes in flue gas composition; while in the late stage of the post-blowing phase or when the operating conditions are unstable, the target prediction value gradually converges to the mechanistic model to prevent prediction divergence caused by data noise.

[0035] In this way, a carbon content prediction result that maintains high accuracy and stability across the entire operating range can be output, providing a reliable decision-making basis for endpoint control.

[0036] Step S104: In response to the target carbon content prediction value meeting the preset control conditions, a gun-lifting control command is sent to the PLC layer to execute the corresponding gun-lifting operation.

[0037] Preset control conditions include the target carbon content prediction value falling within the preset target range, or the trend of the prediction value conforming to the endpoint determination logic, etc.

[0038] For example, the preset target range is set as a dynamic control range: the upper control limit is... The lower control limit is .in, This indicates the target carbon content values ​​for different steel grades; , The tolerance values ​​are dynamically set according to the quality requirements of the steel grade, representing the upper and lower limits.

[0039] The system will only generate and issue control commands when the fusion prediction results clearly indicate that the carbon content of the molten pool has reached or is about to reach the requirements for tapping steel.

[0040] This design ensures that the control action is triggered based on high-confidence predictions rather than instantaneous readings from the original sensors, effectively avoiding false firing or overblowing caused by measurement noise or transient model errors.

[0041] This application uses a digital twin model to perceive the furnace condition in real time and dynamically adjusts the fusion weights of the mechanistic model and the data-driven model accordingly. When the furnace condition is stable and the historical data matches well, the data-driven model's advantage in fitting nonlinear features is fully utilized. When the furnace condition is abnormal or fluctuates drastically, the weight of the mechanistic model is automatically increased, leveraging its physical interpretability and boundary constraint capabilities as a fallback. This adaptive fusion mechanism ensures the stability and accuracy of carbon content prediction results across the entire operating range.

[0042] Based on the above embodiments, the fusion weights of the mechanism model and the data-driven model are dynamically adjusted according to the current furnace condition, including: Step S101A: Construct the current feature vector based on real-time sensing data, and calculate its local density estimate and information entropy in the historical feature space.

[0043] For example, the real-time sensing data is first preprocessed to obtain the current feature vector of the current furnace condition. Current feature vector Input feature vectors of the GBRT model They share the same core gas characteristic dimensions, but Additional features include working condition characteristics (such as gun position height). Oxygen supply intensity (e.g., bottom blowing flow rate) to comprehensively characterize the furnace's operating environment. Construct... and At the same time, data is uniformly extracted from the data pool of the digital twin model to avoid problems such as duplicate collection and inconsistent feature definitions.

[0044] Meanwhile, based on a set of sensing data from historical normal steelmaking cycles, several historical cluster centers representing typical furnace conditions were trained using the K-Means clustering algorithm. .

[0045] Traditional matching methods rely solely on distance, which is prone to misjudgment. This embodiment introduces local density estimation. The current feature vector is calculated. Euclidean distance to its nearest historical cluster center And combined with the number of historical samples covered by the cluster center Calculate the local density index :

[0046] in, To prevent division by zero for the minimum value. The higher the value, the more likely the current operating condition is in a high-density area of ​​historical data, meaning the model is more likely to have "seen" this type of operating condition.

[0047] Simultaneously, calculate the prediction uncertainty entropy. Then, use the current feature vector... The input is fed into multiple weak learners (each tree in GBDT), and the predictions from all trees are collected. Calculate the variance of these results. And convert it into normalized information entropy. :

[0048] in, This approximates the probability distribution of each predicted value. The larger the value, the more divergent the opinions within the model, and the less reliable the prediction.

[0049] Step S101B: Determine the confidence index of the data-driven model based on the local density index and information entropy.

[0050] Specifically, construct confidence indicators :

[0051] in, and This is the adjustment coefficient; Represents the normalization function; the higher the local density ( The larger the confidence level, the greater the prediction uncertainty entropy. (The larger the number), the lower the confidence level.

[0052] This definition takes into account not only familiarity with the working conditions, but also the internal consistency of the model itself.

[0053] Step S101C: In response to the confidence index being higher than the preset confidence threshold, increase the fusion weight of the data-driven model and decrease the fusion weight of the mechanistic model.

[0054] In one specific implementation, the preset confidence threshold can be set to 0.8. When the calculated confidence index is greater than 0.8, it means that the current operating condition is within the historical distribution range. High), and the model's prediction consistency is good ( (Low). At this point, data-driven models typically perform better because they capture subtle non-linear features. Therefore, appropriately increasing the weights of data-driven models can help improve overall prediction accuracy.

[0055] Step S101D: In response to the confidence index being lower than or equal to the preset confidence threshold, increase the fusion weight of the mechanistic model and decrease the fusion weight of the data-driven model.

[0056] Conversely, a low confidence level may indicate the emergence of a new operating condition. Low (OOD situation) or severe internal model conflicts ( (High). At this point, data-driven models may experience overfitting or extrapolation failure. Mechanistic models, based on physicochemical principles, have stronger generalization ability and robustness. Therefore, increasing the weight of mechanistic models can ensure the rationality and safety of prediction results and prevent control errors caused by the model's "blind confidence."

[0057] This dynamic adjustment strategy based on dual-dimensional quantitative evaluation avoids control jitter caused by triggering a single threshold, ensuring the smoothness and rationality of weight switching.

[0058] At the level of the specific implementation of the mechanistic model, the process of the mechanistic model predicting the first carbon content prediction value is as follows: Figure 2 As shown, it includes: Step A1: Obtain the actual cumulative oxygen consumption during the simulated main blowing phase and the real-time total oxygen consumption during the subsequent blowing phase, and add the two together to obtain the total oxygen consumption.

[0059] Step A2: Based on the total oxygen consumption, the preset first amplitude coefficient and the first characteristic oxygen consumption constant, the decarbonization response component in the initial stage of the post-blowing phase is calculated using the exponential decay function.

[0060] This step is used to capture the rapid decrease in carbon content caused by enhanced stirring of the molten pool during the initial stage of post-blowing.

[0061] Step A3: Based on the total oxygen consumption, the preset second amplitude coefficient and the second characteristic oxygen consumption constant, the decarbonization response component in the middle stage of the post-blowing phase is calculated using the exponential decay function.

[0062] Step A4: Based on the total oxygen consumption, the preset third amplitude coefficient and the third characteristic oxygen consumption constant, the decarbonization response component at the end of the post-blowing stage is calculated using the exponential decay function.

[0063] Among them, the first amplitude coefficient, the second amplitude coefficient, and the third amplitude coefficient characterize the intensity of carbon-oxygen reaction in the early, middle, and late stages of the post-blowing phase, respectively; the first characteristic oxygen consumption constant, the second characteristic oxygen consumption constant, and the third characteristic oxygen consumption constant characterize the response sensitivity to oxygen consumption in each stage, respectively.

[0064] Step A5: Weight and fuse each decarbonization response component to obtain the initial first carbon content prediction value.

[0065] Specifically, the mathematical expression of the above mechanism model is as follows:

[0066] in, This represents the initial predicted value of the first carbon content; This represents the actual cumulative oxygen consumption during the simulated main blowing phase, serving as the initial oxygen consumption benchmark for the subsequent blowing phase. This indicates the real-time cumulative oxygen consumption during the post-blowing phase. , , These are the first, second, and third amplitude coefficients, respectively. , , These are the first, second, and third characteristic oxygen consumption constants, representing the dynamic response hysteresis characteristics of each reaction stage to oxygen input, and reflecting the rate-limiting factors in the mass and heat transfer process. .

[0067] In this formula, This is the decarburization response component in the early stage of the post-blowing phase. This component is used to capture the rapid decrease in carbon content caused by enhanced molten pool stirring in the early stage of post-blowing. The second term represents the decarbonization response component in the middle stage of the post-blowing phase. The second term is subtracted from the first term to offset the over-prediction that the initial decarbonization response component may produce in the middle and later stages, thereby more accurately reflecting the actual working condition that the decarbonization rate slows down as the carbon content decreases in the middle stage of the post-blowing phase. The third value is added to the result of the aforementioned subtraction operation to represent the decarbonization response component at the end of the post-blowing stage, in order to smooth the prediction curve at the end of the post-blowing stage and ensure the stability of the model when it approaches the target value in the low-carbon region. Through this structured exponential decay combination, the mechanistic model can still maintain a predictive baseline that conforms to physical trends based on oxygen consumption, the most reliable cumulative amount, even in the absence of real-time flue gas data or when the data is distorted. This provides an indispensable physical constraint anchor for dynamic weight fusion.

[0068] Relying solely on the total macroscopic oxygen consumption as input ignores the nonlinear changes in mass transfer efficiency within the molten pool due to fluctuations in operating conditions, resulting in an inherent lag in the prediction.

[0069] Therefore, in another embodiment, after obtaining the total oxygen consumption, the following steps are also included: The first step involves using the implicit state estimator within the digital twin model to invert the reaction interface area and temperature field distribution on the surface of the molten pool using real-time flue gas composition, and then calculating the effective action coefficient.

[0070] Traditional digital twins only perform data mapping, while this embodiment embeds an implicit state estimator into the digital twin model to estimate implicit physical quantities that cannot be directly measured. The specific implementation is as follows: 1. Flue gas composition to estimate reaction interface area : Using real-time collected carbon monoxide concentration carbon dioxide concentration and real-time measurement of flue gas flow rate , (unit: Nm³ / h).

[0071] According to the kinetics of carbon oxidation reaction, The ratio is positively correlated with the local reactivity of the molten pool surface.

[0072] Establish a reverse model:

[0073] in, is the geometric proportionality constant for gas-liquid mass transfer, s / m (seconds per meter); To prevent small amounts from being divided by zero; among which, This is a temperature correction function, specifically in the form of: , This is the activation energy for the carbon-oxygen reaction. The gas constant is The calculation method is explained below.

[0074] This formula indicates that when the CO content in the flue gas is high and the flow rate is large, it means that the surface of the molten pool boils violently and the reaction interface area increases.

[0075] 2. Inversely elucidating the temperature field distribution: Obtaining multidimensional thermal parameters: Real-time acquisition of infrared radiation intensity at the furnace opening (Obtained by an infrared thermometer) and flue gas outlet temperature At the same time, obtain the current oxygen supply intensity. and gun position height .

[0076] Construct a furnace mouth-molten pool thermal balance correction model: Because the heat inside the molten pool is dissipated to the furnace mouth through conduction, radiation and convection in the slag layer, the furnace mouth temperature is lower than the temperature at the center of the molten pool.

[0077] To accurately assess the effective reaction temperature that has a decisive influence on the carbon-oxygen reaction rate. In this embodiment, the following thermal balance correction equation is established:

[0078] in: It is determined by the radiation intensity at the furnace opening The original surface temperature was initially calculated based on the Stefan-Boltzmann law; This is the heat loss coefficient, a variable that dynamically changes with operating conditions. Specifically, It is positively correlated with the slag layer thickness (which can be estimated from the height of the foamed slag) and the flue gas flow rate. The thicker the slag layer or the more heat is carried away by the flue gas, the better. The smaller the value, the greater the heat loss from the molten pool to the furnace opening; This is a compensation term for the exothermic reaction. Carbon oxidation is a strongly exothermic reaction; the more vigorous the reaction, the higher the local temperature rise in the molten pool. This compensation term is proportional to the instantaneous decarburization rate.

[0079] Will As a representative parameter of temperature field distribution, it can more accurately quantify the contribution of high temperature to reducing the activation energy of carbon-oxygen reaction.

[0080] It should be noted that all the temperature values ​​mentioned above are in K, the thermodynamic temperature scale.

[0081] 3. Calculate the effective action coefficient : By combining the two implicit physical quantities mentioned above, the effective action coefficient is obtained. It characterizes the ratio of the effective driving force actually participating in the carbon oxidation reaction at the current moment to the theoretical maximum driving force:

[0082] in, This represents the maximum possible reaction area of ​​the molten pool. Reference standard temperature; , These are empirical weighting coefficients, with values ​​ranging from 0.3 to 0.7. Control the contribution weight of the reaction interface to the effective action coefficient. The contribution weights of the temperature field are controlled. The calibration of the two coefficients is obtained by nonlinear least squares fitting of historical operating data.

[0083] When the furnace condition is stable and the stirring is moderate A value close to 1 indicates high reaction efficiency; however, when splashing or excessive foam residue leads to gas-liquid separation, Although the contact is large, the effective contact is reduced, or the reaction is inhibited by abnormal temperature. It will dynamically decrease, thereby automatically compensating for the model's over-prediction in subsequent steps.

[0084] Step 2: Based on the baseline total oxygen consumption and the effective action coefficient, calculate the equivalent effective oxygen consumption and reconstruct the input of the mechanism model.

[0085] This embodiment reconstructs it into an equivalent effective oxygen consumption. :

[0086] here, It is the instantaneous oxygen consumption within a tiny time interval. This is calculated by multiplying the instantaneous oxygen consumption by the dynamically changing... Integrating, we obtained an effective energy input that has been corrected by physical mechanisms.

[0087] This step transforms macroscopic oxygen supply data into effective work that truly promotes decarbonization at the microscopic level, eliminating model biases caused by fluctuations in gun position and changes in slag layer thickness.

[0088] The above steps A2-A4 calculate the decarbonization response components for each stage, i.e., using the equivalent effective oxygen consumption. The mathematical expression of the corresponding mechanism model is as follows:

[0089] The meanings of the parameters in the formula are the same as above, and will not be repeated here.

[0090] To further eliminate systematic deviations caused by factors such as batch differences in raw materials and equipment aging, after obtaining the initial predicted value of the first carbon content, the method also includes: Step A6: Identify the type of steel being smelted in the current process.

[0091] Step A7: Retrieve historical furnace data belonging to the same steel type from the historical database.

[0092] Step A8: Calculate the residual between the actual detected carbon content of each historical furnace and the theoretical prediction value of the mechanism model under the same operating conditions.

[0093] After each smelting cycle, the actual endpoint carbon detection results and model prediction trajectory for that furnace are automatically archived, and residual data are stored according to steel type.

[0094] Step A9: Perform an arithmetic mean on the residuals of all historical furnaces, and use the resulting mean as the baseline correction term for carbon content.

[0095] When a new heat cycle begins, the current planned steel grade identifier is first read, and the historical residual sequence of the same steel grade from the most recent N heats (e.g., 50 heats) is retrieved from the database. The arithmetic mean of this sequence is then calculated as the current residual. .

[0096] For example, if a certain steel grade has recently experienced a change in the source of scrap steel, resulting in an actual carbon content that is generally 0.005% higher than the theoretical value, then... It will automatically converge to around +0.005%, thus directly offsetting this systemic shift in the next batch prediction.

[0097] Step A10: Correct the initial first carbon content prediction value based on the benchmark correction term to obtain the final first carbon content prediction value.

[0098] Continuing from the previous example, the mathematical expression of this final predicted first carbon content is as follows:

[0099] in, This is the final predicted value for the first carbon content; This is the baseline correction term for carbon content; the meanings of other parameters are the same as above and will not be repeated here.

[0100] This closed-loop correction mechanism based on historical residuals of the same steel grade enables the mechanistic model to adaptively evolve with the slow drift of production conditions, significantly reducing long-term prediction bias caused by model mismatch.

[0101] It should be understood that although this embodiment uses arithmetic mean as an example of residual aggregation method, in other embodiments, statistical methods such as weighted moving average, exponential smoothing, or outlier removal average based on confidence interval can also be used. As long as their core purpose is to extract systematic deviation compensation from historical data of the same steel grade, they all fall within the protection scope of this application.

[0102] Based on the above embodiments, the process of using a data-driven model to predict the second carbon content prediction value includes: Step B1: Obtain the instantaneous time series values ​​of carbon monoxide concentration, carbon dioxide concentration, and oxygen concentration, as well as their corresponding cumulative values, during the carbon-oxygen reaction process in the target molten pool, and construct an input feature vector based on the collected data.

[0103] The input feature vector used in this embodiment It includes six core dimensions, namely .in, , , These represent the instantaneous volume concentrations of carbon monoxide, carbon dioxide, and oxygen measured by the flue gas analyzer at the current sampling time t. These instantaneous values ​​can sensitively reflect the real-time rate of carbon-oxygen reaction and combustion state inside the molten pool. , , These represent the cumulative emissions or concentration integral values ​​of the three gases from the start of the back-blowing to the current moment, respectively. These cumulative values ​​characterize the material conservation and overall decarbonization progress of the reaction process.

[0104] By simultaneously introducing instantaneous and cumulative features, the model can capture both short-term dynamic disturbances caused by feeding and gun position changes, and grasp the macroscopic quality balance trend throughout the smelting process, thus effectively overcoming the deficiency of insufficient information from single-dimensional features.

[0105] At the data acquisition level, the sampling frequency of flue gas analyzers is usually set to 1Hz to 10Hz. After the original signal is filtered by moving average to remove high-frequency noise, it is resampled and aligned at a fixed time step (e.g., 5 seconds) to ensure the temporal consistency of the input feature vector.

[0106] Step B2: Use a pre-trained gradient boosting regression tree model to perform regression prediction on the input feature vector to obtain the second carbon content prediction value.

[0107] At the model implementation level, Gradient Boosted Regression Trees (GBRT), as an ensemble learning method, iteratively approximates the true carbon content mapping by fitting the residuals of the previous tree. Its predicted output can be expressed as:

[0108] in, This indicates the predicted value for the second carbon content; This represents the gradient boosting regression tree mapping function after training. This represents the model residual term.

[0109] The model's training data comes from a rigorously cleaned paired sample set of "flue gas time series - endpoint carbon content" from a historical production database. For initial deployments or scenarios with insufficient historical data, the system has a built-in simulation data generator based on metallurgical thermodynamics principles. By introducing Gaussian noise and operating condition disturbances, it generates an initial training set covering multiple operating conditions for pre-training the GBRT model. As actual production data accumulates, the pre-trained model gradually migrates to actual operating conditions through incremental learning, achieving a smooth transition from simulation initialization to data-driven evolution.

[0110] Regarding the residual term In this embodiment, the value is not simply ignored as zero, but rather used as a measure of model uncertainty in subsequent fusion decisions. When the model is in a working condition area densely covered by training data, Smaller variance indicates higher prediction confidence; conversely, when the input features fall into a sparse region of the training data, the variance is smaller, resulting in higher prediction confidence; The variance increased significantly.

[0111] In practical engineering applications, methods such as Monte Carlo Dropout or quantile regression can be used to estimate... The distribution range of.

[0112] Furthermore, there is a deep complementary and synergistic relationship between the data-driven model and the mechanistic model, rather than a simple substitution or superposition. This complementarity is reflected in the difference in their adaptability to different operating conditions. When the furnace condition is in a stable operating range, the physicochemical reactions in the molten pool follow a relatively stable nonlinear law. At this time, the data-driven model, with its strong nonlinear fitting ability, can keenly capture the subtle changes in the flue gas composition. Its prediction accuracy is often better than that of the mechanistic model based on simplified assumptions, and therefore it is given a higher fusion weight.

[0113] However, when operating conditions fluctuate drastically or unprecedented abnormal patterns occur (such as severe splashing, back-drying, or equipment failure), the generalization ability of the data-driven model drops sharply, and the prediction results may deviate significantly, because these extreme samples account for very little or even none in the historical training set. In contrast, the mechanistic model, built upon fundamental principles of thermodynamics and kinetics, possesses a natural ability to maintain physical boundary constraints and trends. Even in the absence of real-time, effective data, it can still provide baseline predictions consistent with physical common sense. Therefore, the weight of the data-driven model is automatically reduced, and the mechanistic model is relied upon as a fallback.

[0114] This adaptive division of labor mechanism can maintain high robust predictive performance in complex and ever-changing industrial environments.

[0115] Based on the above embodiments, mapping the real-time sensing data of the target molten pool to a digital twin model includes: Step S101E: Construct a virtual entity of the target molten pool based on 3D modeling technology.

[0116] Specifically, the building elements of virtual entities cover two dimensions: the geometric structure and physical properties of the converter smelting system.

[0117] At the geometric level, 3D modeling technology is used to accurately reconstruct the spatial position and motion constraints of key equipment such as furnace body, oxygen lance, flue, and feeding chute; At the attribute level, each geometric object is bound to a corresponding data container and state variable. For example, the virtual oxygen gun is bound to attributes such as gun position height, oxygen supply flow rate and vibration frequency, and the virtual flue is bound to attributes such as flue gas composition, temperature and flow rate.

[0118] This deep integration of geometry and attributes transforms virtual entities from mere visual display tools into structured carriers of multidimensional heterogeneous data, providing a unified spatial index and semantic framework for subsequent state assessment.

[0119] Step S101F: Acquire multidimensional sensing data collected by the sensing layer.

[0120] Multidimensional sensing data includes furnace schedule information, flue gas data, material data, operating parameters, and process test data.

[0121] Among them, the furnace batch planning information is usually read from the secondary automation system, which includes static set values ​​such as target carbon temperature, steel grade and specifications and raw material ratio; The flue gas data comes from a mass spectrometer or infrared analyzer on the primary dust removal pipeline, providing high-frequency time-series signals of CO, CO2, and O2 concentrations; Material data comes from the weighing system and belt scale, recording the real-time addition of scrap steel, molten iron and various auxiliary materials; The operating parameters are directly collected from the primary basic automation system, covering dynamic control quantities such as oxygen lance position, bottom blowing flow rate, and auxiliary lance detection results; The process test data includes discrete feedback values ​​such as thermocouple temperature measurement and sample analysis of components.

[0122] These data come from diverse sources, have different sampling frequencies, and use various communication protocols. Direct use of such data could easily lead to logical confusion in the state assessment. Therefore, secondary preprocessing is performed through the following steps.

[0123] Step S101G: Preprocess the multidimensional sensing data.

[0124] Preprocessing includes: converting the original data column names into standardized column names within the model, performing effective transformations on non-numerical, infinite, and missing values, and processing key boundary data.

[0125] This embodiment employs a rigorous three-step preprocessing method to ensure that the data input into the digital twin model possesses both physical consistency and computational validity.

[0126] The first step is column name standardization, which involves establishing a standard data dictionary within the model to uniformly map the original field names from different PLC addresses, database headers, or message tags to standardized column names.

[0127] For example, the outputs of "CO_Conc", "CO_Pct" or "Gas_CO" from different manufacturers' analyzers are uniformly converted to "FlueGas_CO_Vol", which eliminates the semantic ambiguity of multi-source data and lays a unified namespace for subsequent fusion computing.

[0128] The second step is outlier validation, which aims to remove or repair dirty data caused by sensor failures or communication packet loss.

[0129] For missing values, linear interpolation or a previous time-hold strategy is used to fill them in to avoid time series breaks; for infinite values ​​or outliers that exceed the physical range, a truncation threshold method is used to limit them to a reasonable range to prevent them from causing numerical overflow or model divergence in subsequent weighted fusion or exponential calculations.

[0130] The third step is critical boundary handling, which involves logically correcting the characteristics of the steelmaking process at different stages. For example, at the start of the post-blowing stage, the cumulative oxygen consumption counter must be forcibly reset to zero to ensure that the oxygen consumption input in the mechanistic model accurately corresponds to the actual reaction process in the post-blowing stage. Another example is that during non-blowing periods, the validity markers of flue gas analysis data are automatically masked to prevent carbon content fluctuations caused by air intake from interfering with the state assessment.

[0131] Step S101H: Map the preprocessed multidimensional sensing data to virtual entities to build the global working condition sensing foundation of the digital twin model.

[0132] In this embodiment, the data sampling frequency of the perception layer is 10Hz. After being filtered by moving average (with a window width of 5 sampling points), the model is updated synchronously at a frequency of 1Hz. The end-to-end data transmission delay is controlled within 200ms to ensure that the state synchronization error between the digital twin model and the physical melt pool is within an acceptable range.

[0133] The multidimensional perception data, after the above preprocessing, is mapped to the corresponding attribute nodes of the virtual entity in real time, achieving millisecond-level synchronization between the physical world and the virtual space.

[0134] Sending the gun-raising control command to the PLC layer is only the initial step in control execution. To ensure operational safety in extreme industrial environments and the long-term accuracy of the model, this embodiment further constructs a dual-verification control command and error closed-loop self-learning mechanism. Specifically, after sending the gun-raising control command to the PLC layer, the method also includes: Step S105: Receive the verification confirmation signal from the PLC layer in response to the control command.

[0135] After generating and issuing the gun-raising control command, the system does not immediately trigger mechanical action, but instead enters a brief handshake waiting window.

[0136] After receiving the control command, the PLC layer first checks in the underlying logic whether the current equipment meets the execution conditions, such as whether the oxygen gun lifting mechanism is in an unlocked state, whether the oxygen supply valve is closed to the safe position, and whether the CRC check of the communication message is correct.

[0137] Only when all these underlying conditions are met will the PLC send a verification and confirmation signal containing a specific status code to the upper-level system.

[0138] Step S106: Perform a validity check on the verification confirmation signal.

[0139] After receiving the signal, the upper-layer system needs to perform a second validity check, including verifying whether the timestamp of the signal is within the valid window, whether the signal format conforms to the predefined protocol, and whether the device ready flag is true.

[0140] Step S107: In response to the successful validity check, a start execution signal is sent to the PLC layer.

[0141] Only after both of these checks pass without error will the final start-up signal be issued, driving the oxygen lance actuator to complete the lance-lifting operation.

[0142] This five-step closed-loop interactive logic of "instruction issuance - status verification - confirmation feedback - secondary verification - execution start" fundamentally eliminates accidents such as accidental gun lifting or over-blowing caused by network jitter, signal crosstalk or momentary equipment failure, and significantly improves the safety and reliability of automated control in the steelmaking process.

[0143] Furthermore, to overcome the problem of performance degradation due to shifts in operating conditions after traditional models are deployed, this embodiment also establishes an online self-learning mechanism based on actual production data. After the gun-raising operation is completed, the method further includes: Step S108: Obtain actual production data, which includes actual carbon composition data, actual oxygen consumption data, and actual flue gas data.

[0144] Once the lance lifting is complete and the molten pool is stationary, the system will automatically trigger the secondary lance detection or sampling and analysis process to obtain the high-precision final carbon composition measurement value for that furnace, and use it as the "true value" label. At the same time, the system extracts the complete actual oxygen consumption curve and flue gas emission time series data for that furnace from the data cache of the digital twin model, forming a complete "input-output" sample pair.

[0145] Step S109: Calculate the prediction error between the actual production data and the predicted target carbon content.

[0146] Step S110: Based on the prediction error, the model parameters of the mechanistic model are dynamically adjusted using the arithmetic mean learning algorithm, and the model parameters of the data-driven model are updated.

[0147] Next, the system calculates the deviation between the measured carbon content and the predicted target carbon content. If the deviation exceeds a preset learning trigger threshold (e.g., 0.005%), the parameter update procedure is initiated.

[0148] For mechanistic models, if the prediction bias is mainly caused by systematic shifts (the biases across multiple furnaces have consistent signs and stable amplitudes), then adjustments should be made first. Among them, Calculated using an arithmetic mean learning algorithm; if the deviation exhibits phased change characteristics (such as increased deviation in the mid-stage of the back-blow), then the corresponding phase is adjusted accordingly. or .

[0149] For example, if the actual carbon content of five consecutive furnaces is 0.008% higher than the predicted value, and the main deviation is concentrated in the middle of the later blowing process, the algorithm will make targeted fine adjustments. The magnitude of the item or The decay rate makes the mid-section of the mechanism curve more closely resemble the actual reaction kinetics under the current raw material conditions.

[0150] For the data-driven model (GBRT) in Example 3, new sample pairs are added to the online learning buffer. When the accumulated sample size reaches the update batch, the weights of the split nodes and leaf nodes of the tree are updated through incremental learning or sliding window retraining, so that it can capture the latest nonlinear perturbation patterns.

[0151] Through the aforementioned mechanism, this application transforms one-time offline modeling into an online evolutionary process that accompanies the entire production lifecycle. In this process, the digital twin model not only serves as a carrier of state awareness but also as a unified container for historical data accumulation and error tracing. Error feedback after each lifting operation becomes nourishment for the model parameters to self-correct, enabling the system to adaptively compensate for process drift caused by refractory erosion, raw material grade fluctuations, or seasonal temperature differences. This ensures that the molten pool carbon content control system maintains high-precision predictive capabilities and high-reliability control levels throughout long-term operation, effectively solving the pain points of high model maintenance costs and poor adaptability in existing technologies.

[0152] Based on the same inventive concept, a molten pool carbon content control device based on digital twins is provided, such as... Figure 3 As shown, it includes: Evaluation unit 301 is used to map real-time sensing data of the target molten pool to a pre-built digital twin model, and evaluate the current furnace condition of the target molten pool based on the digital twin model. The adjustment unit 302 is used to dynamically adjust the fusion weights of the mechanism model and the data-driven model according to the current furnace condition; the mechanism model is an exponential decay empirical model based on the oxygen consumption time series; the data-driven model is a gradient boosting regression tree model. The fusion unit 303 is used to perform weighted fusion of the first carbon content prediction value output by the mechanism model and the second carbon content prediction value output by the data-driven model based on the fusion weight, so as to obtain the target carbon content prediction value. The execution unit 304 is used to send a gun-lifting control command to the PLC layer in response to the target carbon content prediction value meeting the preset control conditions, so as to execute the corresponding gun-lifting operation.

[0153] Based on the same technical concept, embodiments of the present invention also provide an electronic device, such as... Figure 4 As shown, it includes a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404.

[0154] Memory 403 is used to store computer programs; The processor 401, when executing the program stored in the memory 403, implements the steps of the method embodiment.

[0155] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0156] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0157] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0158] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0159] The device for controlling the carbon content of a molten pool based on digital twins provided in this invention embodiment can be specific hardware on the device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this invention embodiment are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0160] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0161] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0162] In addition, the functional units in the embodiments provided by the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0163] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0164] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0165] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for controlling the carbon content of a molten pool based on digital twins, characterized in that, include: The real-time sensing data of the target molten pool is mapped to a pre-constructed digital twin model, and the current furnace condition of the target molten pool is evaluated based on the digital twin model. The fusion weights of the mechanism model and the data-driven model are dynamically adjusted based on the current furnace condition. The mechanistic model is an empirical model of exponential decay based on oxygen consumption time series; the data-driven model is a gradient boosting regression tree model. Based on the fusion weights, the first carbon content prediction value output by the mechanism model and the second carbon content prediction value output by the data-driven model are weighted and fused to obtain the target carbon content prediction value. In response to the target carbon content prediction value meeting the preset control conditions, a gun-raising control command is sent to the PLC layer to execute the corresponding gun-raising operation.

2. The method according to claim 1, characterized in that, The step of dynamically adjusting the fusion weights of the mechanism model and the data-driven model based on the current furnace condition includes: The current feature vector is constructed based on real-time sensing data, and its local density estimate and information entropy in the historical feature space are calculated. The confidence index of the data-driven model is determined based on the local density index and information entropy. In response to the confidence index being higher than a preset confidence threshold, the fusion weight of the data-driven model is increased and the fusion weight of the mechanism model is decreased; In response to the confidence index being lower than or equal to the preset confidence threshold, the fusion weight of the mechanism model is increased and the fusion weight of the data-driven model is decreased.

3. The method according to claim 1, characterized in that, The process by which the mechanistic model predicts the first carbon content prediction value includes: Obtain the actual cumulative oxygen consumption during the simulated main blowing phase and the real-time total oxygen consumption during the subsequent blowing phase, and add the two together to obtain the total oxygen consumption. Based on the total oxygen consumption, the preset first amplitude coefficient and the first characteristic oxygen consumption constant, the decarbonization response component in the initial stage of the post-blowing phase is calculated using an exponential decay function. Based on the total oxygen consumption, the preset second amplitude coefficient and the second characteristic oxygen consumption constant, the decarbonization response component in the middle stage of the post-blowing phase is calculated using the exponential decay function. Based on the total oxygen consumption, the preset third amplitude coefficient, and the third characteristic oxygen consumption constant, the decarbonization response component at the end of the post-blowing stage is calculated using an exponential decay function; wherein, the first amplitude coefficient, the second amplitude coefficient, and the third amplitude coefficient respectively characterize the intensity of carbon-oxygen reaction in the early, middle, and late stages of the post-blowing stage; the first characteristic oxygen consumption constant, the second characteristic oxygen consumption constant, and the third characteristic oxygen consumption constant respectively characterize the response sensitivity to oxygen consumption in each stage; The initial first carbon content prediction value is obtained by weighting and fusing the various decarbonization response components.

4. The method according to claim 3, characterized in that, After obtaining the initial predicted carbon content, the method further includes: Identify the type of steel being smelted in the current process; Retrieve historical furnace data belonging to the same steel grade from the historical database; Calculate the residual between the actual detected carbon content of each historical furnace and the theoretical prediction value of the mechanism model under the same operating conditions for that historical furnace. The residuals of all the aforementioned historical furnaces are arithmetically averaged, and the resulting mean is determined as the baseline correction term for carbon content. The initial first carbon content prediction value is corrected based on the benchmark correction term to obtain the final first carbon content prediction value.

5. The method according to claim 1, characterized in that, The process of the data-driven model predicting the second carbon content prediction value includes: The instantaneous time series values ​​of carbon monoxide concentration, carbon dioxide concentration, and oxygen concentration, as well as their corresponding cumulative values, are obtained in real time during the carbon-oxygen reaction in the target molten pool. An input feature vector is constructed based on the collected data. Using a pre-trained gradient boosting regression tree model, regression prediction is performed on the input feature vector to obtain the second carbon content prediction value.

6. The method according to claim 1, characterized in that, The process of mapping real-time sensing data of the target molten pool to a digital twin model includes: A virtual entity of the target molten pool is constructed based on 3D modeling technology; Acquire multi-dimensional sensing data collected by the sensing layer, including furnace batch planning information, flue gas data, material data, operating parameters, and process test data; The multidimensional sensing data is preprocessed, including: converting the original data column names into standardized column names within the model, performing effective conversion on non-numerical, infinite, and missing values, and processing key boundary data. The preprocessed multidimensional sensing data is mapped to the virtual entity to construct the global working condition sensing foundation of the digital twin model.

7. The method according to claim 1, characterized in that, After sending the gun-raising control command to the PLC layer, the method further includes: Receive the verification and confirmation signal fed back by the PLC layer in response to the control command; The validity of the verification confirmation signal is verified. Upon successful validity verification, a start execution signal is sent to the PLC layer.

8. The method according to claim 7, characterized in that, After the gun-raising operation is completed, the method further includes: Obtain actual production data, which includes actual carbon composition data, actual oxygen consumption data, and actual flue gas data; Calculate the prediction error between the actual production data and the predicted target carbon content; Based on the prediction error, the model parameters of the mechanistic model are dynamically adjusted using an arithmetic mean learning algorithm, and the model parameters of the data-driven model are updated.

9. A device for controlling the carbon content of a molten pool based on digital twins, characterized in that, include: An evaluation unit is used to map real-time sensing data of the target molten pool to a pre-built digital twin model, and evaluate the current furnace condition of the target molten pool based on the digital twin model. The adjustment unit is used to dynamically adjust the fusion weights of the mechanism model and the data-driven model according to the current furnace condition; the mechanism model is an empirical model of exponential decay based on oxygen consumption time series; the data-driven model is a gradient boosting regression tree model. A fusion unit is used to perform weighted fusion of the first carbon content prediction value output by the mechanism model and the second carbon content prediction value output by the data-driven model based on the fusion weight, so as to obtain the target carbon content prediction value. The execution unit is used to send a gun-lifting control command to the PLC layer in response to the target carbon content prediction value meeting the preset control conditions, so as to execute the corresponding gun-lifting operation.

10. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the method described in any one of claims 1-8.

Images