Intelligent control method, device and equipment for steel refining based on machine learning and medium

Abstract

The application discloses a molten steel refining intelligent control method and device based on machine learning, equipment and medium, the method comprises the following steps: collecting the multi-source process data of the whole process of molten steel refining and preprocessing to obtain the whole process data, performing feature labeling on the whole process data to construct a benchmark data set, training a gradient boosting tree and recurrent neural network fusion model using the benchmark data set to obtain a molten steel refining multi-task intelligent control model, inputting real-time whole process data into the molten steel refining multi-task intelligent control model to obtain molten steel state evaluation results, thereby determining key control parameters and generating control instructions to be sent to an industrial control unit to drive the power supply heating system and the alloy feeding system to perform control; since the application accurately evaluates the molten steel state through the fusion model and cooperatively controls power supply and feeding, accurate regulation of composition and temperature is realized, the narrow composition control precision and rhythm stability are improved, the proportion of manual intervention is reduced, and the automation degree is improved.

Description

Technical Field

[0001] This invention relates to the field of automatic control in steel refining, and more particularly to a machine learning-based intelligent control method, apparatus, equipment, and medium for steel refining. Background Technology

[0002] Steel refining is a crucial step in the steelmaking process, situated between the converter or electric furnace primary refining and continuous casting stages. It plays a vital role in adjusting the chemical composition of the molten steel, precisely controlling its temperature, and removing harmful elements and non-metallic inclusions. The level of control during the refining process directly determines the final steel's mechanical properties, surface quality, and internal microstructure, making it a core process for ensuring high-quality steel production. As the steel industry moves towards high-end, green, and intelligent manufacturing, higher demands are being placed on the control precision, production rhythm stability, and automation level of the steel refining process.

[0003] Currently, core operations in the steel refining process, such as composition control and electric heating, still rely primarily on manual intervention. Combined with traditional control methods, the following prominent problems exist: 1. Large error due to manual adjustment: In the traditional steel refining process, parameters such as alloy addition and power of heating mainly rely on the experience of operators. The adjustment standards of different shifts and different operators are not uniform, and it is difficult to adjust quickly according to real-time process data (such as composition deviation and temperature fluctuation). The adjustment lag is obvious, and it is impossible to guarantee the accuracy and stability of steel composition and temperature.

[0004] 2. High labor intensity: The temperature near the ladle is high and the environment is harsh. When the production rhythm is tight, operators need to frequently monitor the state of molten steel, record process data and adjust control parameters on site. The labor intensity is high and it is easily affected by human negligence, which can lead to fluctuations in refining quality.

[0005] 3. Poor process coordination: Steel refining involves multiple stages such as composition control, power supply and heating, and equipment operation. The process parameters of each stage affect each other, but traditional control methods lack a multi-stage coordination and linkage mechanism, which easily leads to unreasonable parameter matching (such as the heating rate not matching the alloy dissolution rhythm), resulting in quality defects such as steel composition segregation and local overheating.

[0006] 4. Low data utilization: The refining process generates a large amount of process data (such as steel composition, temperature, equipment status, smelting results) and MES system production data. However, under traditional methods, these data are not effectively integrated and utilized, and it is impossible to optimize process parameters and improve control accuracy through data mining, making it difficult to meet the needs of large-scale and refined steelmaking production.

[0007] Therefore, how to make full use of multi-source process data to achieve real-time and accurate assessment of the state of molten steel, and automatically generate collaborative control commands to drive the power supply and heating system and alloy feeding system, thereby reducing the intensity of manual intervention and improving the precision of refining control and production stability, has become an urgent technical problem to be solved in the current steel refining field. Summary of the Invention

[0008] The main objective of this invention is to provide a machine learning-based intelligent control method, device, equipment, and medium for steel refining, aiming to solve the technical problems existing in the prior art.

[0009] To achieve the above objectives, the present invention provides a machine learning-based intelligent control method for steel refining, the method comprising the following steps: Collect multi-source process data for the entire steel refining process, including equipment operating status data, production plan data, basic process data, and smelting performance data. The multi-source process data is preprocessed to obtain full-process data; Feature annotation is performed on the entire process data to construct a benchmark dataset, and the pre-constructed original model is trained based on the benchmark dataset to obtain a trained multi-task intelligent control model for molten steel refining. The original model is constructed based on a gradient boosting tree model and a recurrent neural network model. The benchmark dataset includes a training set, a test set, and a validation set. The real-time collected and pre-processed full-process data is input into the steel refining multi-task intelligent control model to obtain the steel condition assessment results. The steel condition assessment results include the steel composition deviation value, the steel temperature deviation value, the preliminary power reference value, and the alloy addition reference value. A control command is sent to the industrial control unit to cause the industrial control unit to drive the power supply heating system and the alloy feeding system to execute the control command. The control command includes key control parameters, which are determined based on the steel condition assessment results.

[0010] Optionally, the preprocessing of the multi-source process data to obtain full-process data includes: Gaussian filtering is applied to the steel temperature data, steel composition data, and oxygen determination data in the multi-source process data to obtain filtered process data. The equipment location data, molten steel weight data, and molten steel temperature data in the filtered process data are normalized to obtain normalized process data. Based on historical data of the same steel grade, the missing data items in the normalized process data are supplemented by linear interpolation to obtain complete process data; The complete process data is associated and bound with the production batch information in the production plan data to obtain associated process data; The associated process data is filtered based on a preset effective sample threshold to remove abnormal operating condition data, thus obtaining the full process data.

[0011] Optionally, the step of annotating the full-process data with features, constructing a benchmark dataset, and training the pre-constructed original model based on the benchmark dataset to obtain the trained multi-task intelligent control model for steel refining includes: The core process features in the entire process data are labeled with multi-label features, and the labeled entire process data is divided into training set, test set and validation set. The training set includes structured process feature data and time-series process feature data. The gradient boosting tree model is iteratively trained based on the structured process feature data in the training set to obtain the composition temperature prediction sub-model. The recurrent neural network model is iteratively trained based on the time-series process feature data in the training set to obtain a dynamic trend prediction sub-model. Based on the validation set, the output weights of the component temperature prediction sub-model and the dynamic trend prediction sub-model are optimized and adjusted to obtain the fusion weight parameters; A multi-task intelligent control model for steel refining is constructed based on the fusion weight parameters, the composition temperature prediction sub-model, and the dynamic trend prediction sub-model.

[0012] Optionally, the step of inputting the real-time collected and preprocessed full-process data into the multi-task intelligent control model for steel refining to obtain steel condition assessment results includes: The real-time process data from the real-time collected and preprocessed full-process data is input into the steel refining multi-task intelligent control model, and the predicted values ​​of steel composition and steel temperature are output. The difference between the predicted steel composition value and the target composition value in the production plan data is calculated to obtain the steel composition deviation value; The difference between the predicted molten steel temperature and the final temperature value in the production plan data is calculated to obtain the molten steel temperature deviation value; Based on the deviation values ​​of the molten steel composition and the deviation values ​​of the molten steel temperature, a preset process parameter mapping table is consulted to obtain the initial power reference value and the alloy addition reference value. Based on the slag turning data and the number of times the large ladle is used in the basic process data, the reference values ​​of the initial power supply and the alloy addition amount are corrected to obtain correction parameters. Based on the correction parameters, the steel refining process and equipment operating status are evaluated to obtain the steel condition evaluation results.

[0013] Optionally, sending control commands to the industrial control unit includes: Acquire steel ladle structure data and energized heating electrode layout data; Based on the steel condition assessment results, the ladle structure data, and the electric heating electrode layout data, the process area is divided into zones to determine the boundary range of the electric heating heat-affected zone and the boundary range of the alloy diffusion zone. Based on the boundary range of the heat-affected zone, the boundary range of the alloy diffusion zone, the steel composition deviation value, and the steel temperature deviation value, key control parameters are calculated. The key control parameters include the type of alloy added, the amount of alloy added, the power of the heating process, and the duration of the heating process. Based on the key control parameters, control commands are generated and sent to the industrial control unit.

[0014] Optionally, the calculation of key control parameters based on the boundary range of the heat-affected zone, the boundary range of the alloy diffusion zone, the steel composition deviation value, and the steel temperature deviation value includes: The types of alloying elements to be added are determined based on the steel composition deviation value and the boundary range of the alloy diffusion zone, thus obtaining the alloy addition types; The amount of alloy added is calculated based on the type of alloy added, the deviation value of the molten steel composition, the weight data of molten steel in the basic data of the process, and the preset alloy yield parameters. Based on the steel temperature deviation value, the boundary range of the heat-affected zone during energization and heating and the preset heating rate parameter, the electrode output power value is calculated to obtain the energization heating power. The duration of energization is calculated based on the steel temperature deviation value, the energizing power, and the preset thermal efficiency parameters to obtain the energization time. The key control parameters are the combination of the type of alloy added, the amount of alloy added, the power of energizing and heating, and the duration of energizing.

[0015] Optionally, the step of determining the process area partitioning based on the steel condition assessment results, the ladle structure data, and the electrode layout data for energized heating, to obtain the boundary range of the heat-affected zone and the alloy diffusion zone for energized heating, includes: Based on the ladle structure data, the geometric center coordinates and inner wall radius of the ladle are determined to obtain the ladle spatial parameters. Based on the electrode layout data for power-on heating, the coordinates of the electrode distribution position are determined, and the electrode spatial parameters are obtained. The boundary range of the heat-affected zone is obtained by defining the area centered on the coordinates of the electrode distribution location and using a preset heat-affected radius as the boundary. The boundary range of the alloy diffusion zone is obtained by defining a region centered on the geometric center coordinates of the ladle and with a preset diffusion radius as the boundary. Based on the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone, a process area partitioning judgment result is generated, which includes the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone.

[0016] Furthermore, to achieve the above objectives, the present invention also proposes a machine learning-based intelligent control device for molten steel refining that applies the machine learning-based intelligent control method for molten steel refining described above. The machine learning-based intelligent control device for molten steel refining includes: The data acquisition module is used to collect multi-source process data for the entire steel refining process. The multi-source process data includes equipment operating status data, production plan data, basic process data, and smelting performance data. The data processing module is used to preprocess the multi-source process data to obtain full-process data; The model training module is used to annotate the features of the entire process data, construct a benchmark dataset, and train the pre-constructed original model based on the benchmark dataset to obtain a trained multi-task intelligent control model for steel refining. The original model is constructed based on a gradient boosting tree model and a recurrent neural network model. The benchmark dataset includes a training set, a test set, and a validation set. The status assessment module is used to input the real-time collected and pre-processed full-process data into the steel refining multi-task intelligent control model to obtain the steel status assessment results. The steel status assessment results include steel composition deviation value, steel temperature deviation value, preliminary power reference value, and alloy addition reference value. The intelligent control module is used to send control commands to the industrial control unit so that the industrial control unit drives the power supply heating system and the alloy feeding system to execute the control commands. The control commands include key control parameters, which are determined based on the steel condition assessment results.

[0017] Furthermore, to achieve the above objectives, this application also proposes a machine learning-based intelligent control device for molten steel refining. The device includes a memory, a processor, and a machine learning-based intelligent control program for molten steel refining stored in the memory. The processor is used to run the machine learning-based intelligent control program for molten steel refining, and the computer program is configured to implement the steps of the machine learning-based intelligent control method for molten steel refining as described above.

[0018] In addition, to achieve the above objectives, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the machine learning-based intelligent control method for molten steel refining described above.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. Increased automation: Eliminates the need for frequent manual intervention in steel condition monitoring and control parameter adjustment, enabling intelligent identification, precise control and real-time monitoring of the entire steel refining process, reducing labor intensity and avoiding human error.

[0020] 2. Improve the stability of molten steel quality: Accurately predict the state of molten steel through machine learning models, and combine multi-process coordinated adjustment and closed-loop control to ensure that the composition, temperature and oxygen value of molten steel are stable within the target range, avoid quality defects such as local overheating and composition segregation, and improve the uniformity and qualification of molten steel.

[0021] 3. Optimize process synergy: Integrate process data from multiple stages, including composition control, power supply and heating, and equipment operation, to achieve synergistic linkage among stages, optimize parameter matching, improve refining efficiency, and reduce energy and alloy consumption.

[0022] 4. Improve data utilization: Integrate MES system data, process data, smelting results and equipment data, and mine data value through data preprocessing and machine learning to support process optimization and model iteration, and adapt to the needs of refined and large-scale steelmaking.

[0023] 5. Achieve full-process traceability: Establish a data traceability system for the entire production process, linking production batches, steel grades, equipment, and control processes to facilitate the identification of quality problems and the review and optimization of process parameters, thereby improving the level of production management. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the structure of a machine learning-based intelligent control device for steel refining, which is part of the hardware operating environment of the embodiment of the present invention. Figure 2 This is a flowchart illustrating the first embodiment of the intelligent control method for steel refining based on machine learning according to the present invention. Figure 3This is a flowchart illustrating the second embodiment of the intelligent control method for steel refining based on machine learning of the present invention. Figure 4 This is a structural block diagram of the first embodiment of the intelligent control device for steel refining based on machine learning of the present invention.

[0026] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0028] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a machine learning-based intelligent control device for steel refining, which is part of the hardware operating environment of an embodiment of the present invention.

[0029] like Figure 1 As shown, the machine learning-based intelligent control device for steel refining may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; the user interface 1003 may also include standard wired and wireless interfaces. The network interface 1004 may optionally include standard wired and wireless interfaces (such as Wireless-Fidelity (Wi-Fi) interfaces). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0030] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on machine learning-based intelligent control equipment for steel refining and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0031] like Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a machine learning-based intelligent control program for steel refining.

[0032] exist Figure 1 In the machine learning-based intelligent control device for molten steel refining shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the machine learning-based intelligent control device for molten steel refining of the present invention can be set in the machine learning-based intelligent control device for molten steel refining. The machine learning-based intelligent control device for molten steel refining calls the machine learning-based intelligent control program for molten steel refining stored in the memory 1005 through the processor 1001 and executes the machine learning-based intelligent control method for molten steel refining provided in the embodiments of the present invention.

[0033] This invention provides a machine learning-based intelligent control method for steel refining, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the intelligent control method for steel refining based on machine learning according to the present invention.

[0034] In this embodiment, the intelligent control method for steel refining based on machine learning includes the following steps: Step S10: Collect multi-source process data for the entire steel refining process.

[0035] It should be understood that the executing entity of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a terminal electronic device capable of performing the above functions. The following description uses a machine learning-based intelligent control device for molten steel refining (hereinafter referred to as the control device) as an example to illustrate this embodiment and the following embodiments.

[0036] It should be noted that the multi-source process data includes equipment operating status data, production plan data, basic process data, and smelting performance data.

[0037] In practical implementation, multi-source process data acquisition is used. Through sensors, equipment status monitoring modules, and MES system interfaces, core process data of the entire process is collected synchronously. This includes: the location and operating status of key equipment such as steel grade, ladle, and continuous casting; production plans and preset process paths in the MES system; basic process data for each process (steel grade specifications, molten steel composition, ladle number, ladle type, ladle feeding temperature, ladle usage frequency, slag turning status, tapping temperature, oxygen determination data, molten steel weight, etc.); and smelting performance data for each process (including the execution status of process parameters at each stage and quality inspection results). All acquisition equipment adopts a high-temperature resistant and anti-interference design, is fixed at designated monitoring points and calibrated regularly, and the MES system data interface is optimized to ensure the accuracy, real-time performance, and completeness of data acquisition, providing basic data support for subsequent molten steel composition regulation and intelligent control of power-on heating.

[0038] Step S20: Preprocess the multi-source process data to obtain full-process data.

[0039] In practical implementation, the control equipment can perform Gaussian filtering on the collected multi-source process data to smooth high-frequency noise in the molten steel temperature and composition data. It uses the minimum-maximum normalization algorithm to map values ​​with different dimensions, such as equipment position, molten steel weight, and temperature, to a unified range. It uses linear interpolation combined with historical data of the same steel grade to supplement missing data items caused by instantaneous sensor failures or packet loss during transmission. It also aligns and verifies all data items according to timestamps, removes outliers that exceed the physical reasonable range, and finally outputs consistent and reliable full-process data.

[0040] Further testing reveals that, in order to improve data quality and consistency and reduce the negative impact of outlier data, step S20 may include: Step S201: Perform Gaussian filtering on the molten steel temperature data, molten steel composition data, and oxygen determination data in the multi-source process data to obtain filtered process data; Step S202: Normalize the equipment location data, molten steel weight data, and molten steel temperature data in the filtered process data to obtain normalized process data; Step S203: Based on historical data of the same steel grade, perform linear interpolation to supplement the missing data items in the normalized process data to obtain complete process data; Step S204: Associate and bind the complete process data with the production batch information in the production plan data to obtain associated process data; Step S205: Based on a preset effective sample threshold, the associated process data is filtered to remove abnormal operating condition data, thus obtaining the full process data.

[0041] In practical implementation, the control equipment performs targeted preprocessing operations on the collected multi-dimensional time-series process data (including equipment status, MES production data, process data, and smelting performance): filtering is performed on core process data such as molten steel temperature, composition, and oxygen content to remove outliers; data of different dimensions such as equipment location and molten steel weight are normalized to unify data standards; missing data is supplemented by interpolation combined with historical data of the same steel grade and operating conditions to ensure data integrity; finally, a unified and standardized input dataset that can be directly used for training machine learning models is formed, preserving the correlation between each data point and the production plan and equipment status.

[0042] It is understood that this embodiment effectively eliminates random noise interference and dimensional differences in the original data, repairs data loss caused by acquisition failure, significantly improves data quality and consistency, reduces the negative impact of abnormal data on subsequent model training, enhances the algorithm's ability to extract key process features and the model's generalization performance under complex working conditions, and provides high-quality data input for intelligent control.

[0043] Step S30: Perform feature annotation on the full-process data, construct a benchmark dataset, and train the pre-constructed original model based on the benchmark dataset to obtain the trained multi-task intelligent control model for molten steel refining.

[0044] It should be noted that the original model is built based on a gradient boosting tree model and a recurrent neural network model. The aforementioned benchmark dataset can be a standardized dataset containing a training set, a validation set, and a test set, where each data sample is associated with a component target, a temperature target, and a process parameter label.

[0045] The training set refers to the subset of sample data used to drive model parameter updates and learn process rules.

[0046] The validation set is a subset of sample data used to evaluate model performance and guide hyperparameter tuning during training.

[0047] The test set refers to a subset of sample data used to independently verify the model's generalization ability and final accuracy after training.

[0048] It should be noted that the multi-task intelligent control model for steel refining in this embodiment refers to an intelligent reasoning model that integrates gradient boosting tree and recurrent neural network algorithms to output reference values ​​for steel composition deviation, temperature deviation, power supply, and alloy addition amount. At the same time, it evaluates the steel refining process and the rationality of equipment operation, providing a basis for subsequent precise control.

[0049] In practical implementation, the control equipment can add labels such as target range of molten steel composition, target temperature value, recommended range of power supply, and alloy addition threshold to each sample in the whole process data according to the endpoint quality standard and process path requirements in the manufacturing execution system. Valid samples are screened and randomly divided into training set, validation set, and test set according to proportion to build a benchmark dataset. The original model architecture including gradient boosting tree branch and recurrent neural network branch is built. The gradient boosting tree branch is iteratively trained using the training set to fit the nonlinear relationship between structured process features and control objectives. At the same time, the recurrent neural network branch is trained using the training set to capture the time-series dependence characteristics of temperature and composition changes. The model hyperparameters and fusion weights are adjusted through the validation set until the prediction error of the model on the test set meets the preset convergence condition. The trained multi-task intelligent control model for molten steel refining is then output.

[0050] Furthermore, in order to screen valid and complete process data samples and improve model performance, step S30 above may include: Step S301: Perform multi-label feature annotation on the core process features in the full process data, and divide the annotated full process data into a training set, a test set and a validation set. The training set includes structured process feature data and time-series process feature data. Step S302: Iteratively train the gradient boosting tree model based on the structured process feature data in the training set to obtain the composition temperature prediction sub-model; Step S303: Iteratively train the recurrent neural network model based on the time-series process feature data in the training set to obtain a dynamic trend prediction sub-model; Step S304: Based on the validation set, optimize and adjust the output weights of the component temperature prediction sub-model and the dynamic trend prediction sub-model to obtain the fusion weight parameters; Step S305: Construct a multi-task intelligent control model for steel refining based on the fusion weight parameters, the composition temperature prediction sub-model, and the dynamic trend prediction sub-model.

[0051] It should be noted that multi-label feature annotation can be a labeling operation that simultaneously associates multiple target attributes such as target values ​​of molten steel composition, target values ​​of temperature, recommended range of power output, and reference values ​​of alloy addition amount for a single data sample.

[0052] Structured process feature data refers to a set of static or semi-static process parameters that have a fixed format and are independent of each other, such as steel grade and specifications, ladle type, molten steel weight, and ladle temperature.

[0053] Time-series process characteristic data can be continuous process data with time-series dependence characteristics, including temperature change curves, dynamic fluctuation sequences of components, and records of constant oxygen value changes over time.

[0054] Understandably, this embodiment uses feature annotation and dataset construction to screen valid and complete process data samples, eliminating invalid data caused by abnormal operating conditions and data acquisition failures. Combined with MES production plans and process path requirements, core process features such as steel composition target range, composition deviation threshold, energizing and heating power range, energizing duration range, and alloy addition threshold are annotated with multiple labels. These labels are then linked to key data such as steel grade specifications, ladle type, and loading temperature to clarify the correlation between process parameters under different steel grades, different equipment states, and different production plans. The annotated multi-dimensional process data is divided into training, validation, and testing sets to construct a benchmark dataset for the intelligent control model of molten steel refining.

[0055] It should be understood that, in this embodiment, the machine learning model training takes into account the complexity of the steel refining process, the variability of operating conditions, and the real-time control requirements, while balancing the model's running speed and control accuracy. An algorithm that integrates gradient boosting trees and recurrent neural networks is used to construct a multi-task intelligent control model for steel refining. The constructed benchmark dataset is used as input to simultaneously train four core tasks: steel composition prediction, temperature prediction, power supply regulation, and alloy ratio optimization. The model's hyperparameters are iteratively adjusted, and the model's accuracy and response speed are verified until the model converges, thus deriving the trained intelligent control model for steel refining.

[0056] In the specific implementation, the control device inputs the structured process feature data from the training set into the initial gradient boosting tree model, calculates the residual between the model output prediction value and the true label, uses the residual data to construct a new decision tree to fit the error distribution and update the model parameters, and repeats the decision tree generation and parameter update operations until the number of iterations reaches a preset threshold or the model loss function converges, and finally outputs a composition and temperature prediction sub-model that can accurately map the nonlinear relationship between structured process features and steel composition and temperature targets.

[0057] The time-series process feature data in the training set are input into the initial recurrent neural network model in time sequence. The memory unit inside the model captures the dynamic dependence of temperature and composition data in the time dimension and outputs the prediction results. The loss value between the prediction results and the true label is calculated, and the network weights and bias parameters are updated using the backpropagation algorithm. The training process is continuously iterated until the loss value stabilizes within the preset range, thereby obtaining a dynamic trend prediction sub-model that can accurately predict the dynamic change trend of molten steel state.

[0058] The validation set data is input into the component temperature prediction sub-model and the dynamic trend prediction sub-model respectively to obtain their respective prediction outputs. The grid search algorithm is used to calculate the comprehensive error index between the fusion prediction result and the true label of the validation set under different weight combinations. By comparing the error magnitudes corresponding to each weight combination, the optimal weight configuration that minimizes the comprehensive error is determined, and the optimal weight configuration is recorded as the fusion weight parameter.

[0059] The composition and temperature prediction sub-model and the dynamic trend prediction sub-model are encapsulated as parallel inference branches. A weighted fusion mechanism for the output results of the two sub-models is established using fusion weight parameters. The model input interface is configured to receive structured process feature data and time-series process feature data. An output layer is set to generate a comprehensive evaluation result that includes steel composition deviation value, temperature deviation value, power reference value and alloy addition reference value. Thus, a multi-task intelligent control model for steel refining with multi-task collaborative inference capability is assembled.

[0060] Step S40: Input the real-time collected and preprocessed full-process data into the steel refining multi-task intelligent control model to obtain the steel condition assessment result.

[0061] It should be noted that the steel condition assessment results include steel composition deviation values, steel temperature deviation values, preliminary power reference values, and alloy addition reference values.

[0062] Among them, the steel composition deviation value refers to the numerical difference between the current predicted value of steel composition and the target composition value, which is used to characterize the degree of need for composition control.

[0063] The steel temperature deviation value refers to the numerical difference between the current predicted steel temperature and the final temperature standard, and is used to characterize the degree of need for heating or cooling.

[0064] The initial power reference value refers to the basic electrode output power value recommended by the model based on the current operating conditions.

[0065] The reference value for alloy addition refers to the basic alloy addition mass value estimated by the model based on composition deviation and molten steel weight.

[0066] In practical implementation, the control equipment inputs the pre-processed full-process data stream, which is collected in real time, into the trained multi-task intelligent control model for steel refining. The model performs parallel inference calculations and outputs the current predicted values ​​of steel composition, steel temperature, preliminary power reference value, and alloy addition reference value. Then, it calculates the difference between the predicted composition value and the target composition value in the production plan to obtain the steel composition deviation value, and calculates the difference between the predicted temperature value and the final temperature standard to obtain the steel temperature deviation value. Finally, it integrates the composition deviation value, temperature deviation value, preliminary power reference value, and alloy addition reference value into the steel condition assessment result.

[0067] It is understood that this embodiment achieves real-time and accurate quantitative assessment of the state of molten steel, quickly identifies the degree of deviation of composition and temperature from target values, and generates preliminary control suggestions based on model reasoning, replacing the traditional judgment method that relies on human experience. This greatly improves the timeliness and objectivity of state perception, provides accurate quantitative basis for the calculation of subsequent key control parameters, and reduces the risk of human misjudgment.

[0068] Furthermore, in order to compensate for the impact of changes in slag thermal resistance and differences in ladle heat dissipation characteristics on the refining process, and to improve the adaptability and accuracy of control parameters to actual operating conditions, step S40 above may include: Step S401: Input the real-time process data from the real-time collected and preprocessed full-process data into the multi-task intelligent control model for steel refining, and output the predicted values ​​of steel composition and steel temperature.

[0069] In the specific implementation, the control equipment transmits the real-time process data extracted from the pre-processed full-process data stream to the input interface of the multi-task intelligent control model for steel refining. The model activates gradient boosting tree branches and recurrent neural network branches in parallel. The gradient boosting tree branches are used to analyze the structured features in the real-time process data to capture nonlinear process laws, while the recurrent neural network branches are used to analyze the temporal dynamic features in the real-time process data to identify the state evolution trend. The reasoning results of the two branches are integrated through a fusion mechanism and the predicted values ​​of steel composition and steel temperature at the current moment are output.

[0070] Step S402: Calculate the difference between the predicted value of the molten steel composition and the target composition value in the production plan data to obtain the molten steel composition deviation value.

[0071] In practice, the control equipment parses the target composition value of the current smelting furnace from the production plan data synchronously obtained from the manufacturing execution system, compares the predicted steel composition value output by the model with the target composition value item by item and performs subtraction operation to calculate the numerical difference between the predicted concentration and the target concentration of each alloy element.

[0072] Step S403: Calculate the difference between the predicted molten steel temperature and the final temperature value in the production plan data to obtain the molten steel temperature deviation value.

[0073] In practice, the control equipment extracts the final temperature value of the current smelting furnace from the production plan data synchronously obtained from the manufacturing execution system, compares the predicted steel temperature value output by the model with the final temperature value, performs a subtraction operation, and calculates the difference between the predicted temperature value and the standard final temperature.

[0074] Step S404: Based on the deviation values ​​of the molten steel composition and the deviation values ​​of the molten steel temperature, query the preset process parameter mapping table to obtain the initial power reference value and the alloy addition reference value.

[0075] In the specific implementation, the deviation values ​​of molten steel composition and molten steel temperature are used as joint index keys to perform retrieval and matching in a pre-stored process parameter mapping table. The process parameter mapping table contains the correspondence between different deviation ranges and recommended control parameters. By looking up the range to which the deviation value belongs, the corresponding recommended power value and recommended alloy addition mass value are obtained. The retrieved recommended power value is determined as the initial power reference value, and the recommended alloy addition mass value is determined as the alloy addition amount reference value.

[0076] Step S405: Based on the slag turning data and the number of times the large ladle is used in the basic process data, the reference value of the initial power supply and the reference value of the alloy addition amount are corrected to obtain the correction parameters. Based on the correction parameters, the steel refining process and equipment operating status are evaluated to obtain the steel status evaluation result.

[0077] In practical implementation, the control equipment reads slag turning data and ladle usage data from the process data. Based on the slag turning data, it determines the slag layer thickness in the furnace and generates a thermal efficiency correction coefficient. Simultaneously, based on the ladle usage data, it assesses the degree of refractory material erosion in the ladle and generates a heat dissipation correction coefficient. The thermal efficiency correction coefficient and the heat dissipation correction coefficient are used to weight and adjust the initial power reference value to obtain the corrected power. The thermal efficiency correction coefficient is used to proportionally adjust the alloy addition reference value to obtain the corrected alloy addition amount. The corrected power and the corrected alloy addition amount are combined into correction parameters. Based on the correction parameters, it analyzes whether the current refining progress meets the preset rhythm requirements and whether the power supply equipment load is within a safe range, generating a molten steel state assessment result that includes molten steel composition deviation value, molten steel temperature deviation value, corrected power, and corrected alloy addition amount.

[0078] Step S50: Send a control command to the industrial control unit so that the industrial control unit drives the power supply heating system and the alloy feeding system to execute the control command.

[0079] It should be noted that the control command refers to a digital signal sequence that includes key control parameters and conforms to the industrial communication protocol format, used to drive the actions of the underlying actuators. The key control parameters are determined based on the steel condition assessment results, and can be specific process values ​​that directly determine the refining operation, such as the type and amount of alloy added, the power of energizing and heating, and the duration of energizing.

[0080] In practical implementation, the control equipment, based on the composition and temperature deviation values ​​in the molten steel condition assessment results, and combined with the molten steel weight, ladle type, and slag overflow situation, corrects and calculates the initial reference values ​​for energizing power and alloy addition amount. It then determines key control parameters including alloy addition type, alloy addition amount, energizing heating power, and energizing duration. These key control parameters are encapsulated into control instructions in a standard communication protocol format and sent to the industrial control unit via an industrial bus. After parsing the control instructions, the industrial control unit generates corresponding drive signals, adjusts the electrode output power of the power supply heating system to control the molten steel heating rate, and simultaneously controls the opening of the hopper valve and the conveying speed of the alloy feeding system to execute the alloy addition operation.

[0081] This embodiment achieves real-time and accurate assessment of the composition and temperature of molten steel refining process and intelligent control of multiple processes by collecting multi-source process data and constructing an intelligent control model that integrates gradient boosting trees and recurrent neural networks. It effectively solves the problems of traditional manual control relying on experience, adjustment lag, and poor coordination, significantly improves the consistency of molten steel quality and the accuracy of narrow composition control, optimizes the matching relationship between electric heating and alloy addition, reduces energy consumption and alloy waste, reduces the intensity of manual intervention, and enhances the stability of production rhythm and the traceability of full-process data.

[0082] refer to Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the intelligent control method for steel refining based on machine learning of the present invention.

[0083] Based on the first embodiment described above, in this embodiment, step S50 further includes: Step S501: Obtain the steel ladle structure data and the electrode layout data for power-on heating.

[0084] It should be noted that ladle structure data refers to a set of equipment parameters that include information on the ladle's geometric dimensions, the thickness of the refractory lining, and the ladle's volume.

[0085] The electrode layout data for power-on heating can be a set of the distribution coordinates of the power-on heating electrodes, the electrode spacing, and the electrode insertion depth parameters.

[0086] In practical implementation, the control equipment can read the ladle geometry, refractory lining thickness, and ladle volume information from the equipment parameter database to obtain ladle structure data. At the same time, it can query the distribution coordinates, electrode spacing, and electrode insertion depth parameters of the current energized heating electrodes through the electrode control system interface to obtain energized heating electrode layout data. The ladle structure data and energized heating electrode layout data are stored in the process area calculation cache for subsequent partition judgment and retrieval.

[0087] Step S502: Based on the steel condition assessment results, the ladle structure data, and the electrode layout data for power-on heating, the process area is divided into zones to determine the boundary range of the heat-affected zone and the alloy diffusion zone.

[0088] It should be noted that the boundary of the heat-affected zone during power-on heating can be the outer edge of the area inside the ladle that is significantly heated by the electrode arc and whose temperature gradient meets the preset requirements.

[0089] The boundary range of the alloy diffusion zone refers to the effective area within which the alloy can achieve uniform diffusion of components and avoid local segregation after being added to molten steel.

[0090] It should be noted that the process area zoning judgment refers to the analysis of dividing the thermal effect area and the alloy diffusion area based on the spatial characteristics of the ladle and the position of the electrodes.

[0091] It is understood that this embodiment can combine the ladle structure, the layout of the energized heating electrodes, the alloy addition landing area, and the temperature field distribution characteristics inside the furnace to perform zoning processing on the output molten steel state evaluation results; clarify the boundary range of the energized heating heat-affected zone and the alloy diffusion zone, judge the uniformity of molten steel temperature and consistency of composition in different areas, avoid process problems such as local overheating, insufficient alloy dissolution, and composition segregation, and ensure the accuracy of energized heating and alloy addition.

[0092] Furthermore, in order to effectively identify and avoid the risk of overlap and interference between the heating area and the feeding area, and to prevent local overheating or alloy burn-off caused by area overlap, step S502 above may include: Step S5021: Determine the geometric center coordinates and inner wall radius of the ladle based on the ladle structure data to obtain the ladle spatial parameters.

[0093] It should be noted that the geometric center coordinates of the ladle refer to the position vector of the center of the ladle's bottom surface in a three-dimensional coordinate system. The inner wall radius of the ladle can be the effective radius of the space inside the ladle that accommodates the molten steel. The spatial parameters of the ladle refer to the set of spatial geometric features that include the geometric center coordinates and the inner wall radius of the ladle.

[0094] Step S5022: Determine the electrode distribution location coordinates based on the energized and heated electrode layout data to obtain the electrode spatial parameters.

[0095] It should be noted that the electrode distribution position coordinates refer to the set of spatial position vectors of each energized heating electrode in a unified three-dimensional spatial coordinate system. Electrode spatial parameters can be a set of electrode spatial distribution characteristics that include the electrode distribution position coordinates.

[0096] In the specific implementation, the lateral offset, longitudinal offset, and vertical height values ​​of each energized heating electrode relative to the ladle reference point are extracted from the energized heating electrode layout data. The relative position information is mapped to a unified three-dimensional spatial coordinate system through a coordinate transformation algorithm to obtain the spatial position vector of each electrode. The set of spatial position vectors of all electrodes is recorded as electrode distribution position coordinates, and the electrode distribution position coordinates are encapsulated as electrode spatial parameters for subsequent area delineation.

[0097] Step S5023: Delineate the region centered on the coordinates of the electrode distribution location and with a preset heat-affected zone as the boundary to obtain the boundary range of the heat-affected zone during power-on heating.

[0098] It should be noted that the preset heat-affected radius refers to the pre-set value representing the effective distance of the electrode arc heat in the horizontal direction of the molten steel.

[0099] In the specific implementation, the control device traverses each electrode position point in the electrode distribution coordinates, takes the position point as the center and extends outwards by a preset heat-affected radius distance on the horizontal section to form a circular projection area. The circular projection areas corresponding to all electrodes are combined to eliminate overlapping parts and form a continuous thermal effect coverage range. At the same time, combined with the heat conduction attenuation characteristics in the depth direction of molten steel, the horizontal coverage range is extended along the vertical direction to generate a three-dimensional columnar region. The outer surface contour of the three-dimensional columnar region is marked as the boundary range of the heat-affected zone when energized and heated.

[0100] Step S5024: Delineate the region with the geometric center coordinates of the ladle as the center and the preset diffusion radius as the boundary to obtain the boundary range of the alloy diffusion zone.

[0101] It should be noted that the preset diffusion radius refers to the pre-set effective horizontal distance value that characterizes the uniform diffusion of alloying elements within molten steel.

[0102] In the specific implementation, the control equipment takes the geometric center coordinates of the ladle as the origin and extends outward by a preset diffusion radius distance on the horizontal section to form a circular diffusion projection area. Based on the convection circulation characteristics of molten steel and the alloy dissolution kinetic model, the effective diffusion depth of alloying elements in the molten steel is determined. The circular diffusion projection area is extended vertically to the effective diffusion depth to construct a three-dimensional diffusion space. The outer edge surface of this three-dimensional diffusion space is defined as the boundary range of the alloy diffusion zone and associated with the spatial configuration record of the current smelting furnace.

[0103] Step S5025: Generate process region partitioning judgment results based on the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone. The process region partitioning judgment results include the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone.

[0104] In practical implementation, the control equipment performs spatial overlay analysis on the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone to detect whether there is overlap or interference between the two regions. It calculates the volume ratio of the overlapping region and evaluates the rationality of the region division. If it is confirmed that the region division meets the preset safety distance requirements, the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone are packaged and combined into a structured data object. The structured data object is then output as a process area partitioning judgment result containing the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone.

[0105] Step S503: Calculate key control parameters based on the boundary range of the heat-affected zone, the boundary range of the alloy diffusion zone, the steel composition deviation value, and the steel temperature deviation value.

[0106] It should be noted that the key control parameters include the type of alloy added, the amount of alloy added, the power of heating, and the duration of heating.

[0107] In practical implementation, the control equipment determines the types of alloying elements that need to be added based on the steel composition deviation value to obtain the alloy addition type. It evaluates the alloy dissolution and diffusion efficiency by combining the boundary range of the alloy diffusion zone and calculates the added mass of each alloying element based on the steel composition deviation value and the weight of the steel to obtain the alloy addition amount. At the same time, it determines the target temperature rise range based on the steel temperature deviation value, analyzes the heat transfer efficiency by combining the boundary range of the heat-affected zone of the energized heating and calculates the electrode output power value with reference to the preset heating rate to obtain the energized heating power. Then, based on the steel temperature deviation value, energized heating power and steel heat capacity parameters, it calculates the duration required to maintain the target temperature rise range to obtain the energization time. Finally, the alloy addition type, alloy addition amount, energized heating power and energization time are integrated into key control parameters.

[0108] Furthermore, in order to reduce alloy waste, achieve precise control of the heating process, and prevent local overheating, step S503 may include: Step S5031: Determine the types of alloying elements to be added based on the steel composition deviation value and the boundary range of the alloy diffusion zone, and obtain the types of alloying elements to be added; Step S5032: Calculate the addition mass of each alloy element based on the alloy addition type, the steel composition deviation value, the steel weight data in the process basic data, and the preset alloy yield parameter to obtain the alloy addition amount.

[0109] In practice, the control equipment reads the steel weight data from the process basic data to determine the total mass of the molten steel for each alloy material among the alloy addition types. It then calculates the target increment of the alloy elements to be added based on the steel composition deviation value. The target increment is converted using preset alloy yield parameters to compensate for element burn-off and slag absorption losses during the smelting process. The actual addition mass required for each alloy material is calculated using the mass balance formula, and the actual addition mass of all alloy materials is summarized as the alloy addition amount.

[0110] Step S5033: Calculate the electrode output power value based on the molten steel temperature deviation value, the boundary range of the heat-affected zone during energization and heating, and the preset heating rate parameter, to obtain the energization and heating power.

[0111] In practical implementation, the control equipment determines the target temperature rise required to reach the final temperature based on the steel temperature deviation. It analyzes the heat transfer efficiency and heat distribution uniformity of the electrode arc within the molten steel by combining the boundary range of the heat-affected zone during energization. Based on the effective thermal conductivity coefficient within the heat-affected zone, it calculates the power requirement for the target temperature rise and limits the maximum temperature rise rate to prevent local overheating by referring to preset heating rate parameters. The power calculation model is used to solve for the electrode output power value that meets the temperature rise requirement and rate constraint. The solved electrode output power value is determined as the energization heating power.

[0112] Step S5034: Calculate the energizing duration based on the molten steel temperature deviation value, the energizing power, and the preset thermal efficiency parameters to obtain the energizing time.

[0113] In practical implementation, the control equipment can use the steel temperature deviation value and the steel heat capacity parameter to calculate the total heat requirement required to heat the steel to the target temperature. Combined with the power supply heating power, the energy value of the steel input per unit time is determined. Preset thermal efficiency parameters are introduced to correct the energy loss caused by ladle heat dissipation, slag thermal resistance and radiation loss. The continuous power supply time required to maintain the target temperature rise is calculated by the ratio between the total heat requirement and the effective input power. The calculated continuous power supply time is marked as the power supply duration.

[0114] Step S5035: Combine the type of alloy added, the amount of alloy added, the power of energizing and heating, and the duration of energizing into the key control parameters.

[0115] In practical implementation, the calculation and deviation analysis of key control parameters can accurately calculate the type and amount of alloy added and correct the alloy ratio parameters based on the steel composition deviation and temperature deviation values ​​output by the model, combined with the requirements of the smelting steel grade, the final quality standard in the MES production plan, and reference to process data such as ladle type, ladle temperature, slag turning situation, and steel weight, as well as previous smelting performance. This ensures that the steel composition and oxygen value reach the target range. Based on the temperature deviation and the preset heating rate requirements, combined with the heat conduction characteristics in the furnace and the equipment operating status, the optimal power supply and heating time are calculated to avoid steel quality defects. Finally, a four-dimensional control parameter system of composition deviation, temperature deviation, power adjustment, and alloy addition amount is formed, providing a quantitative basis for subsequent precise control.

[0116] Step S504: Generate control commands based on the key control parameters and send the control commands to the industrial control unit.

[0117] Understandably, the control equipment encapsulates the alloy type and amount of alloy added in the key control parameters into an alloy feeding instruction frame, and encapsulates the power-on heating power and power-on duration into a power-on control instruction frame. The alloy feeding instruction frame and the power-on control instruction frame are encoded and verified according to the industrial communication protocol standard to generate control instructions. The control instructions are transmitted to the industrial control unit in real time through industrial Ethernet or fieldbus, so that the industrial control unit can parse the control instructions and drive the power supply heating system to adjust the electrode output power and drive the alloy feeding system to perform alloy feeding operations.

[0118] In practical implementation, the control equipment can perform multi-process collaborative intelligent adjustment based on key control parameters. Based on the calculated key control parameters, it can achieve coordinated linkage and precise control of power-on heating and alloy addition: according to the corrected power and duration, the power supply heating system is driven to adjust the electrode output power, smoothly control the heating rate of molten steel, and ensure that the molten steel temperature reaches the final temperature; according to the precisely calculated type and amount of alloy added, the alloy feeding system is controlled to add alloy in batches and at a uniform rate, and the furnace thermal circulation is coordinated to ensure that the alloy is fully dissolved and evenly diffused, thus precisely controlling the composition of molten steel and achieving dual compliance of molten steel composition and temperature.

[0119] The system sends control commands, such as the determined power supply and alloy addition amount, to the industrial control unit to drive the power supply heating system and alloy feeding system. Simultaneously, the system continuously collects process data such as real-time steel temperature, composition changes, and oxygen content, and synchronously acquires updated equipment operating status data, feedback on smelting performance at each process, and MES production plan progress. This data is then input into the intelligent control model for dynamic correction, adjusting the power supply and alloy addition amount to avoid over- or under-adjustment. This forms a closed-loop control system, ensuring that the composition, temperature, and oxygen content remain stable within the target range throughout the steel refining process, guaranteeing consistent refining quality.

[0120] The industrial PC enables integrated real-time monitoring of core data across the entire process, including the location and operating status of equipment such as steel grades, ladles, and continuous casting; the production plan and process path execution progress of the MES system; process data and smelting performance for each step; and control data such as real-time steel temperature, composition data, power supply parameters, and alloy addition amounts. It automatically stores historical control data, process curves, smelting performance, and production batch information, linking all monitoring data with production plans, steel grades, refining quality data, and equipment operation records to achieve traceable management of the entire production process. It supports continuous iterative optimization of model parameters, improving model accuracy and response speed based on feedback from historical smelting performance and operating data. This enables phased automated control of steel refining, improving production efficiency, reducing human intervention errors, and ensuring the stability of steel quality.

[0121] This embodiment acquires data on the ladle structure and electrode layout, and combines this with the steel condition assessment results to determine the process area zoning. This allows for precise identification of the spatial boundaries between the heat-affected zone and the alloy diffusion zone during power-on heating, effectively avoiding the risks of local overheating and compositional segregation. Based on the regional boundary characteristics and state deviation values, key control parameters are calculated collaboratively, achieving precise matching and dynamic optimization of power supply, heating time, and alloy addition strategies. This improves the adaptability and scientific validity of the control parameters to actual operating conditions. Finally, a standardized command generation and distribution mechanism ensures the reliable execution of the control strategy, realizing automated and coordinated control of power supply heating and alloy addition. This significantly improves the control accuracy of steel composition and temperature, ensuring the stability and safety of the refining process, while reducing energy consumption and alloy waste, and improving overall production efficiency.

[0122] In one embodiment, the machine learning-based intelligent control method for steel refining provided by the present invention is applied to the VD furnace steel refining process in a steel plant, adapting to the refining requirements of various steel grades (such as Q235 and HRB400E). The specific steps are as follows: S1. Multi-source process data acquisition: Using high-temperature resistant and anti-interference sensors (temperature sensor measurement range 1500-1700℃, accuracy ±2℃; composition detection sensor can detect core elements such as C, Si, and Mn, accuracy ±0.01%), real-time temperature, composition, and oxygen content data of molten steel are simultaneously collected (oxygen content range 5-50ppm); through the equipment status monitoring module, the position (positioning accuracy ±5mm) and operating status (such as ladle tilt angle and continuous casting machine operating speed) of the ladle and continuous casting equipment are collected; through the MES system interface, production plans (such as the steel grade, batch, and output for the day) and preset process paths (such as...) are obtained. The final composition and final temperature standard for each steel grade; collect basic process data for each process, specifically including: steel grade specification for this furnace (HRB400E), molten steel composition (initial C: 0.18%, Si: 0.45%, Mn: 1.35%), ladle number (G202405), ladle type (ordinary steel ladle), ladle temperature (1580℃), number of times the ladle was used (8 times), slag turning status (no obvious slag turning), tapping temperature standard (1650℃), and molten steel weight (80t); collect smelting performance data for each process (such as the composition test results of the previous process and the duration of power-on heating).

[0123] All sensors are fixed at designated monitoring points on the VD furnace ladle and calibrated once a month. The MES system interface adopts the OPCUA standardized transmission protocol with a data transmission rate of ≥100Mbps to ensure real-time data synchronization, eliminate invalid data during transmission, and ensure the accuracy and integrity of data acquisition.

[0124] S2. Data Preprocessing: For the core data such as molten steel temperature, composition, and oxygen content, Gaussian filtering was used for noise reduction. The standard deviation σ of the Gaussian filter was set to 0.8. Local pixel mean smoothing was used to remove outliers caused by equipment fluctuations and environmental interference (such as data with instantaneous temperature fluctuations exceeding ±10℃). For data with different dimensions such as equipment position (mm), molten steel weight (t), and temperature (℃), a min-max normalization algorithm was used to map them to the [0,1] interval to unify the data standard. For missing data that occurred during the collection process (such as missing oxygen content data at certain moments), linear interpolation was used in combination with historical data of the same steel grade and the same ladle temperature to supplement the dataset and ensure its integrity. Finally, an input dataset with dimensions of 10000×12 (10000 samples, 12 feature dimensions) was formed, preserving the correlation between each data and the production plan and equipment status.

[0125] S3. Feature annotation and dataset construction: Valid data samples were screened, and invalid samples due to equipment failure, slag overflow, and other abnormal operating conditions were removed (a total of 800 invalid samples were removed), retaining 9200 valid samples. Based on the process requirements for HRB400E steel in the MES production plan, the following parameters were defined: target range for molten steel composition (C: 0.20-0.25%, Si: 0.40-0.50%, Mn: 1.40-1.60%), composition deviation threshold (±0.02%), power range for heating (500-800kW), heating time range (30-60min), and alloy addition threshold (S). The core features (i-iron: 0-500kg, Mn-iron: 0-800kg) were labeled with multiple tags; key data such as steel grade specifications, ladle type, and feeding temperature were associated to clarify the relationship between steel composition, temperature, power supply, and alloy addition under the conditions of HRB400E steel grade, ordinary steel ladle, and feeding temperature of 1570-1590℃; the labeled dataset was divided into training set (6440 samples), validation set (1840 samples), and test set (920 samples) in a ratio of 7:2:1 to construct the benchmark dataset for the intelligent control model of molten steel refining.

[0126] S4. Machine Learning Model Training: A multi-task intelligent control model is constructed using an algorithm that integrates Gradient Boosting Tree (XGBoost) and Recurrent Neural Network (LSTM). The XGBoost tree is used to process structured data such as steel composition and temperature, with 1000 iterations, a learning rate of 0.01, and a maximum tree depth of 6. The LSTM is used to process time-series data such as temperature change curves and dynamic changes in composition, with 128 hidden layer nodes, 50 iterations, and a learning rate of 0.005. Using a benchmark dataset as input, four core tasks—composition prediction, temperature prediction, power control, and alloy ratio optimization—are trained simultaneously. The model's hyperparameters are iteratively adjusted using a validation set until convergence (validation set error ≤ 3%). The model accuracy is verified using a test set: composition prediction accuracy ≥ 97%, temperature prediction accuracy ≥ 98%, power control error ≤ 5%, and alloy addition error ≤ 4%. The trained intelligent control model is then exported for subsequent steel state identification and parameter control.

[0127] S5. Real-time identification and assessment of molten steel condition: The pre-processed full-process data (steel grade HRB400E, ladle location, MES production plan, current molten steel composition C: 0.19%, temperature 1585℃, oxygen content 20ppm, smelting performance, etc.) are input into the trained intelligent control model. The model outputs the current molten steel composition deviation (C: -0.01%, Mn: -0.05%), temperature deviation (-65℃), preliminary power reference value of 650kW, and alloy addition reference value (Mn iron: 400kg). The operating condition verification module, combined with the current slag turning situation (no obvious slag turning) and the number of times the ladle has been used (8 times), makes a preliminary correction to the reference values, adjusting the power to 660kW and the Mn iron addition to 420kg to ensure that the results are suitable for the current operating conditions.

[0128] S6. Process area zoning determination: Based on the VD furnace ladle structure, the layout of the electrodes for energizing and heating (3 sets of electrodes, arranged in a triangular pattern), and the alloy addition landing area (center of the ladle), the steel condition assessment results were divided into zones. The boundaries of the heat-affected zone (within a 50cm radius below the electrodes) and the alloy diffusion zone (within a 60cm radius in the center of the ladle) were clearly defined. According to the model, the current temperature of the heat-affected zone is 1590℃, and the compositional uniformity of the alloy diffusion zone is good, with no local overheating or compositional segregation issues, providing a regional basis for subsequent precise control.

[0129] S7. Calculation and Deviation Analysis of Key Control Parameters: Based on the composition and temperature deviations output by the model, combined with the final quality standards for HRB400E steel (C: 0.20-0.25%, Mn: 1.40-1.60%, final temperature 1650℃), and referring to data such as ladle type (ordinary steel ladle), ladle temperature (1580℃), and molten steel weight (80t), the types and amounts of alloys added are accurately calculated: 430kg of Mn iron (to compensate for Mn element deviation) and 50kg of Si iron (to fine-tune the Si element to ensure compliance). Based on the temperature deviation (-65℃) and the preset heating rate (1.2℃ / min), combined with the heat conduction characteristics inside the furnace, the optimal power supply of 660kW and the power supply time of 55min are calculated. A four-dimensional control parameter system of composition deviation, temperature deviation, power adjustment, and alloy addition is formed, clarifying the control range and priority of each parameter.

[0130] S8. Multi-process collaborative intelligent adjustment: Based on four-dimensional control parameters, the coordinated regulation of energization heating and alloy addition is achieved: the energization heating adjustment module drives the power supply system to adjust the electrode output power to 660kW, adopting a segmented heating strategy, heating at 1.3℃ / min for the first 20 minutes (temperature rises to 1606℃), at 1.2℃ / min for the middle 20 minutes (temperature rises to 1630℃), and at 1.3℃ / min for the last 15 minutes (temperature rises to 1650℃) to avoid local overheating; the alloy addition adjustment module controls the feeding system to add the alloy in two batches: the first batch adds 215kg of Mn iron and 25kg of Si iron (added 10 minutes after energization), and the second batch adds 215kg of Mn iron and 25kg of Si iron (added 30 minutes after energization), with each batch adding for 5 minutes, combined with the thermal circulation in the furnace, to ensure that the alloy is fully dissolved and diffused uniformly.

[0131] S9. Closed-loop intelligent control: Control commands such as 660kW power supply and alloy addition (430kg Mn iron, 50kg Si iron) are sent to the PLC industrial control unit to drive the power supply heating system and alloy feeding system. Simultaneously, the system collects real-time data on molten steel temperature, composition changes, and oxygen content every 30 seconds, while also acquiring updated data on the ladle and continuous casting equipment operation status, smelting performance feedback, and MES production plan progress. When the molten steel temperature reaches 1640℃ (temperature deviation -10℃), the model dynamically adjusts the power supply to 620kW to prevent temperature overshoot. When the Mn content reaches 1.42% (deviation +0.02%), Mn iron addition is stopped to ensure the composition meets the standard. This forms a closed-loop control system of "data acquisition - status assessment - parameter calculation - control execution - feedback correction," ensuring that the composition, temperature, and oxygen content remain stable within the target range throughout the entire steel refining process.

[0132] S10. System Integration and End-to-End Monitoring: Through an industrial PC control platform, integrated real-time monitoring of core data across the entire process is achieved. The monitoring interface displays steel grade (HRB400E), ladle location, continuous casting equipment status, MES production plan execution progress, process data for each step (ladle number, casting temperature, etc.), smelting performance, and control data such as real-time steel temperature, composition, power supply, and alloy addition amount. Historical control data, process curves, smelting performance, and production batch information are automatically stored. All data are linked and bound to the production batch (2024050801), steel grade (HRB400E), refining quality data, and equipment operation records to achieve full traceability of the production process. Based on the feedback from the smelting data of this batch, the model hyperparameters are iteratively adjusted (the gradient boosting tree learning rate is adjusted to 0.009) to improve the subsequent control accuracy of the model and achieve phased automated control of steel refining.

[0133] In this embodiment, after adopting the method of the present invention, the steel composition compliance rate increased from 95% to 99%, the temperature control accuracy increased from ±5℃ to ±2℃, the number of manual interventions decreased by 80%, the production efficiency increased by 15%, and the alloy consumption decreased by 8%. This effectively solved many problems in the traditional refining process and significantly improved the level of intelligence and production efficiency of steel refining.

[0134] Furthermore, this embodiment of the invention also proposes a computer-readable storage medium storing a machine learning-based intelligent control program for molten steel refining. When the machine learning-based intelligent control program for molten steel refining is executed by a processor, it implements the steps of the machine learning-based intelligent control method for molten steel refining described above.

[0135] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0136] The aforementioned computer-readable storage medium may be included in a machine learning-based intelligent control device for steel refining; or it may exist independently and not be assembled into the machine learning-based intelligent control device for steel refining.

[0137] Furthermore, this invention also proposes a computer program product, including a machine learning-based intelligent control program for molten steel refining, which, when executed by a processor, implements the steps of the machine learning-based intelligent control method for molten steel refining as described above.

[0138] The specific implementation of the computer program product of the present invention is basically the same as the embodiments of the above-mentioned intelligent control method for steel refining based on machine learning, and will not be repeated here.

[0139] Reference Figure 4 , Figure 4 This is a structural block diagram of the first embodiment of the intelligent control device for steel refining based on machine learning of the present invention.

[0140] like Figure 4 As shown, the intelligent control device for steel refining based on machine learning proposed in this embodiment of the invention includes: Data acquisition module 10 is used to collect multi-source process data of the entire steel refining process. The multi-source process data includes equipment operating status data, production plan data, process basic data and smelting performance data. Data processing module 20 is used to preprocess the multi-source process data to obtain full-process data; The model training module 30 is used to annotate the features of the entire process data, construct a benchmark dataset, and train the pre-constructed original model based on the benchmark dataset to obtain a trained multi-task intelligent control model for steel refining. The original model is constructed based on a gradient boosting tree model and a recurrent neural network model. The benchmark dataset includes a training set, a test set, and a validation set. The status assessment module 40 is used to input the real-time collected and pre-processed full-process data into the steel refining multi-task intelligent control model to obtain the steel status assessment result. The steel status assessment result includes the steel composition deviation value, the steel temperature deviation value, the preliminary power reference value, and the alloy addition reference value. The intelligent control module 50 is used to send control commands to the industrial control unit so that the industrial control unit drives the power supply heating system and the alloy feeding system to execute the control commands. The control commands include key control parameters, which are determined based on the steel condition assessment results.

[0141] Furthermore, in one embodiment, the data acquisition module 10 may include a transmitter: Sensor Module: Employing high-temperature resistant and interference-resistant sensors, including temperature sensors, composition detection sensors, and current / voltage acquisition sensors, these modules are used to accurately acquire core process data such as molten steel temperature, composition, oxygen concentration, and power supply parameters. In principle, the sensors convert physical quantities (temperature, composition concentration, etc.) into transmittable electrical signals through signal conversion. They are also resistant to high-temperature radiation and electromagnetic interference within the furnace, ensuring the accuracy of data acquisition. All sensors are fixed at designated monitoring points on the ladle and undergo regular calibration and maintenance to prevent detection deviations caused by long-term use.

[0142] Equipment Status Monitoring Module: This module collects real-time data on the location and operating status of key equipment such as steel grades, ladles, and continuous casting equipment. It employs monitoring units adaptable to industrial smelting environments and features location positioning and operating parameter feedback. Through a combination of mechanical positioning and signal transmission, it captures abnormal equipment operation (such as ladle misalignment or continuous casting equipment malfunctions) in real time and synchronously transmits the status data to the data acquisition terminal, ensuring coordinated acquisition of equipment status data and process data.

[0143] MES System Interface Module: This module interfaces with the MES system to synchronously acquire management data such as production plans and preset process paths, as well as historical and real-time data on smelting performance at each process stage. The interface uses a standardized data transmission protocol, optimizes data transmission rates, ensures real-time synchronization between MES system data and field-collected data, and also features data verification to eliminate invalid data during transmission, guaranteeing data integrity.

[0144] Furthermore, in one embodiment, the data processing module 20 may include: Filtering module: For core process data such as molten steel temperature, composition, and constant oxygen value, Gaussian filtering and other methods are used to smooth the time-series data by applying local mean smoothing, removing random interference and instantaneous fluctuation noise. Based on the local statistical characteristics of the data, Gaussian filtering smooths and corrects abnormal fluctuations, preserving the core trends of data change while effectively eliminating high-frequency noise and improving data quality.

[0145] Normalization module: Because data such as equipment location, molten steel weight, and temperature have different dimensions, directly inputting them into the model would affect training accuracy. Therefore, a normalization algorithm is used to map data with different dimensions to a unified range. Data scaling eliminates the difference in dimensions, ensuring that process data in each dimension have equal training weights for the model, thus ensuring the rationality and accuracy of model training.

[0146] Missing Data Supplementation Module: This module uses interpolation combined with historical data from the same steel grade and operating conditions to supplement missing data acquired by sensors and transmitted through the MES system. The interpolation method calculates reasonable values ​​for missing data by fitting trends to adjacent valid data, and simultaneously verifies the results with historical data from similar operating conditions to avoid excessive deviations between the supplemented data and the actual process, thus ensuring the completeness and usability of the dataset.

[0147] Furthermore, in one embodiment, the model training module 30 may include: Feature annotation module: Employing a multi-label annotation method, and combining MES production plans and process path requirements, this module annotates core process features such as target ranges for molten steel composition, composition deviation thresholds, power ranges for heating, power duration ranges, and alloy addition thresholds. The annotation process incorporates key data such as steel grade specifications, ladle type, and loading temperature to ensure that the annotation information for each data sample corresponds to the actual smelting conditions. It also links the annotation to equipment status data, clarifying the feature relationships under different operating conditions.

[0148] The benchmark dataset construction module filters valid data samples under different operating conditions, steel grades, and equipment states, covering diverse process scenarios (such as different tank feeding temperatures, different slag turning situations, and different tank types). The labeled multi-dimensional process data is divided into training, validation, and test sets. The training set is used for model training, the validation set for iteratively adjusting hyperparameters, and the test set for verifying the final accuracy of the model, ensuring the diversity and representativeness of the dataset and improving the model's generalization ability.

[0149] Model training module 30 employs an algorithm that combines gradient boosting trees and recurrent neural networks, specifically: This fusion algorithm combines the advantages of two models: gradient boosting tree is used to process structured process data such as steel composition and temperature, and a decision tree is generated iteratively to fit the nonlinear relationship between process parameters and control targets (component compliance and temperature compliance), which has strong feature extraction and fitting capabilities; recurrent neural network is used to process time-series process data (such as temperature change curves and dynamic composition change data), and the memory unit captures the temporal correlation characteristics of the data to accurately predict the dynamic change trends of steel composition and temperature.

[0150] The fusion model integrates the outputs of two algorithms through a weight allocation mechanism. It also employs model pruning and parameter optimization techniques to reduce computational complexity, enabling it to operate efficiently in industrial control environments. This balances model speed and control accuracy, adapting to the changing conditions and real-time control requirements of steel refining.

[0151] Furthermore, in one embodiment, the state assessment module 40 adopts a combination of model reasoning and operating condition verification, specifically as follows: The model inference module inputs the pre-processed full-process data (steel grade, equipment status, MES production plan, process data, etc.) into the trained intelligent control model. Through model inference, it outputs core results such as steel composition deviation, temperature deviation, and preliminary control parameters. The operating condition verification module verifies the model output results in conjunction with the current smelting conditions (such as slag turning situation and number of times the ladle is used) to determine whether the results are suitable for the current operating conditions. If there are deviations, preliminary corrections are made to ensure the rationality and practicality of the output results, providing a reliable basis for subsequent precise control.

[0152] Furthermore, in one embodiment, the intelligent control module 50 may include: The power-on heating regulation module: Based on the calculated power and duration of key control parameters, it drives the power supply heating system to adjust the electrode output power. Employing a segmented heating control strategy, it dynamically adjusts the heating rate according to the steel temperature deviation. When the temperature deviation is large, the heating power is appropriately increased; when the temperature approaches the final temperature, the heating power is reduced to smoothly control the steel temperature to reach the final temperature, avoiding problems such as localized overheating and temperature overshoot.

[0153] Alloy Addition Adjustment Module: Based on precise calculations of the type and amount of alloy to be added, the module controls the alloy feeding system to add alloys in batches and at a uniform rate. Combining the characteristics of the furnace's thermal circulation, the module rationally plans the time intervals between each batch of alloy additions to ensure that the alloys are fully dissolved and evenly diffused after addition, avoiding compositional segregation caused by alloy agglomeration. This precisely controls the composition of the molten steel, achieving dual compliance with standards for both composition and temperature.

[0154] Furthermore, in one embodiment, the machine learning-based intelligent control device for steel refining further includes a closed-loop intelligent control module 60, which comprises: Feedback Control Module: This module collects real-time process data such as molten steel temperature, composition changes, and oxygen content, as well as updated equipment operating status data and smelting performance feedback. It calculates the difference between the current molten steel state (composition, temperature) and the target state, generates an error signal, and feeds it back to the intelligent control model. The model dynamically adjusts control parameters such as power supply and alloy addition based on the error signal, achieving closed-loop correction of control commands and maintaining stable molten steel conditions.

[0155] Real-time verification module: This module integrates data from various sensors and equipment status monitoring units to perform real-time verification of the molten steel condition, equipment operating status, and control command execution, ensuring that all process parameters are within preset ranges. If abnormal parameters are detected (such as excessive temperature deviation or insufficient alloy addition), an early warning signal is immediately issued, and the control system is driven to adjust rapidly to prevent the abnormal conditions from escalating and to ensure refining quality.

[0156] Furthermore, in one embodiment, the machine learning-based intelligent control device for molten steel refining further includes a system integration module 70, which may include: Industrial PC Control Platform: This platform integrates sensor-acquired data, preprocessing results, model output parameters, and control commands, possessing high-performance computing and real-time monitoring capabilities. Equipped with an industrial-grade operating system, the platform supports real-time display of core data throughout the entire process, as well as the storage and retrieval of historical data. A graphical interface presents information such as the state of molten steel, equipment operation status, and the progress of control parameter execution, facilitating real-time monitoring and operational intervention by users.

[0157] The data association and traceability module is used to associate and bind real-time monitoring data, historical control data, process curves, smelting performance data, production batches, steel grades, refining quality data, and equipment operation records to establish a full-process data traceability system. When refining quality problems occur, this module can be used to trace the process parameters and control processes of the corresponding batch, locate the cause of the problem, and provide data support for model parameter optimization.

[0158] Model Iteration and Optimization Module: Based on historical smelting performance and real-time operational data feedback, the module continuously iterates and adjusts the model's hyperparameters to optimize prediction accuracy and response speed. By combining smelting data from different steel grades and operating conditions, the module continuously enriches the benchmark dataset, improves the model's adaptability to complex operating conditions, and gradually optimizes the level of automated control in steel refining.

[0159] This embodiment achieves real-time and accurate assessment of the composition and temperature of molten steel refining process and intelligent control of multiple processes by collecting multi-source process data and constructing an intelligent control model that integrates gradient boosting trees and recurrent neural networks. It effectively solves the problems of traditional manual control relying on experience, adjustment lag, and poor coordination, significantly improves the consistency of molten steel quality and the accuracy of narrow composition control, optimizes the matching relationship between electric heating and alloy addition, reduces energy consumption and alloy waste, reduces the intensity of manual intervention, and enhances the stability of production rhythm and the traceability of full-process data.

[0160] The intelligent control device for molten steel refining based on machine learning provided in this application employs the intelligent control method for molten steel refining based on machine learning in the above embodiments, and can solve the technical problems of intelligent control of molten steel refining based on machine learning. Compared with the prior art, the beneficial effects of the intelligent control device for molten steel refining based on machine learning provided in this application are the same as the beneficial effects of the intelligent control method for molten steel refining based on machine learning provided in the above embodiments, and other technical features in the intelligent control device for molten steel refining based on machine learning are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0161] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0162] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0163] In addition, for technical details not described in detail in this embodiment, please refer to the intelligent control method for steel refining based on machine learning provided in any embodiment of the present invention, which will not be repeated here.

[0164] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0165] It should be noted that the user information (including but not limited to user device information, user personal information, user location information, user behavior information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0166] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0167] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0168] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A machine learning-based intelligent control method for steel refining, characterized in that, The machine learning-based intelligent control method for steel refining includes: Collect multi-source process data for the entire steel refining process, including equipment operating status data, production plan data, basic process data, and smelting performance data. The multi-source process data is preprocessed to obtain full-process data; Feature annotation is performed on the entire process data to construct a benchmark dataset, and the pre-constructed original model is trained based on the benchmark dataset to obtain a trained multi-task intelligent control model for molten steel refining. The original model is constructed based on a gradient boosting tree model and a recurrent neural network model. The benchmark dataset includes a training set, a test set, and a validation set. The real-time collected and pre-processed full-process data is input into the steel refining multi-task intelligent control model to obtain the steel condition assessment results. The steel condition assessment results include the steel composition deviation value, the steel temperature deviation value, the preliminary power reference value, and the alloy addition reference value. A control command is sent to the industrial control unit to cause the industrial control unit to drive the power supply heating system and the alloy feeding system to execute the control command. The control command includes key control parameters, which are determined based on the steel condition assessment results.

2. The intelligent control method for steel refining based on machine learning as described in claim 1, characterized in that, The preprocessing of the multi-source process data to obtain full-process data includes: Gaussian filtering is applied to the steel temperature data, steel composition data, and oxygen determination data in the multi-source process data to obtain filtered process data. The equipment location data, molten steel weight data, and molten steel temperature data in the filtered process data are normalized to obtain normalized process data. Based on historical data of the same steel grade, the missing data items in the normalized process data are supplemented by linear interpolation to obtain complete process data; The complete process data is associated and bound with the production batch information in the production plan data to obtain associated process data; The associated process data is filtered based on a preset effective sample threshold to remove abnormal operating condition data, thus obtaining the full process data.

3. The intelligent control method for steel refining based on machine learning as described in claim 1, characterized in that, The process of annotating the entire process data with features, constructing a benchmark dataset, and training the pre-constructed original model based on the benchmark dataset to obtain a trained multi-task intelligent control model for steel refining includes: The core process features in the entire process data are labeled with multi-label features, and the labeled entire process data is divided into training set, test set and validation set. The training set includes structured process feature data and time-series process feature data. The gradient boosting tree model is iteratively trained based on the structured process feature data in the training set to obtain the composition temperature prediction sub-model. The recurrent neural network model is iteratively trained based on the time-series process feature data in the training set to obtain a dynamic trend prediction sub-model. Based on the validation set, the output weights of the component temperature prediction sub-model and the dynamic trend prediction sub-model are optimized and adjusted to obtain the fusion weight parameters; A multi-task intelligent control model for steel refining is constructed based on the fusion weight parameters, the composition temperature prediction sub-model, and the dynamic trend prediction sub-model.

4. The intelligent control method for steel refining based on machine learning as described in claim 1, characterized in that, The process involves inputting the real-time collected and preprocessed full-process data into the multi-task intelligent control model for steel refining to obtain steel condition assessment results, including: The real-time process data from the real-time collected and preprocessed full-process data is input into the steel refining multi-task intelligent control model, and the predicted values ​​of steel composition and steel temperature are output. The difference between the predicted steel composition value and the target composition value in the production plan data is calculated to obtain the steel composition deviation value; The difference between the predicted molten steel temperature and the final temperature value in the production plan data is calculated to obtain the molten steel temperature deviation value; Based on the deviation values ​​of the molten steel composition and the deviation values ​​of the molten steel temperature, a preset process parameter mapping table is consulted to obtain the initial power reference value and the alloy addition reference value. Based on the slag turning data and the number of times the large ladle is used in the basic process data, the reference values ​​of the initial power supply and the alloy addition amount are corrected to obtain correction parameters. Based on the correction parameters, the steel refining process and equipment operating status are evaluated to obtain the steel condition evaluation results.

5. The intelligent control method for steel refining based on machine learning as described in claim 1, characterized in that, Sending control commands to the industrial control unit includes: Acquire steel ladle structure data and energized heating electrode layout data; Based on the steel condition assessment results, the ladle structure data, and the electric heating electrode layout data, the process area is divided into zones to determine the boundary range of the electric heating heat-affected zone and the boundary range of the alloy diffusion zone. Based on the boundary range of the heat-affected zone, the boundary range of the alloy diffusion zone, the steel composition deviation value, and the steel temperature deviation value, key control parameters are calculated. The key control parameters include the type of alloy added, the amount of alloy added, the power of the heating process, and the duration of the heating process. Based on the key control parameters, control commands are generated and sent to the industrial control unit.

6. The intelligent control method for steel refining based on machine learning as described in claim 5, characterized in that, The calculation of key control parameters based on the boundary range of the heat-affected zone, the boundary range of the alloy diffusion zone, the steel composition deviation, and the steel temperature deviation includes: The types of alloying elements to be added are determined based on the steel composition deviation value and the boundary range of the alloy diffusion zone, thus obtaining the alloy addition types; The amount of alloy added is calculated based on the type of alloy added, the deviation value of the molten steel composition, the weight data of molten steel in the basic data of the process, and the preset alloy yield parameters. Based on the steel temperature deviation value, the boundary range of the heat-affected zone during energization and heating and the preset heating rate parameter, the electrode output power value is calculated to obtain the energization heating power. The duration of energization is calculated based on the steel temperature deviation value, the energizing power, and the preset thermal efficiency parameters to obtain the energization time. The key control parameters are the combination of the type of alloy added, the amount of alloy added, the power of energizing and heating, and the duration of energizing.

7. The intelligent control method for steel refining based on machine learning as described in claim 5, characterized in that, The process area zoning judgment is performed based on the steel condition assessment results, the ladle structure data, and the electrode layout data for energized heating, to obtain the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone for energized heating, including: Based on the ladle structure data, the geometric center coordinates and inner wall radius of the ladle are determined to obtain the ladle spatial parameters. Based on the electrode layout data for power-on heating, the coordinates of the electrode distribution position are determined, and the electrode spatial parameters are obtained. The boundary range of the heat-affected zone is obtained by defining the area centered on the coordinates of the electrode distribution location and using a preset heat-affected radius as the boundary. The boundary range of the alloy diffusion zone is obtained by defining a region centered on the geometric center coordinates of the ladle and using a preset diffusion radius as the boundary. Based on the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone, a process area partitioning judgment result is generated, which includes the boundary range of the heat-affected zone and the boundary range of the alloy diffusion zone.

8. A machine learning-based intelligent control device for molten steel refining, applying the machine learning-based intelligent control method for molten steel refining as described in any one of claims 1 to 7, characterized in that, The device includes: The data acquisition module is used to collect multi-source process data for the entire steel refining process. The multi-source process data includes equipment operating status data, production plan data, basic process data, and smelting performance data. The data processing module is used to preprocess the multi-source process data to obtain full-process data; The model training module is used to annotate the features of the entire process data, construct a benchmark dataset, and train the pre-constructed original model based on the benchmark dataset to obtain a trained multi-task intelligent control model for steel refining. The original model is constructed based on a gradient boosting tree model and a recurrent neural network model. The benchmark dataset includes a training set, a test set, and a validation set. The status assessment module is used to input the real-time collected and pre-processed full-process data into the steel refining multi-task intelligent control model to obtain the steel status assessment results. The steel status assessment results include steel composition deviation value, steel temperature deviation value, preliminary power reference value, and alloy addition reference value. The intelligent control module is used to send control commands to the industrial control unit so that the industrial control unit drives the power supply heating system and the alloy feeding system to execute the control commands. The control commands include key control parameters, which are determined based on the steel condition assessment results.

9. A machine learning-based intelligent control device for steel refining, characterized in that, The machine learning-based intelligent control device for molten steel refining includes: a memory, a processor, and a machine learning-based intelligent control program for molten steel refining stored in the memory. The processor is used to run the machine learning-based intelligent control program for molten steel refining, and the machine learning-based intelligent control program for molten steel refining is configured to implement the machine learning-based intelligent control method for molten steel refining as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a machine learning-based intelligent control program for molten steel refining, which, when executed by a processor, implements the machine learning-based intelligent control method for molten steel refining as described in any one of claims 1 to 7.

Images