Artificial intelligence-based converter steelmaking endpoint control method and system

Abstract

The application discloses a converter steelmaking endpoint control method and system based on artificial intelligence, and relates to the technical field of converter steelmaking. Historical smelting data is acquired, data cleaning and feature engineering processing are performed, key input variables are screened out, and a training sample set is constructed; a endpoint parameter prediction model is constructed by using the training sample set, and the hyperparameters of the model are optimized; during the blowing process, real-time process data is input into the endpoint parameter prediction model in a cycle to obtain predicted carbon content and predicted temperature; at the end of smelting, real-time carbon content, real-time temperature, oxygen blowing amount and coolant amount are calculated in combination with a furnace mouth flame spectrum image and an audio signal to perform a supplementary blowing operation; a molten steel sample is collected, and if the test result does not hit the target range, the supplementary blowing operation is performed again; the above operation is repeated until the test result hits the target range; and the parameters of the endpoint parameter prediction model are updated online according to actual smelting data of the present smelting. The application improves the precision and stability of converter steelmaking endpoint control.

Description

Technical Field

[0001] This invention relates to the field of converter steelmaking technology, specifically to an artificial intelligence-based method and system for controlling the endpoint of converter steelmaking. Background Technology

[0002] Converter steelmaking is one of the most critical processes in modern steel production. Its endpoint control (i.e., the hit rate of endpoint carbon content and endpoint temperature) directly determines the quality of molten steel, smelting cycle, alloy consumption, and production costs. Traditional converter steelmaking endpoint control mainly relies on static model calculations and manual experience judgment.

[0003] In current industrial practice, preliminary calculations are typically performed using static models based on material and heat balance, followed by a single inspection at the end of the blowing process using a secondary lance. Based on the inspection results, manual or semi-automatic supplementary blowing is then performed. However, with increasingly stringent requirements for steel quality and faster production pace, existing control methods have revealed the following significant technical shortcomings:

[0004] First, the models suffer from poor generalization ability and weak anti-interference capability. Specifically, traditional static models are usually based on chemical reaction equations under ideal conditions, with relatively fixed parameter settings. However, in actual production, the composition of molten iron (such as silicon, phosphorus, and sulfur content) fluctuates greatly, the scrap steel ratio changes frequently, and the oxygen supply flow is affected by pipeline pressure fluctuations, resulting in highly nonlinear decarburization and heating rates. Existing BP neural network or standard support vector machine (SVM) models are prone to getting trapped in local optima when processing such multivariate, strongly coupled process data, and their prediction accuracy for abnormal conditions (such as splashing and re-drying) drops sharply, failing to accurately capture subtle changes before the endpoint.

[0005] Secondly, data utilization is low and feature extraction is crude. Although converters are equipped with a large number of sensors (such as oxygen lance flow meters, auxiliary lance temperature and carbon analyzers, and furnace gas analyzers), most existing systems only utilize simple scalar data (such as flow rate and temperature values). For unstructured data containing rich thermodynamic information, such as the spectral images of the furnace flame and the audio signals of the furnace explosions, existing technologies lack effective processing methods. Relying solely on visual observation of flame color or simple threshold judgments cannot quantify the nonlinear mapping relationship between flame spectral characteristics and the carbon content of the molten pool, resulting in a significant waste of valuable process information.

[0006] Secondly, the model parameter optimization is inefficient and lacks adaptability. When constructing the prediction model, the selection of key hyperparameters (such as penalty coefficients and kernel function parameters) often relies on manual trial and error or simple grid search, which is computationally intensive and difficult to find the global optimum. More importantly, converter operating conditions undergo "conceptual drift" with changes in furnace service life, season, and raw material batches. Existing model update mechanisms typically employ "overlay" retraining, that is, directly overwriting the old model with new data. This leads to the model forgetting key knowledge under historical operating conditions (catastrophic forgetting), resulting in poor prediction stability for new furnace cycles.

[0007] The above-mentioned technical defects result in low resource utilization and poor accuracy and stability of endpoint control.

[0008] The information disclosed in this background section is only intended to enhance the understanding of the background technology of this application, and therefore may include prior art that is not known to those skilled in the art. Summary of the Invention

[0009] To address the issues of low resource utilization and poor accuracy and stability of endpoint control in existing technologies, this paper proposes an artificial intelligence-based endpoint control method and system for converter steelmaking. This method integrates heterogeneous data from multiple sources during the converter steelmaking process with artificial intelligence algorithms to improve the utilization rate of various sensors, reduce resource waste, and enhance the accuracy and stability of endpoint control.

[0010] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0011] An AI-based method for controlling the endpoint of converter steelmaking includes the following steps:

[0012] Acquire historical smelting data of the converter steelmaking process, including molten iron composition, scrap steel quantity, oxygen supply flow rate, temperature and carbon content detected by the auxiliary lance, and flame spectrum images and audio signals at the furnace mouth;

[0013] The historical smelting data was cleaned and feature-engineered. Key input variables were selected through independence analysis and correlation analysis to construct a training sample set.

[0014] The non-parallel support vector regression machine algorithm based on wavelet weights uses the training sample set to construct an endpoint parameter prediction model, and introduces a whale swarm optimization algorithm to optimize the hyperparameters of the training model, which is used to predict the carbon content and temperature at the endpoint of converter steelmaking.

[0015] During the blowing process, real-time process data of the converter steelmaking process is acquired and input into the endpoint parameter prediction model to continuously acquire the predicted carbon content and predicted temperature of the smelting endpoint.

[0016] When the smelting process is predicted to be in its final stage, principal component analysis is performed on the flame spectrum image at the furnace opening for dimensionality reduction. Combined with the furnace sound characteristics obtained from the audio signal analysis, the real-time carbon content and real-time temperature are calculated. The material balance calculated by the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the static model is used to calculate the required oxygen and cooling dosage for supplementary blowing using a dynamic programming algorithm, and the supplementary blowing operation is then performed.

[0017] When the replenishment blowing operation is completed, a sample of molten steel is collected for testing. If the test results do not fall within the target range, the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the material balance are obtained to determine the required oxygen and cooling dosage for replenishment blowing. The replenishment blowing operation is then performed again. The above operation is repeated until the test results fall within the target range.

[0018] The actual smelting data for this furnace is obtained, stored in a production database, and the parameters of the endpoint parameter prediction model are updated online using an elastic weight solidification algorithm based on a self-learning module.

[0019] In some specific embodiments, the historical smelting data undergoes data cleaning and feature engineering processing. Key input variables are selected through independence and correlation analysis to construct a training sample set, including:

[0020] use The criteria remove outliers from historical smelting data, including abrupt changes in oxygen supply flow and temperature jumps.

[0021] Missing values ​​are filled in the historical smelting data after removing the outliers. This is done by interpolating the missing non-iron content data using the K-nearest neighbor algorithm to obtain a sample of historical smelting data.

[0022] The oxygen flow rate time series data during the supplementary blowing stage is denoised using wavelet transform theory. Based on the denoising results, the signal confidence of the historical smelting data samples is quantified, and the weight vector of the historical smelting data samples is determined. The weight vector serves as the weight of the historical smelting data samples for training the endpoint parameter prediction model; n is the sample sequence length; weight coefficients. The calculation is performed using the Gaussian function, specifically as follows:

[0023] ;

[0024] in, For the first The Euclidean distance between the original data points and the denoised data; The width of the Gaussian kernel;

[0025] The historical smelting data also includes process parameters such as the amount of molten iron charged, the scrap steel ratio, the total oxygen supply, and the amount of lime added. These, along with the molten iron composition and the oxygen supply flow rate, constitute a candidate input variable set. An independence analysis is then performed on this candidate input variable set.

[0026] The variance inflation factor (VIF) method is used to test the degree of multicollinearity among the process parameters in the candidate input variable set, and the variance inflation factor is calculated.

[0027] ;

[0028] in, To make the first The determination coefficients are obtained by performing linear regression with one of the process parameters as the dependent variable and the remaining process parameters as independent variables; an independence judgment threshold is set to exclude... The process parameters that exceed the independence judgment threshold;

[0029] The correlation between each process parameter and the endpoint carbon content and endpoint temperature was analyzed using the Pearson correlation coefficient matrix. Process parameters with an absolute correlation coefficient greater than 0.6 were selected as the key input variables.

[0030] In some specific embodiments, the wavelet-weighted non-parallel support vector regression algorithm uses the training sample set to construct an endpoint parameter prediction model, and introduces a whale swarm optimization algorithm to optimize the hyperparameters of the training model, including:

[0031] Initialize the parameters of the non-parallel support vector regression machine and Set the kernel function to a radial basis function; split the training sample set into a training subset and a validation subset;

[0032] Construct an optimization solution framework for the non-parallel support vector regression machine. This framework consists of an objective function containing a weighted empirical risk term and a structural risk term, and the deviation between the predicted and actual sample values ​​does not exceed a certain limit. The constraints are constituted; the weighted empirical risk term is introduced into the weight vector, and the fitting slack variables of the samples of each training subset are weighted according to the signal confidence.

[0033] The training subset is input into the optimization framework, and the optimal solution vector is obtained by solving the dual problem. ;

[0034] The regression function is Wherein, ω is the model weight vector of the non-parallel support vector regression machine in the high-dimensional feature space, which is derived from the optimal solution vector; It is a nonlinear mapping function; For bias terms;

[0035] The root mean square error (RMSE) between the predicted and actual values ​​of the endpoint parameter prediction model on the validation subset is used as the fitness function to iteratively update the population position vector. The fitness function is calculated as follows:

[0036] ;

[0037] in, The total number of samples in the verification subset; For the first The actual endpoint carbon content or temperature value of each sample; For the first The model prediction value for each sample;

[0038] The whale swarm optimization algorithm is used to search within the parameter space of the non-parallel support vector regression machine until the fitness function value is less than a preset threshold or the maximum number of iterations is reached. .

[0039] In some specific embodiments, principal component analysis is performed on the flame spectral image at the furnace opening for dimensionality reduction. Combined with the furnace acoustic characteristics obtained from the audio signal analysis, the real-time carbon content and real-time temperature are calculated; including:

[0040] Acquire the flame spectrum image at the furnace opening, with a wavelength range covering 400nm to 1000nm;

[0041] Extract the target regions of the furnace mouth flame core area and furnace mouth flame edge area from the furnace mouth flame spectral image;

[0042] Principal component analysis (PCA) was used to reduce the dimensionality of the target region, and the eigenvalues ​​of its covariance matrix were calculated. Before selection Principal components with a cumulative contribution rate greater than 95% are selected as the spectral images to be processed. The value was determined through cross-validation; the cumulative contribution rate was calculated using the following formula:

[0043] ;

[0044] in, For the front The cumulative contribution rate of each principal component; For the first One eigenvalue; This represents the total number of original spectral bands. For the first One eigenvalue;

[0045] The spectral image to be processed is input into a deep convolutional neural network. Flame texture features are extracted through multiple convolutional layers, and the predicted value of real-time carbon content is output by a fully connected layer. The ReLU function is used as the activation function of the network.

[0046] In some specific embodiments, the material balance calculated from the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the static model is used to calculate the required oxygen and cooling dosage for supplementary blowing using a dynamic programming algorithm, and the supplementary blowing operation is then performed; including:

[0047] Establish supplemental oxygen volume Carbon content detected by the secondary gun Target carbon content and the amount of coolant added The nonlinear mapping relationship between them was fitted using a three-layer BP neural network, and the number of hidden layer nodes was determined to be 7 through trial and error.

[0048] Enter the current molten pool temperature With target temperature The difference Based on the current oxygen content of the molten steel Theoretical oxygen demand is calculated using thermodynamic equilibrium equations. :

[0049] ;

[0050] in, For the quality of molten steel; The measured carbon content of the current secondary gun; The target carbon content; The enthalpy change is due to the carbon oxidation reaction; Oxygen utilization coefficient; Current oxygen content; The decarbonization efficiency coefficient; The coefficient representing the effect of temperature on decarbonization oxygen demand is obtained by regression analysis of historical best furnace data.

[0051] The initial estimated amount of supplemental oxygen is calibrated using the theoretical oxygen demand to obtain the baseline oxygen supply for supplemental blowing.

[0052] An incremental calculation method is introduced, with a control cycle of 3 seconds, to synchronously acquire the real-time carbon content calculated by combining the dimensionality reduction features of the furnace mouth flame spectral image with the sound features inside the furnace. Real-time temperature After dynamically correcting for carbon content and temperature deviations, the oxygen flow correction value for the current cycle is calculated. ;

[0053] First, construct a comprehensive carbon content deviation. Overall temperature deviation :

[0054] ;

[0055] ;

[0056] in, The weighting coefficient for the predicted value ranges from 0.4 to 0.6, gradually decreasing as the reblowing process progresses, while the weighting of the real-time detected value gradually increases with the refining depth.

[0057] The formula for calculating the oxygen flow correction value is:

[0058] ;

[0059] in, To predict carbon content deviation; To predict temperature deviation; The predicted carbon content; To predict temperature; This is the carbon content correction factor. The temperature correction coefficients are all obtained through regression analysis of historical best furnace data.

[0060] The oxygen flow correction value The actual oxygen supply flow command for the current control cycle is generated by superimposing the supplementary oxygen supply benchmark amount in real time.

[0061] Coolant addition amount Based on the heat balance calculation of the molten pool:

[0062] ;

[0063] in, The total amount of heat required to reach the target temperature; This is the measured enthalpy of the current molten pool; Specific heat capacity of the coolant; This is the thermal efficiency coefficient.

[0064] In some specific embodiments, an abnormal operating condition intervention step is also included, which includes:

[0065] Real-time monitoring of converter ring vibration signals Exhaust gas temperature;

[0066] When the vibration signal energy spectral density exceeds the set threshold and the exhaust gas temperature drops by more than 50°C within 10 seconds, it is determined to be a precursor to splashing. In this case, the oxygen lance position is lowered to the ready-to-blown position, and the bottom nitrogen stirring flow rate is increased until the changes in the energy spectral density and the exhaust gas temperature return to normal.

[0067] The vibration signal energy spectral density Calculated using the following formula:

[0068] ;

[0069] in, The result is the Fast Fourier Transform of the vibration signal; and These are the set lower and upper limits for the frequency, respectively.

[0070] Exhaust gas temperature variation range Calculated using the following formula:

[0071] ;

[0072] in, The current exhaust gas temperature; The exhaust gas temperature was 10 seconds ago.

[0073] In some specific embodiments, at the end of the blowing process, a sample of molten steel is collected for testing. If the test results do not fall within the target range, the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the material balance are obtained. The required oxygen and cooling dosage for supplementary blowing are calculated, and the supplementary blowing operation is performed again. This process is repeated until the test results fall within the target range. This includes:

[0074] Obtain laboratory data from the molten steel sample after the blowing process is complete, including the actual carbon content. Actual temperature ;

[0075] Calculate the prediction error of the endpoint parameter prediction model. , ;

[0076] like or If the error is not found, the endpoint parameter prediction model is determined to be mismatched. Based on the test error, the trainable parameters of the endpoint parameter prediction model are corrected online. The trainable parameters include the regression hyperplane weight vector and bias term. Incremental gradient descent is used to adjust the model parameters, and the correction amount is... The calculation formula is:

[0077]

[0078] in, The learning rate ranges from 0.01 to 0.1. For loss function, , The test data for the molten steel sample. Predicted data for the endpoint; This is the current weight coefficient vector.

[0079] In some specific embodiments, the parameters of the endpoint parameter prediction model are updated online using an elastic weight fixation algorithm based on the self-learning module, including:

[0080] The Domain Adversarial Neural Network (ADDA) algorithm in transfer learning is used to align the features of historical smelting data with the target domain data of the current furnace.

[0081] Virtual smelting data is generated using the WGANGP generative adversarial network, and the extreme working condition sample library is expanded by combining feature space interpolation.

[0082] The Fisher information matrix F is calculated based on the Elastic Weight Solidification (EWC) algorithm, and a regularization term is added when updating the parameters of the endpoint parameter prediction model; the objective function... for:

[0083] ;

[0084] in, The loss function for the new furnace data; This is a dynamic regularization coefficient that decreases as the amount of new data samples increases. For the first Fisher information content of each parameter; These are the current model parameters; These are the historically optimal model parameters.

[0085] In some specific embodiments, it also includes:

[0086] After the blowing process is completed, the difference between the target composition and the current composition of the molten steel is obtained;

[0087] The minimum cost of each alloy is calculated using the differences, historical data on alloy yield, and a linear programming model. A heuristic algorithm is used to solve the objective function. The constraints are:

[0088] ;

[0089] in, For the first The unit price of the alloy; For the first The amount of each alloy added; This represents the number of alloy types. For the first The target component content of each element; The allowable deviation range for components; For the current steel molten steel Content of various elements; For the first The first alloy Predicted yield values ​​for each element; For the first The first alloy The inherent content ratio of each element;

[0090] An alloy yield prediction model was constructed using historical data on alloy yield, with input variables including tapping temperature. Oxygen content in molten steel Alloy block size The order in which elements are added determines the output variable, which is the predicted yield value for each element.

[0091] An artificial intelligence-based precision control system for the endpoint of converter steelmaking, used to implement the aforementioned endpoint control method for converter steelmaking, includes:

[0092] The acquisition module acquires real-time process data and historical smelting data of the converter steelmaking process, including molten iron composition, scrap steel quantity, oxygen supply flow rate, molten pool temperature and carbon content detected by the auxiliary lance, furnace mouth flame spectral image and audio signal.

[0093] The data processing module performs data cleaning and feature engineering on historical smelting data, selects key input variables through independence analysis and correlation analysis, and constructs a training sample set.

[0094] The model building module, based on the wavelet weighted non-parallel support vector regression machine algorithm, uses the training sample set to build an endpoint parameter prediction model, and introduces the whale swarm optimization algorithm to optimize the hyperparameters of the endpoint parameter prediction model, which is used to predict the carbon content and temperature at the endpoint of converter steelmaking;

[0095] The prediction module inputs the real-time process data into the endpoint parameter prediction model during the blowing process, and continuously acquires the predicted carbon content and predicted temperature.

[0096] The first calculation module, when predicting that the smelting is entering the final stage, uses an AI visual model to perform principal component analysis and dimensionality reduction on the furnace mouth flame spectrum image, and combines an AI audio model to analyze the sound characteristics inside the furnace, and calculates the real-time carbon content and real-time temperature.

[0097] The second calculation module uses a dynamic programming algorithm to calculate the amount of oxygen and coolant required for the supplementary blowing stage based on the predicted carbon content, predicted temperature, real-time carbon content, real-time temperature, and material balance calculated by the static model.

[0098] The correction module performs supplementary blowing based on the oxygen blowing amount and coolant addition amount obtained from the second calculation module, and collects molten steel samples for testing at the end of the blowing process; if the test results do not hit the target range, the supplementary blowing operation is repeated until the test results hit the target range.

[0099] The model update module acquires the actual smelting data for this furnace, stores it in a production database, and updates the parameters of the endpoint parameter prediction model online using an elastic weight solidification algorithm based on the self-learning module.

[0100] Compared with the prior art, the advantages and positive effects of the present invention are:

[0101] This invention utilizes a wavelet-weighted non-parallel support vector regression machine and a swarm optimization algorithm to construct an endpoint parameter prediction model, thereby improving the accuracy of predicted carbon content and temperature. By processing secondary gun data and flame spectral and audio characteristics using Kalman filtering, high-precision calculations of real-time carbon content and temperature are achieved. Combining the highly accurate predicted carbon content and temperature with high-precision real-time data, a dynamic programming algorithm is used to calculate the supplementary blowing amount and coolant addition amount, improving the accuracy of these amounts and thus enhancing the accuracy and stability of endpoint parameter control. Furthermore, an elastic weighting solidification algorithm and transfer learning are employed to achieve online self-updating of the model, adapting to new operating conditions while avoiding catastrophic forgetting, ensuring the long-term robustness of the endpoint parameter prediction model. Attached Figure Description

[0102] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0103] Figure 1 This is a schematic flowchart of the converter steelmaking endpoint control method according to an embodiment;

[0104] Figure 2 This is a schematic diagram of the functional modules of the converter steelmaking endpoint control system according to an embodiment;

[0105] Figure 3 This is a schematic diagram of the data processing flow according to an embodiment;

[0106] Figure 4 This is a flowchart illustrating the process of constructing an endpoint parameter prediction model according to an embodiment. Detailed Implementation

[0107] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0108] Reference Figure 1 This invention discloses an artificial intelligence-based endpoint control method for converter steelmaking, which includes the following steps:

[0109] S101. Obtain historical smelting data of the converter steelmaking process, including molten iron composition, scrap steel quantity, oxygen supply flow rate, temperature and carbon content detected by the auxiliary lance, and flame spectrum image and audio signal at the furnace mouth; the temperature and carbon content detected by the auxiliary lance are the temperature and carbon content of the molten pool.

[0110] S102. Perform data cleaning and feature engineering on historical smelting data, and select key input variables through independence analysis and correlation analysis to construct a training sample set;

[0111] S103. A non-parallel support vector regression machine algorithm based on wavelet weights is used to construct an endpoint parameter prediction model using a training sample set, and the hyperparameters of the endpoint parameter prediction model are optimized using a whale swarm optimization algorithm. The endpoint parameter prediction model is used to predict the carbon content and temperature of the molten pool at the endpoint of converter steelmaking.

[0112] S104. During the blowing process, real-time process data of the converter steelmaking process is acquired and input into the prediction model, and the predicted carbon content and predicted temperature at the smelting endpoint are acquired in a rolling manner; the real-time process data includes the composition of molten iron, the amount of scrap steel, the oxygen supply flow rate, the temperature and carbon content detected by the auxiliary lance, and the flame spectrum image and audio signal at the furnace mouth.

[0113] S105. When it is predicted that the smelting has entered the final stage, principal component analysis is performed on the flame spectrum image at the furnace mouth for dimensionality reduction. Combined with the furnace sound characteristics obtained from the audio signal analysis, the real-time carbon content and real-time temperature are calculated. The material balance calculated by the predicted carbon content, predicted temperature, real-time carbon content, real-time temperature and static model is used to calculate the oxygen blowing amount and cooling amount required for supplementary blowing using a dynamic programming algorithm, and the supplementary blowing operation is executed. Specifically, the final stage of smelting is determined by setting carbon content threshold and temperature threshold and comparing the predicted carbon content with the carbon content threshold and the predicted temperature with the temperature threshold. That is, when the predicted carbon content is within the carbon content threshold and the predicted temperature is within the temperature threshold, it is determined that the smelting has entered the final stage.

[0114] S106. When the replenishment blowing operation ends, a sample of molten steel is collected for testing. If the test results do not fall within the target range, S105 is executed to obtain the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the material balance. The required amount of oxygen and cooling agent for replenishment blowing is obtained, and the replenishment blowing operation is executed again. The above operation is repeated until the test results fall within the target range.

[0115] S107. Obtain the actual smelting data for this furnace, store it using a production database, and update the parameters of the endpoint parameter prediction model online using an elastic weight solidification algorithm based on the self-learning module.

[0116] Reference Figure 2 The present invention also discloses an artificial intelligence-based converter steelmaking endpoint control system, comprising an acquisition module, a data processing module, a model building module, a prediction module, a first calculation module, a second calculation module, a correction module, and a model update module; used to implement the above-mentioned artificial intelligence-based converter steelmaking endpoint control method.

[0117] The acquisition module is used to acquire real-time process data and historical smelting data of the converter steelmaking process, which include molten iron composition, scrap steel quantity, oxygen supply flow rate, temperature and carbon content detected by the sub-lance, flame spectrum image and audio signal at the furnace mouth; the temperature and carbon content detected by the sub-lance are the temperature and carbon content of the molten pool.

[0118] The data processing module is used to clean and feature-engineer historical smelting data, and to select key input variables through independence analysis and correlation analysis to construct a training sample set.

[0119] The model building module is based on the non-parallel support vector regression machine algorithm with wavelet weights. It uses the training sample set to build the endpoint parameter prediction model and introduces the whale swarm optimization algorithm to optimize the hyperparameters of the endpoint parameter prediction model. The endpoint parameter prediction model is used to predict the carbon content and temperature at the endpoint of converter steelmaking; that is, to obtain the predicted carbon content and predicted temperature.

[0120] During the blowing process, the prediction module inputs real-time process data into the endpoint parameter prediction model and continuously obtains the predicted carbon content and predicted temperature.

[0121] The first calculation module uses an AI visual model to perform principal component analysis and dimensionality reduction on the flame spectrum image at the furnace opening, and combines it with an AI audio model to analyze the sound characteristics inside the furnace, and calculate the real-time carbon content and real-time temperature.

[0122] The second calculation module uses the predicted carbon content, predicted temperature, real-time carbon content, real-time temperature, and static model to calculate the material balance and then uses a dynamic programming algorithm to calculate the amount of oxygen and coolant required for the supplementary blowing stage.

[0123] When the smelting process is predicted to enter its final stage, the second calculation module iteratively calculates the required oxygen blowing amount and coolant addition amount for the supplementary blowing stage using a dynamic programming algorithm based on the predicted carbon content, predicted temperature, real-time carbon content, real-time temperature, and material balance calculated by the static model. The correction module performs supplementary blowing operations based on the oxygen blowing amount and coolant addition amount obtained from the second calculation module, and collects molten steel samples for testing at the end of the blowing process. If the test results do not fall within the target range, the module iteratively obtains the oxygen blowing amount and coolant addition amount calculated by the second calculation module and performs supplementary blowing operations until the test results fall within the target range.

[0124] The model update module acquires the actual smelting data for this furnace, stores it in a production database, and updates the parameters of the endpoint parameter prediction model online using an elastic weight solidification algorithm based on the self-learning module.

[0125] The artificial intelligence-based converter steelmaking endpoint control method and system of this invention improves the accuracy of predicted carbon content and temperature by constructing an endpoint parameter prediction model using wavelet weighted non-parallel support vector regression machine and whale swarm optimization algorithm. Through Kalman filtering of secondary lance data and flame spectrum and audio characteristics, high-precision calculation of real-time carbon content and temperature is achieved. Combining the high-accuracy predicted carbon content and temperature with the high-precision real-time carbon content and temperature, a dynamic programming algorithm is used to calculate the supplementary blowing amount and coolant addition amount, improving the accuracy of supplementary blowing and coolant addition, thereby enhancing the accuracy and stability of endpoint parameter control. Furthermore, an elastic weight solidification algorithm and transfer learning are employed to achieve online self-updating of the model, adapting to new operating conditions while avoiding catastrophic forgetting, ensuring the long-term robustness of the endpoint parameter prediction model.

[0126] The following detailed description of the specific composition, steps, and principles of the AI-based converter steelmaking endpoint control method and system of the present invention will be provided through specific embodiments.

[0127] In some specific embodiments, refer to Figure 1 , Figure 2 , Figure 3 In the endpoint control method for converter steelmaking, historical smelting data undergoes data cleaning and feature engineering. Key input variables are selected through independence and correlation analysis, and a training sample set is constructed. The steps include:

[0128] S201, Adopted The criteria remove outliers from historical smelting data, including abrupt changes in oxygen supply flow rate and temperature jumps; for example, abrupt changes in oxygen supply flow rate and temperature jumps are caused by sensor malfunctions or operational errors.

[0129] S202. Impute missing values ​​in the historical smelting data after removing outliers. This is done by using interpolation based on the K-nearest neighbor algorithm to fill in the missing non-iron content data, thus obtaining a sample of historical smelting data. For example, non-iron can be carbon, phosphorus, sulfur, etc.

[0130] S203. The oxygen flow rate time series data during the supplementary blowing stage is denoised using wavelet transform theory. Based on the denoising results, the signal confidence of each historical smelting data sample is quantified, and the weight vector of the historical smelting data sample is determined. , used for...; where n is the sample sequence length; weight coefficient The calculation is performed using the Gaussian function, specifically as follows:

[0131] ;

[0132] in, For the first The Euclidean distance between the original data points and the denoised data; The width of the Gaussian kernel;

[0133] The weight vector, as the confidence weight of historical smelting data samples, is applied to the training process of the subsequent endpoint parameter prediction model: the smaller the deviation between the original data and the denoising result and the higher the signal confidence, the greater the weight of the sample; the greater the noise disturbance in the original data and the lower the signal confidence, the smaller the weight of the sample. This is used to suppress the interference of random noise collected during the smelting process on the model fitting accuracy.

[0134] S204. Historical smelting data also includes process parameters such as hot metal charge amount, scrap ratio, total oxygen supply, and lime addition amount. These, along with hot metal composition and oxygen supply flow rate, constitute the candidate input variable set. Independence analysis is performed on the candidate input variable set.

[0135] The variance inflation factor (VIF) method was used to test the degree of multicollinearity among variables, and the variance inflation factor corresponding to each process parameter was calculated.

[0136] ;

[0137] in, To make the first The determination coefficients were obtained by performing linear regression with one process parameter as the dependent variable and the remaining process parameters as independent variables; an independence threshold was set to exclude... Redundant variables exceeding the independence threshold are retained only for process parameters that meet the independence requirements; for example, the independence threshold is set to 10.

[0138] S205. The correlation between each process parameter that meets the independence requirement and the endpoint carbon content and endpoint temperature is analyzed using the Pearson correlation coefficient matrix; parameters with an absolute value of correlation coefficient greater than 0.6 are selected as key input variables to construct a training sample set.

[0139] In the converter steelmaking endpoint control system, the data processing module implements the above method.

[0140] In some specific embodiments, refer to Figure 1 , Figure 2 , Figure 4 In the endpoint control method of converter steelmaking, a non-parallel support vector regression machine algorithm based on wavelet weights is used to construct an endpoint parameter prediction model using a training sample set, and a whale swarm optimization algorithm is introduced to optimize the hyperparameters of the training model, including:

[0141] S301. Initialize the parameters of the non-parallel support vector regression machine. and The kernel function is set to a radial basis function; the training sample set is split into a training subset and a validation subset; an optimization solution framework for a non-parallel support vector regression machine is constructed, which consists of an objective function containing a weighted empirical risk term and a structural risk term, and the deviation between the predicted and actual sample values ​​does not exceed a certain limit. The constraints together constitute the objective function; the empirical risk term of the objective function introduces the weights of the corresponding samples into the weight vector. The fitting slack variables of each sample are weighted according to the signal confidence of each sample to reduce the influence of signal confidence noise samples on the model hyperplane; for example, the training sample set is divided by the hold-out method, splitting it into a training subset and a validation subset in an 8:2 ratio; the training subset is used for model fitting of non-parallel support vector regression machine, and the validation subset is used to evaluate the generalization performance of hyperparameter combination.

[0142] S302. Input the training sample set into the optimization solution framework to train the endpoint parameter prediction model, and obtain the optimal solution vector by solving the dual problem. The regression function is: ;in, The model weight vector for the non-parallel support vector regression machine in the high-dimensional feature space regression hyperplane is derived from the optimal solution of the aforementioned dual problem; It is a nonlinear mapping function; For bias terms;

[0143] S303. Using the root mean square error (RMSE) between the predicted and actual values ​​of the endpoint parameter prediction model on the validation subset as the fitness function, iteratively update the population position vector. The fitness function is:

[0144] ;

[0145] in, The total number of samples in the validation set; For the first The actual endpoint carbon content or temperature value of each sample; For the first The model prediction value for each sample;

[0146] S304. Use the whale swarm optimization algorithm to search within the parameter space of the non-parallel support vector regression machine until the fitness function value is less than a preset threshold or the maximum number of iterations is reached. .

[0147] In the converter steelmaking endpoint control system, the model building module implements the above method.

[0148] In some specific embodiments, refer to Figure 1 , Figure 2 In the converter steelmaking endpoint control method, principal component analysis is performed on the furnace mouth flame spectrum image for dimensionality reduction. Combined with the furnace sound characteristics obtained from audio signal analysis, the real-time carbon content and real-time temperature are calculated. Specifically, the following steps are included:

[0149] Obtain the flame spectrum image of the converter furnace opening, which is a hyperspectral image data with a wavelength range covering 400nm to 1000nm;

[0150] Extract the target region from the spectral image of the furnace flame, which is the core region and edge region of the furnace flame, and remove background noise interference;

[0151] Principal component analysis (PCA) was used to reduce the dimensionality of the target region in the furnace flame spectral image, and the eigenvalues ​​of its covariance matrix were calculated. Before selection Principal components with a cumulative contribution rate greater than 95% are selected as the spectral images to be processed. The value was determined through cross-validation; the cumulative contribution rate was calculated using the following formula:

[0152] ;

[0153] in, For the front The cumulative contribution rate of each principal component; For the first One eigenvalue; This represents the total number of original spectral bands. For the first One eigenvalue;

[0154] The data of the spectral image to be processed is input into a deep convolutional neural network; flame texture features are extracted through multiple convolutional layers; the fully connected layer outputs the real-time predicted value of carbon content; the ReLU function is used as the network activation function.

[0155] In the converter steelmaking endpoint control system, the first calculation module implements the above method.

[0156] In some specific embodiments, refer to Figure 1 , Figure 2 In the converter steelmaking endpoint control method, the material balance calculated by the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the static model is used to calculate the required oxygen and cooling dosage for supplementary blowing using a dynamic programming algorithm, and then the supplementary blowing operation is performed; specifically including:

[0157] Establish supplemental oxygen volume Carbon content detection with secondary gun Target carbon content and the amount of coolant added The nonlinear mapping relationship between them was fitted using a three-layer BP neural network, with the number of hidden layer nodes determined to be 7 through trial and error. The three-layer BP neural network used samples of the supplementary blowing stage working conditions in historical smelting data as training data, and the trained network served as the carrier of the nonlinear mapping relationship. When the supplementary blowing process was started, the current carbon content detected by the secondary lance, the target carbon content, and the planned amount of coolant added were input into the network, and the initial estimated value of the supplementary blowing oxygen amount was output as the benchmark set value for supplementary blowing oxygen supply control.

[0158] Enter the current molten pool temperature With target temperature The difference Based on the current oxygen content of the molten steel Theoretical oxygen demand is calculated using thermodynamic equilibrium equations. :

[0159] ;

[0160] in, For the quality of molten steel; The carbon content of the current secondary gun is being tested; The target carbon content; The enthalpy change is due to the carbon oxidation reaction; Oxygen utilization coefficient; Current oxygen content; The decarbonization efficiency coefficient; , The coefficient of temperature on decarbonization oxygen demand is obtained by regression analysis of historical best furnace data; the initial estimated value of supplementary oxygen supply is calibrated by theoretical oxygen demand to obtain the baseline oxygen supply for supplementary blowing.

[0161] An incremental calculation method is introduced, with a control cycle of 3 seconds, to synchronously acquire the real-time carbon content calculated by combining the dimensionality reduction features of the furnace mouth flame spectral image with the sound features inside the furnace. Real-time temperature After dynamically correcting for carbon content and temperature deviations, the oxygen flow correction value for the current cycle is calculated. ;

[0162] First, construct a comprehensive carbon content deviation. Overall temperature deviation :

[0163] ;

[0164] ;

[0165] The formula for calculating the oxygen flow correction value is:

[0166] ;

[0167] in, The weighting coefficient for the predicted value ranges from 0.4 to 0.6, gradually decreasing as the reblowing process progresses, while the weighting of the real-time detected value gradually increases with the refining depth. To predict carbon content deviation; To predict temperature deviation; To predict carbon content; To predict temperature; This is the carbon content correction factor. The temperature correction coefficients are all obtained through regression analysis of historical best furnace data.

[0168] oxygen flow correction value The actual oxygen supply flow command for the current control cycle is generated by superimposing the supplementary oxygen supply benchmark amount onto the real-time oxygen supply.

[0169] Coolant addition amount Based on the heat balance calculation of the molten pool:

[0170] ;

[0171] in, The total amount of heat required to reach the target temperature; This is the measured enthalpy of the current molten pool; Specific heat capacity of the coolant; This is the thermal efficiency coefficient.

[0172] In the endpoint control of converter steelmaking, the second calculation module implements the above method.

[0173] In some specific embodiments, the converter steelmaking endpoint control method also includes abnormal operating condition intervention steps:

[0174] Real-time monitoring of converter ring vibration signals Related to the rate of change in CO concentration in the exhaust gas composition and the exhaust gas temperature;

[0175] When the vibration signal energy spectral density exceeds the set threshold and the exhaust gas temperature drops by more than 50°C within 10 seconds, it is determined to be a precursor to splashing. The oxygen gun position is then lowered to the waiting position to stop blowing oxygen, and the bottom nitrogen stirring flow rate is increased until the changes in energy spectral density and exhaust gas temperature return to normal.

[0176] Vibration signal energy spectral density Calculated using the following formula:

[0177] ;

[0178] in, The result is the Fast Fourier Transform of the vibration signal; and These are the set lower and upper limits for the frequency, respectively.

[0179] Exhaust gas temperature change rate Calculated using the following formula:

[0180] ;

[0181] in, The current exhaust gas temperature; The exhaust gas temperature was 10 seconds ago.

[0182] The converter steelmaking endpoint control system also includes an abnormal operating condition handling module, which implements the above-mentioned methods.

[0183] The converter steelmaking endpoint control method and system in this embodiment calculates the amount of supplementary blowing and coolant addition based on dynamic programming algorithm, and introduces an abnormal intervention mechanism based on vibration and exhaust gas analysis to effectively avoid splashing and back-drying, and significantly improve the endpoint hit rate.

[0184] In some specific embodiments, refer to Figure 1 , Figure 2 In the converter steelmaking endpoint control method, at the end of blowing, a sample of molten steel is collected for testing. If the test results do not fall within the target range, the predicted carbon content, predicted temperature, real-time carbon content, real-time temperature, and material balance are obtained. The required oxygen and cooling dosage for supplementary blowing are calculated, and the supplementary blowing operation is performed again. The above operation is repeated until the test results fall within the target range. Specifically, this includes:

[0185] Obtain laboratory data from the molten steel sample after the blowing process is complete, including the actual carbon content. Actual temperature and phosphorus content ;

[0186] Calculate the prediction error of the endpoint parameter prediction model , ;

[0187] like or If the error is not specified, it is determined to be a mismatch in the endpoint parameter prediction model. Based on the test error, the trainable parameters of the model are corrected online. The trainable parameters include the regression hyperplane weight vector and bias terms. The model parameters are adjusted using incremental gradient descent, with the correction amount... Calculated using the following formula:

[0188]

[0189] in, The learning rate ranges from 0.01 to 0.1. For loss function, ; This is the current weight coefficient vector;

[0190] The actual measured values ​​of the test results obtained from the analysis of molten steel samples, including the actual carbon content. Actual temperature and phosphorus content .

[0191] The predicted values ​​output by the endpoint parameter prediction model include predicted carbon content, predicted temperature, and predicted phosphorus content.

[0192] In the converter steelmaking endpoint control system, the Xu Zheng module implements the above method.

[0193] In some specific embodiments, the converter steelmaking endpoint control method uses an elastic weight solidification algorithm based on a self-learning module to update the parameters of the endpoint parameter prediction model online, specifically including:

[0194] By utilizing the Domain Adversarial Neural Network (ADDA) algorithm in transfer learning, historical smelting data and target domain data of the current furnace are feature-aligned to solve the problem of inconsistent sample distribution.

[0195] Virtual smelting data is generated using the WGANGP generative adversarial network, and the extreme working condition sample library is expanded by combining feature space interpolation.

[0196] The Fisher information matrix F is calculated based on the Elastic Weight Solidification EWC algorithm; a regularization term is added when updating model parameters to prevent catastrophic forgetting of old knowledge, and the objective function... The formula is:

[0197] ;

[0198] in, The loss function for the new furnace data; This is a dynamic regularization coefficient that decreases as the amount of new data samples increases. For the first Fisher information content of each parameter; These are the current model parameters; These are the historically optimal model parameters.

[0199] It should be noted that historical smelting data includes the characteristics of each furnace's blowing process (such as oxygen supply, amount added, time series, etc.) and the corresponding endpoint test results (carbon content, temperature, phosphorus content, etc.). This data forms the basis for the initial training of the endpoint parameter prediction model. It is abundant and fully labeled, but it reflects past raw material conditions, equipment status, and operating habits.

[0200] The target domain data refers to the data from the current heat being smelted. The raw material batches for the current heat may differ from historical data, the composition of the molten iron may fluctuate, the scrap steel ratio may change, and even seasonal temperature and humidity can affect the smelting process. These factors cause a shift between the data distribution of the current heat and historical data; this is known as domain discrepancy. By aligning historical smelting data with the target domain data of the current heat, the problem of inconsistent sample distribution is resolved, improving the availability and accuracy of training data, and consequently, enhancing the accuracy of the endpoint parameter prediction model.

[0201] In the converter steelmaking endpoint control system, the model update module implements the above method.

[0202] In some specific embodiments, the converter steelmaking endpoint control method also includes narrow composition control during the alloying stage, which includes the following steps:

[0203] After the blowing process is completed, the difference between the target composition and the current composition of the molten steel is obtained;

[0204] The minimum cost of each alloy is calculated using the differences, historical data on alloy yield, and a linear programming model. A heuristic algorithm is used to solve the objective function. The constraints are:

[0205] ;

[0206] in, For the first The unit price of the alloy; For the first The amount of each alloy added; This represents the number of alloy types. For the first The target component content of each element; The allowable deviation range for components; For the current steel molten steel Content of various elements; For the first The first alloy Predicted yield values ​​for each element; For the first The first alloy The inherent content ratio of each element;

[0207] An alloy yield prediction model was constructed using historical data on alloy yield, with input variables including tapping temperature. Oxygen content in molten steel Alloy block size The order in which elements are added determines the output variable, which is the predicted yield value for each element.

[0208] The converter steelmaking endpoint control system also includes an alloy composition control module, which is used to implement the above methods.

[0209] The converter steelmaking endpoint control method and system in this embodiment achieves precise control of narrow composition during the alloying stage by predicting alloy yield and optimizing alloy addition amount through linear programming, thereby reducing production costs and having extremely high industrial application value.

[0210] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0211] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0212] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0213] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An artificial intelligence-based endpoint control method for converter steelmaking, characterized in that, Includes the following steps: Acquire historical smelting data of the converter steelmaking process, including molten iron composition, scrap steel quantity, oxygen supply flow rate, temperature and carbon content detected by the auxiliary lance, and flame spectrum images and audio signals at the furnace mouth; The historical smelting data was cleaned and feature-engineered. Key input variables were selected through independence analysis and correlation analysis to construct a training sample set. The non-parallel support vector regression machine algorithm based on wavelet weights uses the training sample set to construct an endpoint parameter prediction model, and introduces a whale swarm optimization algorithm to optimize the hyperparameters of the training model, which is used to predict the carbon content and temperature at the endpoint of converter steelmaking. During the blowing process, real-time process data of the converter steelmaking process is acquired and input into the endpoint parameter prediction model to continuously acquire the predicted carbon content and predicted temperature of the smelting endpoint. When the smelting process is predicted to be in its final stage, principal component analysis is performed on the flame spectrum image at the furnace opening for dimensionality reduction. Combined with the furnace sound characteristics obtained from the audio signal analysis, the real-time carbon content and real-time temperature are calculated. The material balance calculated by the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the static model is used to calculate the required oxygen and cooling dosage for supplementary blowing using a dynamic programming algorithm, and the supplementary blowing operation is then performed. When the replenishment blowing operation is completed, a sample of molten steel is collected for testing. If the test results do not fall within the target range, the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the material balance are obtained to determine the required oxygen and cooling dosage for replenishment blowing. The replenishment blowing operation is then performed again. The above operation is repeated until the test results fall within the target range. The actual smelting data for this furnace is obtained, stored in a production database, and the parameters of the endpoint parameter prediction model are updated online using an elastic weight solidification algorithm based on a self-learning module.

2. The converter steelmaking endpoint control method according to claim 1, characterized in that, The historical smelting data was cleaned and feature-engineered. Key input variables were selected through independence and correlation analysis to construct a training sample set, including: use The criteria remove outliers from historical smelting data, including abrupt changes in oxygen supply flow and temperature jumps. Missing values ​​are filled in the historical smelting data after removing the outliers. This is done by interpolating the missing non-iron content data using the K-nearest neighbor algorithm to obtain a sample of historical smelting data. The oxygen flow rate time series data during the supplementary blowing stage is denoised using wavelet transform theory. Based on the denoising results, the signal confidence of the historical smelting data samples is quantified, and the weight vector of the historical smelting data samples is determined. The weight vector serves as the weight of the historical smelting data samples for training the endpoint parameter prediction model; n is the sample sequence length; weight coefficients. The calculation is performed using the Gaussian function, specifically as follows: ； in, For the first The Euclidean distance between the original data points and the denoised data; The width of the Gaussian kernel; The historical smelting data also includes process parameters such as the amount of molten iron charged, the scrap steel ratio, the total oxygen supply, and the amount of lime added. These, along with the molten iron composition and the oxygen supply flow rate, constitute a candidate input variable set. An independence analysis is then performed on this candidate input variable set. The variance inflation factor (VIF) method is used to test the degree of multicollinearity among the process parameters in the candidate input variable set, and the variance inflation factor is calculated. ； in, To make the first The determination coefficients are obtained by performing linear regression with one of the process parameters as the dependent variable and the remaining process parameters as independent variables; an independence judgment threshold is set to exclude... The process parameters that exceed the independence judgment threshold; The correlation between each process parameter and the endpoint carbon content and endpoint temperature was analyzed using the Pearson correlation coefficient matrix. Process parameters with an absolute correlation coefficient greater than 0.6 were selected as the key input variables.

3. The converter steelmaking endpoint control method according to claim 2, characterized in that, The non-parallel support vector regression algorithm based on wavelet weights constructs an endpoint parameter prediction model using the training sample set, and introduces a whale swarm optimization algorithm to optimize the hyperparameters of the training model, including: Initialize the parameters of the non-parallel support vector regression machine and Set the kernel function to a radial basis function; split the training sample set into a training subset and a validation subset; Construct an optimization solution framework for the non-parallel support vector regression machine. This framework consists of an objective function containing a weighted empirical risk term and a structural risk term, and the deviation between the predicted and actual sample values ​​does not exceed a certain limit. The constraints are constituted; the weighted empirical risk term is introduced into the weight vector, and the fitting slack variables of the samples of each training subset are weighted according to the signal confidence. The training subset is input into the optimization framework, and the optimal solution vector is obtained by solving the dual problem. ; The regression function is Wherein, ω is the model weight vector of the non-parallel support vector regression machine in the high-dimensional feature space, which is derived from the optimal solution vector; It is a nonlinear mapping function; For bias terms; The root mean square error (RMSE) between the predicted and actual values ​​of the endpoint parameter prediction model on the validation subset is used as the fitness function to iteratively update the population position vector. The fitness function is calculated as follows: ； in, The total number of samples in the verification subset; For the first The actual endpoint carbon content or temperature value of each sample; For the first The model prediction value for each sample; The whale swarm optimization algorithm is used to search within the parameter space of the non-parallel support vector regression machine until the fitness function value is less than a preset threshold or the maximum number of iterations is reached. .

4. The converter steelmaking endpoint control method according to claim 3, characterized in that, Principal component analysis (PCA) is performed on the flame spectral image at the furnace opening for dimensionality reduction. Combined with the furnace acoustic characteristics obtained from the audio signal analysis, the real-time carbon content and real-time temperature are calculated; including: Acquire the flame spectrum image at the furnace opening, with a wavelength range covering 400nm to 1000nm; Extract the target regions of the furnace mouth flame core area and furnace mouth flame edge area from the furnace mouth flame spectral image; Principal component analysis (PCA) was used to reduce the dimensionality of the target region, and the eigenvalues ​​of its covariance matrix were calculated. Before selection Principal components with a cumulative contribution rate greater than 95% are selected as the spectral images to be processed. The value was determined through cross-validation; the cumulative contribution rate was calculated using the following formula: ； in, For the front The cumulative contribution rate of each principal component; For the first One eigenvalue; This represents the total number of original spectral bands. For the first One eigenvalue; The spectral image to be processed is input into a deep convolutional neural network. Flame texture features are extracted through multiple convolutional layers, and the predicted value of real-time carbon content is output by a fully connected layer. The ReLU function is used as the activation function of the network.

5. The converter steelmaking endpoint control method according to claim 4, characterized in that, The material balance calculated from the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the static model is used to calculate the required oxygen and cooling dosage for supplementary blowing using a dynamic programming algorithm, and then the supplementary blowing operation is performed. include: Establish supplemental oxygen volume Carbon content detection with secondary gun Target carbon content and the amount of coolant added The nonlinear mapping relationship between them was fitted using a three-layer BP neural network, and the number of hidden layer nodes was determined to be 7 through trial and error. Enter the current molten pool temperature With target temperature The difference Based on the current oxygen content of the molten steel Theoretical oxygen demand is calculated using thermodynamic equilibrium equations. : ； in, For the quality of molten steel; The carbon content of the current secondary gun is being tested; The target carbon content; The enthalpy change is due to the carbon oxidation reaction; Oxygen utilization coefficient; Current oxygen content; The decarbonization efficiency coefficient; , The coefficient representing the effect of temperature on decarbonization oxygen demand is obtained by regression analysis of historical best furnace data. The initial estimated amount of supplemental oxygen is calibrated using the theoretical oxygen demand to obtain the baseline oxygen supply for supplemental blowing. An incremental calculation method is introduced, with a control cycle of 3 seconds, to synchronously acquire the real-time carbon content calculated by combining the dimensionality reduction features of the furnace mouth flame spectral image with the sound features inside the furnace. Real-time temperature After dynamically correcting for carbon content and temperature deviations, the oxygen flow correction value for the current cycle is calculated. ; First, construct a comprehensive carbon content deviation. Overall temperature deviation : ； ； in, The weighting coefficient for the predicted value ranges from 0.4 to 0.6, gradually decreasing as the reblowing process progresses, while the weighting of the real-time detected value gradually increases with the refining depth. The formula for calculating the oxygen flow correction value is: ； in, To predict carbon content deviation; To predict temperature deviation; The predicted carbon content; To predict temperature; This is the carbon content correction factor. The temperature correction coefficients are all obtained through regression analysis of historical best furnace data. The oxygen flow correction value The actual oxygen supply flow command for the current control cycle is generated by superimposing the supplementary oxygen supply benchmark amount in real time. Coolant addition amount Based on the heat balance calculation of the molten pool: ； in, The total amount of heat required to reach the target temperature; This is the measured enthalpy of the current molten pool; Specific heat capacity of the coolant; This is the thermal efficiency coefficient.

6. The converter steelmaking endpoint control method according to claim 5, characterized in that, It also includes abnormal operating condition intervention steps, which include: Real-time monitoring of converter ring vibration signals Exhaust gas temperature; When the vibration signal energy spectral density exceeds the set threshold and the exhaust gas temperature drops by more than 50°C within 10 seconds, it is determined to be a precursor to splashing. In this case, the oxygen lance position is lowered to the ready-to-blown position, and the bottom nitrogen stirring flow rate is increased until the changes in the energy spectral density and the exhaust gas temperature return to normal. The vibration signal energy spectral density Calculated using the following formula: ； in, The result is the Fast Fourier Transform of the vibration signal; and These are the set lower and upper limits for the frequency, respectively. Exhaust gas temperature variation range Calculated using the following formula: ； in, The current exhaust gas temperature; The exhaust gas temperature was 10 seconds ago.

7. The converter steelmaking endpoint control method according to claim 6, characterized in that, At the end of the blowing process, a sample of molten steel is collected for testing. If the test results do not fall within the target range, the predicted carbon content, the predicted temperature, the real-time carbon content, the real-time temperature, and the material balance are obtained. The required oxygen and cooling dosage for replenishment blowing are calculated, and the replenishment blowing operation is performed again. This process is repeated until the test results fall within the target range. This includes: Obtain laboratory data from the molten steel sample after the blowing process is complete, including the actual carbon content. Actual temperature ; Calculate the prediction error of the endpoint parameter prediction model. , ; like or If the error is not found, the endpoint parameter prediction model is determined to be mismatched. Based on the test error, the trainable parameters of the endpoint parameter prediction model are corrected online. The trainable parameters include the regression hyperplane weight vector and bias term. Incremental gradient descent is used to adjust the model parameters, and the correction amount is... The calculation formula is: ； in, The learning rate ranges from 0.01 to 0.

1. For loss function, , The test data for the molten steel sample. Predicted data for the endpoint; This is the current weight coefficient vector.

8. The converter steelmaking endpoint control method according to claim 7, characterized in that, The self-learning module employs an elastic weight fixation algorithm to update the parameters of the endpoint parameter prediction model online, including: The Domain Adversarial Neural Network (ADDA) algorithm in transfer learning is used to align the features of historical smelting data with the target domain data of the current furnace. Virtual smelting data is generated using the WGANGP generative adversarial network, and the extreme working condition sample library is expanded by combining feature space interpolation. The Fisher information matrix F is calculated based on the Elastic Weight Solidification (EWC) algorithm, and a regularization term is added when updating the parameters of the endpoint parameter prediction model; the objective function... for: ； in, The loss function for the new furnace data; This is a dynamic regularization coefficient that decreases as the amount of new data samples increases. For the first Fisher information content of each parameter; These are the current model parameters; These are the historically optimal model parameters.

9. The converter steelmaking endpoint control method according to claim 8, characterized in that, Also includes: After the blowing process is completed, the difference between the target composition and the current composition of the molten steel is obtained; The minimum cost of each alloy is calculated using the differences, historical data on alloy yield, and a linear programming model. A heuristic algorithm is used to solve the objective function. The constraints are: ； in, For the first The unit price of the alloy; For the first The amount of each alloy added; This represents the number of alloy types. For the first The target component content of each element; The allowable deviation range for components; For the current steel molten steel Content of various elements; For the first The first alloy Predicted yield values ​​for each element; For the first The first alloy The inherent content ratio of each element; An alloy yield prediction model was constructed using historical data on alloy yield, with input variables including tapping temperature. Oxygen content in molten steel Alloy block size The order in which elements are added determines the output variable, which is the predicted yield value for each element.

10. An artificial intelligence-based converter steelmaking endpoint precision control system, used to implement the artificial intelligence-based converter steelmaking endpoint control method as described in any one of claims 1-9, comprising: The acquisition module acquires real-time process data and historical smelting data of the converter steelmaking process, including molten iron composition, scrap steel quantity, oxygen supply flow rate, molten pool temperature and carbon content detected by the auxiliary lance, furnace mouth flame spectral image and audio signal. The data processing module performs data cleaning and feature engineering on historical smelting data, selects key input variables through independence analysis and correlation analysis, and constructs a training sample set. The model building module, based on the wavelet weighted non-parallel support vector regression machine algorithm, uses the training sample set to build an endpoint parameter prediction model, and introduces the whale swarm optimization algorithm to optimize the hyperparameters of the endpoint parameter prediction model, which is used to predict the carbon content and temperature at the endpoint of converter steelmaking; The prediction module inputs the real-time process data into the endpoint parameter prediction model during the blowing process, and continuously acquires the predicted carbon content and predicted temperature. The first calculation module, when predicting that the smelting is entering the final stage, uses an AI visual model to perform principal component analysis and dimensionality reduction on the furnace mouth flame spectrum image, and combines an AI audio model to analyze the sound characteristics inside the furnace, and calculates the real-time carbon content and real-time temperature. The second calculation module uses a dynamic programming algorithm to calculate the amount of oxygen and coolant required for the supplementary blowing stage based on the predicted carbon content, predicted temperature, real-time carbon content, real-time temperature, and material balance calculated by the static model. The correction module performs supplementary blowing based on the oxygen blowing amount and coolant addition amount obtained from the second calculation module, and collects molten steel samples for testing at the end of the blowing process; if the test results do not hit the target range, the supplementary blowing operation is repeated until the test results hit the target range. The model update module acquires the actual smelting data for this furnace, stores it in a production database, and updates the parameters of the endpoint parameter prediction model online using an elastic weight solidification algorithm based on the self-learning module.

Images