Converter steelmaking end point intelligent control method based on artificial intelligence

Abstract

The invention provides a converter steelmaking end point intelligent control method based on artificial intelligence. Aiming at the problems of scarcity of extreme working condition samples, poor model generalization, lack of a real-time data enhancement mechanism and the like in the prior art, the method comprises the following steps: acquiring conventional working condition data to construct a feature library for abnormal recognition, constructing a basic model and embedding a metallurgical law, and solving the problem of insufficient samples by utilizing an ADDA algorithm; virtual data are generated through WGAN-GP, the data are perfected in combination with feature space interpolation and an EWC algorithm, regulation and control are performed based on a prediction result, and model parameters are updated; the method can improve the prediction accuracy of the model under the extreme working condition, adapts to the dynamic change of the furnace condition, and improves the molten steel quality and the smelting efficiency.

Description

Technical Field

[0001] The invention belongs to the technical field of converter steelmaking, and in particular relates to an intelligent control method for converter steelmaking endpoint based on artificial intelligence. Background Art

[0002] In the converter steelmaking process, the core of endpoint control is to precisely control the temperature and carbon content of the molten steel to meet the standards. Current AI-based control methods generally face the problem of a shortage of samples of extreme working conditions. Traditional deep learning models suffer from an uneven distribution of training data, and their prediction errors increase significantly when dealing with such working conditions. For example, when the phosphorus content of molten iron is greater than 0.3%, the endpoint carbon prediction deviation of the existing model increases by 0.02% compared with conventional working conditions, resulting in delayed adjustment of the supplementary blowing, which seriously affects the quality of the molten steel and smelting efficiency.

[0003] Chinese patent publication number CN112668234B provides an intelligent control method for converter steelmaking endpoint. It is implemented by the following subsystems: 1) Data preprocessing subsystem: This subsystem collects and preprocesses data from a database. Through independence and correlation analysis, it determines the input variables for the endpoint carbon content and temperature prediction subsystem models, ensuring model accuracy. 2) Molten steel endpoint prediction subsystem: This subsystem uses a non-parallel support vector regression algorithm based on wavelet weights to predict the endpoint carbon content and endpoint temperature of converter steelmaking. 3) Oxygen blowing and auxiliary material calculation subsystem: This subsystem combines a whale swarm optimization algorithm with an incremental calculation method to calculate the optimization error based on the output feedback of the prediction model. While minimizing the optimization error, it calculates the required oxygen blowing amount and the amount of auxiliary materials such as lime and light-burned dolomite to be added during the blowing stage. 4) Model update subsystem: This subsystem is regularly updated and upgraded based on actual production conditions. This method enables one-click steelmaking in converters.

[0004] The main defects of the existing technology are: first, the sample size of extreme working conditions is insufficient, making it difficult to build a reliable model through traditional supervised learning; second, the ability to transfer data features across working conditions is weak, and the model generalization is poor; third, there is a lack of real-time data enhancement mechanism, which cannot adapt to dynamic changes in furnace conditions.

[0005] To this end, the present invention provides an artificial intelligence-based intelligent control method for converter steelmaking endpoint. Summary of the Invention

[0006] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background technology is solved.

[0007] The technical solution adopted by the present invention to solve its technical problem is: In a first aspect, the present invention provides an artificial intelligence-based converter steelmaking endpoint intelligent control method, comprising: S1: Collect basic data of normal working conditions to construct working condition characteristics and abnormal events, and build a multimodal feature library based on abnormal events for abnormal event identification and early warning; S2: Build a basic model using basic data from conventional working conditions. Through cross-domain feature alignment and target domain adaptive processing, train the basic model to adapt from conventional working conditions to extreme working conditions. Metallurgical laws are embedded in the basic model training to constrain the basic model. The ADDA algorithm is used to address the problem of insufficient samples from extreme working conditions. S3: Combine a small number of extreme working condition samples with physical constraints to generate an adversarial network, expand these small number of extreme working condition samples into a large amount of virtual data, use WGAN-GP to improve the loss function and embed metallurgical constraints, let the basic model learn extreme working condition data, and use feature space interpolation technology to improve the extreme working condition data; S4: Based on the endpoint carbon temperature predicted by the basic model, the predicted endpoint carbon temperature is regulated, and the basic model parameters are updated using the predicted endpoint carbon temperature and the actual endpoint carbon temperature.

[0008] As a further improvement of the present invention: the basic data of conventional working conditions is collected as follows: Collect basic data of routine operating conditions during the converter steelmaking process. The basic data of routine operating conditions are the basic parameters of the steelmaking process, which directly affect the composition and temperature changes of the molten steel. Among them, the basic data of routine operating conditions include: the content of C carbon, Si silicon, P phosphorus, and S sulfur elements.

[0009] As a further improvement of the present invention, the specific process of constructing a multimodal feature library based on abnormal events is as follows: For abnormal events such as splashing and backflow during the steelmaking process, multimodal data is collected and a feature library is established to identify and warn of abnormal events. Multimodal data including flame images and acoustic signals is collected. When splashing occurs, the brightness of the furnace mouth flame will suddenly change, and the high-frequency components in the acoustic signal will increase sharply. During backflow, the sound and flame shape in the furnace will also change specifically. The collected multimodal data are analyzed to extract key features that can characterize abnormal events, namely the brightness mutation characteristics of flame images and the high-frequency component surge characteristics of acoustic signals, and a multimodal feature library of abnormal events is established.

[0010] As a further improvement of the present invention, the specific process of constructing the basic model using basic data of conventional working conditions is as follows: A three-layer LSTM stack with 256 neurons in each layer is used to capture the long-term and short-term dependencies of the decarbonization process; A self-attention layer is added before the output layer to calculate the weight of the features at each time step and focus on the key operation stage. The weight calculation formula is: ,in, is the similarity between the current hidden state and the hidden state of all time steps, calculated by dot product, is the hidden state at the t-th time step, is the set of hidden states at all time steps, is the normalized denominator to ensure that the sum of the attention weights of all time steps is 1.

[0011] As a further improvement of the present invention: the specific process of the feature alignment is: Normalize and align the conventional working condition data in time series, where the normalization process scales the numerical features to Interval,temporal alignment means that the blowing time of different furnaces is uniformly interpolated into 300 time steps.

[0012] As a further improvement of the present invention, the specific process of constraining the basic model is as follows: Based on the double-film theory, the decarburization rate satisfy: ,in, is the mass transfer coefficient, is the gas-liquid interface area, To balance the carbon concentration, is the current carbon content of molten steel; during training, the mean square error between the decarburization rate predicted by the model and the value calculated by this formula is used as the auxiliary loss: , is the dynamic auxiliary loss, is the number of time steps, is the decarbonization rate at time t predicted by the model, is the decarburization rate at time t calculated by the mechanism formula; Constraints are based on the heat balance equation, and the constraints are: ,in, To blow oxygen and release heat, Release heat for decarburization, The slag evaporates and absorbs heat. For thermal radiation, Because heat is required to heat up the molten steel, the temperature predicted by the basic model must deviate from the value calculated by this equation by less than 10°C.

[0013] As a further improvement of the present invention: the specific process of expanding into a large amount of virtual data is: Based on the WGAN-GP generative adversarial network, a small number of real samples under extreme working conditions are used to generate virtual heat data with physical rationality; When the furnace condition is detected to be in an extreme state, the data enhancement module is automatically started to perform time series perturbation on the sensor signal of the current furnace, i.e., adding 5% noise and feature space interpolation; An elastic weight consolidation algorithm is used to retain the conventional knowledge learned by the basic model when adding new extreme working condition samples, and virtual samples are created using physically reasonable data driven by WGAN-GP.

[0014] As a further improvement of the present invention, the specific process of embedding metallurgical constraints is as follows: Add a constraint layer after the generator output layer to ensure that the generated data meets the following conditions: The decarburization rate decreases monotonically, which is consistent with the law of decreasing carbon concentration in the later stage of blowing; The temperature change rate is less than 5℃ / min, which means avoiding violation of the thermal equilibrium principle; Basicity R∈[2.5,3.5], i.e., the physical property constraint of slag; Generating data x must satisfy: ; If it is not satisfied, the generated sample is corrected by the projected gradient method, where is the change in carbon content, that is, the reduction in carbon content during blowing. is the time variation, i.e. the time step of blowing, is the temperature change, The calcium oxide content in the slag, is the silicon dioxide content in the slag.

[0015] As a further improvement of the present invention, the specific process of regulating the predicted endpoint carbon temperature is as follows: When the predicted deviation is greater than 10℃, the dynamic supplementary blowing strategy is triggered; Metallurgical mechanism constraints are introduced to verify the rationality of the decisions of the basic model. When the deviation between the endpoint carbon temperature time predicted by the basic model and the endpoint carbon temperature obtained by the mechanism calculation is greater than 15%, the manual collaborative correction mechanism is activated.

[0016] As a further improvement of the present invention, the specific process of updating the basic model parameters using the predicted endpoint carbon temperature and the actual endpoint carbon temperature is as follows: After each heat is completed, the actual endpoint carbon temperature is compared with the predicted endpoint carbon temperature, and the basic model parameters are updated through TD time difference learning to form an iterative optimization loop of control and feedback.

[0017] In a second aspect, the present invention provides an artificial intelligence-based converter steelmaking endpoint intelligent control system, comprising: Multimodal feature library module: collects basic data of normal working conditions to construct working condition characteristics and abnormal events, and builds a multimodal feature library based on abnormal events for abnormal event identification and early warning; Basic model training module: This module uses basic data from conventional working conditions to build a basic model. Through cross-domain feature alignment and target domain adaptive processing, the basic model is trained to adapt from conventional working conditions to extreme working conditions. Metallurgical laws are embedded in the basic model training to constrain the basic model. The ADDA algorithm is used to address the problem of insufficient samples from extreme working conditions. Physical and Metallurgical Constraint Module: This module combines a small number of extreme working condition samples with physical constraints to generate an adversarial network, expanding these samples into a large amount of virtual data. WGAN-GP is used to improve the loss function and embed metallurgical constraints, allowing the basic model to learn extreme working condition data. Feature space interpolation technology is then used to improve the extreme working condition data. Control module: Based on the endpoint carbon temperature predicted by the basic model, the predicted endpoint carbon temperature is controlled, and the basic model parameters are updated using the predicted endpoint carbon temperature and the actual endpoint carbon temperature.

[0018] The beneficial effects of the present invention are as follows: 1. Cross-domain feature migration is achieved through the ADDA algorithm. Combining the WGAN-GP generative adversarial network with physical constraints, a small number of extreme working condition samples are expanded into a large amount of virtual data, effectively solving the problem of insufficient extreme working condition samples. Through cross-domain feature alignment and target domain adaptive processing, the basic model is trained to adapt from conventional working conditions to extreme working conditions. Combined with the elastic weight consolidation algorithm (EWC), this prevents the model from forgetting previously learned conventional knowledge when learning new working conditions, enhancing model generalization. Metallurgical laws (such as double-film theory and heat balance equations) are embedded in the basic model training and the model is constrained to ensure that the generated data conforms to physical laws such as the monotonically decreasing decarburization rate and the temperature change rate of less than 5°C / min, making the model prediction more accurate.

[0019] 2. Control is performed based on the model's predicted endpoint carbon temperature. When the predicted deviation exceeds the limit, a dynamic re-blowing strategy or a manual collaborative correction mechanism is triggered. By comparing the actual and predicted endpoint carbon temperatures, TD time difference learning is used to update the model parameters, forming a control feedback iterative optimization loop, improving smelting efficiency and molten steel quality. A multimodal feature library is constructed to identify and warn abnormal events such as splashing and back-drying during the steelmaking process, providing a basis for intelligent control. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be further described below with reference to the accompanying drawings.

[0021] Figure 1 is a flow chart of the steps of the present invention; Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0022] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0023] Example 1:

[0024] like Figure 1 As shown, an artificial intelligence-based converter steelmaking endpoint intelligent control method according to an embodiment of the present invention includes: S1: Collect basic data of normal working conditions to construct working condition characteristics and abnormal events, and build a multimodal feature library based on abnormal events for abnormal event identification and early warning; The basic data of conventional working conditions are specifically as follows: Basic data of conventional working conditions: collect basic data of conventional working conditions such as molten iron composition (C, Si, P, S), scrap steel ratio, oxygen blowing parameters (flow rate, pressure, gun position trajectory), etc. Specifically, the basic data of conventional working conditions in the converter steelmaking process are collected. The basic data of conventional working conditions are the basic parameters of the steelmaking process, which directly affect the composition and temperature changes of the molten steel. Among them, the basic data of conventional working conditions include: the content of C (carbon), Si (silicon), P (phosphorus), and S (sulfur); The contents of C (carbon), Si (silicon), P (phosphorus), and S (sulfur) directly affect the decarburization and slag formation reactions as well as the quality of the final molten steel. The specific process of constructing the working condition characteristics is as follows: The lining erosion degree characteristics are extracted using an infrared thermal imager. When the erosion depth is greater than 5 cm, it is marked as an extreme condition. The slag layer fluctuation spectrum characteristics are obtained using a sonar sensor. When the standard deviation is greater than 0.5 dB, it is determined to be an extreme condition. Specifically, the abnormal events are: For abnormal events that occur during the steelmaking process, such as splashing and back drying, multimodal data is collected and a feature library is established to identify and warn of abnormal events. Collect multimodal data including flame images and acoustic signals; For example, when splashing occurs, the brightness of the furnace flame will change suddenly, and the high-frequency components in the acoustic signal will surge. When back-drying occurs, the sound and flame shape in the furnace will also change specifically. Feature library establishment: Analyze the collected multimodal data to extract key features that can characterize abnormal events, such as brightness mutation characteristics of flame images and high-frequency component surge characteristics of acoustic signals, and establish a multimodal feature library of abnormal events. This feature library is used for subsequent analysis of real-time collected data to promptly identify and warn of abnormal events, providing a basis for intelligent control. Specifically, the sonar sensor is used to obtain the slag layer fluctuation spectrum characteristics, and the field programmable gate array (FPGA) is used to perform real-time spectrum analysis. The bandpass filter method is used to remove the environmental noise, and the time domain normalization method is used to normalize the acoustic signal energy of each detection to interval; The short-time Fourier transform is used to generate the time-frequency diagram. At the same time, the time-frequency diagram is generated by calculation, focusing on the high-frequency band of 20~80kHz. The high-frequency band of 20~80kHz is mainly the energy range of the slag layer fluctuation; Establish a multimodal feature library for abnormal events such as splashing and backdrying, which includes the characteristics of sudden changes in flame image brightness and surges in high-frequency components of acoustic signals; At the same time, the multimodal feature library also includes the following data information: Record the scrap ratio, i.e. the proportion of scrap steel in the raw materials. The amount of scrap steel added affects the heat balance in the furnace and the composition of the molten steel. Therefore, the scrap steel ratio is also an indicator of the quality of the final molten steel. Record oxygen blowing parameters, including oxygen flow rate, pressure, and lance position trajectory. Oxygen blowing is a key operation in converter steelmaking. Control of oxygen blowing parameters directly affects decarburization rate, molten steel temperature, and the progress of furnace reactions. Record the amount of slag agent added. Different slag agents have significant differences in slag basicity, desulfurization and dephosphorization efficiency, and furnace lining protection. Recording furnace gas composition, which reflects in real time the carbon-oxygen reaction process in the furnace, whether the furnace gas escape path is unobstructed, and the interaction between the oxygen lance jet and the molten pool; S2: Build a basic model using basic data from conventional working conditions. Through cross-domain feature alignment and target domain adaptive processing, train the basic model to adapt from conventional working conditions to extreme working conditions. Metallurgical laws are embedded in the basic model training to constrain the basic model. The ADDA algorithm is used to address the problem of insufficient samples from extreme working conditions. Basic model pre-training uses basic data from conventional working conditions to build an LSTM-Attention basic model. This model learns the general metallurgical laws of decarburization dynamics and thermal balance, enabling it to understand the mapping between processes and results in conventional steelmaking scenarios, laying the foundation for adaptation to complex working conditions. Target domain adaptation: For abnormal events, the adversarial transfer learning (ADDA) algorithm is used to reduce the feature distribution difference between the source domain (non-abnormal events) and the target domain (abnormal events) through the gradient reversal layer; Meta-knowledge extraction: extracting key features from historical abnormal event samples; For example, the prior knowledge that the decarburization rate decreases by 8% for every 0.1% increase in phosphorus content is embedded in the model loss function as a regularization term; Specifically, 80% of the conventional operating condition data was selected and divided into time series feature data and static data. The time series feature data included oxygen blowing flow rate, gun position trajectory, and furnace gas composition. The static data included initial C / Si / P / S content of molten iron, scrap steel ratio, and slag-forming agent addition. Normalize and align the conventional working condition data in time series, where the normalization process scales the numerical features to Interval,temporal alignment means that the blowing time of different heats is uniformly interpolated into 300 time steps; A three-layer LSTM stack with 256 neurons in each layer is used to capture the long-term and short-term dependencies of the decarbonization process; For example, the change in decarburization rate is strongly correlated with the oxygen lance position 3 minutes ago, and LSTM memorizes this association through cell state; A self-attention layer is added before the output layer to calculate the weight of the features at each time step and focus on the key operation stage. The weight calculation formula is: ,in, is the similarity between the current hidden state and the hidden state of all time steps, calculated by dot product, is the hidden state at the t-th time step, is the set of hidden states at all time steps, To normalize the denominator, ensure that the sum of the attention weights of all time steps is 1; Metallurgical laws are embedded in model training to constrain the model. The constraint rules are as follows: Based on the double-film theory, the decarburization rate satisfy: ,in, is the mass transfer coefficient, is the gas-liquid interface area, To balance the carbon concentration, is the current carbon content of molten steel; during training, the mean square error between the decarburization rate predicted by the model and the value calculated by this formula is used as the auxiliary loss: , is the dynamic auxiliary loss, is the number of time steps, is the decarbonization rate at time t predicted by the model, is the decarburization rate at time t calculated by the mechanism formula; Constraints are based on the heat balance equation, and the constraints are: ,in, To blow oxygen and release heat, Release heat for decarburization, The slag evaporates and absorbs heat. For thermal radiation, Because heat is required to heat up the molten steel, the temperature predicted by the basic model must deviate from the value calculated by this equation by less than 10°C; 20% of conventional operating data were selected to verify the accuracy of the model. If the endpoint carbon prediction error is less than or equal to 0.01%, the temperature error is less than 10°C, and the decarburization rate curve is more than 90% consistent with the mechanism model, it means that the basic model training is complete and the metallurgical law embedding is effective. However, in actual production, raw material fluctuations and equipment aging can lead to frequent extreme operating conditions, necessitating further expansion of the model's cross-operating condition adaptability. The ADDA algorithm is used to realize cross-condition feature migration. The generator G maps the source domain (normal working condition data) and the target domain (extreme working condition data) to a shared feature space. ,Right now: , Discriminator D: judge the feature From the source domain or the target domain, output probability ; Use confrontation training thinking to and In the feature space, it is indistinguishable and deceives the discriminator, which accurately distinguishes the source domain and target domain features, improving the domain classification accuracy. Represents the generator G on the source domain data The feature mapping result is Represents the generator G for the target domain data The feature mapping result of Using the Gradient Reversal Layer (GRL) technique, the forward propagation is set as: , that is, without changing the eigenvalue, negative weights are added to the gradient through back propagation, that is, ,in, is the domain adaptation strength, which encourages the generator to learn domain-invariant features and suppress domain-related features. GRL is the gradient reversal layer. The input feature vector can be the data feature of normal working conditions or extreme working conditions. is the identity matrix, Refers to the partial derivative of GRL with respect to x; Loss function and optimization, generator loss: , the goal is to make the discriminator misjudge the source domain features as the target domain, where is the generator loss, which measures the generator’s ability to deceive the discriminator. is the expectation, averaged over all samples in the dataset, middle, is a data sample of normal working conditions, It is a normal working condition data set. middle, is a data sample of extreme working conditions, is an extreme working condition data set, and D is a probability value output by the discriminator; Discriminator loss: ,in, is the discriminator loss, is the mathematical expectation, middle, Generate data for normal working conditions. It is a normal working condition data set. middle, Generate data for extreme conditions, is an extreme working condition data set, 、 is the output of the generator, The discriminator is trained to measure the ability of the discriminator to distinguish between normal and extreme working conditions, with the goal of maximizing the domain classification accuracy. Alternating training is used: first, G is fixed to optimize D, then D is fixed to optimize G, until the discriminator classification accuracy drops to 50%; Meta-knowledge extraction: extract key features from historical extreme operating condition samples to create decarburization rate constraints. The decarburization rate constraints are: ,in, is the decarburization rate change of the i-th sample, unit (% / min), is the change in phosphorus content of the i-th extreme working condition sample, is the meta-knowledge coefficient, To measure the deviation between the effect of phosphorus content change on decarburization rate predicted by the basic model and the metallurgical prior knowledge; The adversarial diversion (ADDA) solves the problem of few samples under extreme working conditions, but pure data-driven methods may deviate from metallurgical mechanisms. For example, if the basicity is predicted to exceed [2.5, 3.5], the decarburization rate may violate the phosphorus content inhibition rule. Therefore, it is necessary to embed metallurgical prior knowledge to make the basic model adaptable across working conditions and conform to physical laws. Create an alkalinity constraint term, which is: , ensuring that the model predicts the slag basicity , falls on Interval, when out of range, penalty is imposed, where T is the number of time steps, which in this invention is equal to 300. The slag basicity at time t predicted by the basic model, the center value of the reasonable interval of 3.0 basicity, the half width of the reasonable interval of 0.5 basicity, is the alkalinity constraint; Loss function integration, final loss function: ,in, To predict losses, is the dynamic constraint loss, which measures the fit between the basic model and the metallurgical dynamics. In order to measure the deviation between the effect of phosphorus content change on decarburization rate predicted by the basic model and the metallurgical prior knowledge, is the alkalinity constraint, 、 、 is the weight coefficient, determined by cross-validation; S3: Combine a small number of extreme working condition samples with physical constraints to generate an adversarial network, expand these small number of extreme working condition samples into a large amount of virtual data, use WGAN-GP to improve the loss function and embed metallurgical constraints, let the basic model learn extreme working condition data, and use feature space interpolation technology to improve the extreme working condition data; Virtual sample generation: Based on the WGAN-GP generative adversarial network, a small number of real samples under extreme working conditions are used to generate physically reasonable virtual heat data; For example, the temperature-carbon content change curve when the phosphorus content is 0.4% is simulated; When the furnace condition is detected to be in an extreme state, the data enhancement module is automatically started to perform time series perturbation on the sensor signal of the current furnace, i.e., adding 5% noise and feature space interpolation; Adopting the EWC (Elastic Weight Consolidation) algorithm, when adding new extreme working condition samples, it retains the conventional knowledge learned by the basic model to avoid catastrophic forgetting; WGAN-GP uses physically plausible data driven by WGAN-GP to create virtual samples. Traditional GANs suffer from the problems of single-data generation and unstable training. WGAN-GP addresses these issues through the following improvements: Use Wasserstein distance instead of KL divergence to measure the difference between the generated distribution and the true distribution, and the gradient is more stable; Introducing a gradient penalty term to ensure the discriminator is Lipschitz continuous and avoid gradient disappearance; Generator G inputs random noise z and condition vector c, and generates heat data through multiple layers of fully connected layers; Discriminator D, which outputs the Wasserstein distance estimate and does not use the sigmoid activation function; Add a constraint layer after the generator output layer to ensure that the generated data meets the following conditions: The decarburization rate decreases monotonically, which is consistent with the law of decreasing carbon concentration in the later stage of blowing; The temperature change rate is less than 5℃ / min, which means avoiding violation of the thermal equilibrium principle; Basicity R∈[2.5,3.5], i.e., the physical property constraint of slag; Generating data x must satisfy: ; If it is not satisfied, the generated sample is corrected by the projected gradient method, where is the change in carbon content, that is, the reduction in carbon content during blowing. Time variation, i.e. the time step of blowing, The change in temperature, The content of calcium oxide (CaO) in the slag, Silicon dioxide in slag content; To address the issues of traditional GANs, such as limited data generation and unstable training, WGAN-GP improves the loss function and embeds metallurgical constraints, allowing the basic model to learn physically plausible extreme working condition data. Training process and parameters, loss function: ,in, is the real data distribution, is the sample distribution sampled from the line connecting the real and generated data, is the conditional vector, , is the gradient penalty weight, middle, For real data, middle, is random noise, is the distribution of random noise, is the expectation, averaged over all samples in the dataset to ensure that the loss is a global statistic, D is the output of the discriminator, middle, is the interpolation sample between the real data and the generated data, is the distribution of samples of real data and generated data, used to calculate the gradient penalty, Interpolate samples for the discriminator D The gradient, When a phosphorus content of 0.4% is input, WGAN-GP can generate 300 sets of virtual heat data, where the Pearson correlation coefficient between the temperature-carbon content curve and the real sample is greater than 0.92, and all of them meet the metallurgical constraints. Using time series perturbation technology, when extreme working conditions are detected, dynamic noise is added to the sensor signal: ,in, is the disturbance coefficient, is the signal standard deviation, ensuring that the disturbance amplitude is within 5%, is the sensor signal after disturbance, Raw sensor signal; It is still difficult to cover the diversity of characteristics of extreme working conditions by relying solely on sensor signal disturbances; For example, the complex slag changes of high silicon hot metal; Therefore, feature space interpolation technology is introduced to improve the information of the current heat with the characteristics of similar historical working conditions, broadening the learning boundary of the basic model for extreme working conditions: The feature space interpolation technology is specifically as follows: For the current heat feature vector Similar working conditions in history and Interpolate between to generate new eigenvectors: ,in, is the interpolation coefficient, ensuring that the new feature is within the physically feasible domain; Real-time enhanced trigger mechanism, multi-condition joint triggering, starts the enhanced module when one of the following conditions is met: Abnormal molten iron composition: Si>1.2% or P>0.3%; Abnormal condition of furnace lining: erosion depth > 5cm; Abnormal slag condition: Sonar SD>0.5dB and HFR>30%; Incremental learning update: EWC algorithm prevents catastrophic forgetting; The core principle of EWC, the key assumption is that some weights are very important for old tasks (normal working conditions), and these weights need to be preserved when learning new tasks (extreme working conditions); In order for the basic model to accurately identify important weights, that is, weights that are critical for normal working conditions (old tasks), it is necessary to quantify the importance of weights through the Fisher information matrix and then apply it to the training constraints of extreme working conditions; Fisher information matrix: Calculates the importance of each weight w to the old task loss: in, middle, is a normal working condition data set. It is a normal working condition data set. For old mission loss, is the mathematical expectation, is a single weight parameter of the model, Loss of old tasks The gradient of the weights; When training a new task, add a regularization term to penalize changes in important weights: in, The weights trained for the old task, is the weight coefficient, Loss for new missions, is the total loss of the new task, are the elements of the Fisher information matrix, is the current weight of the model; Calculate the Fisher information matrix F using normal working condition data, and calculate the new sample of extreme working condition And update the weights. During back propagation, constraints are imposed on important weights according to F. For example, the update step size of the weights related to furnace temperature prediction is reduced by 50%; S4: Based on the endpoint carbon temperature predicted by the basic model, the predicted endpoint carbon temperature is regulated, and the basic model parameters are updated using the predicted endpoint carbon temperature and the actual endpoint carbon temperature; Predict the endpoint carbon temperature in real time based on the basic model and regulate the predicted endpoint carbon temperature; When the predicted deviation is greater than 10℃, the dynamic supplementary blowing strategy is triggered; Metallurgical mechanism constraints are introduced to verify the rationality of the basic model's decisions. When the deviation between the endpoint carbon temperature time predicted by the basic model and the endpoint carbon temperature calculated by the mechanism is greater than 15%, the manual collaborative correction mechanism is activated. After each heat is completed, the actual endpoint carbon temperature is compared with the predicted endpoint carbon temperature, and the basic model parameters are updated through TD time difference learning to form an iterative optimization loop of control and feedback.

[0025] Example 2:

[0026] like Figure 2 As shown, based on Example 1, the present invention provides an artificial intelligence-based converter steelmaking endpoint intelligent control system, comprising: Multimodal feature library module: collects basic data of normal working conditions to construct working condition characteristics and abnormal events, and builds a multimodal feature library based on abnormal events for abnormal event identification and early warning; Basic model training module: This module uses basic data from conventional working conditions to build a basic model. Through cross-domain feature alignment and target domain adaptive processing, the basic model is trained to adapt from conventional working conditions to extreme working conditions. Metallurgical laws are embedded in the basic model training to constrain the basic model. The ADDA algorithm is used to address the problem of insufficient samples from extreme working conditions. Physical and Metallurgical Constraint Module: This module combines a small number of extreme working condition samples with physical constraints to generate an adversarial network, expanding these samples into a large amount of virtual data. WGAN-GP is used to improve the loss function and embed metallurgical constraints, allowing the basic model to learn extreme working condition data. Feature space interpolation technology is then used to improve the extreme working condition data. Control module: Based on the endpoint carbon temperature predicted by the basic model, the predicted endpoint carbon temperature is controlled, and the basic model parameters are updated using the predicted endpoint carbon temperature and the actual endpoint carbon temperature.

[0027] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based intelligent control method for converter steelmaking endpoint, characterized by: include: S1: Collect basic data of normal working conditions to construct working condition characteristics and abnormal events, and build a multimodal feature library based on abnormal events for abnormal event identification and early warning; S2: Build a basic model using basic data from conventional working conditions. Through cross-domain feature alignment and target domain adaptive processing, train the basic model to adapt from conventional working conditions to extreme working conditions. Metallurgical laws are embedded in the basic model training to constrain the basic model. The ADDA algorithm is used to address the problem of insufficient samples from extreme working conditions. S3: Combine a small number of extreme working condition samples with physical constraints to generate an adversarial network, expand these small number of extreme working condition samples into a large amount of virtual data, use WGAN-GP to improve the loss function and embed metallurgical constraints, let the basic model learn extreme working condition data, and use feature space interpolation technology to improve the extreme working condition data; S4: Based on the endpoint carbon temperature predicted by the basic model, the predicted endpoint carbon temperature is regulated, and the basic model parameters are updated using the predicted endpoint carbon temperature and the actual endpoint carbon temperature.

2. The method for intelligently controlling converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific data collected for conventional working conditions are: Collect basic data of routine operating conditions during the converter steelmaking process. The basic data of routine operating conditions are the basic parameters of the steelmaking process, which directly affect the composition and temperature changes of the molten steel. Among them, the basic data of routine operating conditions include: the content of C carbon, Si silicon, P phosphorus, and S sulfur elements.

3. The method for intelligently controlling converter steelmaking endpoint based on artificial intelligence according to claim 1, wherein: The specific process of constructing a multimodal feature library based on abnormal events is as follows: For abnormal events such as splashing and backflow during the steelmaking process, multimodal data is collected and a feature library is established to identify and warn of abnormal events. Multimodal data including flame images and acoustic signals is collected. When splashing occurs, the brightness of the furnace mouth flame will suddenly change, and the high-frequency components in the acoustic signal will increase sharply. During backflow, the sound and flame shape in the furnace will also change specifically. The collected multimodal data are analyzed to extract key features that can characterize abnormal events, namely the brightness mutation characteristics of flame images and the high-frequency component surge characteristics of acoustic signals, and a multimodal feature library of abnormal events is established.

4. The method for intelligently controlling converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of constructing the basic model using basic data of conventional working conditions is as follows: A three-layer LSTM stack with 256 neurons in each layer is used to capture the long-term and short-term dependencies of the decarbonization process; A self-attention layer is added before the output layer to calculate the weight of the features at each time step and focus on the key operation stage. The weight calculation formula is: ,in, is the similarity between the current hidden state and the hidden state of all time steps, calculated by dot product, is the hidden state at the t-th time step, is the set of hidden states at all time steps, is the normalized denominator to ensure that the sum of the attention weights of all time steps is 1.

5. The method for intelligently controlling converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of feature alignment is as follows: Normalize and align the conventional working condition data in time series, where the normalization process scales the numerical features to Interval,temporal alignment means that the blowing time of different furnaces is uniformly interpolated into 300 time steps.

6. The method for intelligently controlling converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of constraining the basic model is as follows: Based on the double-film theory, the decarburization rate satisfy: ,in, is the mass transfer coefficient, is the gas-liquid interface area, To balance the carbon concentration, is the current carbon content of molten steel; during training, the mean square error between the decarburization rate predicted by the model and the value calculated by this formula is used as the auxiliary loss: , is the dynamic auxiliary loss, is the number of time steps, is the decarbonization rate at time t predicted by the model, is the decarburization rate at time t calculated by the mechanism formula; Constraints are based on the heat balance equation, and the constraints are: ,in, To blow oxygen and release heat, Release heat for decarburization, The slag evaporates and absorbs heat. For thermal radiation, Because heat is required to heat up the molten steel, the temperature predicted by the basic model must deviate from the value calculated by this equation by less than 10°C.

7. The method for intelligently controlling converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of expanding into a large amount of virtual data is as follows: Based on the WGAN-GP generative adversarial network, a small number of real samples under extreme working conditions are used to generate virtual heat data with physical rationality; When the furnace condition is detected to be in an extreme state, the data enhancement module is automatically started to perform time series perturbation on the sensor signal of the current furnace, i.e., adding 5% noise and feature space interpolation; An elastic weight consolidation algorithm is used to retain the conventional knowledge learned by the basic model when adding new extreme working condition samples, and virtual samples are created using physically reasonable data driven by WGAN-GP.

8. The method for intelligently controlling converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of embedding metallurgical constraints is as follows: Add a constraint layer after the generator output layer to ensure that the generated data meets the following conditions: The decarburization rate decreases monotonically, which is consistent with the law of decreasing carbon concentration in the later stage of blowing; The temperature change rate is less than 5℃ / min, which means avoiding violation of the thermal equilibrium principle; Basicity R∈[2.5,3.5], i.e., the physical property constraint of slag; Generating data x must satisfy: ; If it is not satisfied, the generated sample is corrected by the projected gradient method, where is the change in carbon content, that is, the reduction in carbon content during blowing. is the time variation, i.e. the time step of blowing, is the temperature change, The calcium oxide content in the slag, is the silicon dioxide content in the slag.

9. The method for intelligently controlling converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of regulating the predicted endpoint carbon temperature is as follows: When the predicted deviation is greater than 10℃, the dynamic supplementary blowing strategy is triggered; Metallurgical mechanism constraints are introduced to verify the rationality of the decisions of the basic model. When the deviation between the endpoint carbon temperature time predicted by the basic model and the endpoint carbon temperature obtained by the mechanism calculation is greater than 15%, the manual collaborative correction mechanism is activated.

10. The method for intelligently controlling converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of updating the basic model parameters using the predicted endpoint carbon temperature and the actual endpoint carbon temperature is as follows: After each heat is completed, the actual endpoint carbon temperature is compared with the predicted endpoint carbon temperature, and the basic model parameters are updated through TD time difference learning to form an iterative optimization loop of control and feedback.

Images