A converter steelmaking endpoint intelligent control method based on artificial intelligence

Abstract

The application provides a converter steelmaking endpoint intelligent control method based on artificial intelligence; in view of the problems of extreme working condition sample scarcity, poor model generalization, lack of real-time data enhancement mechanism and the like in the prior art, the method collects conventional working condition data to construct a feature library for abnormality identification, constructs a basic model and embeds metallurgical rules, solves the sample shortage problem by using an ADDA algorithm, generates virtual data by using a WGAN-GP, combines feature space interpolation and an EWC algorithm to improve the data, and regulates and controls based on the prediction result and updates the model parameters; the method can improve the prediction accuracy of the model under extreme working conditions, adapt to dynamic changes of the furnace condition, and improve the molten steel quality and smelting efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of converter steelmaking, and in particular relates to a converter steelmaking endpoint intelligent control method based on artificial intelligence. BACKGROUND

[0002] In the converter steelmaking process, the core of endpoint control is to accurately regulate the steel temperature and carbon content to meet the standard. Current control methods based on artificial intelligence generally face the problem of a lack of extreme working condition samples. Traditional deep learning models have significantly increased prediction errors when dealing with such working conditions due to uneven distribution of training data. For example, when the phosphorus content of molten iron is >0.3%, the existing model's endpoint carbon prediction deviation increases by 0.02% compared to the regular working condition, leading to a lag in supplementary blowing adjustment and seriously affecting the quality of molten steel and smelting efficiency.

[0003] A converter steelmaking endpoint intelligent control method provided by Chinese Patent Publication No. CN112668234B is implemented by the following subsystems: 1) a data preprocessing subsystem: data is collected from a database and preprocessed, and through independence and correlation analysis, the input variables of the endpoint carbon content and temperature prediction subsystem model are determined to ensure model accuracy; 2) a molten steel endpoint prediction subsystem: a non-parallel support vector regression machine algorithm based on wavelet weights is used to predict the endpoint carbon content and endpoint temperature of converter steelmaking; 3) an oxygen blowing amount and auxiliary material calculation subsystem: combining the whale optimization algorithm and the incremental calculation method, the optimization error is calculated based on the output feedback of the prediction model, and under the premise of ensuring the minimum optimization error, the required oxygen blowing amount, lime and light-burned dolomite addition amount in the blowing stage are calculated; 4) a model updating subsystem: the prediction subsystem is regularly updated and upgraded according to actual production conditions. One-key steelmaking of the converter can be realized.

[0004] The main defects of the prior art are: first, the sample amount of extreme working conditions is insufficient, making it difficult to construct a reliable model through traditional supervised learning; second, the cross-condition data feature migration ability 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] Therefore, the application provides a converter steelmaking endpoint intelligent control method based on artificial intelligence. SUMMARY

[0006] In order to make up for the deficiencies of the prior art and solve at least one technical problem proposed in the background art.

[0007] The technical solution adopted by the application to solve its technical problems is:

[0008] In a first aspect, the application provides a converter steelmaking endpoint intelligent control method based on artificial intelligence, comprising:

[0009] S1: Collecting conventional working condition basic data to construct working condition characteristics and abnormal events, and constructing a multi-modal feature library based on the abnormal events for identification and early warning of abnormal events;

[0010] S2: Constructing a basic model using conventional working condition basic data, training the basic model to migrate and adapt from conventional working conditions to extreme working conditions through cross-domain feature alignment and target domain self-adaptive processing, and embedding metallurgical laws into the basic model training to constrain the basic model, and using the ADDA algorithm to solve the problem of insufficient extreme working condition samples;

[0011] S3: Generating an adversarial network by combining a small amount of extreme working condition samples with physical constraints, expanding a small amount of extreme working condition samples into a large amount of virtual data, using WGAN-GP to improve the loss function combined with embedded metallurgical constraints, and letting the basic model learn extreme working condition data, and using feature space interpolation technology to perfect the extreme working condition data;

[0012] S4: Based on the predicted end-point carbon temperature, the predicted end-point carbon temperature is regulated, and the predicted end-point carbon temperature and the actual end-point carbon temperature are used to update the basic model parameters.

[0013] As a further improved scheme of the present application: the collected conventional working condition basic data is specifically:

[0014] Collecting conventional working condition basic data in the converter steelmaking process, the conventional working condition basic data is the basic parameter of the steelmaking process, which directly affects the composition and temperature change of the molten steel; wherein the conventional working condition basic data includes the content of C carbon, Si silicon, P phosphorus and S sulfur elements.

[0015] As a further improved scheme of the present application: the specific process of constructing a multi-modal feature library based on abnormal events is:

[0016] For abnormal events occurring in the steelmaking process, spitting and dry return, multi-modal data is collected and a feature library is established for identification and early warning of abnormal events, multi-modal data including flame images and acoustic signals are collected, when spitting occurs, the brightness of the flame at the furnace mouth will change suddenly, and the high frequency component in the acoustic signal will surge; when dry return occurs, the sound and flame shape in the furnace will also have specific changes;

[0017] The collected multi-modal data is analyzed to extract key features that can represent abnormal events, i.e. brightness mutation features of flame images and high frequency component surge features of acoustic signals, and a multi-modal feature library of abnormal events is established.

[0018] As a further improved scheme of the present application: the specific process of constructing a basic model using conventional working condition basic data is:

[0019] A three-layer LSTM stack with 256 neurons in each layer is used to capture the long and short-term dependencies of the decarburization process.

[0020] A self-attention layer is added before the output layer to calculate the weight of each time step feature and focus on the key operation stage, wherein the calculation formula of the weight is: wherein, is the similarity between the current hidden state and all time step hidden states, which is calculated by dot product, is the hidden state of the t-th time step, is the hidden state set of all time steps, is the normalization denominator, which ensures that the sum of the attention weights of all time steps is 1.

[0021] As a further improvement of the present application: the specific process of the feature alignment is:

[0022] The conventional working condition data is standardized and time series aligned, wherein the standardization processing scales the numerical features to interval, and the time series alignment refers to that the blowing time of different heats is uniformly interpolated to 300 time steps.

[0023] As a further improvement of the present application: the specific process of the constraint on the base model is:

[0024] Based on the double film theory, the decarburization rate satisfies: wherein, is the mass transfer coefficient, is the gas-liquid interface area, is the equilibrium carbon concentration, is the current carbon content of the molten steel; the mean square error between the decarburization rate predicted by the model and the calculated value of the formula is used as an auxiliary loss during training: , is the kinetic auxiliary loss, is the number of time steps, is the decarburization rate predicted by the model at time t, is the decarburization rate calculated by the mechanism formula at time t;

[0025] Based on the heat balance equation, the constraint condition is: wherein, is the oxygen blowing heat release, is the decarburization heat release, is the slag evaporation heat absorption, is the heat radiation, is the heat required for molten steel temperature rise, and the deviation between the temperature predicted by the base model and the calculated value of the equation is less than 10℃.

[0026] As a further improvement of the present application: the specific process of expanding a large number of virtual data is:

[0027] Based on the WGAN-GP generative adversarial network, a virtual furnace data with physical rationality is generated under the condition of a small amount of real samples in extreme working conditions.

[0028] When the furnace condition enters the extreme working condition, the data enhancement module is automatically started, and the time sequence disturbance of the sensor signal of the current furnace is performed, that is, 5% noise and feature space interpolation are added.

[0029] The elastic weight consolidation algorithm is used, the normal knowledge learned by the basic model is retained when the new extreme working condition sample is added, and the virtual sample is created by using the WGAN-GP driven physical reasonable data.

[0030] As a further improvement of the present application: the specific process of embedding the metallurgical constraint is:

[0031] The constraint layer is added after the generator output layer to ensure that the generated data meets the following conditions:

[0032] The decarburization rate is monotonically decreasing, that is, it meets the carbon concentration reduction rule in the later blowing period;

[0033] The temperature change rate is less than 5℃ / min, that is, the violation of the heat balance principle is avoided;

[0034] The basicity R is in the range of [2.5, 3.5], that is, the physical property constraint of the slag is met;

[0035] The generated data x needs to meet: If not, the generated sample is corrected by the projection gradient method, wherein, is the change amount of carbon content, that is, the reduction value of carbon content in the blowing process, is the time change amount, that is, the time step of blowing, is the change amount of temperature, is the content of calcium oxide in the slag, is the content of silicon dioxide in the slag.

[0036] As a further improvement of the present application: the specific process of regulating the predicted end-point carbon temperature is:

[0037] When the prediction deviation is greater than 10℃, the dynamic blow strategy is triggered;

[0038] The metallurgical mechanism constraint is introduced to reasonably check the decision of the basic model, and when the deviation between the end-point carbon temperature time predicted by the basic model and the end-point carbon temperature calculated by the mechanism is greater than 15%, the artificial collaborative correction mechanism is started.

[0039] As a further improvement of the present application: the specific process of updating the basic model parameters by the predicted end-point carbon temperature and the actual end-point carbon temperature is:

[0040] After each furnace, the actual end-point carbon temperature is compared with the predicted end-point carbon temperature, and the basic model parameters are updated by TD time difference learning to form an iterative optimization loop of control and feedback.

[0041] In the second aspect, the present application provides a converter steelmaking end-point intelligent control system based on artificial intelligence, comprising:

[0042] The multi-modal feature library module: collects conventional working condition basic data to construct working condition features and abnormal events, and constructs a multi-modal feature library based on abnormal events for identification and early warning of abnormal events;

[0043] The basic model training module: uses conventional working condition basic data to construct a basic model, trains the basic model to migrate and adapt from conventional working conditions to extreme working conditions through cross-domain feature alignment and target domain adaptive processing, and embeds metallurgical laws into the basic model training to constrain the basic model, and uses the ADDA algorithm to solve the problem of insufficient extreme working condition samples;

[0044] The physical and metallurgical constraint module: combines a small amount of extreme working condition samples with physical constraints to generate a generative adversarial network, expands a small amount of extreme working condition samples into a large amount of virtual data, uses a WGAN-GP improved loss function combined with embedded metallurgical constraints to make the basic model learn extreme working condition data, and uses feature space interpolation technology to perfect the extreme working condition data;

[0045] The regulation and control module: based on the predicted end-point carbon temperature of the basic model, regulates and controls the predicted end-point carbon temperature, and updates the basic model parameters by the predicted end-point carbon temperature and the actual end-point carbon temperature.

[0046] The beneficial effects of the present application are as follows:

[0047] 1. Cross-domain feature migration is realized by the ADDA algorithm, a small amount of extreme working condition samples is expanded into a large amount of virtual data by combining a WGAN-GP generative adversarial network with physical constraints, effectively solving the problem of insufficient extreme working condition samples, the basic model is trained to migrate and adapt from conventional working conditions to extreme working conditions through cross-domain feature alignment and target domain adaptive processing, the model is prevented from forgetting learned conventional knowledge when learning new working conditions by combining elastic weight consolidation (EWC), the model generalization is enhanced, the metallurgical laws (double membrane theory, heat balance equation, etc.) are embedded into the basic model training to constrain the model, ensuring that the generated data conforms to the physical laws of monotonically decreasing decarburization rate and temperature change rate <5℃ / min, and the model prediction is more accurate.

[0048] 2. The model controls the final carbon temperature based on the predicted value. When the prediction deviation exceeds the limit, a dynamic supplementary blowing strategy or a manual collaborative correction mechanism is triggered. The model parameters are updated by comparing the actual and predicted final carbon temperatures and using TD time difference learning to form a control feedback iterative optimization loop. This improves smelting efficiency and steel quality. A multimodal feature library is constructed to identify and warn of abnormal events such as splashing and re-drying during the steelmaking process, providing a basis for intelligent control. Attached Figure Description

[0049] The invention will now be further described with reference to the accompanying drawings.

[0050] Figure 1 This is a flowchart of the steps of the present invention;

[0051] Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0052] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0053] Example 1:

[0054] like Figure 1 As shown in the embodiment of the present invention, an intelligent control method for the endpoint of converter steelmaking based on artificial intelligence includes:

[0055] S1: Collect basic data on normal operating conditions to construct operating condition characteristics and abnormal events, and build a multimodal feature library based on abnormal events for the identification and early warning of abnormal events;

[0056] The specific basic data for normal operating conditions are as follows:

[0057] Basic data for routine operating conditions: Collect basic data for routine operating conditions such as molten iron composition (C, Si, P, S), scrap steel ratio, and oxygen blowing parameters (flow rate, pressure, lance position trajectory);

[0058] Specifically, basic data on conventional operating conditions during the converter steelmaking process are collected. These basic data are fundamental parameters of the steelmaking process, directly affecting the composition and temperature changes of molten steel. Among them, the basic data on conventional operating conditions include the content of elements such as C (carbon), Si (silicon), P (phosphorus), and S (sulfur).

[0059] The contents of C (carbon), Si (silicon), P (phosphorus), and S (sulfur) elements directly affect the decarburization and slag-forming reactions and the final quality of molten steel.

[0060] The specific process for constructing the operating condition features is as follows:

[0061] The erosion degree feature of the furnace lining is extracted by an infrared thermal imager, and when the erosion depth is greater than 5 cm, it is marked as an extreme working condition, and the slag layer fluctuation frequency spectrum feature is obtained by a sonar sensor, and when the standard deviation is greater than 0.5 dB, it is determined as an extreme working condition;

[0062] Specifically, the abnormal event is specifically:

[0063] For the abnormal events occurring in the steelmaking process, such as spitting and dry return, multi-modal data is collected and a feature library is established for identification and early warning of abnormal events;

[0064] Multi-modal data including flame images and acoustic signals are collected;

[0065] For example, when spitting occurs, the brightness of the flame at the furnace mouth will change suddenly, and the high-frequency component in the acoustic signal will increase sharply; when dry return occurs, the sound in the furnace and the flame shape will also change in a specific way;

[0066] Feature library establishment: analyze the collected multi-modal data, extract key features that can represent abnormal events, such as brightness mutation features of flame images and high-frequency component increase features of acoustic signals, and establish a multi-modal feature library of abnormal events; this feature library is used for subsequent analysis of real-time collected data to identify and warn abnormal events in time, providing a basis for intelligent control;

[0067] Specifically, the slag layer fluctuation frequency spectrum feature is obtained by using a sonar sensor, and at the same time, a field programmable gate array (FPGA) is used for real-time frequency spectrum analysis, a band-pass filtering method is used to remove environmental noise, and a time domain normalization method is used to normalize the energy of the acoustic signal detected each time to The interval;

[0068] Short-time Fourier transform is used to generate a time-frequency graph, and the time-frequency graph is calculated to focus on the 20-80 kHz high-frequency band, which is the energy interval of the slag layer fluctuation;

[0069] A multi-modal feature library of spitting and dry return abnormal events is established, which includes flame image brightness mutation and acoustic signal high-frequency component increase features;

[0070] At the same time, the multi-modal feature library also includes the following data information:

[0071] The scrap steel ratio, i.e., the proportion of scrap steel in the raw materials, is recorded; the addition amount of the scrap steel ratio will affect the heat balance in the furnace and the composition of the molten steel, so the scrap steel ratio is also an index for measuring the quality of the final molten steel;

[0072] Record the oxygen blowing parameters, wherein the oxygen blowing parameters include: oxygen blowing flow, pressure and gun position trajectory; oxygen blowing is a key operation of converter steelmaking, and the control of oxygen blowing parameters directly affects the decarburization rate, the temperature of molten steel and the progress of the reaction in the furnace;

[0073] Record the amount of slag agent added, wherein different slag agents have significant differences in slag basicity, desulfurization and dephosphorization efficiency, and furnace lining protection;

[0074] Record the furnace gas composition, wherein the furnace gas composition reflects the progress of the carbon-oxygen reaction in the furnace, whether the escape path of the furnace gas is smooth, and the interaction state of the oxygen lance jet and the molten pool;

[0075] S2: Use conventional working condition basic data to build a basic model, train the basic model to migrate and adapt from conventional working conditions to extreme working conditions through cross-domain feature alignment and target domain adaptive processing, and embed the metallurgical law into the basic model training. The basic model is constrained, and the ADDA algorithm is used to solve the problem of insufficient extreme working condition samples;

[0076] Basic model pre-training, using conventional working condition basic data to build an LSTM-Attention basic model, learning decarburization dynamics, thermal balance general metallurgical law, let the basic model master the process and result mapping relationship under the conventional steelmaking scene, lay the foundation for complex working condition adaptation;

[0077] Target domain adaptation, for abnormal events, use the adversarial transfer learning (ADDA) algorithm, reduce the feature distribution difference between the source domain (non-abnormal events) and the target domain (abnormal events) through the gradient reversal layer;

[0078] Extracting meta-knowledge from historical abnormal event samples;

[0079] 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;

[0080] Specifically, 80% of the conventional working condition data is selected, and the conventional working condition data is divided into: time series feature data and static data; wherein, the time series feature data includes: oxygen blowing flow, gun position trajectory, furnace gas composition; static data includes: initial C / Si / P / S content of molten iron, scrap steel ratio, slag agent addition amount;

[0081] Standardize the conventional working condition data and align the time series, wherein the standardization processing scales the numerical features to interval, and the time series alignment means that the blowing time of different heats is uniformly interpolated to 300 time steps;

[0082] Use a three-layer LSTM stack with 256 neurons per layer to capture long and short-term dependencies in the decarburization process;

[0083] For example, the change of decarburization rate is strongly related to the lance position 3 minutes ago, and LSTM memorizes this correlation through cell state;

[0084] A self-attention layer is added before the output layer to calculate the weight of each time step feature and focus on the key operation stage. The calculation formula of the weight is: wherein, is the similarity between the current hidden state and all time step hidden states, which is calculated by dot product, is the hidden state of the t-th time step, is the hidden state set of all time steps, is the normalization denominator, ensuring that the sum of the attention weights of all time steps is 1;

[0085] Metallurgical laws are embedded in the model training to constrain the model. The constraint rules are as follows:

[0086] Based on the double film theory, the decarburization rate satisfies: wherein, is the mass transfer coefficient, is the gas-liquid interface area, is the equilibrium carbon concentration, is the current carbon content of the molten steel; During training, the mean square error between the decarburization rate predicted by the model and the value calculated by the formula is used as an auxiliary loss: , is the kinetic auxiliary loss, is the number of time steps, is the decarburization rate predicted by the model at time t, is the decarburization rate calculated by the mechanism formula at time t;

[0087] Based on the heat balance equation, the constraint condition is: wherein, is the heat released by oxygen blowing, is the heat released by decarburization, is the heat absorbed by slag evaporation, is the heat radiation, is the heat required for molten steel to rise in temperature. The deviation between the temperature predicted by the basic model and the value calculated by the equation should be less than 10°C;

[0088] Select 20% of the conventional working condition data to verify the accuracy of the model. If the prediction error of the end-point carbon is less than or equal to 0.01%, the temperature error is less than 10°C, and the decarburization rate curve is consistent with the mechanism model with a degree of more than 90%, it means that the basic model training is completed and the metallurgical laws are embedded effectively.

[0089] However, in actual production, raw material fluctuations and equipment aging will lead to frequent extreme working conditions, which requires further expansion of the model's cross-condition adaptation ability;

[0090] ADDA algorithm is used to realize cross-condition feature migration, and the generator G maps the source domain (normal condition data) and the target domain (extreme condition data) to a shared feature space , that is, , the discriminator D judges whether the feature from the source domain or the target domain, and outputs a probability ;

[0091] Using the idea of adversarial training, let and be indistinguishable in the feature space, deceive the discriminator, and the discriminator accurately distinguishes the source domain and the target domain features to improve the domain classification accuracy, where represents the feature mapping result of the generator G to the source domain data , represents the feature mapping result of the generator G to the target domain data ;

[0092] Using the gradient reversal layer (GRL) technology, the forward propagation is set as: , that is, the feature value is not changed, and the negative weight is added to the gradient through the backward propagation, that is , where 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, is the input feature vector, which can be the data features of the normal condition or the extreme condition, is the unit matrix, means the partial derivative of GRL with respect to x;

[0093] Loss function and optimization, generator loss: , the goal is to let the discriminator misjudge the source domain features as the target domain, where is the generator loss, which measures the ability of the generator to deceive the discriminator, is the expectation, which takes the average of all samples in the data set, , where is the data sample of the normal condition, is the normal condition data set, , where is the data sample of the extreme condition, is the extreme condition data set, and D is a probability value output by the discriminator;

[0094] Discriminator loss: , where is the discriminator loss, is the mathematical expectation, , where generated data for normal working conditions, is a normal working condition data set, , generated data for extreme working conditions, is an extreme working condition data set, , is the output of the generator, is the discriminator; the ability of the discriminator to distinguish between normal working conditions and extreme working conditions is measured, and the goal is to maximize the domain classification accuracy; alternating training is adopted: first, fix G to optimize D, then fix D to optimize G, until the classification accuracy of the discriminator falls to 50%;

[0095] meta-knowledge extraction, key features are extracted from historical extreme working condition samples, and a decarburization rate constraint term is manufactured, and the decarburization rate constraint term is: , wherein, is the decarburization rate change of the ith sample, unit (% / min), is the phosphorus content change of the ith extreme working condition sample, is the meta-knowledge coefficient, is the deviation between the predicted phosphorus content change of the base model and the influence of the decarburization rate and the metallurgical priori knowledge;

[0096] Adversarial transfer (ADDA) solves the problem of few extreme working condition samples, but pure data-driven may deviate from the metallurgical mechanism;

[0097] For example, if the predicted basicity exceeds [2.5, 3.5] and the decarburization rate violates the phosphorus content inhibition rule; therefore, it is necessary to embed metallurgical priori knowledge to make the base model both cross-condition adaptive and comply with physical laws;

[0098] Manufacture of basicity constraint term, basicity constraint term is: , to ensure that the predicted slag basicity , falls in interval, and is punished when it exceeds the range, wherein T is the time step, and in the present application, the time step is equal to 300, is the predicted slag basicity at time t, the center value of the 3.0 basicity reasonable interval, and the half-width of the 0.5 basicity reasonable interval, is the basicity constraint term;

[0099] Loss function integration, the final loss function is: , wherein, is the prediction loss, is the kinetic constraint loss, which measures the fitness of the base model to the metallurgical kinetics, is the deviation between the predicted phosphorus content change of the base model and the influence of the decarburization rate and the metallurgical priori knowledge, is the basicity constraint term, , , are weight coefficients, determined by cross-validation;

[0100] S3: Combine a small number of extreme working condition samples with physical constraints to generate a generative adversarial network, expand a small number of extreme working condition samples into a large number of virtual data, use WGAN-GP to improve the loss function, and combine the embedding metallurgical constraints to make the basic model learn the extreme working condition data, and use the feature space interpolation technology to perfect the extreme working condition data;

[0101] Virtual sample generation: based on WGAN-GP generative adversarial network, generate virtual heat data with physical rationality based on a small number of real samples in extreme working conditions;

[0102] For example, simulate the temperature-carbon content change curve when the phosphorus content is 0.4%;

[0103] When it is detected that the furnace condition enters the extreme working condition, the data enhancement module is automatically started, and the time sequence disturbance of the sensor signal of the current heat is performed, that is, 5% noise and feature space interpolation are added;

[0104] EWC (Elastic Weight Consolidation) algorithm is used to retain the conventional knowledge learned by the basic model when new extreme working condition samples are added, and to avoid catastrophic forgetting;

[0105] WGAN-GP driven physical reasonable data is used to create virtual samples, and traditional GAN has the problems of single generated data and unstable training, and WGAN-GP solves the problems through the following improvements:

[0106] Wasserstein distance is used to replace KL divergence to measure the difference between generated distribution and real distribution, and the gradient is more stable;

[0107] A gradient penalty term is introduced to ensure that the discriminator is Lipschitz continuous and avoid gradient disappearance;

[0108] The generator G inputs random noise z and condition vector c, and generates heat data through multiple fully connected layers;

[0109] The discriminator D outputs the Wasserstein distance estimate value, and does not use the sigmoid activation function;

[0110] A constraint layer is added after the output layer of the generator to ensure that the generated data meets the following conditions:

[0111] The decarburization rate is monotonically decreasing, that is, it conforms to the carbon concentration reduction rule in the later blowing period;

[0112] The temperature change rate is <5℃ / min, that is, to avoid violating the heat balance principle;

[0113] R∈[2.5,3.5], i.e. the slag physical property constraint;

[0114] The generated data x needs to satisfy: If not, the generated sample is corrected by the projection gradient method, wherein, is the change of carbon content, i.e. the reduction of carbon content in the blowing process, is the time change, i.e. the time step of blowing, is the change of temperature, is the content of calcium oxide (CaO) in the slag, is the content of silicon dioxide in the slag ;

[0115] To solve the problems of single generated data and unstable training of traditional GAN, WGAN-GP improves the loss function and combines the embedded metallurgical constraint to make the basic model learn physically reasonable extreme working condition data;

[0116] Training process and parameters, loss function: wherein, is the real data distribution, is the sample distribution sampled on the line connecting the real and generated data, is the conditional vector, is the gradient penalty weight, In the above, is the real data, In the above, is random noise, is the distribution of random noise, is the expectation, which is the average of all samples in the data set, ensuring that the loss is a global statistic, and D is the output of the discriminator, In the above, is the interpolation sample between the real data and the generated data, is the distribution of the real data and the generated data sample, used to calculate the gradient penalty, is the gradient of the interpolation sample to the discriminator D,

[0117] Under the condition of input phosphorus content of 0.4%, WGAN-GP can generate 300 groups of virtual furnace data, wherein the Pearson correlation coefficient of the temperature-carbon content curve and the real sample is >0.92, and all of them meet the metallurgical constraint;

[0118] Using time series disturbance technology, when an extreme working condition is detected, dynamic noise is added to the sensor signal: wherein, is the disturbance coefficient, is the signal standard deviation, ensuring that the disturbance amplitude is within 5%, for the disturbed sensor signals, original sensor signals;

[0119] It is still difficult to cover the feature diversity of extreme conditions only by sensor signal disturbance;

[0120] Exemplary, complex slag changes of high-silicon hot metal;

[0121] Therefore, the feature space interpolation technique is introduced to improve the information of the current heat with similar historical conditions, and to expand the learning boundary of the basic model for extreme conditions:

[0122] The feature space interpolation technique is as follows:

[0123] For the current heat feature vector Interpolate between the historical similar working condition features and Generate a new feature vector: where, is the interpolation coefficient, which ensures that the new feature is within the physically feasible range;

[0124] Real-time enhancement trigger mechanism, multi-condition joint trigger, when one of the following conditions is met, start the enhancement module:

[0125] Abnormal hot metal composition: Si>1.2% or P>0.3%;

[0126] Abnormal lining state: erosion depth>5cm;

[0127] Abnormal slag condition: sonar SD>0.5dB and HFR>30%;

[0128] Incremental learning update: EWC algorithm prevents catastrophic forgetting;

[0129] The core principle of EWC is that some weights are very important for old tasks (regular working conditions), and these weights need to be preserved when learning new tasks (extreme working conditions).

[0130] In order to make the basic model accurately identify the important weights, that is, the weights that are crucial to the regular working conditions (old tasks), the importance of the weights needs to be quantified through the Fisher information matrix, and then applied to the training constraints of the extreme working conditions.

[0131] Fisher information matrix: calculate the importance of each weight w to the old task loss: where, where, is the regular working condition data set, is the regular working condition data set, is the old task loss, is the mathematical expectation, for the individual weight parameters of the model, for the old task loss gradient of the weight;

[0132] During new task training, a regularization term is added to penalize the change of important weights: where, is the weight after old task training, is the weight coefficient, is the new task loss, is the total loss of the new task, is the element of the Fisher information matrix, is the current weight of the model;

[0133] The Fisher information matrix F is calculated using normal operating condition data, and for extreme operating condition new samples, and the weight is updated, and during back propagation, constraints are imposed on important weights according to F, such as reducing the update step of the relevant weight of the furnace temperature prediction by 50%;

[0134] S4: Based on the predicted end-point carbon temperature of the base model, the predicted end-point carbon temperature is regulated, and the base model parameters are updated using the predicted end-point carbon temperature and the actual end-point carbon temperature;

[0135] Based on the real-time prediction of the end-point carbon temperature of the base model, the predicted end-point carbon temperature is regulated;

[0136] When the prediction deviation is greater than 10℃, the dynamic blow strategy is triggered;

[0137] And a metallurgical mechanism constraint is introduced to reasonably verify the decision of the base model, and when the deviation between the predicted end-point carbon temperature of the base model and the end-point carbon temperature calculated by the mechanism is greater than 15%, an artificial collaborative correction mechanism is started;

[0138] After each heat, the actual end-point carbon temperature is compared with the predicted end-point carbon temperature, and the base model parameters are updated through TD time difference learning to form an iterative optimization loop of control and feedback.

[0139] Embodiment 2:

[0140] As Figure 2 shown, based on embodiment 1, the present application provides a converter steelmaking end-point intelligent control system based on artificial intelligence, comprising:

[0141] A multi-modal feature library module: collect normal operating condition basic data to construct operating condition features and abnormal events, and construct a multi-modal feature library based on abnormal events for identification and early warning of abnormal events;

[0142] The basic model training module: a basic model is constructed by using the basic data of the conventional working condition, the migration and adaptation ability of the basic model from the conventional working condition to the extreme working condition is trained through cross-domain feature alignment and target domain self-adaptation processing, the metallurgical law is embedded in the basic model training, the basic model is constrained, and the ADDA algorithm is used to solve the problem of insufficient extreme working condition samples;

[0143] The physical and metallurgical constraint module: a small amount of extreme working condition samples is combined with physical constraints to generate a generative adversarial network, the small amount of extreme working condition samples is expanded into a large amount of virtual data, the WGAN-GP improved loss function is used in combination with embedded metallurgical constraints, the basic model learns extreme working condition data, and the feature space interpolation technology is used to perfect the extreme working condition data;

[0144] The regulation module: based on the predicted end-point carbon temperature, the predicted end-point carbon temperature is regulated, and the predicted end-point carbon temperature and the actual end-point carbon temperature are used to update the basic model parameters.

[0145] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application, and various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An intelligent control method for converter steelmaking endpoint based on artificial intelligence, characterized in that: Comprise: S1: Collecting conventional working condition basic data to construct working condition characteristics and abnormal events, and constructing a multi-modal feature library based on abnormal events for identification and early warning of abnormal events; The specific process of constructing a multi-modal feature library based on abnormal events is: For abnormal events occurring in the steelmaking process, spitting and dry return, collect multi-modal data and establish a feature library for identification and early warning of abnormal events. Collect multi-modal data including flame images and sound signals. When spitting occurs, the brightness of the flame at the furnace mouth will change suddenly, and the high-frequency component of the sound signal will surge. When dry return occurs, the sound and flame shape in the furnace will also have specific changes. Analyze the collected multi-modal data and extract key features that can represent abnormal events, i.e. brightness mutation features of flame images and high-frequency component surge features of sound signals, to establish a multi-modal feature library of abnormal events. S2: Construct a basic model using conventional working condition basic data, align cross-domain features, and adaptively process target domains to train the basic model to migrate and adapt from conventional working conditions to extreme working conditions, and embed metallurgical laws into the basic model training to constrain the basic model, and use ADDA algorithm to solve the problem of insufficient extreme working condition samples; The specific process of constructing a basic model using conventional working condition basic data is: A three-layer LSTM stack is used, with 256 neurons in each layer, to capture the long and short term dependencies of the decarburization process. An attention layer is added before the output layer to calculate the weight of each time step feature, and the calculation formula of the weight is: wherein, is the similarity between the current hidden state and all time step hidden states, which is calculated by dot product, is the hidden state of the tth time step, is the hidden state set of all time steps, is the normalization denominator, which ensures that the sum of the attention weights of all time steps is 1; The specific process of constraining the basic model is: Based on the double film theory, the decarburization rate satisfies: wherein, is the mass transfer coefficient, is the gas-liquid interface area, is the equilibrium carbon concentration, is the current molten steel carbon content; during training, the mean square error between the decarburization rate predicted by the model and the calculated value of the formula is taken as an auxiliary loss: , is the kinetic auxiliary loss, is the time step, is the decarburization rate at time t predicted by the model, is the decarburization rate at time t calculated by the mechanism formula; Based on the constraint of heat balance equation, the constraint condition is: Wherein, is the oxygen blowing heat release, is the decarburization heat release, is the slag evaporation heat absorption, is the heat radiation, is the heat required for molten steel temperature rise, the basic model predicts that the temperature required deviation is less than 10℃ from the calculated value of the heat balance equation; S3: Combine a small amount of extreme working condition samples with physical constraints to generate a generative adversarial network, expand a small amount of extreme working condition samples into a large amount of virtual data, use WGAN-GP to improve the loss function, combine embedded metallurgical constraints, let the basic model learn extreme working condition data, and use feature space interpolation technology to improve extreme working condition data; The specific process of embedding metallurgical constraints is: Add a constraint layer after the generator output layer to ensure that the generated data meets the following conditions: The decarburization rate is monotonically decreasing, i.e. it complies with the carbon concentration reduction law in the later blowing period; The temperature change rate is <5℃ / min, i.e. it avoids violating the heat balance principle; The basicity R is in [2.5, 3.5], i.e. it satisfies the physical property constraints of the slag; The generated data x needs to satisfy: If not, the generated sample is corrected by a projection gradient method, wherein, is a change in carbon content, i.e., a reduction in carbon content in the blowing process, is a time change, i.e., a time step of blowing, is a change in temperature, is a content of calcium oxide in the slag, is a content of silicon dioxide in the slag; When the erosion depth is >5cm, mark it as an extreme working condition, use a sonar sensor to obtain the fluctuation frequency spectrum features of the slag layer, and determine it as an extreme working condition when the standard deviation is >0.5dB; S4: Based on the predicted end-point carbon temperature, regulate the predicted end-point carbon temperature, and update the basic model parameters using the predicted end-point carbon temperature and the actual end-point carbon temperature.

2. The intelligent control method for the converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of collecting conventional working condition basic data is: Collecting conventional working condition basic data in the converter steelmaking process, the conventional working condition basic data is the basic parameter of the steelmaking process, which directly affects the composition and temperature change of the molten steel; Among them, the conventional working condition basic data includes: the content of C carbon, Si silicon, P phosphorus and S sulfur elements.

3. The intelligent control method for the converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of feature alignment is: The conventional working condition data is standardized and time-aligned, wherein the standardization processing scales the numerical characteristics to the interval, and the time alignment refers to that the converting times of different furnace batches are uniformly interpolated as 300 time steps.

4. The intelligent control method for the 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: Based on the WGAN-GP generative adversarial network, generate virtual furnace data with physical rationality based on a small amount of real extreme working condition samples; When the furnace condition is detected to enter an extreme working condition, an automatic data enhancement module is started, and time series disturbance is performed on the sensor signals of the current heat, that is, 5% noise and feature space interpolation are added; An elastic weight consolidation algorithm is used, when new extreme working condition samples are added, the conventional knowledge learned by the basic model is retained, and virtual samples are created by using WGAN-GP driven physical reasonable data.

5. The intelligent control method for the converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of regulating the predicted end-point carbon temperature is: When the prediction deviation is greater than 10℃, a dynamic blow strategy is triggered; And a metallurgical mechanism constraint is introduced, the rationality of the decision of the basic model is checked, and when the time deviation of the end-point carbon temperature predicted by the basic model and the end-point carbon temperature calculated by the mechanism is greater than 15%, an artificial collaborative correction mechanism is started.

6. The intelligent control method for the converter steelmaking endpoint based on artificial intelligence according to claim 1, characterized in that: The specific process of updating the basic model parameters by using the predicted end-point carbon temperature and the actual end-point carbon temperature is: After each heat, the actual end-point carbon temperature is compared with the predicted end-point carbon temperature, the basic model parameters are updated by TD time difference learning, and an iterative optimization loop of control and feedback is formed.

Images