Process parameter optimization method and device based on migration condition generative adversarial network, medium and equipment

Abstract

The invention relates to a process parameter optimization method and device based on a migration condition generative adversarial network, a medium and equipment, and the method comprises the steps: obtaining the production data of a plurality of historical heats of a first steel type and a second steel type, and selecting the scalar data with raw material information as the main part and the time sequence data with process parameters as the main part in the production data; constructing a migration condition generative adversarial network model, inputting the source domain data and the target domain data into the model at the same time, generating process parameters matched with corresponding steel grades through a feature migration method, and training the model to obtain a trained model; and obtaining information of new sample data of the target domain, and inputting the information into the trained model to obtain process parameters of a new sample. According to the method, the defects of a traditional model in generalization performance are effectively overcome for the problems in small sample steel production, and a more efficient solution is provided for small sample steel production.

Description

Technical Field

[0001] The present invention belongs to the field of metallurgical technology, and particularly relates to a process parameter optimization method, device, medium and equipment based on a migration conditional generative adversarial network. Background Art

[0002] Converter steelmaking is an important part of the iron and steel production process, and its purpose is to obtain molten steel with end-point composition and temperature meeting the target requirements. The converter steelmaking end-point control technology includes static control and dynamic control. Static control is the basis of dynamic control. If the static control method is too simple, a large amount of manual experience is required to participate in the control process of process parameters, which not only brings great challenges to the dynamic control in the later stage of smelting, but also easily leads to non-compliance of the end-point composition and temperature.

[0003] When producing new steel grades, in order to meet the needs of specific industries or fields, there are usually relatively strict requirements for material properties, and at the same time, its production cost is high, resulting in a small market demand. Since manufacturers often only produce in small batches according to customer needs, large-scale production tests cannot be carried out or sufficient production data cannot be accumulated, resulting in scarce sample data of new steel grades. In the production scenario of small-sample steel grades, the generalization performance of the model for new steel grades is poor, and how to effectively optimize the process parameters of small-sample steel grades is particularly important.

[0004] The converter steelmaking static control model usually includes the following models: mechanism model, incremental model, statistical model and artificial intelligence model. The mechanism model is based on various physical and chemical reactions occurring in the converter process, and uses the principles of material balance and heat balance to derive relevant formulas and calculate the values of various variables. In 2023, Tang disclosed a method for analyzing the quality of molten steel in the converter steelmaking process based on the fusion of data and mechanism, established a system model for the quality of molten steel in the steelmaking process by fusing data and mechanism, and proposed an adaptive filtering algorithm based on neural network for online analysis of the quality of molten steel to obtain a real-time estimated value of the quality of molten steel.

[0005] The statistical model is a mathematical description method based on the relationship between input and output variables in the production process. The construction of this model depends on the collection and analysis of a large amount of on-site production data. By applying the technical means of mathematical statistics, researchers can determine and quantify the influence of input variables on the output results and their internal relationships. In 2023, Chen et al. established a multiple linear regression model to calculate the lime addition amount. The verification results showed that the calculation accuracy was 93%, and this model was successfully applied to a 120t converter in a steel plant, reducing the consumption of lime.

[0006] The incremental model is a calculation method that predicts the process parameters required for the current heat by referring to historical heat data. It predicts key process variables by analyzing the process pattern of the previous heat or selected reference heats and combining the raw material characteristics and conditions of the current heat. In 2018, Li et al. applied the incremental model to the converter steelmaking process to control the end point of the molten steel. The entire campaign of the converter was regarded as continuous, while ignoring the changes in the molten pool and the influence of slag-making materials on blowing in adjacent heats, and only referring to the smelting results of the previous heat to examine the influence on the relative changes in the process parameters of the current heat. The static model includes a slag-making model, a coolant addition model, and an oxygen consumption addition model. Production through this model has greatly improved the smelting effect, and the splashing rate has been reduced by 1.24%.

[0007] Compared with traditional algorithms, the artificial intelligence model has significant advantages in dealing with complex and changeable process, and can effectively cope with the randomness and uncertainty in the process production. In 2022, He et al. disclosed a method for controlling the oxygen blowing amount in the TSC stage of a large converter, which uses a random forest artificial intelligence algorithm with a targeted oxygen blowing amount model to construct the internal relationship between various influencing factors of the converter, and realizes the accurate prediction and control of the oxygen blowing amount in the TSC stage of the large converter. In 2024, Sun et al. based on the prediction model of the improved neural network learning extreme machine (IELM), combined with the improved particle swarm optimization algorithm (PSO) as the control model, obtained the optimized oxygen consumption and scrap addition amount. 280 samples were used for training and 142 samples were used for testing. The temperature was considered to be on target when the temperature deviation was within the range of ±15°C, and the carbon content was considered to be on target when the carbon content deviation was within the range of ±0.015%. The hit rate of the temperature was 72.676%, and the hit rate of the carbon content was 82.112%.

[0008] However, due to the complex composition of the charged raw materials and auxiliary materials and the variety of steel grades in the converter steelmaking process, traditional static control models are often difficult to adapt to such highly variable process requirements in actual steelmaking production. Facing the production scenarios of small-sample steel grades, the existing static control models have poor generalization performance, which leads to insufficient accuracy when optimizing process parameters, thus affecting the control effect of the end point composition and temperature of small-sample steel grades.

[0009] In summary, when producing new steel grades, it is usually to meet the needs of specific industries or fields. Their production costs are high and the output is limited, resulting in less sample data. For the optimization of process parameters for the production of these steel grades, more adaptable and flexible control methods need to be introduced to improve production efficiency and end point hit rate. Summary of the Invention

[0010] In order to overcome the problem that it is difficult for static models in the prior art to comprehensively capture the variation laws of key process parameters for small-sample steel grades, resulting in the inability to accurately establish the mapping relationship between raw material characteristics and process parameters, and thus causing a low hit rate of the end-point temperature, the present invention provides a process parameter optimization method, device, medium and equipment based on a migration conditional generative adversarial network to solve the above problems existing in the prior art.

[0011] A process parameter optimization method based on a migration conditional generative adversarial network, the method specifically includes the following steps:

[0012] S1) Obtain the production data of a number of historical heats of the first steel grade and the second steel grade, and select the scalar data mainly based on raw material information and the time-series data mainly based on process parameters in the production data. Among them, the number of production data of the first steel grade is greater than that of the second steel grade. The production data of the first steel grade is used as the source domain data, and the production data of the second steel grade is used as the target domain data;

[0013] S2) Construct a migration conditional generative adversarial network model, input the normalized source domain data and target domain data into the model at the same time, train the model, and after training is completed, generate the process parameters matching the first steel grade and the second steel grade respectively and the trained model;

[0014] S3) Obtain the information of the new sample data of the target domain, and input it into the trained model to obtain the process parameters of the new sample.

[0015] In the above aspect and any possible implementation manner, a further implementation manner is provided. The scalar data includes hot metal temperature, hot metal addition amount, hot metal Si content, hot metal Mn content, hot metal P content, hot metal S content, scrap addition amount and target end-point temperature, and this part is raw material information; the time-series data includes oxygen lance height, oxygen flow rate, bottom-blowing N2 flow rate and bottom-blowing Ar flow rate, and this part is process parameters.

[0016] In the above aspect and any possible implementation manner, a further implementation manner is provided. The migration conditional generative adversarial network model includes three connected parts: a feature migration layer, a generator and a discriminator;

[0017] The feature migration layer is used to align the features of the target domain data and the source domain data, and output consistent feature inputs;

[0018] The generator deeply learns the feature inputs to generate process parameters matching the raw material data;

[0019] The discriminator is used to determine the approximation degree between the generated process parameters and the actual process parameters.

[0020] For the aspects and any possible implementation manners described above, a further implementation manner is provided. Specifically, S2 includes:

[0021] S21) Input the source domain data and the target domain data into the feature transfer layer simultaneously. By minimizing the difference in the feature distributions of the source domain and the target domain data, the two are made closer in the feature space, and the data with aligned features is generated;

[0022] S22) Input the data with aligned features and the generated random noise into the generator to generate corresponding process parameters;

[0023] S23) Input the generated process parameters and the real process parameters into the discriminator to obtain the output score of the generated process parameters;

[0024] S24) The generator and the discriminator perform adversarial training so that the generator gradually generates process parameters approaching the actual ones.

[0025] For the aspects and any possible implementation manners described above, a further implementation manner is provided. The feature transfer layer uses the Maximum Mean Discrepancy (MMD) and the Covariance Alignment (CORAL) loss functions to calculate the mean difference and covariance difference between the source domain samples and the target domain samples, which are respectively:

[0026] The Maximum Mean Discrepancy (MMD) loss function quantifies the distribution distance between the source domain data and the target domain data by calculating the mean difference in the high-dimensional feature space, thereby realizing the distribution alignment between the source domain data and the target domain data. Its expression is:

[0027]

[0028] where E[f(H1)] represents the feature mean of all source domain samples, E[f(H2)] represents the feature mean of all target domain samples, represents the mean difference between the source domain samples and the target domain samples;

[0029] The Covariance Alignment (CORAL) loss function reduces the difference in the variance distribution between the two by minimizing the covariance difference between the source domain data and the target domain data. Its expression is:

[0030]

[0031] where, is the feature covariance matrix of the source domain data, is the feature covariance matrix of the target domain data, |||| F represents the Frobenius norm, that is, the square root of the sum of the squares of the matrix elements, represents the covariance difference between the source domain samples and the target domain samples.

[0032] For the aspects and any possible implementation described above, a further implementation is provided. The model adopts a loss function L, and its value is the sum of the maximum mean discrepancy (MMD) loss L MMD , the covariance alignment (CORAL) loss L CORAL , and the generative adversarial loss L GD , and its expression is:

[0033] L = L MMD + L CORAL + L GD

[0034] wherein, the generative adversarial loss L GD includes the generator loss L G and the discriminator loss L D .

[0035] For the aspects and any possible implementation described above, a further implementation is provided. The random noise vector is generated in S1), specifically: random noise vectors consistent with the dimension of the time-series data variables are generated respectively according to the sample numbers of the first steel type and the second steel type.

[0036] The present invention also provides a process parameter optimization device based on a transfer conditional generative adversarial network. The device is used to implement the method described above, and the device includes:

[0037] An acquisition module, configured to acquire production data of a number of historical furnace batches of the first steel type and the second steel type, and select scalar-type data mainly based on raw material information and time-series data mainly based on process parameters in the production data. Among them, the number of production data of the first steel type is greater than that of the second steel type. The production data of the first steel type is used as source domain data, and the production data of the second steel type is used as target domain data;

[0038] A construction module, configured to construct a transfer conditional generative adversarial network model, input the normalized source domain data and target domain data into the model at the same time, train the model, and after the training is completed, generate process parameters matching the first steel type and the second steel type and the trained model respectively;

[0039] An obtaining module, configured to obtain information on new sample data in the target domain, and input it into the trained model to obtain process parameters of the new sample.

[0040] The present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by a processor to implement the method described above.

[0041] The present invention also provides an electronic device, and the electronic device includes:

[0042] A memory that stores executable instructions;

[0043] A processor that runs the executable instructions in the memory to implement the described method.

[0044] Advantages of the present invention

[0045] 1. The present invention proposes a process parameter optimization method based on a transfer conditional generative adversarial network, which effectively overcomes the deficiencies of traditional models in generalization performance for the problems in the production of small-sample steel grades, and provides a more efficient solution for the production of small-sample steel grades;

[0046] 2. The present invention introduces a feature transfer learning method. By combining the maximum mean discrepancy loss and the covariance alignment loss function, it minimizes the distribution difference between the source-domain steel grade data and the target-domain steel grade data, so that the model can better adapt to the data characteristics of the target-domain steel grades, and improves the accuracy of the process parameter optimization for small-sample steel grades;

[0047] 3. The present invention improves the network structure of the generator, adopts a hybrid model of one-dimensional convolution and long short-term memory network, fully explores the strong coupling relationship between scalar data mainly based on raw material information and time-series data mainly based on process parameters, and improves the stability of the production process;

[0048] 4. The present invention is applied to the process parameter optimization of converter steelmaking. Through the existing raw material information, it can directly generate the parameters of the oxygen lance height, oxygen flow rate, bottom blowing N2 flow rate, and bottom blowing Ar flow rate that match the raw material information, which helps iron and steel enterprises improve production efficiency, realize steelmaking intelligence, and significantly promote the improvement of economic benefits and the improvement of the quality of steel products. Brief description of the drawings

[0049] Figure 1 is a schematic flowchart of the process parameter optimization method based on a transfer conditional generative adversarial network provided by an embodiment of the present invention;

[0050] Figure 2 is a comparison diagram of the probability density distributions of the source-domain data and the target-domain data along the x-axis and y-axis after dimensionality reduction processing before and after transfer learning in an embodiment of the present invention. Among them, (a) represents the comparison diagram of the probability density distribution of the normalized source-domain and target-domain scalar data along the x-axis, (b) represents the comparison diagram of the probability density distribution of the feature-aligned source-domain and target-domain scalar data along the x-axis, (c) represents the comparison diagram of the probability density distribution of the normalized source-domain and target-domain scalar data along the y-axis, and (d) represents the comparison diagram of the probability density distribution of the feature-aligned source-domain and target-domain scalar data along the y-axis;

[0051] Figure 3It is a schematic diagram for comparing and analyzing the process parameters of the new samples in the target domain and the original process parameters provided by the embodiments of the present invention;

[0052] Figure 4 It is a schematic diagram of the predicted end temperature of the new samples in the target domain provided by the embodiments of the present invention. Specific embodiments

[0053] For a better understanding of the technical solution of the present invention, the content of the present invention includes but is not limited to the specific embodiments hereinafter. Similar technologies and methods should be regarded as within the scope of protection of the present invention. To make the technical problems to be solved, technical solutions and advantages of the present invention clearer, the following will be described in detail with reference to the drawings and specific embodiments.

[0054] It should be clear that the embodiments described in the present invention are only a part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0055] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "the" and "said" used in the embodiments of the present invention and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0056] The present invention provides a process parameter optimization method based on a transfer conditional generative adversarial network. The method specifically includes the following steps:

[0057] S1) Obtain the production data of several historical heats of the first steel grade and the second steel grade, and respectively select the scalar data mainly based on raw material information and the time series data mainly based on process parameters from the production data of the two steel grades. Among them, the number of production data of the first steel grade is greater than the number of samples of the second steel grade. Take the selected production data of the first steel grade as the source domain data, and the production data of the second steel grade as the target domain data;

[0058] S2) Construct a transfer conditional generative adversarial network model, and input the source domain data and the target domain data into the model at the same time. Train the model. After the training is completed, generate the process parameters matching the first steel grade and the second steel grade respectively;

[0059] S3) Obtain the information of the production data of the new samples in the target domain, and input it into the trained model to obtain the process parameters of the new samples.

[0060] Further, the scalar data includes hot metal temperature, hot metal addition amount, hot metal Si content, hot metal Mn content, hot metal P content, hot metal S content, scrap addition amount, and target end temperature; the time-series data includes lance height, oxygen flow rate, bottom blowing N2 flow rate, and bottom blowing Ar flow rate.

[0061] Further, the migration conditional generative adversarial network model includes three connected parts: a feature migration layer, a generator, and a discriminator;

[0062] The feature migration layer is used to align the features of the target domain data and the source domain data and output consistent features;

[0063] The generator deeply learns the features to generate process parameters that match the raw material data;

[0064] The discriminator is used to determine the approximation degree between the generated process parameters and the actual process parameters.

[0065] Further, the S2 specifically includes:

[0066] S21) Input the source domain data and the target domain data into the feature migration layer at the same time. By minimizing the difference in the feature distributions of the source domain and the target domain data, the two are made closer in the feature space to generate aligned feature data;

[0067] S22) Input the aligned feature data and the generated random noise into the generator to generate corresponding process parameters;

[0068] S23) Input the generated process parameters and the real process parameters into the discriminator to obtain the output score of the generated process parameters;

[0069] S24) The generator and the discriminator perform adversarial training so that the generator gradually generates process parameters approaching the actual ones.

[0070] Further, the feature migration layer uses the maximum mean discrepancy MMD and the covariance alignment CORAL loss function to calculate the mean difference and covariance difference between the source domain samples and the target domain samples, which are respectively:

[0071] The maximum mean discrepancy MMD loss function quantifies the distribution distance between the source domain data and the target domain data by calculating the mean difference in the high-dimensional feature space, so as to realize the distribution alignment of the source domain data and the target domain data. Its expression is:

[0072]

[0073] Among them, E[f(H1)] represents the feature mean of all source domain samples, and E[f(H2)] represents the feature mean of all target domain samples, Indicates the mean difference between source domain samples and target domain samples;

[0074] The covariance alignment CORAL loss function reduces the difference in variance distribution between the source domain data and the target domain data by minimizing the covariance difference between them. Its expression is:

[0075]

[0076] Where, is the feature covariance matrix of the source domain data, is the feature covariance matrix of the target domain data, |||| F represents the Frobenius norm, that is, the square root of the sum of the squares of the matrix elements, represents the covariance difference between the source domain samples and the target domain samples.

[0077] Furthermore, the model adopts a loss function L, whose value is the sum of the maximum mean difference MMD loss L MMD , the covariance alignment CORAL loss L CORAL , the generative adversarial loss L GD and the generation loss L R . Its expression is:

[0078] L = L MMD + L CORAL + L GD + L R

[0079] Where, the generative adversarial loss L GD includes the generator loss L G and the discriminator loss L D .

[0080] Specifically, the specific process of the present invention is as follows:

[0081] As Figure 1 shown, a process parameter optimization method based on a transfer conditional generative adversarial network of the present invention specifically includes the following steps:

[0082] S1) Obtain the production data of several historical heats of steel grade 1 and steel grade 2, and select scalar data mainly based on raw material information and time series data mainly based on process parameters from them. Among them, steel grade 1 is used as the first steel grade, and the number of production data of the first steel grade is greater than the number of production data of the second steel grade. The selected production data of the first steel grade is used as the source domain data, and steel grade 2 is used as the second steel grade, and the production data of the second steel grade is used as the target domain data. The production data includes scalar data and time series data, and the number of samples referred to is the number of production data samples, which is numerically consistent with the number of scalar data and time series data;

[0083] S11) Collect the production data of several historical heats of Steel Grade 1 and Steel Grade 2 from the converter steelmaking site. Both the first Steel Grade 1 and the second Steel Grade 2 are low-C and low-P steel grades, but their grades are different: Steel Grade 1 has a larger production quantity and a greater market demand, serving as the source domain data; Steel Grade 2 is a newly developed steel grade with a smaller production quantity, serving as the target domain data. The production data participating in the modeling mainly includes two categories: scalar data and time-series data. Among them, the scalar data includes hot metal temperature, hot metal addition amount, hot metal Si content, hot metal Mn content, hot metal P content, hot metal S content, scrap addition amount, and target end temperature. This part is raw material information; the time-series data includes lance height, oxygen flow rate, bottom blowing N2 flow rate, and bottom blowing Ar flow rate. This part is process parameters; the process flows of Steel Grade 1 and Steel Grade 2 are the same. Steel Grade 1 is currently being produced and sold, and Steel Grade 2 is a new product. Optimizing the process parameters can be regarded as the process parameter generation process. This model is trained by putting Steel Grade 1 and Steel Grade 2 together, making the generated process parameters of Steel Grade 2 better.

[0084] S12) Before inputting the data into the model, perform data preprocessing; among them, for scalar data, for the outlier data in the production data samples, the 3σ criterion is adopted for judgment. If the data is abnormal, it is directly removed; for the missing value data in the production data samples, the method of mean filling is adopted; for time-series data, since the blowing time of each heat is of unequal length, interpolation processing is carried out to align the sampling points; for the preprocessed scalar data and time-series data collection, normalization processing is carried out in the variable dimension, mapping the data into the interval [0, 1]. The function is to eliminate the dimensional differences of all the data input into the model;

[0085] S13) Generate random noise vectors: Generate random noise vectors with the same dimension as the time-series data variables respectively according to the sample quantities of the first steel grade and the second steel grade. They are generated according to the standard normal distribution, with an average value of 0 and a standard deviation of 1. The purpose is to input them into the generator to generate corresponding process parameters.

[0086] S2) Construct a migration conditional generative adversarial network model, input the normalized source domain data and target domain data into the model at the same time, train the model. After training is completed, generate the process parameters matching the first steel grade and the second steel grade respectively, and the trained model;

[0087] The adversarial network model includes three parts: a feature migration layer, a generator, and a discriminator;

[0088] The said feature migration layer is used to align the features of the target domain data and the source domain data and output consistent input features;

[0089] The generator deeply learns the input features to generate process parameters matching the raw material data;

[0090] The discriminator is used to determine the approximation degree between the generated process parameters and the actual process parameters.

[0091] This step specifically includes: S21) Inputting the preprocessed source domain data and target domain data into the feature migration layer simultaneously, and minimizing the difference in the feature distributions of the source domain and target domain data to make them closer in the feature space;

[0092] The feature migration layer adopts a combination of a two-layer fully connected network structure and the nonlinear activation function Sigmoid to map the normalized source domain data and target domain data into a high-dimensional feature space, reduce the distribution difference between the source domain and the target domain, achieve the feature alignment of the preprocessed source domain data and target domain data, thereby eliminating the distribution difference between the source domain and the target domain, and providing consistent input features for the subsequent generator. The normalized source domain data and target domain data are both samples with the end point temperature hitting.

[0093] The input of the feature migration layer is divided into two parts, one part is the normalized source domain scalar data, and the other part is the normalized target domain scalar data; the output of the feature migration layer is the feature-aligned source domain and target domain scalar data.

[0094] Regarding the two-layer fully connected network structure and the nonlinear activation function Sigmoid as a whole, the following description reflects this whole.

[0095] For the normalized scalar data X1 in the source domain and the normalized scalar data X2 in the target domain, the source domain data output H1 and the target domain output H2 are respectively expressed as:

[0096] H1 = W2(Sigmoid(W1X1 + b1)) + b2(1)

[0097] H2 = W2(Sigmoid(W1X2 + b1)) + b2(2)

[0098] Among them, W1 and b1 are unknowns, representing the weight matrix and bias coefficient of the first-layer fully connected network structure respectively, W2 and b2 are unknowns, representing the weight matrix and bias coefficient of the second-layer fully connected network structure respectively, and Sigmoid() represents the nonlinear activation function.

[0099] The feature transfer layer introduces the Maximum Mean Discrepancy (MMD) and Covariance Alignment (CORAL) loss functions after the second fully connected network. By minimizing the distribution difference between the source domain output H1 and the target domain output H2 through these two loss functions, feature-aligned source domain data and feature-aligned target domain data are obtained, thus significantly enhancing the similarity between the data, completing the feature transfer between the scalar data of the source domain and the target domain after normalization, and solving the drawback of poor model generalization ability caused by the small number of samples of the small-sample steel grades.

[0100] The Maximum Mean Discrepancy (MMD) loss measures and minimizes the distribution difference between the source domain data and the target domain data. It quantifies the distribution distance between them by calculating the mean difference between the source domain data output H1 and the target domain output H2 in the high-dimensional feature space, thereby achieving the distribution alignment between the source domain data and the target domain data. Its expression is:

[0101]

[0102] where E[f(H1)] represents the feature mean of all source domain data, and E[f(H2)] represents the feature mean of all target domain data. These two means can be obtained using known or existing calculation methods, which will not be elaborated in this invention. The smaller it is, the smaller the mean difference between the source domain output H1 and the target domain output H2, and thus the better the model transfer effect.

[0103] The Covariance Alignment (CORAL) loss aims to reduce the difference in variance distribution between the source domain data and the target domain data by minimizing the covariance difference between them, thereby improving the performance of the model in the target domain. Its expression is:

[0104]

[0105] where is the feature covariance matrix of the source domain data, is the feature covariance matrix of the target domain data. These two matrices can be obtained using known or existing methods, which will not be elaborated in this invention. || || F represents the Frobenius norm, that is, the square root of the sum of the squares of the matrix elements. The smaller it is, the smaller the covariance difference between the source domain data and the target domain data, and thus the better the model transfer effect.

[0106]

[0107] where H 1i is the variable vector of the i-th sample in the scalar data of the source domain, is the mean of each sample in H1, and n is the number of samples in the scalar data of the source domain.

[0108]

[0109] Among them, H 2i' is the variable vector of the i'-th sample, is the mean value of each sample in H2, and n' is the number of samples in the target domain data.

[0110] The migration conditional generative adversarial network model introduces a loss function L, which is used during model training to guide the optimization of model parameter adjustment. The maximum mean discrepancy (MMD) loss L MMD , covariance alignment (CORAL) loss L CORAL , and generative adversarial loss L GD are combined in the following way, and its expression is:

[0111] L = L MMD + L CORAL + L GD

[0112] The generative adversarial loss L GD includes the generator loss L G , discriminator loss L D .

[0113] The generator loss L G is defined as follows:

[0114] L G = E z:p(z) [log(1 - D(G(z)))]

[0115] where G and D are the generator and discriminator respectively, z is the generated random noise vector, and E z～p(z) is the expectation following the probability distribution p of z.

[0116] The discriminator loss L D is defined as follows:

[0117]

[0118] where z is the generated random noise vector, is the generated random noise vector, E z～p(z) is the expectation following the probability distribution p of z, is the expectation following the probability distribution q, is the existing algorithm formula, and both p and q are numerical values in [0, 1].

[0119] Furthermore, by comparing the distribution characteristics of the two, it can be intuitively seen that there are differences in the mean and variance between the normalized source-domain and target-domain scalar data and the source-domain and target-domain scalar data with feature alignment, thus verifying the effect of transfer learning on data distribution. As Figure 2 (a) represents the comparison diagram of the probability density distribution of the normalized source-domain and target-domain scalar data along the x-axis, Figure 2 (b) represents the comparison diagram of the probability density distribution of the source-domain and target-domain scalar data with feature alignment along the x-axis, Figure 2 (c) represents the comparison diagram of the probability density distribution of the normalized source-domain and target-domain scalar data along the y-axis, Figure 2 (d) represents the comparison diagram of the probability density distribution of the source-domain and target-domain scalar data with feature alignment along the y-axis. Figure 2 (a) and Figure 2 (b) verify that the means between the source-domain and target-domain scalar data with feature alignment after transfer learning are more similar, Figure 2 (c) and Figure 2 (d) verify that the variances between the source-domain and target-domain scalar data with feature alignment after transfer learning are more similar.

[0120] S22) Input the source-domain and target-domain scalar data with feature alignment and the generated random noise vector into the generator to generate corresponding process parameters;

[0121] The generator adopts a combination of a one-dimensional convolutional module and a long short-term memory module;

[0122] Specifically, the one-dimensional convolutional module extracts local features of the generated random noise by moving the convolutional kernel, and can capture the change law of the random noise within the time window;

[0123] The long short-term memory module extracts global features of the generated random noise through memory units and gate mechanisms, and can capture the dependency relationship of the random noise changing over time.

[0124] The source-domain and target-domain scalar data with feature alignment, and the random noise are processed to obtain global and local features, which are all fused and mapped into the shared feature space through a fully connected layer to generate source-domain process parameters and target-domain process parameters.

[0125] S23) Input the generated source-domain process parameters and target-domain process parameters, and the normalized source-domain time-series data and target-domain time-series data into the discriminator to distinguish the differences between the real time-series data and the generated process parameters, and obtain the output score of the generated process parameters;

[0126] The output score is a probability value between 0 and 1. The closer this value is to 1, the closer the generated process parameters are to the true process parameter score. The closer this value is to 0, the greater the difference between the generated process parameters and the true process parameters.

[0127] S24) Start training the migration conditional generative adversarial network model. First, fix the feature migration layer and the generator, and train the discriminator to distinguish between true process parameters and generated process parameters. Then, fix the discriminator, and train the feature migration layer and the generator to generate more realistic data to deceive the discriminator. When the discriminator can no longer effectively distinguish between true process parameters and generated process parameters, the training terminates. At this time, the process parameters generated by the generator are used as the process parameters approximating the actual ones, thereby improving the quality and accuracy of the generated process parameters and solving the traditional static control.

[0128] S3) Input the new sample raw material information data of the target domain and the generated random noise vector into the feature migration layer and the generator in the trained migration conditional generative adversarial network model. After being processed by the model, the process parameters corresponding to the new sample are generated. This method directly generates process parameters matching the raw material information. Compared with the original process parameters, when the new parameters are used and processed by the trained model, the end-point hit rate is significantly improved after production.

[0129] Furthermore, as Figure 3 Process parameter curves of the lance height, oxygen flow rate, bottom blowing N2 flow rate, and bottom blowing Ar flow rate are respectively generated for the new samples in the target domain. This sample is a furnace with an inaccurate end-point temperature, indicating that the newly generated process parameters have completed the process parameter optimization compared with the original process parameters.

[0130] Using the historical production data of a 260t converter in a steel plant to test this method, given the raw material information of the new samples in the target domain, and using the migration conditional generative adversarial network model, the process parameters of the new samples in the target domain are obtained. The original process parameters of the new samples with a hit end-point temperature have an end-point temperature hit rate of 87.5% through the prediction model, and the generated process parameters have an end-point hit rate of 84.3% through the prediction model; the original process parameters of the new samples with an inaccurate end-point temperature have an end-point temperature hit rate of 21.8% through the prediction model, and the generated process parameters have an end-point hit rate of 84.8% through the prediction model. As Figure 4 (a) shows the end-point temperature prediction results of the new samples in the target domain using the original process parameters and the generated process parameters. This sample is a furnace with a hit end-point temperature using the original process parameters; as Figure 4 (b) shows the end-point temperature prediction results of the new samples in the target domain using the original process parameters and the generated process parameters. This sample is a furnace with an inaccurate end-point temperature using the original process parameters.

[0131] As an embodiment disclosed by the present invention, the present invention further provides a process parameter optimization device based on a transfer conditional generative adversarial network. The device is used to implement the method described above. The device includes: an acquisition module, configured to acquire production data of a number of historical furnace batches of a first steel type and a second steel type, and select scalar-type data mainly based on raw material information and time-series data mainly based on process parameters from the production data. Among them, the number of production data of the first steel type is greater than that of the second steel type. The production data of the first steel type is used as source domain data, and the production data of the second steel type is used as target domain data;

[0132] A construction module, configured to construct a transfer conditional generative adversarial network model, input the normalized source domain data and target domain data into the model at the same time, train the model, and after the training is completed, generate process parameters matching the first steel type and the second steel type respectively and the trained model;

[0133] An obtaining module, configured to obtain information of new sample data of the target domain, input it into the trained model, and obtain the process parameters of the new sample.

[0134] As an embodiment disclosed by the present invention, the present invention further provides a computer storage medium, on which a computer program is stored, and the computer program is executed by a processor to implement the method.

[0135] As an embodiment disclosed by the present invention, the present invention further provides an electronic device, and the electronic device includes:

[0136] A memory, storing executable instructions;

[0137] A processor, where the processor runs the executable instructions in the memory to implement the method.

[0138] The above description shows and describes several preferred embodiments of the present invention. However, as mentioned above, it should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be within the scope of the application concept described herein, and can be modified through the above teachings or the technology or knowledge in related fields. And any changes and modifications made by those skilled in the art without departing from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. A process parameter optimization method based on migration condition generative adversarial network, characterized in that: The method specifically comprises the following steps: S1) obtaining production data of several historical heats of the first steel grade and the second steel grade, selecting scalar data mainly based on raw material information and time series data mainly based on process parameters from the production data, wherein the amount of production data of the first steel grade is greater than the amount of production data of the second steel grade, taking the production data of the first steel grade as source domain data, and taking the production data of the second steel grade as target domain data; S2) constructing a migration conditional generative adversarial network model, inputting the normalized source domain data and target domain data into the model at the same time, training the model, and after the training is completed, generating process parameters matching the first steel grade and the second steel grade and the trained model respectively; S3) Obtain information of new sample data and random noise vectors in the target domain, input them into the trained model, and obtain process parameters of the new samples.

2. The method according to claim 1, characterized in that The scalar data include molten iron temperature, molten iron addition amount, molten iron Si content, molten iron Mn content, molten iron P content, molten iron S content, scrap steel addition amount and target endpoint temperature; the time series data include oxygen lance height, oxygen flow rate, bottom blowing N2 flow rate and bottom blowing Ar flow rate.

3. The method according to claim 1, characterized in that The transfer conditional generative adversarial network model includes three parts: a feature transfer layer, a generator and a discriminator connected to each other; The feature migration layer is used to align the features of the target domain data and the source domain data and output consistent input features; The generator performs in-depth learning on the input features to generate process parameters matching the raw material data; The discriminator is used to determine the degree of approximation between the generated process parameters and the actual process parameters.

4. The method according to claim 3, characterized in that The S2 specifically includes: S21) inputting the source domain data and the target domain data into the feature migration layer at the same time, minimizing the difference in feature distribution of the source domain data and the target domain data so that the two are closer in feature space, thereby generating feature-aligned data; S22) inputting the feature-aligned data and the corresponding random noise into a generator to generate corresponding process parameters; S23) inputting the generated process parameters and the true process parameters into the discriminator to obtain an output score of the generated process parameters; S24) The generator and the discriminator are trained adversarially so that the generator gradually generates process parameters that are close to the actual ones.

5. The method according to claim 3, characterized in that: The feature migration layer uses the maximum mean difference MMD and covariance alignment CORAL loss function to calculate the mean difference and covariance difference between the source domain samples and the target domain samples, which are: The maximum mean difference MMD loss function quantifies the distribution distance between the source domain data and the target domain data by calculating the mean difference between them in the high-dimensional feature space, thereby achieving distribution alignment between the source domain data and the target domain data. Its expression is: Among them, E[f(H1)] represents the feature mean of all source domain samples, and E[f(H2)] represents the feature mean of all target domain samples. Represents the mean difference between source domain samples and target domain samples; The covariance alignment CORAL loss function reduces the difference in variance distribution between the source domain data and the target domain data by minimizing the covariance difference between the two. Its expression is: in, is the feature covariance matrix of the source domain data, is the feature covariance matrix of the target domain data, |||| F represents the Frobenius norm, which is the square root of the sum of the squares of the matrix elements. Represents the covariance difference between source domain samples and target domain samples.

6. The method according to claim 5, characterized in that The model uses a loss function L, whose value is the maximum mean difference MMD loss L MMD , covariance alignment CORAL loss L CORAL , Generate adversarial loss L GD The sum of is expressed as: L=L MMD +L CORAL +L GD Among them, the generated adversarial loss L GD Including the generator loss L G and the discriminator loss L D .

7. The method according to claim 1, characterized in that The random noise vector is generated in S1), specifically: a random noise vector consistent with the dimension of the time series data variable is generated according to the sample numbers of the first steel grade and the second steel grade respectively.

8. A process parameter optimization device based on migration condition generation adversarial network, characterized in that: The device is used to implement the method according to any one of claims 1 to 7, and the device comprises: an acquisition module, used to acquire production data of several historical heats of the first steel grade and the second steel grade, select scalar data mainly based on raw material information and time series data mainly based on process parameters from the production data, wherein the amount of production data of the first steel grade is greater than the amount of production data of the second steel grade, and use the production data of the first steel grade as source domain data and the production data of the second steel grade as target domain data; A construction module is used to construct a transfer conditional generative adversarial network model, input the normalized source domain data and target domain data into the model at the same time, train the model, and after the training is completed, generate process parameters matching the first steel grade and the second steel grade and the trained model respectively; The acquisition module is used to obtain the information of new sample data in the target domain, input it into the trained model, and obtain the process parameters of the new sample.

9. A computer storage medium, characterized in that The medium stores a computer program, and the computer program is executed by a processor to implement the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The electronic device comprises: A memory storing executable instructions; A processor, wherein the processor runs the executable instructions in the memory to implement the method according to any one of claims 1 to 7.

Images