Converter batching data generation and optimization method

Abstract

The invention discloses a converter batching data generation and optimization method, which comprises the following steps: acquiring an original industrial data set in a converter batching process, preprocessing the original industrial data set, inputting preprocessed training data into an encoder, generating a mean vector and a variance vector of potential variable distribution, and optimizing the mean vector and the variance vector; constraining the similarity between the potential variable distribution and the standard Gaussian distribution through KL divergence loss; inputting the potential variable and the random noise vector into a generator to generate a simulation sample, and distinguishing a real sample from the generated sample through a discriminator; a forced discriminator module is introduced, and a discrimination result is converted into an additional loss item; and carrying out data generation on the missing batching parameters, outputting a complete batching data set conforming to process constraints, constructing a cost objective function, a quality objective function and a resource consumption objective function, constructing constraint conditions of chemical element content of a target finished product, optimizing the converter batching data, and generating optimized converter batching data.

Description

Technical Field

[0001] The present invention relates to the field of industrial Internet of Things, and particularly to a method for generating and optimizing converter batching data. Background Art

[0002] With the rapid development of Internet of Things technology, a large amount of data can be collected in real time at the industrial site. During the process of converter steelmaking, since the data of various batching materials have a great impact on the final product, it is necessary to collect the data to facilitate researchers to make timely adjustments according to the data. However, these data often fail to meet the completeness condition. Specifically, the auxiliary variables (such as temperature, pressure, flow rate, etc.) used as the input of the soft measurement model usually have a high sampling frequency, while the key quality variables (label variables) as the output of the model often have a low sampling frequency or even serious missing due to the difficulty or high cost of measurement. This data imbalance phenomenon leads to a scarcity of label samples in the industrial process, severely restricting the training effect of the soft measurement model and the performance of online update. Therefore, how to effectively fill the missing label samples has become one of the research hotspots in the field of industrial process soft measurement. The generative model, with its powerful data generation ability, provides a new idea for solving the problem of missing label samples. By learning the internal distribution structure of the data, this type of model can generate complementary data highly consistent with the real samples, thus providing sufficient data support for the training and update of the downstream soft measurement model. However, in the actual industrial scenario, due to the scarcity and discontinuous distribution of label samples, the training effect of the generative model is often restricted.

[0003] There are few comparative documents in this field. In the patent with the application number 202011352099.3 in the related field, an industrial fault diagnosis unbalanced time-series data augmentation method is disclosed. Step 1: Prepare the training data set; Step 2: Construct the network structure of GRUBEGAN; Step 3: Train the constructed GRUBEGAN network model; Step 4: Use the trained GRUBEGAN generative adversarial network model to generate artificial data of the small sample type. After training, the model inputs a simple random variable z|t to generate time-series data conforming to time t, and expands the generated data set into the small sample type of the original data. The industrial site environment is complex, and the data often contains noise and outliers, resulting in unstable sample quality, making it difficult for the generator to accurately capture key data features, and there may be a significant deviation between the generated samples and the real values. Summary of the Invention

[0004] To achieve the above object and other related objects, the present invention discloses a method for generating and optimizing converter batching data, including: Step S1: Obtain the original industrial dataset of the converter batching process, and preprocess the original industrial dataset to obtain the preprocessed training data. The original industrial dataset includes multi-dimensional sensor data and the corresponding missing batching quality parameters; Step S2: Input the preprocessed training data into the encoder to generate the mean vector and variance vector of the latent variable distribution, and use the KL divergence loss to constrain the similarity between the latent variable distribution and the standard Gaussian distribution; Step S3: Input the latent variable and the random noise vector into the generator to generate simulated samples, and use the discriminator to distinguish between real samples and generated samples; Introduce a forced discriminator module, and the forced discriminator performs compliance discrimination on the generated samples based on the prior constraint conditions defined by the converter metallurgy process knowledge base, and converts the discrimination result into an additional loss term; Step S4: Construct a loss function, including the generative adversarial loss, KL divergence loss, additional loss of the forced discriminator, and reconstruction loss, and jointly optimize the parameters of the generator and discriminator through backpropagation; Step S5: Use the trained FD-VEGAN model to generate data for the missing batching parameters, and output a complete batching dataset that meets the process constraints; Step S6: Construct a cost objective function, a quality objective function, and a resource consumption objective function, and construct the constraint conditions for the chemical element composition content of the target finished product, and use the improved TD3 algorithm to solve the problem, optimize the converter batching data, and generate the optimized converter batching data.

[0005] Further, Step S2 includes: The encoder encodes the input real samples into a mean vector and a variance vector; On the one hand, combine the above vectors into latent variables and input them into the generator. On the other hand, compare the KL divergence with the mean and variance vectors of the standard Gaussian distribution to generate a regularization loss; Quantify the difference between the reconstructed samples and the real samples through the reconstruction loss to generate a reconstruction loss; The encoder optimizes its own performance according to the regularization loss and the reconstruction loss.

[0006] Further, Step S3 includes: The generator receives the random noise and the latent variable output by the encoder as inputs, generates corresponding noise samples and reconstructed samples, and uses the discriminator to distinguish the generated samples from the real samples to generate a discriminator loss; Introduce a forced discriminator based on prior knowledge, and use its discrimination result as an additional loss; The generator performs adversarial training according to the discriminator loss and the additional loss to generate high-quality samples.

[0007] Furthermore, the regularization loss includes: ; ; Among them, is the probability distribution characteristic of the latent variable z obtained based on the specific input x, is the posterior distribution given by the encoder, is the prior distribution of the latent space, is the definition of KL divergence, and the subscripts prior and post respectively refer to the prior distribution and posterior distribution generated from the encoder, and respectively represent the mean and standard deviation of the prior distribution of the standard Gaussian distribution, and respectively represent the mean and standard deviation of the posterior distribution output by the encoder.

[0008] Furthermore, the reconstruction loss includes: ; is a multivariate Gaussian distribution.

[0009] Furthermore, in the basic VAE model, the total loss of the regularization loss and the reconstruction loss is: ; Among them is the reconstruction loss, is the regularization loss.

[0010] Furthermore, in the dynamic β-VAE model, the total loss of the regularization loss and the reconstruction loss is: ; Among them is the reconstruction loss in the nth round, is the regularization loss in the nth round; is a parameter that changes dynamically according to the round.

[0011] Furthermore, the loss of the generator in step S3 is: ; Among them, and are weighting coefficients, which are respectively used to adjust the losses of the samples generated from the latent variable and the noise , is the weight coefficient of the reconstruction loss, is the reconstruction loss, ensuring that the generator can reconstruct the input data according to the latent variable. The generator not only needs to generate noise samples to create new data and also generate potential variables from the encoder to reconstruct real samples.

[0012] Furthermore, the loss function of the discriminator is: ; where is the score of the discriminator for real samples, is the score of the discriminator for the samples generated from the potential variables generated by the encoder, is the score of the discriminator for the samples generated from the noise .

[0013] Furthermore, among a batch of M samples, a certain generated virtual sample is , for the i-th element therein, the size of the additional penalty generated by it is set as follows: ; In the formula, is the element involved in forced discrimination in the i-th sample, is the upper limit of the value of the element in the sample set according to prior knowledge in the forced discriminator, is the lower limit, is an adjustable parameter, and the sum of the additional penalties composed of this batch of samples is: .

[0014] Furthermore, step S6 includes: Real-time monitor the progress of each objective function, and dynamically adjust the weights according to the relative change rate of the objective function to avoid over-optimization of the objective function; Add historical experience samples, calculate the cosine similarity between the current state and historical experience, preferentially select relevant experiences, and maintain the diversity of training by combining the selection of random experiences, so as to accelerate the convergence of the strategy and avoid overfitting; Add noise during the solution process, and adaptively adjust the noise scale according to the reward variance and the number of training steps, so that the solution converges to the optimal strategy.

[0015] Furthermore, adaptively adjusting the noise scale according to the reward variance and the number of training steps includes: Every time several rounds of training are performed, calculate the average reward variance of the recent rollbacks. When the variance is lower than the preset value, increase the noise, and when the variance is higher than the preset value, reduce the noise; The noise also decays exponentially with the number of training steps.

[0016] Second aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, characterized in that the method described above when the program is executed by a processor.

[0017] Third aspect, the present invention provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the method described above.

[0018] Fourth aspect, the present invention provides a converter burden data generation and optimization system, including: A data acquisition module, configured to acquire an original industrial data set of the converter burden process, and preprocess the original industrial data set to obtain preprocessed training data, where the original industrial data set includes multi-dimensional sensor data and corresponding missing burden quality parameters; A pre-training module, configured to input the preprocessed training data into an encoder to generate a mean vector and a variance vector of a latent variable distribution, and constrain the similarity between the latent variable distribution and a standard Gaussian distribution through a KL divergence loss; A sample generation module, configured to input the latent variable and a random noise vector into a generator to generate simulated samples, and distinguish real samples from generated samples through a discriminator; introduce a forced discriminator module, where the forced discriminator performs compliance discrimination on the generated samples based on prior constraint conditions defined by a converter metallurgy process knowledge base, converts the discrimination result into an additional loss term, and jointly optimizes the parameters of the generator and the discriminator; A data generation module, configured to use the trained FD-VEGAN model to generate data for the missing burden parameters, and output a complete burden data set that meets the process constraints; A burden data optimization module, configured to construct a cost objective function, a quality objective function, and a resource consumption objective function, and construct constraint conditions for the chemical element composition content of the target finished product, and use an improved TD3 algorithm for solution to optimize the converter burden data and generate optimized converter burden data.

[0019] Further, in the pre-training module: The encoder encodes the input real samples into a mean vector and a variance vector; On the one hand, the above vectors are combined into latent variables and input into the generator, and on the other hand, they are compared with the mean and variance vectors of the standard Gaussian distribution through KL divergence to generate a regularization loss; Quantify the difference between the reconstructed samples and the real samples through a reconstruction loss to generate a reconstruction loss; The encoder optimizes its own performance according to the regularization loss and the reconstruction loss.

[0020] Further, in the sample generation module: The generator takes random noise and the latent variables output by the encoder as inputs, generates corresponding noise samples and reconstructed samples, and distinguishes the generated samples from the real samples through the discriminator to generate the discriminator loss. A forced discriminator based on prior knowledge is introduced, and its discrimination result is used as an additional loss. The generator performs adversarial training according to the discriminator loss and the additional loss to generate high-quality samples.

[0021] Through the above technical solutions, first, the VAE model is introduced to construct the latent space. The existing data is encoded by VAE to capture the global features of the data and construct a structured latent space. Second, after obtaining the latent vectors generated by the VAE encoder, the generator of GAN assigns different generation weights to different vectors in the latent space, thereby generating more representative samples. In addition, the real samples are mapped to the latent space through the encoder, forming a closed-loop bidirectional mapping mechanism with the samples generated by the generator, further enhancing the similarity and distribution consistency between the generated samples and the real data. Combined with the forced discriminator module, the method can effectively guide the generator to more accurately simulate the key data features and generate high-quality samples. Then, the improved TD3 algorithm is used for solution to optimize the converter burdening data and generate the optimized converter burdening data. Description of the Drawings

[0022] In combination with the drawings and with reference to the following detailed description, the above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent. The drawings are used to better understand the solution and do not constitute a limitation to the present disclosure. In the drawings, the same or similar reference numerals represent the same or similar elements, where: Figure 1 is the flow chart of the present invention; Figure 2 is the system block diagram of the present invention; Figure 3 is the framework diagram of the FD-VEGAN method; Figure 4 is the data enhancement module based on VAE; Figure 5 is the sample generation module based on FD-GAN; Figure 6 is the framework diagram of the converter alloy burdening optimization method based on the improved TD3 algorithm Figure 7 is the comparison diagram of the experimental results of each comparison model; Figure 8 is the comparison diagram of the experimental results of each ablation model; Figure 9 is the change diagram of the total amount of raw materials G, the raw material cost F and the product quality Z. Detailed Embodiments

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present invention.

[0024] Referring to Figure 1 , an oxygen converter burden data generation and optimization method provided by an embodiment of the present invention includes: Step S1: Obtain the original industrial data set of the oxygen converter burden process, and preprocess the original industrial data set to obtain the preprocessed training data. The original industrial data set includes multi-dimensional sensor data and corresponding missing burden quality parameters; Step S2: Input the preprocessed training data into the encoder to generate the mean vector and variance vector of the latent variable distribution, and constrain the similarity between the latent variable distribution and the standard Gaussian distribution through the KL divergence loss; Step S3: Input the latent variable and the random noise vector into the generator to generate simulated samples, and use the discriminator to distinguish between real samples and generated samples; introduce a forced discriminator module, and the forced discriminator performs compliance discrimination on the generated samples based on the prior constraint conditions defined by the oxygen converter metallurgy process knowledge base, and converts the discrimination result into an additional loss term; Step S4: Construct a loss function, including the generative adversarial loss, the KL divergence loss, the additional loss of the forced discriminator, and the reconstruction loss, and jointly optimize the parameters of the generator and the discriminator through backpropagation; Step S5: Use the trained FD-VEGAN model to generate data for the missing burden parameters and output a complete burden data set that meets the process constraints.

[0025] Step S6: Construct a cost objective function, a quality objective function, and a resource consumption objective function, and construct the constraint conditions for the chemical element composition content of the target finished product, and use the improved TD3 algorithm to solve the problem, optimize the oxygen converter burden data, and generate the optimized oxygen converter burden data.

[0026] Further, step S2 includes: The encoder encodes the input real sample into a mean vector and a variance vector; On the one hand, the above vectors are combined into a latent variable and input into the generator, and on the other hand, they are compared with the mean and variance vectors of the standard Gaussian distribution through KL divergence to generate a regularization loss; Quantify the difference between the reconstructed sample and the real sample through the reconstruction loss to generate the reconstruction loss; The encoder optimizes its own performance according to the regularization loss and the reconstruction loss.

[0027] Further, step S3 includes: The generator receives random noise and the latent variable output by the encoder as inputs, generates corresponding noise samples and reconstructed samples, and distinguishes the generated samples from the real samples through the discriminator to generate the discriminator loss; Introduce a forced discriminator based on prior knowledge and use its discrimination result as an additional loss; The generator performs adversarial training according to the discriminator loss and the additional loss to generate high-quality samples.

[0028] Further, the regularization loss includes: ; ; wherein, is the posterior distribution given by the encoder, is the prior distribution of the latent space (usually a standard normal distribution), is the definition of KL divergence, and the subscripts prior and post respectively refer to the prior distribution and posterior distribution generated from the encoder.

[0029] Further, the reconstruction loss includes: ; is a multivariate Gaussian distribution.

[0030] Further, the total loss of the regularization loss and the reconstruction loss is: ; where is the reconstruction loss, is the regularization loss.

[0031] Further, the total loss of the regularization loss and the reconstruction loss is: ; where is the reconstruction loss, is the regularization loss; The dynamically changing is expressed as: ); wherein, the sign function returns the sign of any real number ; ; ; ; ; Among them, the hyperparameters are all previous rounds, is the last round of change. Among them, is the weight decay rate for controlling the loss of the regularization term, is the weight enhancement rate for controlling the loss of the reconstruction term, is the momentum factor, used to stabilize the mutation of the β value, are respectively the smoothing coefficients of the historical loss reference window; Furthermore, the loss of the generator in step S3 is: ; Among them, and are weighting coefficients, used to adjust the losses of the samples generated from the latent variable and the noise respectively, is the reconstruction loss, ensuring that the generator can reconstruct the input data according to the latent variable. The generator not only needs to create new data by generating noise samples , but also reconstruct the real samples by generating the latent variable generated by the encoder.

[0032] Furthermore, the loss function of the discriminator is: ; Among them, is the score of the discriminator for the real sample, is the score of the discriminator for the sample generated from the latent variable generated by the encoder, is the score of the discriminator for the sample generated from the noise .

[0033] Furthermore, among a batch of M samples, a certain generated virtual sample is , for the i-th element in it, the size of the additional penalty generated by it is set as follows: ; In the formula, is the element involved in the forced discrimination in the i-th sample, is the upper limit of the value of the element in the sample set according to prior knowledge in the forced discriminator, is the lower limit, is an adjustable parameter, and the sum of the additional penalties composed of this batch of samples is: 。

[0034] Step S6 specifically includes: Monitor the progress of each objective function in real time, and dynamically adjust the weights according to the relative change rate of the objective function to avoid over-optimization of the objective function; Add historical experience samples. By calculating the cosine similarity between the current state and historical experience, preferentially select relevant experiences, and combine the selection of random experiences to maintain the diversity of training, thereby accelerating the convergence of the policy and avoiding overfitting; Add noise during the solution process, and adaptively adjust the noise scale according to the reward variance and the number of training steps, so that the solution converges to the optimal policy.

[0035] Next is an embodiment of the present invention to explain the above steps.

[0036] To solve the problem that the scarcity of label samples affects the training and updating effect of the downstream soft sensor model in the actual industrial scenario, this embodiment proposes a sample filling method based on FD-VEGAN. First, a pre-training module based on VAE is proposed to construct a latent space, capture the global features and hidden structures of the input data, and generate latent vectors; secondly, a sample generation model based on FD-GAN is proposed. After obtaining the latent vectors generated by VAE, GAN further generates complementary samples according to the sampling of the latent space, combines the discriminator and the forced discriminator module to guide the generator to more accurately simulate the key data features, solves the problems of unstable model training and mode collapse, and ensures the high quality and stability of the generated samples.

[0037] Generating samples using a generative model usually depends on a large number of high-quality training samples, that is, when the label samples are scarce, the high-quality available samples are scarce, which affects the training effect of the generative model. To address this problem, referring to Figure 3 and Figure 4 , this embodiment proposes a pre-training module based on VAE. First, the encoder encodes the input real samples into a mean vector and a variance vector. On the one hand, these variables are combined into latent variables as inputs to the generator, and on the other hand, they are compared with the mean and variance vectors of the standard Gaussian distribution for KL divergence to regularize the encoder; referring to Figure 4 and Figure 5 , secondly, the generator takes the latent variables as inputs and generates reconstructed samples similar to the original samples through the decoding process. The difference between the reconstructed samples and the real samples is quantified by the reconstruction loss to optimize the performance of the encoder, ensure the reasonable distribution of the latent variables, and prompt the generator to generate high-quality samples.

[0038] In the input layer, VAE encodes the input into a multivariate latent distribution to improve the structure of the latent space. The encoder maps the input to a multivariate Gaussian distribution Approximately real, and then the decoder reconstructs the input data from the latent variables, which is obtained from the density function The encoder and decoder neural networks are respectively and parameterized. The loss function includes a reconstruction term and a regularization term. The reconstruction term measures the difference between the input samples and the generated samples, and the regularization term is used to constrain the structure of the latent space to ensure that the posterior distribution of the latent variables is close to the prior distribution. Usually, the Kullback-Leibler (KL) divergence is used to calculate the difference between the posterior distribution and the prior distribution. The loss function is shown in Equation 1.1: (1.1) (1.2) (1.3) where, is the probability distribution characteristic of the latent variable z obtained based on the specific input x, is the posterior distribution given by the encoder, is the prior distribution of the latent space (usually a standard normal distribution), is the definition of the KL divergence, and the subscripts prior and post respectively refer to the prior distribution and the posterior distribution generated from the encoder. The total loss function of the VAE is shown in Equation 1.4: (1.4) (1.5) The quality of the auto-encoded reconstruction is controlled by , is the regularization loss. At a higher level, the regularization term controls the smoothness and regularity of the latent space. The trade-off between these two losses affects the performance of the VAE. This embodiment proposes to apply a scaling trick to the regularization loss, resulting in the objective function of the modified β-VAE: (1.6) The role of is to balance the reconstruction and regularization term losses. A lower value will result in better reconstruction, while a higher value will lead to posterior collapse. To solve this problem, -annealing is adopted, that is, gradually increases from a lower value to a fixed point. At any given epoch n, the objective function of the dynamic -VAE is: The dynamically changing is expressed as: (1.8) Among them, the sign function returns the sign of any real number . . Formula 1.8 aims to optimize the reconstruction term and the regularization term, and respectively increase and decrease the terms used for .

[0039] (1.9) (1.10) (1.11) Among them, the hyperparameter is all the previous rounds, is the last round of change.

[0040] When is positive, due to the term in formula 1.8, decreases, which means that the reconstruction loss increases compared to the historical minimum reconstruction loss according to formula 1.9. When is positive, similarly, it means that the regularization loss is increasing according to formula 1.10. The last term of formula 1.11 adds a loss constraint to the entire model, and slows down the change trend of through this constraint, making the value change slowly. value.

[0041] The dynamic β-VAE changes the value dynamically through the loss difference, and the change of the value will affect the loss, and they affect each other until they reach equilibrium. During the training process of the feature mapping part, the value will change according to the change of the loss value in formulas 1.7 - 1.11. During the training process, starts from 0, and when and

[0042] maintain equilibrium without increasing the total loss, the model converges.There are complex correlations among various variables of industrial process data, and the data quality is unstable. Traditional generation models are difficult to simulate the true distribution of the data, resulting in distorted generated samples. To address this problem, this embodiment proposes a sample generation module based on FD-GAN. First, the generator receives the latent variables and random noise output by the encoder as inputs, generates corresponding samples, and the discriminator distinguishes the generated samples from the real samples. Second, a forced discriminator based on prior knowledge is introduced, and its discrimination result is added as an additional loss term to the optimization objective function to guide the generator to avoid generating samples that do not conform to the prior knowledge. Finally, the generator and the discriminator optimize the generator through adversarial training to generate high-quality samples, making the samples more suitable for the training of downstream models.

[0043] The loss functions of the generator and the discriminator are derived from the classic GAN framework. The goal of the generator is to generate samples that are as real as possible, so that the discriminator cannot distinguish the generated samples from the real samples. The loss function of the generator is defined by the following formula: (1.12) where, and are weighting coefficients, which are used to adjust the losses of the samples generated from the latent variable and the noise respectively. is the reconstruction loss, which ensures that the generator can reconstruct the input data according to the latent variables. The generator not only needs to create new data by generating noise samples , but also reconstruct the real samples by generating the latent variables generated by the encoder. Different from traditional GANs, the generator here not only acts as a generator, but also serves as the decoder in VAE.

[0044] The task of the discriminator is to distinguish real data from generated data (samples from the generator). The loss function of the discriminator is: (1.13) where, is the score of the discriminator for real samples, is the score of the discriminator for the samples generated from the latent variables generated by the encoder, is the score of the discriminator for the samples generated from the noise .

[0045] Next, prior knowledge is introduced to strengthen the ability of the discriminator. By adding corresponding terms to the loss function, appropriate additional losses are given when the triggering limit conditions are met to guide the generator to generate valuable samples. Suppose among a batch of M samples, a certain generated virtual sample is , for the i-th element among them , the additional penalty generated by it is set as follows: (1.14) In the formula, is the element involved in forced discrimination in the i-th sample. is the upper limit of the value of the element in the sample set according to prior knowledge in the forced discriminator, and is the lower limit. (1.15) Add the additional penalty term to the loss of the generator: (1.16) The loss of the discriminator is still , and the objective function of the GAN with a forced discriminator is finally: (1.17) When using the generative model to fill the label samples of industrial production data, due to the scarcity and discontinuous distribution of label samples, the problem of low-quality generated samples occurs. To address the above problems, this embodiment proposes a sample filling method based on FD-VEGAN. First, introduce the VAE model to construct the latent space, encode the existing data through VAE, capture the global features of the data and construct a structured latent space; second, after obtaining the latent vectors generated by the VAE encoder, the generator of the GAN assigns different generation weights to different vectors in the latent space, thereby generating more representative samples; in addition, map the real samples to the latent space through the encoder, forming a closed-loop bidirectional mapping mechanism with the samples generated by the generator, further enhancing the similarity and distribution consistency between the generated samples and the real data. Combining the forced discriminator module enables the method to effectively guide the generator to more accurately simulate the key data features and generate high-quality samples.

[0046] Converter batching refers to selecting appropriate raw materials and additives according to the specific requirements of steelmaking during the converter steelmaking process to ensure the smooth progress of the steelmaking process and the quality of the final product. The main raw materials for converter steelmaking include: Hot metal and scrap iron: This is one of the main raw materials for converter steelmaking, providing the iron element 1 required for steelmaking.

[0047] Pig iron for steelmaking: Used to increase the carbon content and impurity elements in hot metal.

[0048] Alloy return steel: Used to adjust the chemical composition of steel.

[0049] Slag-making materials: such as limestone, dolomite, etc., are used to control the composition and properties of the molten slag.

[0050] Oxidants: such as oxygen, are used to oxidize impurity elements.

[0051] Deoxidants and alloying agents: are used to adjust the chemical composition of steel and remove oxides.

[0052] Carburizers: are used to increase the carbon content in steel.

[0053] The selection and use of these raw materials need to be precisely calculated and adjusted according to specific steelmaking requirements and the chemical composition of the target product. By reasonably selecting and using these raw materials, the smooth progress of the converter steelmaking process can be ensured, and high-quality steel meeting the requirements can be obtained.

[0054] After that, referring to Figure 6 , for the batching optimization process, multiple objectives usually need to be considered simultaneously, including cost minimization, quality optimization, resource consumption minimization, etc., and there are usually conflicts among these objectives. First, a target function weight adjustment module is proposed. By monitoring the progress of each target function in real time and dynamically adjusting the weights according to the relative change rate of the target function, over-optimization of certain target functions is avoided, and at the same time, the progress of other targets is ensured to be fully concerned; second, a similarity experience replay module is proposed. By calculating the cosine similarity between the current state and historical experiences, relevant experiences are preferentially selected, and the diversity of training is maintained by combining the selection of random experiences, so as to accelerate the convergence of the strategy and avoid overfitting; finally, a noise adjustment module is proposed. The scale of the noise is adaptively adjusted according to the reward variance and the number of training steps to ensure sufficient exploration in the initial stage of training and gradually converge to the optimal strategy in the later stage. Through the above methods, the multi-objective optimization process in alloy batching is optimized, and a better alloy batching solution is provided.

[0055] Specifically include: First, construct the target function. To reduce the cost in the actual production process of the factory, increase resource utilization rate, optimize environmental benefits, and ensure the quality of products at the same time, establish the target function formulas (1.18), (1.19), and (1.20).

[0056] (1.18) Among them, F represents the cost of using raw materials, nc represents there are nc kinds of raw materials, represents the unit price (yuan / kg) of the i-th raw material, represents the usage amount of the i-th added raw material.

[0057] (1.19) (1.19) Among them, G represents the total amount of raw materials used, nc represents there are nc kinds of raw materials, Represents the dosage of the i-th added raw material.

[0058] (1.20) Among them, represents the weight of the molten steel in the ladle during tapping, Z represents the quality of the produced product, mc represents having mc elements, represents the dosage of the i-th added raw material. represents the element content of the j-th element contained in the i-th raw material. represents the element recovery rate of the i-th raw material. represents the element content of the j-th element in the molten steel. represents the optimal control point of the j-th element content in the product. The smaller the value of Z, the better the product quality.

[0059] Establish the following constraints according to the requirements of chemical element content, quality requirements, etc. in the actual production process.

[0060] The chemical element composition content of the target finished product can meet within a certain range, that is, the content of each element should be greater than the lower limit value and less than the upper limit value: (1.21) Among them, is the lower limit of the composition requirement of the j-th element of the finished product. is the upper limit of the composition requirement of the j-th element of the finished product. is the content of the j-th element in the i-th raw material. represents the element recovery rate of the i-th raw material.

[0061] For the dosage of each raw material should meet the following non-negative constraint: .

[0062] Aiming at the conflicts and weight imbalances between different objectives in the converter burden optimization, dynamically adjust the weights of each objective function to ensure the balance between objectives. This module can monitor the optimization progress of each objective function in real time and dynamically adjust the weights of each objective function according to the actual progress. When the progress of a certain objective function is large, its weight will be correspondingly reduced to avoid over-optimization on this objective function; on the contrary, when the progress of a certain objective function is slow, its weight will be enhanced to ensure that this objective function receives sufficient attention in the subsequent optimization process. Through this dynamic weight adjustment method, the importance of each objective function in the multi-objective optimization process can be effectively balanced, avoiding the imbalance of importance between objective functions, thereby improving the quality of the optimization results and ensuring the rationality of the final solution.

[0063] The rate of change of the objective function reflects the progress of the objective function in the current iteration. Let the objective function in the current iteration and the previous iteration have values of and , respectively. Then its relative rate of change can be expressed as: (1.22) This formula measures the magnitude of the change of the objective function in the current iteration. If the relative rate of change is large, it indicates that the objective function has made significant progress in the current iteration; otherwise, it means that the progress of the objective function is less. For each objective function , if its rate of change in the current iteration is large, its weight should be reduced to avoid over-optimizing the objective function that has already made great progress in the subsequent optimization process. Conversely, if the rate of change of the objective function is small, indicating less progress, a larger weight needs to be assigned. Based on this, a dynamic weight adjustment formula is proposed: (1.23) where is a small constant used to prevent division by zero errors. This formula shows that when the relative rate of change of the objective function is large, the weight will be reduced; conversely, when the rate of change is small, the weight increases. To ensure that the sum of the weights is 1, the weights of all objective functions need to be normalized: (1.24) This weight normalization ensures that after each optimization iteration, the sum of the weights of all objective functions is 1, and at the same time dynamically adjusts their attention levels during the optimization process according to the progress of the objective functions.

[0064] Finally, the adjusted weights are used to construct the reward function. The form of the reward function is: (1.25) where is the dynamically adjusted objective weight, is the normalized objective function value, is the final comprehensive reward function.

[0065] (1.26) where is the value of the objective function at the current solution , and are the objective functions Mean and standard deviation in all historical iterations: (1.27) (1.28) Regarding the problems of low sample efficiency and the possible entrapment in local optima during the training process in the bof burden optimization, by replaying representative and similar experience samples, the policy convergence is accelerated and the exploration is enhanced to avoid local optima. The learning efficiency is improved by preferentially selecting experiences highly relevant to the current state, while the diversity of training is ensured by combining the selection of random experiences, thus avoiding model overfitting and accelerating the policy convergence.

[0066] The experience replay buffer is used to store the experiences obtained by the agent during the interaction with the environment. Each experience is usually represented as a quadruple, including the current state, the action taken by the agent, the reward obtained, and the new state transferred to. Specifically, each experience is represented as: (1.29) where, represents the current state, indicating the environmental information of the agent network at the current moment, represents the action selected by the agent network in state , is the immediate reward obtained by the agent network from the environment after executing action , is the new state reached after executing action .

[0067] In the similarity experience replay mechanism, by calculating the cosine similarity between the current state and historical experiences, the experiences are divided into two categories: similar experiences and random experiences. Similar experiences refer to those similar to the current state , which helps the agent network quickly learn in the context of the current task and optimize its policy; random experiences refer to the experiences randomly selected from the replay buffer, aiming to maintain the diversity of training and avoid the overfitting of the model only focusing on the current situation.

[0068] To select the most relevant experience to the current state from the experience replay buffer, the cosine similarity is used to measure the similarity between the current state and historical experiences. By calculating the included angle between two vectors to judge their similarity degree, the calculation formula is: (1.30) where, and respectively represent the vector representations of the current state and the historical experience state , and represent their L2 norms respectively. The value range of cosine similarity is from -1 to 1, and the closer the value is to 1, the more similar the two are.

[0069] The selection numbers of similar experiences and random experiences are controlled by the similarity of experiences and a certain ratio. In each training step, a hyperparameter is set to represent the ratio of similar experiences. Specifically, assume that the current replay buffer contains experiences, and the agent network hopes to select experiences from the buffer ( can be greater than or less than the total number of experiences in the buffer). According to the ratio , a certain proportion of similar experiences and random experiences are selected.

[0070] The number of selected similar experiences is: (1.31) where is the selection ratio of similar experiences, is the minimum number of similar experiences to prevent too few similar experiences from being selected.

[0071] The number of selected random experiences is: (1.32) where represents the number of randomly selected experiences, is the minimum number of random experiences to ensure that there are sufficiently diverse experiences in each training.

[0072] After each selection of experiences from the replay buffer, the agent network uses these experiences for policy update. The update process is driven by the temporal difference (TD) error. The TD error is used to evaluate the difference between the predicted value of the agent and the actually obtained reward, and the agent's policy is adjusted according to this difference. The calculation formula of the TD error is: (1.33) where is the reward function, is the Q value corresponding to the current experience, indicating the value of taking action in state , is the discount factor, which is used to control the influence of future rewards, is the next state and

[0073] Traditional converter burdening optimization methods usually have the problem of being prone to falling into local optimal solutions and lack sufficient exploration to avoid this situation. When TD-3 selects actions, it directly adds Gaussian noise with a fixed scale to the output of the Actor network (1.34) to ensure a certain degree of exploration. However, a fixed noise scale is difficult to adapt to the entire training process and may lead to insufficient exploration in the initial stage of training or interference with convergence in the later stage. To address this problem, this section proposes a noise adjustment module (NA) that enables the noise scale to be adaptively adjusted according to policy stability and the number of training steps. This mechanism mainly includes the following two steps: Adaptive adjustment based on reward variance: After the agent is trained for several episodes (set to 5 episodes), calculate the average reward variance of the recent several episodes and use it to measure the stability of the policy. Increase the noise when the variance is low to avoid falling into local optima, and decrease the noise when the variance is high to enhance exploitation: (1.35) where, represents the average reward over several episodes.

[0074] Dynamic decay of noise with the number of training steps: To ensure that the agent network converges to a stable optimal policy in the later stage of training, the noise also needs to decay exponentially with the number of training steps . Finally, the policy expression after noise adjustment is: (1.36) (1.37) where, is the noise decay rate, is the number of training steps. In this way, in the initial stage of training, due to the small number of steps, the noise is large, which helps with sufficient exploration; while in the later stage of training, the noise gradually decays, making the policy tend to be stable.

[0075] Next, compare the results of the burdening generated by the method of the present invention with the results of the burdening generated by other methods.

[0076] The experimental data is from the real data of the factory. To verify the effectiveness of the method in this chapter, taking the production of HRB400E steel as an example, a total of 8 raw materials are required, and the specific parameters are shown in Tables 1 and 2. Table 1 shows the upper and lower limits and control points of the element contents of the finished product process. Table 2 shows the raw material prices of each raw material and the contents of each element.

[0077] Table 1 Upper and lower limits of various element contents in HRB400E Element symbol Element name Lower limit (%) Upper limit (%) Optimal value (control point) C Carbon 0.21 0.25 0.23 Si Silicon 0.15 0.40 0.17 Mn Manganese 1.35 1.45 1.37 P Phosphorus 0 0.045 - S Sulfur 0 0.045 - Table 2 Content of Various Elements in Raw Materials

[0078] To verify the effectiveness of the I-TD3 method, the evaluation indicators include raw material cost, the dosage of each raw material, the total dosage of raw materials, and product quality. The average value of the experimental results of 100 furnace batches is taken as the final result of each indicator. Product quality reflects the sum of the absolute values of the differences between the content of each element in the finished product and the optimal control point. The smaller the product quality, the better the product quality.

[0079] The experimental parameters are set as follows: the number of training rounds is 1×10 5 , the maximum number of iterations per round is 500, the batch size is 512, the soft update smoothing factor is 0.005, the discount factor is 0.7, and the learning rates of the evaluation network and the policy network are both 3×10 -4 , the activation function is the ReLU function, the number of neurons in the input layer is 256, and a total of three hidden layers are used with 128 neurons in each layer.

[0080] To verify the effectiveness of the I-TD3 model, this paper conducts experimental comparisons with multiple comparison models. The brief descriptions of each model are as follows: MOEDO model: This model is a multi-objective optimization algorithm based on exponential distribution optimization. It combines non-dominated sorting and crowding distance to ensure the diversity and uniform distribution of solutions, and introduces an integrated information feedback mechanism to dynamically balance the global exploration and local development capabilities of the model.

[0081] DDQN model: This model reduces the overestimation problem of Q values in DQN by introducing two Q networks, which are used for action selection and Q value estimation respectively.

[0082] A-TD3 model: This model introduces an asynchronous parallel mechanism on the basis of TD3, optimizes the learning efficiency of the continuous action space through an adaptive update strategy, uses the parallel mechanism to improve the convergence speed, and designs two adaptive weight functions based on Off-policy learning to dynamically adjust the weights of the local agent network.

[0083] From Table 3 and Figure 7It can be seen that the I-TD3 model is superior to other models in terms of total raw material consumption, raw material cost control, and product quality. Compared with MOEDO, DDQN, and A-TD3, the batching scheme generated by I-TD3 can significantly reduce raw material consumption while optimizing the overall cost expenditure. This optimization effect indicates that the model is more accurate in batching strategies and can adaptively adjust the feeding ratios of various raw materials according to historical data and the requirements of the current furnace, ensuring the reasonable allocation of raw materials and avoiding unnecessary waste. This is due to the introduction of the similarity experience replay module and the objective function weight adjustment module in I-TD3. By selecting experiences related to the current state, the agent can better learn the optimal batching scheme, thereby reducing raw material usage. At the same time, it monitors the change rate of each objective function in real time, avoids over-optimization of certain objective functions, and ensures that other objectives receive sufficient attention, thus balancing the relationship between cost and quality. Through this method, I-TD3 can significantly reduce the batching cost while ensuring product quality. In terms of product quality, I-TD3 has the best quality performance, reaching 0.0883, which is significantly better than other models. This is due to the noise adjustment module in I-TD3, which adaptively adjusts the noise scale, enabling the agent to conduct sufficient exploration in the initial stage of training and gradually converging to the optimal strategy in the later stage of training. The gradual attenuation of the noise helps the model avoid the dilemma of local optimal solutions and ultimately improves the batching quality.

[0084] In summary, the I-TD3 model significantly improves the efficiency and quality of alloy batching optimization by introducing the objective function weight adjustment, similarity experience replay, and noise adjustment modules. Compared with other traditional optimization algorithms, I-TD3 demonstrates excellent performance in optimizing raw material usage, reducing costs, and improving product quality.

[0085] Table 3 Comparison of experimental results of each method Index MOEDO DDQN A-TD3 I-TD3 (Ours) Silicomanganese alloy (kg) 3185.4 2960.5 2801.5 2743.8 Medium-carbon ferromanganese (kg) 142.3 115.7 58.2 34.7 Vanadium-nitrogen alloy (kg) 24.8 18.2 15.4 14.5 Aluminum wire (kg) 32.7 27.5 14.1 9.2 High-carbon ferrochrome (kg) 0 48.3 20.1 18.9 High-nitrogen ferrovanadium (kg) 0 0 0 0 Carburizer (kg) 18.6 105.9 93.4 98.6 Silicomanganese alloy balls (kg) 0 62.4 70.2 78.3 Total raw material consumption G (kg) 3403.8 3338.5 3063.6 2998 Raw material cost F (yuan) 27289.44 25506.87 22479.34 21783.33 Product quality Z 0.1792 0.1522 0.0912 0.0883 To further explore the contributions of each module in the I-TD3 model, ablation experiments were designed and conducted in this section. Each key module was gradually removed from the complete model, and its impact on the model performance was observed. Five model versions were designed for the experiment: the TD3 model, the I-TD3 w / o OFWA model, the I-TD3 w / o SER model, the I-TD3 w / o NA model, and the complete I-TD3 model. The introductions of each ablation model are as follows: TD3 model: This model is a commonly used deep reinforcement learning method. All modules are removed, the weights of each objective function in the reward function are all 1 / 3, the experience pool sampling adopts prioritized experience replay, there is no selection of similar experiences, and the noise selects a fixed-scale Gaussian noise.

[0086] I-TD3w / oOFWA model: This model removes the objective function weight adjustment module and uses static weights. The weights of each objective function in the reward function are all set to 1 / 3, and only the noise mechanism and experience replay are improved based on the TD3 algorithm.

[0087] I-TD3w / oSER model: This model removes the similarity experience replay module. The experience pool sampling adopts prioritized experience replay, without selecting similar experiences, and only the noise mechanism and reward function weights are improved based on the TD3 algorithm.

[0088] I-TD3w / oNA model: This model removes the noise adjustment module and selects fixed-scale Gaussian noise to achieve exploration and exploitation, and only the experience replay and reward function weights are improved based on the TD3 algorithm.

[0089] Table 4 Comparison of experimental results of each method Index TD3 I-TD3 w / o OFWA I-TD3 w / o SER I-TD3 w / o NA I-TD3 (Ours) Silicomanganese alloy (kg) 2855.2 2790.1 2805.3 2747.6 2743.8 Medium-carbon ferromanganese (kg) 87.6 56.3 65.4 49.2 34.7 Vanadium-nitrogen alloy (kg) 15.8 15.6 17.2 15.1 14.5 Aluminum wire (kg) 21.3 15.4 19.1 13.6 9.2 High-carbon ferrochrome (kg) 29.7 20.3 25.1 22.4 18.9 High-nitrogen ferrovanadium (kg) 0 0 0 0 0 Carburizer (kg) 101.2 95.4 102.3 98.1 98.6 Silicomanganese alloy balls (kg) 45.1 65.2 72.1 69.3 78.3 Total raw material consumption G (kg) 3155.9 3034.5 3146.7 3009.8 2998 Raw material cost F (yuan) 23677.16 22472.56 22951.88 22017.15 21783.33 Product quality Z 0.1049 0.0965 0.1027 0.097 0.0883 From Table 4 and Figure 8It can be seen that I-TD3 is superior to other ablation models in terms of raw material consumption, cost control, and product quality. Compared with the batching schemes generated by other ablation models, the total amount of raw materials is effectively reduced, and the cost expenditure is significantly decreased. This optimization effect indicates that the complete I-TD3 model can fully exert the synergistic effect of each module, enabling the agent network to more accurately adjust the feeding ratios of different raw materials and ensuring the balance between economy and quality. As the baseline model, the TD3 model adopts conventional deep reinforcement learning methods, with the weights of all objective functions in the reward function being 1, using prioritized experience replay, and without introducing any improvement modules. In contrast, the I-TD3w / oOFWA model removes the objective function weight adjustment module and adopts static weights (all weights are 1), making the model unable to flexibly adjust the weights of objective functions. Although the usage of some raw materials (such as ferrosilicon manganese alloy and aluminum wire) of this model has decreased, its overall performance is not significantly superior to the TD3 model. The usage of ferrosilicon manganese alloy has decreased from 2855.2 kg to 2790.1 kg, but the overall raw material cost is still high (22472.56 yuan), and the product quality is 0.0965, failing to achieve the best performance. The I-TD3w / oSER model removes the similarity experience replay module and adopts traditional prioritized experience replay. This model performs relatively poorly in terms of total raw material usage and cost. Although the usage of ferrosilicon manganese alloy (2805.3 kg) is close to that of the I-TD3w / oOFWA model, the improvement in the usage of other raw materials, such as medium-carbon ferromanganese and aluminum wire, is not significant. Especially in terms of total raw material usage, it is slightly reduced compared with the TD3 model, but compared with other models, it still fails to reach the optimal state. The I-TD3w / oNA model removes the noise adjustment module and uses Gaussian noise with a fixed scale to achieve the balance between exploration and exploitation. This model has relatively significant improvements in raw material usage and cost. The total raw material usage drops to 3009.8 kg, and the raw material cost is 22017.15 yuan, indicating that the introduction of the noise adjustment module contributes greatly to the model performance, especially in terms of resource usage and cost optimization. The complete I-TD3 model combines the objective function weight adjustment module, the similarity experience replay module, and the noise adjustment module. This model shows the optimal performance in all key indicators. The total raw material usage is 2998 kg, the raw material cost is 21783.33 yuan, and the product quality is 0.0883. This result indicates that by introducing these three modules, the I-TD3 model significantly optimizes raw material usage, cost control, and product quality, and performs better than other ablation versions.

[0090] In summary, the objective function weight adjustment module, the similarity experience replay module, and the noise adjustment module play important roles in improving the performance of the I-TD3 model. The objective function weight adjustment module can flexibly adjust the weights of each objective and optimize the allocation of resources; the similarity experience replay module improves the learning efficiency by preferentially selecting experiences similar to the current task; while the noise adjustment module effectively balances exploration and exploitation, enabling the agent to be more stable and efficient when exploring new strategies. Therefore, the various modules of the I-TD3 model work together to jointly improve the overall performance of the model.

[0091] To explore the impact of the discount factor on the model, a parameter sensitivity analysis experiment was conducted on it.

[0092] In the formula of TD error, the discount factor acts on the Q value of the future state to adjust the dependence of the current decision on future rewards. The larger the discount factor, the more the agent network focuses on future returns; while a smaller discount factor indicates that the agent network is more inclined to focus on immediate rewards. As shown in Table 5 and Figure 9 as follows, the value of the discount factor affects the training results of the model. The total raw material consumption, raw material cost, and product quality show a trend of first increasing and then decreasing with the increase of the discount factor. For example, the total raw material consumption reaches the optimal value when the discount factor is 0.7 and the sub-optimal value when it is 0.6. This is because in the case of a smaller discount factor, the agent network pays more attention to immediate feedback, which may lead to local optimal solutions or overfitting in policy learning. Although a larger discount factor can pay more attention to long-term rewards, in a complex dynamic environment, relying on future rewards for a long time may cause the policy update to be too slow and even lead to an unstable learning process. Therefore, the discount factor of 0.7 was finally selected as the experimental parameter.

[0093] Table 5 Impact of the discount factor on the model

[0094] In the second aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the above method is implemented.

[0095] In the third aspect, the present invention provides an electronic device, including a memory and a processor, characterized in that the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method described in any one of the above.

[0096] In the fourth aspect, referring to Figure 2 , the present invention discloses a converter burden data generation and optimization system, characterized in that it includes: A data acquisition module, which is used to acquire the original industrial data set in the converter burdening process, preprocess the original industrial data set to obtain the preprocessed training data, and the original industrial data set includes multi-dimensional sensor data and corresponding missing burdening quality parameters; A pre-training module, which is used to input the preprocessed training data into an encoder to generate a mean vector and a variance vector of the latent variable distribution, and use the KL divergence loss to constrain the similarity between the latent variable distribution and the standard Gaussian distribution; A sample generation module, which is used to input the latent variable and a random noise vector into a generator to generate simulated samples, and use a discriminator to distinguish between real samples and generated samples; a forced discriminator module is introduced, and the forced discriminator is based on the prior constraint conditions defined by the converter metallurgy process knowledge base to perform compliance discrimination on the generated samples, convert the discrimination result into an additional loss term, and jointly optimize the parameters of the generator and the discriminator; A data generation module, which is used to use the trained FD-VEGAN model to generate data for the missing burdening parameters and output a complete burdening data set that meets the process constraints; A burdening data optimization module, which is used to construct a cost objective function, a quality objective function, and a resource consumption objective function, and construct the constraint conditions for the chemical element composition content of the target finished product, and use an improved TD3 algorithm to solve the problem, optimize the converter burdening data, and generate optimized converter burdening data.

[0097] Further, in the pre-training module: The encoder encodes the input real samples into a mean vector and a variance vector; On the one hand, the above vectors are combined into latent variables and input into the generator, and on the other hand, they are compared with the mean and variance vectors of the standard Gaussian distribution by KL divergence to generate a regularization loss; Quantify the difference between the reconstructed samples and the real samples through the reconstruction loss to generate a reconstruction loss; The encoder optimizes its own performance according to the regularization loss and the reconstruction loss.

[0098] Further, in the sample generation module: The generator receives the random noise and the latent variable output by the encoder as inputs, generates corresponding noise samples and reconstructed samples, uses the discriminator to distinguish the generated samples from the real samples, and generates a discriminator loss; Introduce a forced discriminator based on prior knowledge and use its discrimination result as an additional loss; The generator performs adversarial training according to the discriminator loss and the additional loss to generate high-quality samples.

[0099] Those skilled in the art can understand that, unless otherwise defined, all terms used herein (including technical terms and scientific terms) have the same meaning as the general understanding of those of ordinary skill in the field to which the present invention pertains. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted in an idealized or overly formal sense unless specifically defined.

[0100] For method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the embodiments of the present invention are not limited by the described action sequence, because according to the embodiments of the present invention, certain steps can be carried out in other sequences or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential for the embodiments of the present invention.

[0101] From the description of the above embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present application.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating and optimizing converter batching data, characterized in that Including: Step S1: Obtain the original industrial dataset of the converter batching process, and preprocess the original industrial dataset to obtain the preprocessed training data. The original industrial dataset includes multi-dimensional sensor data and the corresponding missing batching quality parameters; Step S2: Input the preprocessed training data into the encoder to generate the mean vector and variance vector of the latent variable distribution, and use the KL divergence loss to constrain the similarity between the latent variable distribution and the standard Gaussian distribution; Step S3: Input the latent variable and the random noise vector into the generator to generate simulated samples, and use the discriminator to distinguish between real samples and generated samples; introduce a forced discriminator module, and the forced discriminator performs compliance discrimination on the generated samples based on the prior constraint conditions defined by the converter metallurgy process knowledge base, and convert the discrimination result into an additional loss term; Step S4: Construct a loss function, and jointly optimize the parameters of the generator and the discriminator through backpropagation to train the generation model; Step S5: Use the model trained in Step S4 to generate data for the missing batching parameters, and output a complete batching dataset that meets the process constraints; Step S6: Construct a cost objective function, a quality objective function, and a resource consumption objective function, and construct the constraint conditions for the chemical element composition content of the target finished product, and use the improved TD3 algorithm to solve, optimize the converter batching data, and generate the optimized converter batching data.

2. The method according to claim 1, wherein Step S2 includes: The encoder encodes the input real samples into a mean vector and a variance vector; On the one hand, the above vectors are combined into latent variables and input into the generator, and on the other hand, they are compared with the mean and variance vectors of the standard Gaussian distribution through KL divergence to generate a regularization loss; Quantify the difference between the reconstructed samples and the real samples through the reconstruction loss to generate a reconstruction loss; The encoder optimizes its own performance according to the regularization loss and the reconstruction loss.

3. The method according to claim 1, wherein Step S3 includes: The generator receives the random noise and the latent variable output by the encoder as inputs, generates the corresponding noise samples and reconstructed samples, and uses the discriminator to distinguish the generated samples from the real samples to generate the discriminator loss; Introduce a forced discriminator based on prior knowledge, and use its discrimination result as an additional loss; The generator performs adversarial training according to the discriminator loss and the additional loss.

4. The method according to claim 2, characterized in that, The regularization loss includes: ； ； Among them, is the probability distribution characteristic of the latent variable z obtained based on the specific input x, is the posterior distribution given by the encoder, is the prior distribution of the latent space, is the definition of the KL divergence, where the subscripts prior and post refer to the prior distribution and the posterior distribution generated from the encoder respectively, and represent the mean and standard deviation of the prior distribution of the standard Gaussian distribution respectively, and represent the mean and standard deviation of the posterior distribution output by the encoder respectively.

5. The method according to claim 4, wherein The reconstruction loss includes: ； is a multivariate Gaussian distribution.

6. The method according to claim 4, wherein In the basic VAE model, the total loss of the regularization loss and the reconstruction loss is: ； wherein is the reconstruction loss, is the regularization loss.

7. The method according to claim 4, characterized in that In the dynamic β-VAE model, the total loss of the regularization loss and the reconstruction loss is: ； where is the reconstruction loss of the n-th round, is the regularization loss of the n-th round; is a parameter that changes dynamically according to the round.

8. The method according to claim 1, wherein The loss of the generator in Step S3 is: ； Among them, and are weighting coefficients, which are respectively used to adjust the losses of samples generated from the latent variable and the noise The generated samples, is the weight coefficient of the reconstruction loss, is the reconstruction loss, is the score given by the discriminator to the samples generated from the latent variable generated by the encoder, is the score given by the discriminator to the samples generated from the noise generated.

9. The method according to claim 8, wherein The loss function of the discriminator is: ； Among them, is the score of the discriminator for real samples, is the score of the discriminator for the samples generated from the latent variables generated by the encoder and is the score of the discriminator for the samples generated from noise and 10. The method according to claim 1, characterized in that, Step S6 includes: Monitor the progress of each objective function in real time, and dynamically adjust the weights according to the relative change rate of the objective function; Add historical experience samples by calculating the cosine similarity between the current state and the historical experience; Add noise during the solution process, and adaptively adjust the noise scale according to the reward variance and the number of training steps.

11. The method according to claim 10, characterized in that, Adapting the noise scale according to the reward variance and the number of training steps includes: Every few training rounds, calculate the average reward variance of the recent rounds. When the variance is lower than the preset value, increase the noise, and when the variance is higher than the preset value, decrease the noise; The noise also decays exponentially with the number of training steps.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method described in any one of claims 1-11.

13. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory. When the computer program is executed by the processor, the processor is caused to execute the steps of the method described in any one of claims 1 to 11.

14. A converter batching data generation and optimization system, characterized in that, Comprising: A data acquisition module for acquiring an original industrial data set of the converter batching process, preprocessing the original industrial data set to obtain preprocessed training data, where the original industrial data set includes multi-dimensional sensor data and corresponding missing batching quality parameters; A pre-training module for inputting the preprocessed training data into an encoder to generate a mean vector and a variance vector of the latent variable distribution, and constraining the similarity between the latent variable distribution and the standard Gaussian distribution through the KL divergence loss; A sample generation module for inputting the latent variable and a random noise vector into a generator to generate simulated samples, and distinguishing between real samples and generated samples through a discriminator; introducing a forced discriminator module, where the forced discriminator performs compliance discrimination on the generated samples based on the prior constraint conditions defined by the converter metallurgy process knowledge base, converts the discrimination result into an additional loss term, and jointly optimizes the parameters of the generator and the discriminator; A data generation module for using the trained FD-VEGAN model to generate data for the missing batching parameters and output a complete batching data set that meets the process constraints; A batching data optimization module for constructing a cost objective function, a quality objective function, and a resource consumption objective function, and constructing constraint conditions for the chemical element composition content of the target finished product, and using an improved TD3 algorithm for solution to optimize the converter batching data and generate optimized converter batching data.

15. The system according to claim 14, wherein In the pre-training module: The encoder encodes the input real samples into a mean vector and a variance vector; On the one hand, the above vectors are combined into a latent variable and input into the generator, and on the other hand, they are compared with the mean and variance vectors of the standard Gaussian distribution through KL divergence to generate a regularization loss; Quantify the difference between the reconstructed samples and the real samples through the reconstruction loss to generate a reconstruction loss; The encoder optimizes its own performance according to the regularization loss and the reconstruction loss.

16. The system according to claim 14, wherein In the sample generation module, the generator receives the random noise and the latent variable output by the encoder as inputs, generates corresponding noise samples and reconstructed samples, and distinguishes the generated samples from the real samples through the discriminator to generate a discriminator loss; Introduce a forced discriminator based on prior knowledge and use its discrimination result as an additional loss; The generator performs adversarial training according to the discriminator loss and the additional loss to generate high-quality samples.

Images