Converter end point carbon temperature dynamic self-adaptive prediction method based on meta-learning SAE

Abstract

The invention discloses a converter endpoint carbon temperature dynamic adaptive prediction method based on meta-learning SAE, and the method comprises the steps: firstly carrying out the clustering of historical steelmaking data through employing a Wasserstein distance weighted Dirichlet process Gaussian mixture model (WDPGMM), and automatically dividing a plurality of working condition modes; then, constructing a MetaSAE model for the data of each mode, and training the MetaSAE model to improve the generalization ability of the model to different working conditions; and finally, according to the posterior probability of a new sample of a to-be-detected heat, judging the working condition mode to which the new sample belongs, selecting the most similar sample subset from the historical data of the corresponding mode by using the mutual information weighted JS divergence, and performing real-time fine adjustment on the MetaSAE model, thereby realizing the dynamic prediction of the molten steel end point carbon content and temperature of the new sample. The method can adapt to data distribution changes caused by multiple working conditions and sensor drifting in the steelmaking process, and continuous real-time accurate prediction of the carbon content and temperature of the molten steel is achieved.

Description

Technical Field

[0001] This invention relates to the field of metallurgical automation control technology, specifically to a dynamic adaptive prediction method for converter endpoint carbon temperature based on meta-learning SAE. Background Technology

[0002] Converter steelmaking is a crucial process for producing high-quality steel, accounting for a major share of global steel production. During converter blowing, the exothermic oxidation reaction of carbon directly affects the temperature of the molten steel. The final carbon content and temperature of the molten steel must be strictly controlled within target ranges to ensure the smooth progress of subsequent processes and product quality. If the final carbon content is too high, it will hinder desulfurization and dephosphorization refining of the molten steel; if it is too low, it will lead to an increase in oxygen and nitrogen content, affecting steel quality. Therefore, accurate detection and control of the final carbon content and temperature in converter steelmaking are of great significance for improving steel quality, steelmaking efficiency, and achieving energy conservation and emission reduction.

[0003] Existing technologies for predicting the final carbon content and temperature in converter steelmaking have the following shortcomings: First, they lack a modeling mechanism for multiple operating conditions, making it difficult for a single model to adapt to frequently changing operating modes; second, they lack a model update strategy to address process concept drift, leading to a decrease in model accuracy over time. To address these technical problems, it is necessary to provide a new soft-sensoring method that can automatically identify multiple operating conditions in the steelmaking process and adjust the model in real time based on new data, thereby improving the stability and accuracy of final carbon content and temperature predictions.

[0004] Therefore, in order to solve the above problems, this paper proposes a dynamic adaptive prediction method for converter endpoint carbon temperature based on meta-learning SAE. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of poor adaptability of existing soft measurement models in converter steelmaking endpoint control under multiple operating conditions and the decrease in prediction accuracy caused by data distribution drift. This invention provides a dynamic adaptive prediction method for converter endpoint carbon temperature based on meta-learning SAE to improve the accuracy and stability of endpoint carbon content and temperature prediction.

[0006] To achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: a dynamic adaptive prediction method for converter endpoint carbon temperature based on meta-learning SAE, characterized by comprising the following steps:

[0007] S1. Based on historical steelmaking process data, collect initial molten iron composition, furnace charge input, blowing parameters, furnace gas composition and endpoint measurement data to construct the original dataset for the soft measurement model;

[0008] S2. The Dirichlet Process Gaussian Mixture Model (DPGMM) is used to cluster the data, automatically grouping the data according to the feature distribution. At the same time, Wasserstein distance is introduced as a measure of distribution difference, and the traditional DPGMM is improved to form the Wasserstein distance-weighted Dirichlet Process Gaussian Mixture Model (WDPGMM) to improve the adaptability of clustering to non-Gaussian complex distributions. Unsupervised clustering is performed on historical data to automatically divide multiple working condition pattern clusters and determine the data distribution of each pattern.

[0009] S3. Train the stacked autoencoder (SAE) model using historical data from all operating conditions. Use unsupervised layer-by-layer pre-training to extract deep features from the steelmaking process data, construct a multi-layer nonlinear feature representation structure, and extract the latent representation of the input variables for subsequent soft measurement modeling of the endpoint carbon content and molten steel temperature.

[0010] S4. The stacked autoencoder model is meta-trained using a model-independent meta-learning algorithm. The inner loop is fast-updated and the outer loop is global-optimized on the data corresponding to multiple working conditions to obtain the initial model parameters that can quickly adapt to each mode, thus obtaining the meta-learning stacked autoencoder (MetaSAE) soft measurement model.

[0011] S5. For a new furnace sample to be predicted, calculate its posterior probability under each working condition mode of the clustering model, and determine the working condition mode to which the sample belongs.

[0012] S6. Based on the working condition mode of the sample to be tested, select several samples that are most similar to the sample from the historical data using mutual information weighted JS divergence as training data for model micro-calls.

[0013] S7. Use the selected similar samples to quickly fine-tune the MetaSAE soft measurement model, and output the predicted values ​​of the carbon content and temperature of the molten steel corresponding to the sample to be tested.

[0014] Furthermore, the historical steelmaking process data refers to industrial process data collected during the actual converter steelmaking production process of the steel plant, including the raw material composition of each heat, operating parameters, and the final carbon content and molten steel temperature measurements.

[0015] Furthermore, the clustering process in S2 iteratively performs variational E-steps and M-steps:

[0016] In step E, the augmentation responsibility of each sample belonging to each Gaussian component is calculated according to formula (1):

[0017]

[0018] In the formula, π kLet the prior probability of the k-th Gaussian component be such that it satisfies the following condition: And π k ≥0, Indicates sample x i With a mean of μ k The covariance is Σ k The probability density function under a multivariate Gaussian distribution, where λ is an adjustable penalty coefficient. Let the second-order Wasserstein distance between the i-th sample and the k-th Gaussian distribution be represented by Equation (2):

[0019]

[0020] In the formula, It is a set of joint distributions, and its marginal distributions are respectively and Integral term ||xy|| 2 dγ(x,y) represents the transmission cost under the joint distribution γ;

[0021] Formula (2) measures the distance from sample point x i Distribution to Gaussian distribution The optimal transmission cost is the minimum geometric matching cost between the two in the probability space.

[0022] In the M-step, the mixed weights π are re-estimated based on the degree of responsibility. k and the mean μ of each Gaussian component k Covariance Σ k The parameters are adjusted, and the posterior concentration parameters of the Dirichlet process are updated to adjust the number of clusters. Training is terminated when ELBO converges, and the trained WDPGMM model and the pattern partitioning results of the data are obtained.

[0023] Furthermore, in S3, the unsupervised pre-training of SAE layer by layer is performed as follows:

[0024] With all layers except the current one fixed, train only the current encoder-decoder pair subnetwork, optimizing its parameters to minimize the reconstruction error loss function with the goal of reconstructing the input data.

[0025]

[0026] In the formula, x n For encoder input, This is the decoder output, where N represents the total number of samples in the training set.

[0027] Deep feature extraction refers to the greedy, layer-by-layer pre-training of stacked autoencoders (SAEs) to extract features with the objective function of minimizing reconstruction error. This results in a well-initialized global SAE model with preliminary feature extraction capabilities, which serves as the initial parameters for subsequent meta-learning.

[0028] Furthermore, in S4, the meta-learning training employs the Model Independent Meta-Learning Algorithm (MAML), which optimizes the initial parameters of the SAE model by performing inner and outer loop training on multiple meta-tasks composed of the operating conditions. Specifically, the inner loop performs a finite-step gradient descent on the support set data of each meta-task to update the model parameters, while the outer loop summarizes the error gradients on the query sets of each meta-task to update the global initial parameters, thereby obtaining model parameter initialization adapted to any operating condition.

[0029] Furthermore, in S5, the posterior probability is used to characterize the probability that the test sample belongs to each Gaussian mixture model operating mode; the formula for calculating the posterior probability is as shown in equation (4):

[0030]

[0031] In the formula, x i Let P(k|x) represent the feature vector of the i-th sample to be tested. i ) represents a given sample x i The posterior probability of belonging to the k-th Gaussian distribution, π k Let the prior probability (mixture coefficient) of the k-th Gaussian component satisfy the following condition: And π k ≥0. Indicates sample x i With a mean of μ k The covariance is Σ k The probability density function under a multivariate Gaussian distribution.

[0032] Based on the posterior probability, the operating mode with the highest probability is selected as the mode to which the sample to be tested belongs; this realizes the automatic identification of the current operating condition of the furnace to be tested, and prepares for subsequent targeted model fine-tuning.

[0033] Furthermore, in S6, the step of selecting several samples most similar to the current sample from historical data using mutual information-weighted JS divergence involves calculating the similarity between the test sample and historical samples using mutual information-weighted JS divergence; this measures the distance between the test sample and each candidate historical sample, and selects the set of samples with the smallest distance as the similar sample set; the specific method is as follows:

[0034] By introducing the mutual information weights between the test sample and historical samples on the feature variables, the calculation formula of JS divergence is modified, and the formula of JS divergence weighted by mutual information is obtained, as shown in equation (5):

[0035]

[0036] In the formula, p(x1) and p(x2) are the normalized probability distributions of the two samples. This is the transpose of the mutual information weight vector, used to weight the divergence contribution in each dimension.

[0037] Furthermore, in S7, the model fine-tuning employs an instant learning strategy to rapidly update the MetaSAE model; iterates the MetaSAE model using the similar sample set to adjust the model parameters to adapt to the characteristics of the working conditions of the test sample; and then uses the fine-tuned model to predict the endpoint carbon content and temperature of the test sample.

[0038] The beneficial effects of this invention are:

[0039] This invention proposes using a Wasserstein distance-weighted Dirichlet process Gaussian mixture model for cluster analysis of historical steelmaking data. This model can automatically determine the optimal number of patterns and accurately classify different operating modes, providing fundamental data support for soft sensor models to model under different operating conditions. A meta-learning strategy is introduced to train a stacked autoencoder model, combining the feature extraction capabilities of deep learning with the rapid adaptive capabilities of meta-learning to fully address the multi-modal distribution characteristics of steelmaking process data. A mutual information-weighted JS divergence strategy is used to select similar samples and fine-tune the model online, achieving real-time adaptive adjustment of model parameters and effectively reducing the impact of data distribution drift on prediction accuracy. Test results based on actual converter steelmaking production data show that the proposed soft sensor model can maintain stable prediction performance when steelmaking operating conditions change, making it suitable for the real-time control needs of the metallurgical industry. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is an overall flowchart of the converter steelmaking endpoint carbon temperature soft measurement method according to an embodiment of the present invention;

[0042] Figure 2 The endpoint data of converter steelmaking in this embodiment of the invention is visualized by TSNE dimensionality reduction after clustering using the Wasserstein distance-weighted Dirichlet process Gaussian mixture model;

[0043] Figure 3 This invention provides a comparison of the converter end-point carbon temperature prediction method with various mature soft measurement methods in terms of the accuracy of predicting the carbon content of converter end-point molten steel.

[0044] Figure 4 This invention provides a comparison of the converter end-point steel carbon temperature prediction method with various mature soft measurement methods in terms of the accuracy of converter end-point steel temperature prediction.

[0045] Figure 5 This invention provides a comparison of the converter end-point carbon temperature prediction method with various mature soft measurement methods in terms of RMSE prediction of converter end-point carbon content in molten steel.

[0046] Figure 6 This invention provides a comparison of the converter end-point steel carbon temperature prediction method with various mature soft measurement methods in terms of converter end-point steel temperature prediction RMSE.

[0047] Figure 7 This is a fitting effect diagram of the predicted carbon content of molten steel at the converter endpoint in an embodiment of the present invention;

[0048] Figure 8 This is a fitting effect diagram of the converter final steel temperature prediction in an embodiment of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] The implementation process of the converter endpoint carbon temperature dynamic adaptive prediction method based on meta-learning SAE is as follows:

[0052] Step 1: Obtain a large amount of historical process data from the actual converter steelmaking production process, including raw material composition, charge rate, blowing time, temperature and carbon content measured by the auxiliary lance, and the final carbon content and molten steel temperature at tapping. Preprocess the collected data, such as removing outliers, filling missing values, and normalizing the data, to ensure data quality.

[0053] Step 2: Since the original data does not explicitly label the operating condition category of each furnace batch, this embodiment first performs unsupervised clustering analysis on the training set data to identify the implicit operating condition patterns.

[0054] Specifically, the Dirichlet Process Gaussian Mixture Model (DPGMM) is used to cluster the data, automatically grouping the data according to the characteristic distribution; at the same time, the Wasserstein distance is introduced as a measure of distribution difference, and the traditional DPGMM is improved to form the WDPGMM model, so as to improve the adaptability of clustering to non-Gaussian complex distributions.

[0055] The clustering process iteratively performs variational E-steps and M-steps: In the E-step, the enhanced responsibility (i.e., posterior probability) of each sample belonging to each Gaussian component is calculated according to formula (1):

[0056]

[0057] In the formula, π k Let the prior probability of the k-th Gaussian component be such that it satisfies the following condition: And π k ≥0, Indicates sample x i With a mean of μ k The covariance is Σ k The probability density function under a multivariate Gaussian distribution, where λ is an adjustable penalty coefficient. Let the second-order Wasserstein distance between the i-th sample and the k-th Gaussian distribution be represented by Equation (2):

[0058]

[0059] In the formula, It is a set of joint distributions, and its marginal distributions are respectively and Integral term ||xy|| 2 dγ(x,y) represents the transmission cost under the joint distribution γ. This formula measures the cost from sample point x. i Distribution to Gaussian distribution The optimal transmission cost is the minimum geometric matching cost between the two in the probability space.

[0060] In the M-step, the mixed weights π are re-estimated based on the degree of responsibility. k and the mean μ of each Gaussian component k Covariance Σ k The parameters are adjusted, and the posterior concentration parameters of the Dirichlet process are updated to adjust the number of clusters. Training is terminated when ELBO converges, and the trained WDPGMM model and the pattern partitioning results of the data are obtained.

[0061] Step 3: Input all the training data obtained in Step 2 into the Stacked Autoencoder (SAE) model for offline pre-training. The SAE model consists of multiple layers of autoencoders stacked layer by layer. The encoder part contains multiple layers of nonlinear neuron mappings, and the output of each layer serves as the input of the next layer. The structure of the decoder part is symmetrical to the encoder. Perform unsupervised pre-training of the SAE layer by layer: fix the parameters of all layers except the current layer, and train only the current encoder-decoder pair sub-network, so that it optimizes the parameters with the goal of reconstructing the input data and minimizes the reconstruction error loss function.

[0062]

[0063] In the formula, x n For encoder input, This is the decoder output, where N represents the total number of samples in the training set.

[0064] Through the above steps, a well-initialized SAE global model is obtained, which has preliminary feature extraction capabilities and can be used as initial parameters for subsequent meta-learning.

[0065] Step 4: After completing the initial training of SAE, the Model-Independent Meta-Learning (MAML) algorithm is introduced to enhance the training of the soft sensor model, enabling it to better adapt to the data characteristics under different working conditions. The specific process is as follows: The datasets corresponding to the K working condition modes identified in Step 2 are constructed into K meta-tasks, denoted as {D1, D2, ..., D...}. K For each meta-task D i The dataset is divided into a support set and a query set, where i = (1, 2, ..., K). The support set is used to simulate model training under this working condition, and the query set is used to evaluate the model's generalization performance under this condition. Subsequently, meta-training of the model's initial parameters is achieved by performing double-loop optimization on these meta-tasks.

[0066] Inner Loop: For the i-th meta-task, after unsupervised pre-training initialization using SAE, let the current model parameters be θ. Its support set data is denoted as... Where N S To support a set number of samples, x n For the input features, y n This corresponds to its target output. The model takes the input sample x as input. n Perform a nonlinear mapping to obtain the reconstructed output. and regression prediction results The total loss function on the support set is composed of the reconstruction error and the regression error, specifically expressed as:

[0067]

[0068] Next, the inner loop performs multiple gradient descent operations on the support set loss to obtain the local parameters θ after the task has been adapted. i :

[0069]

[0070] Where α is the inner loop learning rate, θ is the current model parameters, and θ i To enable the model to quickly learn updated parameters for task i, This indicates that the gradient is calculated with respect to the parameter θ.

[0071] b) Outer Loop:

[0072] After completing the inner loop updates for all tasks, the MetaSAE model calculates the loss after the inner loop updates on the query set for each task. Optimize the global initialization parameters of the model to improve its generalization ability. Let the query set of the i-th task be... (sample size is N) Q Then we have equation (6):

[0073]

[0074] In the formula, N Q x is the number of samples in the query set. n For input features, For x n The reconstructed output obtained by performing nonlinear mapping, It is the regression prediction result of the model on the input sample, y n It is the true label value of the sample.

[0075] Then, the outer loop weights and sums the query set losses of all tasks to form the meta-loss, as shown in equation (7):

[0076]

[0077] And with this as the target, the initial parameter θ is updated by gradient, as shown in equation (8):

[0078]

[0079] Where β is the outer loop learning rate. Through the outer loop update, θ will be optimized in the direction of reducing the validation error of all tasks, thereby gradually forming a model parameter initialization that performs well on various tasks; after multiple nested training loops, the optimal initial parameters with fast adaptability can be obtained; through the outer loop update, the initial parameters of the model are gradually optimized so that they can reduce the query error of all tasks at the same time.

[0080] Step 5: Enter the online prediction stage. When a new converter blowing furnace is about to end, obtain the real-time process data of the current furnace from the actual production unit; after performing the same preprocessing as in the training stage on the sample to be tested, calculate its posterior probability of belonging to each mode data, as shown in equation (9):

[0081]

[0082] In the formula, x i Let P(k|x) represent the feature vector of the i-th sample to be tested. i ) represents a given sample x i The posterior probability of belonging to the k-th Gaussian distribution, π k Let the prior probability (mixture coefficient) of the k-th Gaussian component satisfy the following condition: And π k ≥0. Indicates sample x i With a mean of μ k The covariance is Σ k The probability density function under a multivariate Gaussian distribution.

[0083] This enables automatic identification of the current operating condition of the furnace under test, preparing for subsequent targeted model fine-tuning.

[0084] Step Six: From the historical database corresponding to the pattern determined in Step Five, select a batch of historical samples most similar to the sample to be tested for model fine-tuning. In this embodiment, mutual information-weighted JS divergence is used as the similarity measure. Mutual information (MI) is a fundamental quantity in information theory used to measure the statistical dependency between two random variables, and its definition is as shown in equation (10):

[0085]

[0086] In the formula, (x,y) represents the joint probability distribution of variables x and y, and (x) and (y) are their marginal probability distributions, respectively.

[0087] For a given sample, in order to assign differentiated weights to the feature dimensions and thus measure the similarity between samples more precisely, a mutual information weighting mechanism is adopted. The formula for calculating the weight of the d-th feature is as shown in equation (11):

[0088]

[0089] In the formula, x d Let y be the d-th feature component of the sample, and MI(x) be the target predictor variable. d (,y) represents feature x d Mutual information between the target y and the target y, where D is the total dimension of the sample features.

[0090] To meet the probability distribution input requirements of JS divergence, the samples need to be normalized. The normalization method is shown in equation (12):

[0091]

[0092] In the formula, x represents the original sample vector, x d Let p(x) represent its d-th feature component, and p(x) be the normalized sample representation, which can be regarded as a discrete probability distribution.

[0093] Then, the JS divergence formula weighted by mutual information is given by the following equation (13):

[0094]

[0095] In the formula, p(x1) and p(x2) are the normalized probability distributions of the two samples. This is the transpose of the mutual information weight vector, used to weight the divergence contribution in each dimension.

[0096] Compared to traditional indifferent weighting, this weighted JS divergence better reflects the differences between the feature dimensions of interest to the target variable, significantly improving the accuracy and rationality of similar sample selection and providing more representative data support for rapid model fine-tuning. The samples selected in this way are similarly distributed in the feature space to the test samples, and the importance of the features is considered, making them highly representative and informative of the current working conditions.

[0097] Step Seven: Incorporate the similar sample set obtained in Step Six into the training process of the MetaSAE model to quickly fine-tune the model parameters. Specifically, using the current parameters of the MetaSAE model as initial values, fine-tuning is performed using the similar sample set: the prediction error of these samples is calculated, and the model parameters are adjusted along the gradient direction. Finally, the test samples for the current heat are input into the fine-tuned model, and the predicted values ​​of carbon content and molten steel temperature at the end of the heat are output. Thus, dynamic online prediction of the carbon temperature at the end of converter steelmaking is achieved.

[0098] In this embodiment, the process of the soft measurement method described in this invention is fully realized through steps one to seven above. This method makes full use of the multi-mode information hidden in historical big data, firstly trains the MetaSAE model with good initial parameters offline, and then combines the real-time data of the new furnace to perform online pattern recognition and targeted fine-tuning, thereby significantly improving the accuracy and robustness of the endpoint carbon content and temperature prediction.

[0099] This method has been thoroughly validated experimentally on a steel production site dataset. During the experiment, a total of 3300 samples under normal operating conditions were collected. To simulate real industrial processes, 3000 process samples were used as the training set, and 300 samples were used as the test set for online performance evaluation of the model. To verify the effectiveness of the proposed method, the root mean square error (RMSE) and prediction accuracy (PA) were used to evaluate the performance of the prediction model, and their calculation methods are shown in equations (14) and (15) below:

[0100]

[0101] Where, N test y represents the number of test samples. prediction The PE value represents the predicted value. When the prediction error is within the hit range, it is considered a hit and is denoted as 1. Otherwise, it is denoted as 0.

[0102] The prediction error range for temperature is Th = 10℃, and the prediction error range for carbon content is Th = 0.02%. PA represents the prediction accuracy of carbon temperature within the error range; the larger the value, the better the prediction performance of the model.

[0103] Figure 2 This shows the distribution of the clustered data. Figures 3-6 The differences between our method and other basic software testing methods in converter steelmaking process data were compared: Among them, Figure 3 and Figure 4 It is the prediction accuracy PA. Figure 5 and Figure 6 It is RMSE.

[0104] This application compares the predictive performance of 12 soft sensor models—PLS, SVR, BP, CNN, LSTM, VAE, VW-SAE, Transformer, SSPAE, GSTAE, GMVAE, and ICF-SAE—on the final carbon content and temperature of converter steelmaking.

[0105] Figure 7 and Figure 8 This demonstrates the fitting between the predicted carbon content and temperature of the converter at the end point of the embodiment of the present invention and the actual values.

[0106] Experimental results show that the method of the present invention achieves prediction hit rates of 84.67% (carbon content) and 85.00% (temperature) within error tolerance ranges of ±0.02% carbon content and ±10℃ temperature, respectively, demonstrating high accuracy and stability. In practical deployment, the system can continuously and in real-time predict the endpoint parameters of molten steel and adaptively adjust model parameters according to changes in operating conditions, thereby providing scientific, accurate, and efficient decision support for production control and process optimization.

[0107] This document illustrates the principles and implementation methods of the present invention through specific examples. These embodiments are intended to aid in understanding the method and core concepts of the present invention. Those skilled in the art will readily recognize that, guided by the principles of this invention, appropriate adjustments can be made to the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as limiting the present invention.

Claims

1. A dynamic adaptive prediction method for converter endpoint carbon temperature based on meta-learning SAE, characterized in that, Includes the following steps: S1. Based on historical steelmaking process data, collect initial molten iron composition, furnace charge input, blowing parameters, furnace gas composition and endpoint measurement data to construct the original dataset for the soft measurement model; S2. The Dirichlet Process Gaussian Mixture Model (DPGMM) is used to cluster the data, automatically grouping the data according to the characteristic distribution. At the same time, Wasserstein distance is introduced as a measure of distribution difference, and the traditional DPGMM is improved to form the Wasserstein distance-weighted Dirichlet Process Gaussian Mixture Model (WDPGMM) to improve the adaptability of clustering to non-Gaussian complex distributions. S3. Train the stacked autoencoder (SAE) model using historical data from all operating conditions. Use unsupervised layer-by-layer pre-training to extract deep features from the steelmaking process data, construct a multi-layer nonlinear feature representation structure, and extract the latent representation of the input variables for subsequent soft measurement modeling of the endpoint carbon content and molten steel temperature. S4. The stacked autoencoder model is meta-trained using a model-independent meta-learning algorithm. The inner loop local update and outer loop global optimization are performed on the data corresponding to multiple working conditions to obtain the initial model parameters that can quickly adapt to each mode, thus obtaining the meta-learning stacked autoencoder (MetaSAE) soft measurement model. S5. For a new furnace sample to be predicted, calculate its posterior probability under each working condition mode of the clustering model, and determine the working condition mode to which the sample belongs. S6. Based on the working condition mode of the sample to be tested, select several samples that are most similar to the sample from the historical data using mutual information weighted JS divergence as training data for model micro-calls. S7. Use the selected similar samples to quickly fine-tune the MetaSAE soft measurement model, and output the predicted values ​​of the carbon content and temperature of the molten steel corresponding to the sample to be tested.

2. The converter endpoint carbon temperature dynamic adaptive prediction method based on meta-learning SAE according to claim 1, characterized in that, The historical steelmaking process data refers to industrial process data collected during the actual converter steelmaking production process in steel plants, including raw material composition, operating parameters, and final carbon content and molten steel temperature measurements for each heat.

3. The converter endpoint carbon temperature dynamic adaptive prediction method based on meta-learning SAE according to claim 1, characterized in that, The clustering process in S2 iteratively performs variational E-steps and M-steps: In step E, the augmentation responsibility of each sample belonging to each Gaussian component is calculated according to formula (1): In the formula, π k Let the prior probability of the k-th Gaussian component be such that it satisfies the following condition: And π k ≥0, Indicates sample x i With a mean of μ k The covariance is Σ k The probability density function under a multivariate Gaussian distribution, where λ is an adjustable penalty coefficient. Let the second-order Wasserstein distance between the i-th sample and the k-th Gaussian distribution be represented by Equation (2): In the formula, It is a set of joint distributions, and its marginal distributions are respectively and Integral term ||xy|| 2 dγ(x,y) represents the transmission cost under the joint distribution γ; Formula (2) measures the distance from sample point x i Distribution to Gaussian distribution The optimal transmission cost is the minimum geometric matching cost between the two in the probability space. In the M-step, the mixed weights π are re-estimated based on the degree of responsibility. k and the mean μ of each Gaussian component k Covariance Σ k The parameters are adjusted, and the posterior concentration parameters of the Dirichlet process are updated to adjust the number of clusters. Training is terminated when ELBO converges, and the trained WDPGMM model and the pattern partitioning results of the data are obtained.

4. The converter endpoint carbon temperature dynamic adaptive prediction method based on meta-learning SAE according to claim 1, characterized in that, In S3, the unsupervised pre-training of SAE layer by layer is performed as follows: With the parameters of all layers except the current layer fixed, train only the current encoder-decoder pair subnetwork, optimizing its parameters with the goal of reconstructing the input data and minimizing the reconstruction error loss function, as shown in equation (3): In the formula, x n For encoder input, This is the decoder output, where N represents the total number of samples in the training set. Deep feature extraction refers to the greedy layer-by-layer pre-training of stacked autoencoders (SAEs) to extract features with the objective function of minimizing reconstruction error. We obtain a well-initialized SAE global model with preliminary feature extraction capabilities, which can be used as initial parameters for subsequent meta-learning.

5. The converter endpoint carbon temperature dynamic adaptive prediction method based on meta-learning SAE according to claim 1, characterized in that, In S4, the meta-learning training adopts the Model Independent Meta-Learning Algorithm (MAML), which optimizes the initial parameters of the SAE model by performing inner and outer loop training on multiple meta-tasks composed of the operating conditions. The inner loop performs a finite-step gradient descent on the support set data of each meta-task to update the model parameters, and the outer loop summarizes the error gradients on the query set of each meta-task to update the global initial parameters, thereby obtaining model parameter initialization that is adapted to any operating condition.

6. The converter endpoint carbon temperature dynamic adaptive prediction method based on meta-learning SAE according to claim 1, characterized in that, In S5, the posterior probability is used to characterize the probability that the test sample belongs to each Gaussian mixture model operating mode. The formula for calculating the posterior probability is shown in equation (4): In the formula, x i Let P(k|x) represent the feature vector of the i-th sample to be tested. i ) represents a given sample x i The posterior probability of belonging to the k-th Gaussian distribution, π k Let the prior probability (mixture coefficient) of the k-th Gaussian component satisfy the following condition: And π k ≥0; Indicates sample x i With a mean of μ k The covariance is Σ k The probability density function under a multivariate Gaussian distribution; Based on the posterior probability, the operating mode with the highest probability is selected as the mode to which the sample to be tested belongs, thereby realizing the automatic identification of the current operating condition of the furnace to be tested, and preparing for subsequent targeted model fine-tuning.

7. The converter endpoint carbon temperature dynamic adaptive prediction method based on meta-learning SAE according to claim 1, characterized in that, In S6, the step of selecting several samples most similar to the current sample from historical data using mutual information-weighted JS divergence involves calculating the similarity between the test sample and historical samples using mutual information-weighted JS divergence; this measures the distance between the test sample and each candidate historical sample, and selects the set of samples with the smallest distance as the similar sample set; the specific method is as follows: By introducing the mutual information weights between the test sample and historical samples on the feature variables, the calculation formula of JS divergence is modified, and the formula of JS divergence weighted by mutual information is obtained, as shown in equation (5): In the formula, p(x1) and p(x2) are the normalized probability distributions of the two samples, and w T This is the transpose of the mutual information weight vector, used to weight the divergence contribution in each dimension.

8. The converter endpoint carbon temperature dynamic adaptive prediction method based on meta-learning SAE according to claim 1, characterized in that, In S7, the model fine-tuning adopts an instant learning strategy to quickly update the MetaSAE model; the MetaSAE model is iterated using the similar sample set to adjust the model parameters to adapt to the characteristics of the working conditions of the test sample; and then the fine-tuned model is used to predict the endpoint carbon content and temperature of the test sample.

Images