Distillation SAE and dynamic integrated converter steelmaking carbon temperature soft measurement method

Abstract

The invention discloses a distillation SAE and dynamic integration converter steelmaking carbon temperature soft measurement method, which comprises the following steps of: training a teacher model on a large-scale data set of a conventional production working condition, and compressing the model by adopting a knowledge distillation technology; then, a small amount of data from unconventional production working conditions is used for fine tuning to obtain a plurality of expert-type student stacking auto-encoder models SAE; in the prediction stage, a to-be-tested sample is mapped to respective feature space through a plurality of SAEs, and the posterior probability of the sample is calculated in each feature space based on a Gaussian distribution model; and finally, respectively inputting a to-be-measured sample into a plurality of corresponding regression devices to obtain a predicted value, and carrying out weighted fusion on the output of each regression device according to a posterior probability to realize dynamic integration selection of the soft measurement model. The method can effectively adapt to complex data distribution of multi-working-condition changes in converter steelmaking, the carbon temperature prediction precision under the unconventional working condition is improved, and the method has high robustness and engineering practicability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metallurgical automation control, and particularly relates to a distillation SAE and dynamic integrated converter steelmaking carbon temperature soft measurement method. BACKGROUND

[0002] Converter steelmaking is one of the key processes in modern steel metallurgy, and the accuracy of its endpoint control is of great significance to the stability of molten steel composition, the control of smelting cost and the improvement of product quality. Carbon content and steel temperature are two core parameters for endpoint control of converter steelmaking. If not properly controlled, they can easily lead to problems such as "back blowing" or "over blowing", which seriously affect smelting efficiency and product consistency.

[0003] At present, the mainstream converter steelmaking endpoint measurement and control methods in industrial field mainly include the following types:

[0004] 1) Manual experience and sampling analysis method: the operator judges the endpoint according to historical experience and process parameters, supplemented by manual sampling analysis. This method relies on human experience, has strong hysteresis, poor data real-time performance, and is difficult to meet the needs of modern efficient steelmaking.

[0005] 2) Sublance detection method: using a sublance probe to measure the temperature and carbon content of molten steel in the converter in real time, which is one of the most widely used online detection methods. However, the sublance probe has high cost, limited service life, and is easily damaged in high-temperature smelting environment, affecting the continuity and accuracy of detection.

[0006] 3) Furnace gas analysis method: by analyzing the changes of CO, CO2 and other gas components in the exhaust gas at the furnace mouth, the degree of carbon-oxygen reaction is indirectly inferred to estimate the endpoint carbon temperature state. This method has certain real-time performance, but the sensitivity of the gas analyzer and the robustness of the algorithm model are required to be high, and the accuracy is easily affected in multiple working condition switching.

[0007] 4) Data-driven soft measurement method: with the development of artificial intelligence and big data technology, soft measurement models based on historical process data have gradually become the mainstream research direction. Typical methods include BP neural network, support vector regression (SVR), random forest (RF) and deep neural network (such as stacked autoencoder SAE). Although these methods can learn complex nonlinear mapping relationships from data, in the context of multiple working conditions and unbalanced data distribution, the model generalization ability and prediction stability still have obvious shortcomings.

[0008] Especially in actual industrial production, affected by factors such as raw material fluctuations, equipment aging, and operator differences, the converter steelmaking process shows obvious diversity in working conditions, that is, the data has multimodal prior distribution characteristics; and the data of unconventional working conditions often face the problems of insufficient samples and data imbalance due to their low frequency and short duration, resulting in a significant decline in the performance of soft measurement models under such working conditions.

[0009] Therefore, existing technologies urgently need a modeling approach that can fully leverage the advantages of big data under conventional operating conditions while also taking into account adaptability to unconventional conditions. To this end, a soft sensing modeling approach combining knowledge distillation, stacked autoencoders, posterior probability modeling, and dynamic weighted integration is proposed. This approach enables high-precision prediction and dynamic control of the endpoint carbon temperature in converter steelmaking, suitable for complex and changing industrial environments. Summary of the Invention

[0010] The present invention aims to solve technical problems such as poor adaptability of existing soft measurement models in converter steelmaking endpoint control under multiple working conditions, insufficient samples of unconventional working conditions, and unbalanced data distribution leading to decreased prediction accuracy. A converter steelmaking carbon temperature soft measurement method combining distillation SAE and dynamic integration is proposed to achieve accurate prediction of carbon content and molten steel temperature, thereby improving the intelligence and stability of converter steelmaking endpoint control.

[0011] In order to achieve the above technical effects, the present invention is implemented by the following technical solutions: a method for measuring carbon temperature and softness of converter steelmaking with distillation SAE and dynamic integration, characterized by comprising the following steps:

[0012] S1. Collect data on initial molten iron and blowing consumables;

[0013] S2. Use clustering method to classify the collected data into conventional working conditions and unconventional working conditions;

[0014] S3. Utilize conventional operating data and build a teacher model based on a deep neural network structure. Then, use a stacked autoencoder (SAE) with label supervision to train the teacher model to obtain a carbon temperature prediction model with global modeling capabilities.

[0015] S4. Perform knowledge distillation on the trained teacher model and transfer its knowledge to multiple lightweight student models through soft labels and hard labels to improve the model's compressibility and generalization ability, thus obtaining the student SAE model.

[0016] S5. For different unconventional production conditions, the student model is fine-tuned using the corresponding small sample data to obtain multiple expert-level student SAE models that are capable of modeling specific conditions. The mapping space of the unconventional condition data on the feature space of each expert student SAE is obtained, and multiple corresponding regressors are trained based on this.

[0017] S6. Input the test sample into multiple student SAE models and extract their feature representations respectively; introduce Gaussian mixture models in each feature space to model their respective feature distributions; calculate the posterior probability of the test sample in the feature distribution of different models to reflect its attribution under different working conditions;

[0018] S7. Input the above-extracted features into the corresponding regressors to obtain multiple carbon temperature prediction results; dynamically weight and fuse each regression result according to the posterior probability, and output the final prediction value.

[0019] Furthermore, the historical process data is actual converter steelmaking production data of the steel plant.

[0020] Furthermore, in S2, the clustering method refers to clustering the original data using a Gaussian mixture model, and then obtaining the distribution of the original data under the Gaussian mixture model so as to divide the data.

[0021] Furthermore, in S3, the training teacher model is divided into layer-by-layer unsupervised training and overall supervised training;

[0022] The layer-by-layer unsupervised training refers to the network parameters θ of each autoencoder (AE) in the stacked autoencoder (SAE) AE The reconstruction loss formula (1) is used as the objective function for layer-by-layer greedy unsupervised pre-training. Formula (1) is as follows:

[0023]

[0024] Where: x i is the encoder input, Reconstruct the output for the decoder;

[0025] The overall supervised training refers to adding a regression layer at the end of the network as the regression prediction output of the SAE, and supervising the entire SAE with the root mean square error (RMSE) as the loss function.

[0026] Furthermore, in S4, the knowledge distillation of the trained teacher model is performed by jointly training a smaller student model using soft labels from the teacher model and hard labels from the true labels.

[0027] Furthermore, in S6, the Gaussian mixture model is used to describe sample data with multiple Gaussian distributions; the probability density function of the Gaussian mixture model is shown in the following formula (2):

[0028]

[0029] in,

[0030]

[0031] In the formula, d is the data vector dimension, T is the transpose symbol, each Gaussian distribution represents a cluster, each Gaussian distribution is represented by a mean vector μ k , a covariance matrix Σ k , and a mixed weight π k represents the probability that the vector belongs to this Gaussian distribution, satisfying Where K is the number of Gaussian distributions.

[0032] For multiple multi-condition data with prior distribution knowledge, after SAE feature extraction, a Gaussian mixture model is fitted on each independent feature mapping space, and multiple Gaussian distributions with prior knowledge are obtained.

[0033] Further, in S7, the posterior probability represents the probability that each data point in the Gaussian mixture model belongs to a certain Gaussian distribution; specifically, the output results of different expert regressors are dynamically weighted and integrated according to the posterior probability, which can determine the ability area to which the to-be-tested sample belongs, and selectively mix the output results of the regressors of multiple ability areas to improve the prediction effect.

[0034] The beneficial effects of the present application are:

[0035] Compared with the prior art, the present application introduces a knowledge distillation mechanism, which significantly compresses the model size while maintaining high prediction accuracy, significantly improving the adaptability and generalization ability of the model under variable conditions. Secondly, by constructing multiple expert student models and combining a dynamic integration strategy based on posterior probability, the modeling difficulty caused by sample imbalance under unconventional conditions is effectively alleviated. Thirdly, the multi-model feature mapping and dynamic fusion mechanism enhances the ability of the model to describe complex nonlinear relationships, significantly improving the accuracy and stability of end-point carbon temperature prediction. Finally, this method does not rely on expensive or easily damaged physical sensors, has good engineering practicability and on-site deployment value, and is suitable for real-time control requirements of the metallurgical industry. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used for the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0037] Figure 1 The step flow chart of the distillation SAE and dynamic integration converter steelmaking carbon temperature soft measurement method described in the present application;

[0038] Figure 2This is a system structure diagram of the distillation SAE and dynamically integrated converter steelmaking carbon temperature soft measurement method of the present invention;

[0039] Figure 3 The amount of data of different categories after the data collected for the present invention are clustered using the Gaussian mixture model;

[0040] Figure 4 This is a comparison chart of the distillation SAE and dynamic integrated converter steelmaking carbon temperature soft measurement method of the present invention and several mature and classic soft measurement methods in terms of the prediction accuracy of the converter endpoint molten steel carbon content and temperature;

[0041] Figure 5 This is a diagram of the fitting effect of the present invention on the test set. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] Example 1

[0044] A method for measuring carbon temperature in converter steelmaking by distillation SAE and dynamic integration, comprising the following steps:

[0045] Step 1: Based on a large amount of historical process data collected from the actual industrial converter steelmaking process, including information such as molten iron composition, scrap steel addition amount, oxygen blowing time, feed amount, auxiliary lance temperature measurement data, endpoint carbon content, and molten steel temperature, collect initial molten iron and blowing consumables data;

[0046] Step 2: Use the Gaussian mixture model to cluster the original data, and then obtain the distribution of the original data under the Gaussian mixture model. The collected data is divided into conventional working conditions and unconventional working conditions. The conventional working condition data has a large sample size and a stable distribution, representing conventional production conditions; the unconventional working condition data has a small sample size and a sparse distribution, representing unconventional working conditions (such as raw material changes, equipment maintenance, extreme operations, etc.);

[0047] Step 3: Use normal working condition data to train a stacked autoencoder (SAE) model as the teacher model. The model structure is as follows: the input layer dimension is the original feature dimension; the encoder part contains multiple layers of linear transformation, each layer is followed by an activation function; the decoder structure is symmetrical; the reconstruction loss function is used. Layer-by-layer pre-training and RMSE (2) are jointly trained; finally, deep feature representation is extracted in the penultimate layer:

[0048]

[0049]

[0050] where x i is the encoder input, is the decoder reconstruction output, is the model prediction, y i is the sample true value.

[0051] Step four, the trained teacher model is used to guide the training process of multiple lightweight student SAE models, a knowledge distillation strategy is adopted, the teacher model output is used as a soft label, and the true label is a hard label, which guides the student model learning; the loss function of this step is shown in formula (3):

[0052]

[0053] where N is the number of training samples; for a certain input x i , the corresponding teacher model prediction is v i , the student model prediction is z i , and the true label is y i ; T is the distillation temperature; a is an adjustable hyperparameter, which is used to balance the influence of soft label and hard label on training.

[0054] Step five, each student model is fine-tuned on different unconventional working condition data; each student model has modeling advantages under a specific working condition;

[0055] Step six, the to-be-tested sample is input into each student stacked autoencoder model (SAE), and the feature space of each model is modeled by a Gaussian mixture model to calculate the posterior probability γ ik of the to-be-tested sample under each model; in the Gaussian mixture model (GMM), the posterior probability is used to represent the probability of each data point belonging to a certain Gaussian distribution. Given a sample x i and the parameters of the Gaussian model, the posterior probability of the sample x i belonging to the kth Gaussian distribution is shown in formula (4):

[0056]

[0057] where μ k is the mean vector of the kth Gaussian component; Σ k is the covariance matrix of the kth Gaussian component; π k is the prior probability of the kth Gaussian component; (x i | μ k , Σ k ) represents the kth Gaussian distribution at point x ithe probability density value of the component k; (z k |x i ) represents the probability of belonging to component k given x i .

[0058] Step seven: according to the posterior probability γ ik obtained in step six, the outputs of the regressors corresponding to the SAE model of each student are dynamically weighted and fused to obtain the carbon content and temperature prediction values considering the prediction results of different working condition models. Assuming that there are K working conditions, for each sample x i , the posterior probability of belonging to the kth working condition is γ ik , and the output weight w k of the kth working condition is as shown in formula (5).

[0059]

[0060] In the formula, k is the number of Gaussian distribution; γ ik is the probability of the ith sample x i belonging to the kth Gaussian distribution,

[0061] If the prediction value of the corresponding regressor is y k , then the weighted integrated prediction output Y of multiple Gaussian distributions is as shown in formula (6).

[0062]

[0063] In the formula, k is the number of Gaussian distribution; y k is the prediction value of regressor k; w ik represents the weight of the kth regressor for x i .

[0064] The greater the posterior probability of the sample x i belonging to the kth working condition, the greater the weight on the total output, and the tan function is used to avoid the influence of the size of y ik on the overall output. When γ ik is closer to 1, the weight tends to +∞. The output results of different expert regressors are selected by dynamically weighted integration of the posterior probability.

[0065] Example 2

[0066] Since the original data does not contain explicit annotations about the production working condition category, the samples cannot be directly divided according to the working condition, so in this embodiment, an unsupervised learning method is used to model the prior distribution of the data, as follows:

[0067] The Gaussian mixture model (GMM) is used to cluster the original data, thereby realizing automatic grouping and working condition division of the data in the feature space; the clustering results are used to distinguish the distribution characteristics of the data under different potential working conditions and provide a classification basis for subsequent fine-tuning of the student model.

[0068] The number of data samples corresponding to each cluster is as follows Figure 3 As shown, it can be seen that the original data category 3 has the largest amount of data, followed by category 2 and category 4, and category 1 and category 5 have less data. The data distribution conforms to the characteristics of normal distribution, which means that the situation where the amount of data under unconventional working conditions is small actually exists.

[0069] To ensure the feasibility of the experiment, this example randomly selected 1,600 data points from category 3 as normal operating condition data, of which 1,200 were used as a training set and 400 as a test set. 800 data points were randomly selected from categories 2 and 4 as abnormal operating condition data, of which 400 were used as a training set and 400 as a test set. Through this clustering process, the implicit operating conditions under different raw material compositions, equipment states, or operating conditions can be effectively distinguished, providing effective data support for the multi-model dynamic integrated prediction mechanism proposed in this application.

[0070] In the prediction of carbon content and temperature at the end of converter steelmaking, regression accuracy (RA) is also an important indicator for measuring model performance. This indicator is used to evaluate the proportion of predicted values ​​on the test set that fall within the set error tolerance, reflecting the practical application reliability of the model. The RA calculation formula is shown in formula (7):

[0071]

[0072] Where N test The number of samples in the test set, Represents the sample to be tested x i The predicted value, y i Represents the sample to be tested x i The true value of

[0073] Is a conditional function, if the difference between the two inputs is less than T e , then output 1, indicating that the prediction error is within the allowable range; otherwise, output 0. It is defined as shown in formula (8).

[0074]

[0075] Where, is the model prediction value, y i is the true value of the sample.

[0076] The model size of the teacher SAE in this embodiment is [40-20-10-8-4-1].

[0077] The model size of the student SAE in this example is [20-8-4-1].

[0078] In this embodiment, the distillation temperature is 1 and the α parameter is 1.

[0079] In this embodiment, T e =0.02.

[0080] Figure 4 The differences between this method and other basic soft-sensing methods on converter steelmaking process data were compared. Figure 4 The middle figure (a) shows the carbon content, and the figure (b) shows the temperature; Figure 5 The left picture in the middle shows carbon content, and the right picture shows temperature.

[0081] When testing other soft sensing methods, the working conditions are no longer distinguished. Instead, all 2000 training data are used as the training set and 1200 test data are used as the test set. This application uses traditional machine learning models such as partial least squares regression (PLS), support vector machine regression (SVR), random forest regression (RFR) and deep learning models such as BP neural network (BPNN) method to compare with this method.

[0082] The above embodiments are only used to illustrate the principles and embodiments of the present invention and are intended to help understand the technical concept and core ideas of the present invention. For those of ordinary skill in the art, without departing from the spirit and essence of the present invention, various modifications, improvements or replacements can be made to specific embodiments and application ranges, and these should all be considered to fall within the scope of protection of the present invention. Therefore, the contents in this specification should not be construed as limiting the present invention, and the scope defined by the claims should be based on the content of its claims.

Claims

1. A method for measuring carbon temperature in converter steelmaking by distillation SAE and dynamic integration, characterized in that: The following steps are involved: S1. Collecting initial molten iron and blowing consumables data based on historical process data; S2. Use clustering method to classify the collected data into conventional working conditions and unconventional working conditions; S3. Utilize conventional operating data and build a teacher model based on a deep neural network structure. Then, use a stacked autoencoder (SAE) with label supervision to train the teacher model to obtain a carbon temperature prediction model with global modeling capabilities. S4. Perform knowledge distillation on the trained teacher model and transfer its knowledge to multiple lightweight student models through soft labels and hard labels to improve the model's compressibility and generalization ability, thus obtaining the student SAE model. S5. For different unconventional production conditions, the student model is fine-tuned using the corresponding small sample data to obtain multiple expert-level student SAE models that are capable of modeling specific conditions, and the mapping space of the unconventional condition data on the feature space of each expert student SAE is obtained; S6. Input the sample to be tested into multiple student SAE models and extract their feature representations respectively; A Gaussian mixture model is introduced into each feature space to model the respective feature distributions. The posterior probability of the tested sample in the feature distribution of different models is calculated to reflect its degree of belonging under different working conditions. S7. Input the above extracted features into the corresponding teacher model to obtain multiple carbon temperature prediction results; dynamically weight and fuse each regression result according to the posterior probability, and output the final prediction value.

2. The method for measuring carbon temperature of converter steelmaking by distillation SAE and dynamic integration according to claim 1, characterized in that: The historical process data are actual converter steelmaking production data of the steel plant.

3. The method for measuring carbon temperature of converter steelmaking by distillation SAE and dynamic integration according to claim 1, characterized in that: In S2, the clustering method refers to clustering the original data using a Gaussian mixture model, and then obtaining the distribution of the original data under the Gaussian mixture model so as to divide the data.

4. The method for measuring carbon temperature of converter steelmaking by distillation SAE and dynamic integration according to claim 1, characterized in that: In S3, the training teacher model is divided into layer-by-layer unsupervised training and overall supervised training; The layer-by-layer unsupervised training refers to the network parameters θ of each autoencoder (AE) in the stacked autoencoder (SAE) AE The reconstruction loss formula (1) is used as the objective function for layer-by-layer greedy unsupervised pre-training. Formula (1) is as follows: Where N is the number of training samples, x i is the encoder input, Output for the decoder; The overall supervised training refers to adding a regression layer at the end of the network as the regression prediction output of the SAE, and supervising the entire SAE with the root mean square error (RMSE) as the loss function.

5. The method for measuring carbon temperature of converter steelmaking by distillation SAE and dynamic integration according to claim 1, characterized in that: In S4, the knowledge distillation of the trained teacher model is specifically performed by jointly training a smaller student model through the soft labels from the teacher model and the hard labels from the true labels.

6. The method for measuring carbon temperature in converter steelmaking by combining distillation SAE and dynamic integration according to claim 1, wherein: In S6, the Gaussian mixture model is used to describe sample data with multiple Gaussian distributions; the probability density function of the Gaussian mixture model is shown in the following formula (2): in, Where d is the dimension of the data vector, T is the transposed symbol, each Gaussian distribution represents a cluster, and each Gaussian distribution is represented by the mean vector μ k , covariance matrix Σ k is formed, and the mixed weight π k To express the probability that the vector belongs to this Gaussian distribution, satisfying Where K is the number of Gaussian distributions; For multiple working condition data with prior distribution knowledge, after SAE feature extraction, a Gaussian mixture model is fitted on each independent feature mapping space to obtain multiple Gaussian distributions with prior knowledge.

7. The method for measuring carbon temperature of converter steelmaking by distillation SAE and dynamic integration according to claim 1, characterized in that: In S7, the posterior probability represents the probability that each data point in the Gaussian mixture model belongs to a certain Gaussian distribution; specifically, the output results of different expert regressors are selected by dynamically weighted integration of the posterior probability, which can determine the ability area to which the sample to be tested belongs, and selectively mix the output results of regressors of multiple ability areas to improve the prediction effect.

Images