A method for predicting continuous casting slab quality based on Bayesian network

Abstract

The present invention discloses a continuous casting billet quality prediction method based on a Bayesian network, which relates to the technical field of steel production. By analyzing the causal relationship between process parameters and billet quality, evaluation indicators of parameters and quality are determined, and an evaluation system for the production process is constructed; the topological structure of the Bayesian network is determined to achieve qualitative analysis of each node in the continuous casting production process; parameters of the Bayesian network are learned to achieve qualitative analysis of each node in the continuous casting production process; an inference engine is used to perform inference analysis on the Bayesian network to obtain evaluation results of production parameters and quality defects; each production parameter (or part of the production parameters) is input into the model as an evidence variable, and through forward inference analysis of the Bayesian network, prediction of billet quality can be achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of steel production, and in particular to a continuous casting billet quality prediction method based on a Bayesian network. Background Art

[0002] Predicting the quality of continuously cast ingots and identifying the core factors influencing quality defects can provide an important basis for ingot quality control. Continuous casting is the process in the steel production process where molten steel solidifies into ingots. This process includes pouring molten steel from a ladle into a tundish, casting it from the tundish into a mold, and then cooling it through the mold and secondary cooling zone. Each step involves multiple production parameters. Furthermore, there are many types of quality defects, such as center segregation, center porosity, shrinkage, center cracks, center cracks, corner cracks, and more. Furthermore, not only are the relationships between production parameters and quality defects in continuously cast ingots complex, but the various production parameters and quality defects also interact due to non-independent causal relationships. Therefore, the key issue in ingot quality prediction is how to accurately identify the source of continuous cast ingot quality problems and the core indicators of production parameters by describing the qualitative and quantitative relationships between various parameters in the continuous casting process, and accurately and effectively optimize the continuous casting process.

[0003] Currently, there are three main methods for predicting the quality of continuous casting billets based on Bayesian networks: mechanism modeling, expert system methods, and data-driven intelligent methods. The mechanism modeling method analyzes the impact of production parameters on billet defects or the causes of quality defects by establishing a solidification heat transfer model. However, there is currently a lack of verification methods for the accuracy of the mechanism model. The expert system-based quality prediction method does not explore the quantitative relationship between continuous casting production parameters and quality defects, making it difficult to identify the dominant and core factors. Intelligent methods represented by data-driven methods mostly use deep learning methods or mathematical statistics methods. This method lacks causal explanations and is not conducive to tracing the causes of quality. Summary of the Invention

[0004] The present invention addresses the above technical problems, overcomes the shortcomings of the prior art, and provides a continuous casting billet quality prediction method based on a Bayesian network, comprising:

[0005] (1) Sorting out the relationship between continuous casting production parameters and continuous casting billet quality

[0006] Analyze the influence of various process parameters on the quality defects of casting billets based on solidification mechanism and metallurgical knowledge;

[0007] (2) Establish a continuous casting process evaluation system

[0008] The continuous casting process evaluation system is usually composed of the selected evaluation indicators and their own characteristics at various levels, and is divided into two levels: quality defect level and production parameter level;

[0009] Considering the characteristics of Bayesian networks, data preprocessing and normalization are performed on each evaluation index. The evaluation indexes in the original data are all "continuous" data, so they need to be discretized and preprocessed.

[0010] (3) Billet quality prediction model based on Bayesian network

[0011] 1) A scoring search-based structure learning algorithm is used to learn the Bayesian network structure of the prediction model. All possible Bayesian network structures are searched, each of which is scored. The scoring results guide the next round of search until the scores converge. The scoring function uses the Bayesian formula to solve the posterior probability of the network structure under a given data set D.

[0012] 2) Parameter learning of Bayesian networks in the slab quality prediction model;

[0013] 3) Logical reasoning of the Bayesian network in the billet quality prediction model, using a joint tree inference engine for logical reasoning;

[0014] 4) Actual production parameters are input into the model as evidence variables, and through forward reasoning analysis using the Bayesian network, the quality of the ingot is predicted;

[0015] (4) Reverse tracing of billet quality issues based on Bayesian networks

[0016] Similar to the slab quality prediction model, after completing structural and parameter learning, the complete Bayesian network is reverse traced using a joint tree inference engine to construct a Bayesian network-based slab quality reverse traceability model. Using slab quality defects as evidence variables, the Bayesian network is reverse traced for the continuous casting production process, determining the discrete states and marginal probability distributions of each production parameter node. This results in a reverse traceability evaluation of the production parameters in the continuous casting process, identifying key parameters that influence product quality and providing a basis for product quality control.

[0017] The technical solution further defined in the present invention is:

[0018] In the aforementioned continuous casting billet quality prediction method based on Bayesian network, in step (1), the process parameters include molten steel superheat, billet drawing speed, crystallizer cooling water temperature difference, secondary cooling parameters, and steel type. The quality defects of the billet mainly include: central segregation, porosity, shrinkage, and cracks.

[0019] In the aforementioned continuous casting billet quality prediction method based on Bayesian network, in step (2), first, the steel grade composition is analyzed based on actual production data; second, the evaluation index of the quality defect layer is determined based on the quality defect data; finally, combined with the research and analysis of the industrial background, the evaluation index of the production parameter layer is determined based on the mutual influence relationship between various production parameters.

[0020] In the aforementioned continuous casting billet quality prediction method based on Bayesian network, in step (2), the discretization preprocessing is specifically as follows: first, the original data is subjected to data pivoting processing, and the discretization values ​​of each evaluation index are determined in combination with the research and analysis of the industrial background; second, based on the results of the data pivoting processing and the determined discretization values, the evaluation indexes are further analyzed to determine the discretization standards of each evaluation index; finally, the original data is normalized and integrated into a normalized data set.

[0021] In the aforementioned continuous casting billet quality prediction method based on Bayesian network, in step (3), the scoring formula is as follows:

[0022] Formula ①.

[0023] In the aforementioned continuous casting billet quality prediction method based on Bayesian network, in step (3), the network structure learning process is as follows:

[0024] Step 1: Determine the initial network structure model;

[0025] Step 2: Score each possible network structure change based on the scoring function and find the change with the highest joint probability of the dataset D and the network structure Ni;

[0026] Step 3: Modify the network structure model according to step 2;

[0027] Step 4: Repeat steps 2 and 3 until the network structure with the maximum joint probability of the data set D and the network structure is found.

[0028] In the aforementioned continuous casting billet quality prediction method based on Bayesian network, in step (3), the main input items of structural learning include: normalized discrete data set, the number of possible discrete state values ​​of each node, node order, and the maximum number of parent nodes, wherein the normalized discrete data set is obtained after processing the actual continuous casting production data; the number of possible discrete state values ​​of each node is determined by the evaluation system; the node order is determined according to the order of each production parameter in the continuous casting process and the correlation of quality defects, and the production parameters are sorted from the tundish to the crystallizer and then to the secondary cooling zone, and the quality defects are ranked after all the production parameters.

[0029] In the aforementioned continuous casting billet quality prediction method based on Bayesian network, in step (3), based on the given data set D and the obtained Bayesian network topology structure, the main steps of parameter learning are:

[0030] Step 1: Determine the prior distribution of network parameters;

[0031] Step 2: Based on the given data set D, calculate the posterior distribution of the node according to the Bayesian formula:

[0032] Formula②

[0033] Considering the complexity of the continuous casting process and the unknown prior distributions of production parameters and quality defects, maximum likelihood estimation is used to obtain an unbiased estimate of the original distribution.

[0034] In the aforementioned continuous casting billet quality prediction method based on Bayesian network, in step (3), the first step of the joint tree reasoning is to transform the Bayesian network into a connection tree. The connection tree is essentially an undirected tree, and the maximum complete subtree of the undirected tree is transformed into each tree node of the connection tree. The second step is to calculate the probability of each node according to the message passing protocol defined in the connection tree.

[0035] The beneficial effects of the present invention are as follows: by analyzing the causal relationship between process parameters and billet quality, evaluation indicators of parameters and quality are determined, and an evaluation system for the production process is constructed; the topological structure of the Bayesian network is determined, and qualitative analysis of each node in the continuous casting production process is realized; parameter learning of the Bayesian network is realized, and qualitative analysis of each node in the continuous casting production process is realized; the Bayesian network is reasoned and analyzed using an inference engine to obtain evaluation results of production parameters and quality defects; each production parameter (or part of the production parameters) is input into the model as an evidence variable, and through forward reasoning analysis of the Bayesian network, prediction of billet quality can be realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A billet quality prediction process based on Bayesian network;

[0037] Figure 2 Flowchart for Bayesian network structure learning. DETAILED DESCRIPTION

[0038] This embodiment provides a continuous casting billet quality prediction method based on Bayesian network, such as Figure 1 , including the following steps:

[0039] (1) Sorting out the relationship between continuous casting production parameters and continuous casting billet quality

[0040] Based on solidification mechanism and metallurgical knowledge, the effects of various process parameters such as molten steel superheat, casting speed, mold cooling water temperature difference, secondary cooling parameters, and steel grade on the quality defects of the casting are analyzed. The main quality defects of the casting include: center segregation, porosity, shrinkage, and cracks.

[0041] (2) Establish a continuous casting process evaluation system

[0042] The five principles followed in selecting indicators and constructing the system are as follows: scientific principle, comprehensive and systematic principle, purpose principle, principle of organic combination of qualitative and quantitative analysis, and principle of steel grade difference. The scientific principle means that the continuous casting process can be evaluated accurately and effectively; the comprehensive and systematic principle means that both the correlation between various indicators and the causal relationship between indicators and parameters should be considered, and the economic benefits of the steel enterprise should be reflected; the purpose principle means that the goal is to be guided by the purpose and the goal of optimizing the continuous casting process and improving the economic benefits of the enterprise should be comprehensively considered; the principle of organic combination of qualitative and quantitative analysis means that the research focuses on whether there is a causal relationship of mutual influence and mutual connection between different evaluation indicators, as well as the extent of their mutual influence; the principle of steel grade difference means that when selecting evaluation indicators for steel grades with similar compositions, it is ensured that the different cooling and solidification characteristics of each steel grade are not confused;

[0043] The continuous casting process evaluation system is usually composed of the selected evaluation indicators and their own characteristics at various levels. It is divided into two levels: the quality defect level and the production parameter level. First, the steel grade composition is analyzed based on actual production data. Second, the evaluation indicators of the quality defect level are determined based on the quality defect data. Finally, combined with the research and analysis of the industrial background, the evaluation indicators of the production parameter level are determined based on the mutual influence relationship between various production parameters.

[0044] Considering the characteristics of Bayesian networks, data preprocessing and normalization are performed on each evaluation indicator. The evaluation indicators in the original data are all "continuous" data, so they need to be discretized. First, the original data is pivoted, and the discretized values ​​of each evaluation indicator are determined in combination with research and analysis of the industrial background. Secondly, based on the results of the pivot processing and the determined discretized values, the evaluation indicators are further analyzed to determine the discretization standards of each evaluation indicator. Finally, the original data is normalized and integrated into a normalized data set.

[0045] (3) Billet quality prediction model based on Bayesian network

[0046] 1) A scoring search-based structure learning algorithm is used to learn the Bayesian network structure of the prediction model. All possible Bayesian network structures are searched, each possible structure is scored, and the next round of search is guided by the scoring results until the scoring converges. The scoring function uses the Bayesian formula to solve the posterior probability of the network structure under a given data set D. The scoring formula is as follows:

[0047] Formula ①;

[0048] The network structure learning process is as follows Figure 2 :

[0049] Step 1: Determine the initial network structure model;

[0050] Step 2: Score each possible network structure change based on the scoring function and find the change with the highest joint probability of the dataset D and the network structure Ni;

[0051] Step 3: Modify the network structure model according to step 2;

[0052] Step 4: Repeat steps 2 and 3 until the network structure with the maximum joint probability of the data set D and the network structure is found;

[0053] The main inputs for structural learning include: a normalized discrete data set, the number of possible discrete state values ​​for each node, the node order, and the maximum number of parent nodes. The normalized discrete data set is obtained by processing actual continuous casting production data; the number of possible discrete state values ​​for each node is determined by the evaluation system; and the node order is determined based on the sequence of various production parameters in the continuous casting process and the correlation between quality defects. Production parameters are sorted from the tundish to the crystallizer to the secondary cooling zone, with quality defects ranked after all production parameters.

[0054] 2) Parameter learning of the Bayesian network in the billet quality prediction model. The purpose of parameter learning is to obtain the marginal probability distribution of each node in the Bayesian network - the network parameters, to achieve quantitative analysis of the Bayesian network, and to obtain a quantitative expression of the mutual influence and interdependence between nodes. Based on the given data set D and the obtained Bayesian network topology, the main steps of parameter learning are:

[0055] Step 1: Determine the prior distribution of network parameters;

[0056] Step 2: Based on the given data set D, calculate the posterior distribution of the node according to the Bayesian formula:

[0057] Formula②

[0058] Considering the complexity of the continuous casting process, the prior distribution of production parameters and quality defects is unknown, and the maximum likelihood estimation is used to obtain an unbiased estimate of the original distribution;

[0059] 3) Logical reasoning of the Bayesian network in the billet quality prediction model. The reasoning is to query the state of each node on the basis of obtaining a complete Bayesian network, perform inference calculations on the probability distribution of each node, and output the maximum possible discrete state of each node. The logical reasoning is performed using a joint tree inference engine. The first step of the joint tree inference is to convert the Bayesian network into a joint tree. The joint tree is essentially an undirected tree. The maximum complete subtree of the undirected tree is converted into each tree node of the joint tree. The second step is to calculate the probability of each node according to the message passing protocol defined in the joint tree.

[0060] 4) Actual production parameters are input into the model as evidence variables, and through forward reasoning analysis using the Bayesian network, the quality of the ingot is predicted;

[0061] (4) Reverse tracing of billet quality issues based on Bayesian networks

[0062] Similar to the ingot quality prediction model, after completing structure learning and parameter learning, the complete Bayesian network is reverse traced based on the joint tree inference engine to construct a ingot quality reverse traceability model based on the Bayesian network; using ingot quality defects as evidence variables, the Bayesian network of the continuous casting production process is reverse traced to obtain the discrete state of each production parameter node and its marginal probability distribution, thereby obtaining the reverse traceability evaluation results of the production parameters in the continuous casting process engineering, finding the key parameters affecting product quality, and providing a basis for product quality control.

[0063] In addition to the above embodiments, the present invention may also have other implementations. Any technical solution formed by equivalent replacement or equivalent transformation falls within the scope of protection required by the present invention.

Claims

1. A continuous casting slab quality prediction method based on Bayesian network, characterized by: include: (1) Sorting out the relationship between continuous casting production parameters and continuous casting billet quality Analyze the influence of various process parameters on the quality defects of casting billets based on solidification mechanism and metallurgical knowledge; (2) Establish a continuous casting process evaluation system The continuous casting process evaluation system is usually composed of the selected evaluation indicators and their own characteristics at various levels, and is divided into two levels: quality defect level and production parameter level; Considering the characteristics of Bayesian networks, data preprocessing and normalization are performed on each evaluation indicator. The evaluation indicators in the original data are all "continuous" data, so they need to be discretized and preprocessed. The discretization preprocessing is as follows: first, the original data is pivoted, and the discretized values ​​of each evaluation index are determined based on the research and analysis of the industrial background; second, based on the results of the pivoting and the determined discretized values, the evaluation indexes are further analyzed to determine the discretization standards of each evaluation index; finally, the original data is normalized and integrated into a normalized data set; (3) Billet quality prediction model based on Bayesian network 1) A scoring search-based structure learning algorithm is used to learn the Bayesian network structure of the prediction model. All possible Bayesian network structures are searched, each of which is scored. The scoring results guide the next round of search until the scores converge. The scoring function uses the Bayesian formula to solve the posterior probability of the network structure under the given data set D. The network structure learning process is as follows: Step 1: Determine the initial network structure model; Step 2: Score each possible network structure change based on the scoring function and find the change with the highest joint probability of the dataset D and the network structure Ni; Step 3: Modify the network structure model according to step 2; Step 4: Repeat steps 2 and 3 until the network structure with the maximum joint probability of the data set D and the network structure is found; The main inputs for structural learning include: a normalized discrete data set, the number of possible discrete state values ​​for each node, the node order, and the maximum number of parent nodes. The normalized discrete data set is obtained by processing actual continuous casting production data; the number of possible discrete state values ​​for each node is determined by the evaluation system; and the node order is determined based on the sequence of various production parameters in the continuous casting process and the correlation between quality defects. Production parameters are sorted from the tundish to the crystallizer to the secondary cooling zone, with quality defects ranked after all production parameters. 2) Parameter learning of Bayesian networks in the slab quality prediction model; Given a dataset D and obtaining the Bayesian network topology, the main steps of parameter learning are: Step I: Determine the prior distribution of network parameters; Step II: Based on the given data set D, calculate the posterior distribution of the node according to the Bayesian formula: Considering the complexity of the continuous casting process, the prior distribution of production parameters and quality defects is unknown, and the maximum likelihood estimation is used to obtain an unbiased estimate of the original distribution; 3) Logical reasoning of the Bayesian network in the billet quality prediction model, using a joint tree inference engine for logical reasoning; 4) Actual production parameters are input into the model as evidence variables, and the quality of the ingot is predicted through forward reasoning analysis using the Bayesian network; (4) Reverse tracing of billet quality issues based on Bayesian networks Similar to the ingot quality prediction model, after completing structure learning and parameter learning, the complete Bayesian network is reverse traced based on the joint tree inference engine to construct a ingot quality reverse traceability model based on the Bayesian network; using ingot quality defects as evidence variables, the Bayesian network of the continuous casting production process is reverse traced to obtain the discrete state of each production parameter node and its marginal probability distribution, thereby obtaining the reverse traceability evaluation results of the production parameters in the continuous casting process engineering, finding the key parameters affecting product quality, and providing a basis for product quality control.

2. The method for predicting continuous casting slab quality based on Bayesian network according to claim 1, characterized in that: In (1), the process parameters include molten steel superheat, casting speed, mold cooling water temperature difference, secondary cooling parameters, and steel type. The quality defects of the ingot mainly include: center segregation, porosity, shrinkage, and cracks.

3. The method for predicting continuous casting slab quality based on Bayesian network according to claim 1, characterized in that: In (2), first, the steel grade composition is analyzed based on actual production data; second, the evaluation indicators of the quality defect layer are determined based on the quality defect data; finally, combined with the research and analysis of the industrial background, the evaluation indicators of the production parameter layer are determined based on the mutual influence relationship between various production parameters.

4. The method for predicting continuous casting slab quality based on Bayesian network according to claim 1, characterized in that: In (3), the scoring formula is as follows:

5. The method for predicting continuous casting slab quality based on Bayesian network according to claim 1, characterized in that: In (3), the first step of the joint tree reasoning is to transform the Bayesian network into a joint tree. The joint tree is essentially an undirected tree. The maximum complete subtree of the undirected tree is transformed into each tree node of the joint tree. The second step is to calculate the probability of each node according to the message passing protocol defined in the joint tree.

Images