Casting blank bonding judgment method capable of realizing online self-learning

By constructing a random forest model and dynamically adjusting the weights of the decision trees, the problem of reduced adaptability of the crystallizer leakage prediction system was solved, and self-learning and dynamic optimization of billet adhesion determination were achieved, thereby improving the accuracy of determination and production stability.

CN119026025APending Publication Date: 2024-11-26HEBEI DAHE MATERIAL TECH CO LTD +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410936359.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing expert system for predicting steel leakage from crystallizers becomes less adaptable and its overall performance declines over time, requiring frequent updates and affecting production stability.

Method used

A random forest is constructed by using the bagging method to build multiple decision tree models. The weights of the decision tree models are dynamically adjusted through online data labeling to achieve self-learning and iterative updates.

Benefits of technology

The accuracy of billet adhesion determination has been improved, the model can adapt to production changes in a timely manner, reduce manual intervention, and improve the stability and accuracy of production.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026025A_ABST
    Figure CN119026025A_ABST
Patent Text Reader

Abstract

The invention relates to an online self-learning casting blank bonding judgment method, which comprises the following steps of: collecting thermocouple data when a casting blank is bonded and not bonded, randomly sampling the data by adopting a bagging method, and constructing a plurality of different decision tree models to form a random forest; and by using the mark of the online data, the weight of the decision tree model is dynamically adjusted, and the decision tree model can be updated and iterated. According to the method, the prediction precision of the integrated model is improved, the formed random forest is high in anti-noise capacity and not prone to being caught in overfitting, the independent decision tree in the model can conveniently adjust the weight and self iteration updating, model adjustment can be rapidly carried out in real time according to online feedback data, and the model is made to adapt to online changes in time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This patent application belongs to the field of continuous casting technology in the metallurgical industry, and more specifically, it relates to a method for determining the bonding of ingots that can be self-learned online. Background Art

[0002] As one of the most important links in steel production, continuous casting can greatly improve production efficiency. However, it may often lead to production interruptions due to steel leakage accidents, which will not only cause serious economic losses, but the molten steel flowing out of the shell will also damage the continuous casting equipment, block the guide rollers or foot rollers, and affect the normal production rhythm of the continuous casting machine.

[0003] The mold breakout prediction expert system is the most effective means to prevent the occurrence of steel breakout accidents caused by the adhesion of the ingot. Since the large-scale industrial production of slab continuous casting in the 1970s, foreign countries have been committed to studying the causes of steel breakout caused by the adhesion of the ingot, and have carried out in-depth research on the detection and prediction of steel breakout. In the decades of research on the formation mechanism of steel breakout, a variety of steel breakout prediction models have been established, and artificial intelligence technology and visualization technology have been used to jointly detect steel breakout characteristics and issue early warnings.

[0004] However, during the use of the crystallizer steel leakage prediction expert system, as time goes by and production conditions gradually change, the original crystallizer steel leakage prediction expert system may face the phenomenon of reduced adaptability and decreased comprehensive performance. Therefore, system developers are required to continuously modify and improve the system according to the on-site conditions and update and optimize the system.

[0005] The system developers constantly make manual modifications and improvements and update and optimize the system based on on-site conditions. This approach has the following problems: first, the developers need to have a full understanding of the current production process of the production line; second, frequent updates will affect the stability of production; third, as time goes by, the connection between equipment manufacturers and production lines may no longer be close. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a method for determining ingot bonding that can be self-learned online, which can effectively solve the problems existing in the background technology that the original crystallizer leakage prediction expert system may face over time, such as reduced adaptability and decreased comprehensive performance, and effectively improves the accuracy of ingot bonding determination, and can realize dynamic optimization and adaptation of the model according to new production data.

[0007] In order to solve the above problems, the technical solution adopted by the present invention is:

[0008] An online self-learning method for determining billet adhesion is proposed. The thermocouple data of billets with and without adhesion are collected, and the data are randomly sampled using the bagging method to construct multiple different decision tree models to form a random forest. The weight of the decision tree model is dynamically adjusted and the decision tree model can be updated and iterated by using the marking of online data.

[0009] The specific steps include:

[0010] S1. Collect the historical data of the thermocouple temperature of the copper plate of the crystallizer when the billet is bonded and when it is not bonded: The thermocouple data of the billet bonding includes the characteristic vector X composed of the temperature data of the first row, the second row, and the third row of thermocouples corresponding to the bonding position in the last 60 seconds after the bonding occurs. N The thermocouple data of the ingot without bonding includes the characteristic vector X composed of the temperature data of the first row, the second row, and the third row of thermocouples in a normal thermocouple column without bonding for the last 60 seconds. Z , X N With X Z The format is [T 1_1 、T 1_2 ,……T 1_60 、T 2_1 、T 2_2 ,……T 2_60 、T 3_1 、T 3_2 ,……T 3_60 ], mark the bonding data as (X N , 1), unbonded data is marked as (X Z , 0), each set of bonded data and each set of unbonded data are merged into a data set Q, and the total number of data sets in the data set Q is recorded as n;

[0011] S2. Randomly sample the data: repeatedly randomly select a set of data from the data set Q and put it into the new data set Q 新 Until the new data set Q 新 The amount of data in the new dataset Q reaches n. 1 Duplicate data is allowed in the dataset Q; data is randomly selected from the dataset Q and put into the new dataset Q. 新 The process is repeated 100 times, and 100 data sets are obtained and recorded as {Q 1 , Q 2 , Q 3 ...Q 100};

[0012] S3, training data set: use decision tree model to train 100 data sets {Q 1 , Q 2 , Q 3 ...Q 100} are trained separately, and the set of 100 decision tree models is recorded as {Tree 1 、Tree 2 、Tree 3 ...Tree 100};

[0013] S4. Label weights: The weights of the 100 decision tree models are recorded as {Weight 1 、Weight 2 、Weight 3 ……Weight 100}, and initialize the weights of 100 decision tree models to 1;

[0014] S5. Real-time output of weighted prediction results: 100 decision tree models are arranged online, and weighted prediction results are output in real time, which are: bonding occurs or no bonding occurs;

[0015] S6. Data marking: After the weighted prediction result is that bonding occurs, observe whether the actual billet bonding occurs. If the billet bonding actually occurs, the data is marked as (X N , 1) If the billet does not actually bond, the data is marked as (X Z ,0), add the labeled data to the data set Q;

[0016] S7. Modify the weight of each decision tree model according to the actual ingot bonding observation results.

[0017] Furthermore, in S2, the bagging method is used to randomly sample the data.

[0018] Furthermore, the decision tree model in S3 adopts the C4.5 algorithm.

[0019] Furthermore, in S5, the method of outputting the weighted prediction result is to calculate the sum of the weights of the decision tree models whose judgment results are that bonding occurs. 粘结 The sum of the weights of the decision tree model whose result is no bonding 未粘结 , when Weight 粘结 ≥Weight 未粘结 When , the weighted prediction result is that bonding occurs; when Weight 粘结 <Weight 未粘结 When , the weighted prediction result is that no bonding occurs.

[0020] Further, in S7, “correcting the weights of each decision tree model according to actual slab bonding observation results” means:

[0021] If the prediction result of the decision tree model is consistent with the actual observation result of the bonding of the ingot, its weight is increased by 0.05, and its weight is not greater than the maximum weight threshold;

[0022] If the prediction result of the decision tree model does not conform to the actual observation result of the bonding of the ingot, its weight is reduced by 0.1. When the weight of the decision tree model is less than the minimum weight threshold, the decision tree model is deleted from the decision tree model set;

[0023] When a decision tree model is deleted, a set of data is randomly selected from the data set Q and put into the new data set Q new Middle, to Q new The amount of data in the new Duplicate data is allowed in the data set; the decision tree model is used to analyze the data set Q new Train and get a new decision tree model denoted as Tree m , the new decision tree model Tree m Add it to the original decision tree model set and initialize the weight of the decision tree model to 1. Here Q new It is a similar new dataset generated by the same method as the previous Q1 dataset.

[0024] Furthermore, the maximum weight threshold is 1.2 and the minimum weight threshold is 0.5.

[0025] Due to the adoption of the above technical solution, the beneficial effects achieved by the present invention are:

[0026] The invention collects thermocouple data when the ingot is bonded and when it is not bonded, uses bagging method to randomly sample the data, and constructs multiple different decision tree models to form a random forest. By marking the online data, the weight of the decision tree model is dynamically adjusted and the decision tree model can be updated and iterated.

[0027] By using the random forest method to determine adhesion, while providing accurate determination results in a modeled and automated manner, this model has the following characteristics, which are different from traditional machine learning models: First, it uses an integrated learning method to combine weak classifiers into strong classifiers, thereby improving the prediction accuracy of the integrated model; second, randomness is introduced into the model data, and the resulting random forest has strong noise resistance and is not prone to overfitting; third, the independent decision trees in the model can easily adjust weights and iteratively update themselves, and can quickly adjust the model in real time based on online feedback data, so that the model can adapt to online changes in a timely manner.

[0028] The present invention effectively solves the problems existing in the background technology, such as reduced adaptability and decreased comprehensive performance of the original crystallizer leakage prediction expert system over time, effectively improves the accuracy of billet bonding judgment, and can realize dynamic optimization and adaptation of the model according to new production data. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a process flow chart of the present invention. DETAILED DESCRIPTION

[0030] The present invention is further described in detail below with reference to the embodiments.

[0031] A method for determining the bonding of a billet that can be self-learned online, the specific steps of the invention are: by collecting thermocouple data when the billet is bonded and when it is not bonded, the data is randomly sampled using the bagging method, and multiple different decision tree models are constructed to form a random forest. By using the marking of online data, the weight of the decision tree model is dynamically adjusted and the decision tree model can be updated and iterated.

[0032] like Figure 1 The example shown:

[0033] The first step is to collect the historical data of the thermocouple temperature of the copper plate of the crystallizer when the billet is bonded and when it is not bonded. The thermocouple data of the billet bonding includes the characteristic vector X composed of the temperature data of the first row, the second row, and the third row of thermocouples corresponding to the bonding position in the last 60 seconds after the bonding occurs. N The thermocouple data of the ingot without bonding includes the characteristic vector X composed of the temperature data of the first row, the second row, and the third row of thermocouples in a normal thermocouple column without bonding for the last 60 seconds. Z , X N With X Z The format is [T 1_1 、T 1_2 ,……T 1_60 、T 2_1 、T 2_2 ,……T 2_60 、T 3_1 、T 3_2 ,……T 3_60 ], mark the bonding data as (X N , 1), unbonded data is marked as (X Z , 0), each set of bonding data and each set of unbonded data are combined into a data set Q, and the total number of data sets in the data set Q is recorded as n. Each set of bonding data here refers to the characteristic vector X composed of the temperature data of the first row, second row, and third row of thermocouples in the column corresponding to the bonding position in the last 60 seconds after bonding occurs. N, and then mark it as (X N , 1) is recorded as a set (or one time) of bonding data.

[0034] Step 2: Repeatedly randomly select a set of data from the data set Q and put it into the new data set Q 新 In, to the new data set Q 新 The amount of data in the new dataset Q reaches n. 新 Duplicate data is allowed in Q. Repeatedly randomly select data from data set Q and put it into the new data set Q. 新 The process is repeated 100 times, and 100 data sets are obtained and recorded as {Q 1 , Q 2 , Q 3 ...Q 100}.

[0035] Here I want to say that a new data set Q is randomly generated from data set Q by repeating it 100 times 1 , Q 2 , Q 3 ...Q 100 A total of 100 new data sets Q 新 , these 100 new data sets Q 新 As the data set used by S3 later.

[0036] The third step is to use the decision tree model to analyze the 100 data sets {Q 1 , Q 2 , Q 3 ...Q 100} are trained separately, and the set of 100 decision tree models is recorded as {Tree 1 、Tree 2 、Tree 3 ...Tree 100}. The decision tree model uses the C4.5 algorithm.

[0037] Step 4: The weights of the 100 decision tree models are recorded as {Weight 1 、Weight 2 、Weight 3 ……Weight 100}, and initialize the weights of 100 decision tree models to 1.

[0038] Step 5: Arrange 100 decision tree models online and output weighted prediction results in real time. The method for outputting weighted prediction results is to calculate the sum of weights of the decision tree models whose judgment results are that bonding occurs. 粘结 The sum of the weights of the decision tree model whose result is no bonding 未粘结 , when Weight 粘结≥Weight 未粘结 When , the weighted prediction result is that bonding occurs. 粘结 <Weight 未粘结 When , the weighted prediction result is that no bonding occurs. For example, at a certain moment, 85 decision trees predict that bonding occurs. Among the 85 decision trees, 20 weights are 1.2, 40 are 1, and 25 are 0.8. Therefore, Weight 粘结 =20*1.2+40*1+25*0.8=84, the remaining 15 decision trees predict that no bonding occurs, of which 10 decision trees have a weight of 1, 4 decision trees have a weight of 0.8, and 1 decision tree has a weight of 0.5. Therefore, Weight 未粘结 =10*1+4*0.8+1*0.5=13.7. Because Weight 粘结 >Weight 未粘结 , so the weighted prediction of the model is that bonding occurs.

[0039] Step 6: After the weighted prediction result is that bonding occurs, observe whether the actual billet bonding occurs. If the billet bonding actually occurs, the data is marked as (X N , 1) If the billet does not actually bond, the data is marked as (X Z , 0), and add the marked data to the data set Q. If the prediction results are observed in the actual casting, it is found that bonding occurs on the casting, so the data is marked as (X N ,1),Add the labeled data to the dataset Q.

[0040] Step 7. Correct the weight of each decision tree model according to the actual ingot bonding observation results. If the prediction result of the decision tree model is consistent with the actual ingot bonding observation results, its weight is increased by 0.05, and its weight is not greater than the maximum weight threshold of 1.2. If the prediction result of the decision tree model does not conform to the actual ingot bonding observation results, its weight is reduced by 0.1. When the weight of the decision tree model is less than the minimum weight threshold of 0.5, the decision tree model is deleted from the decision tree model set.

[0041] Therefore, in this prediction result, the weights of the 85 decision tree models that predicted correctly were increased, and the weights after the increase were 20 weights of 1.2, 40 weights of 1.05, and 25 weights of 0.85. The weights of the 15 decision tree models that predicted incorrectly were reduced, and the weights after the reduction were 10 decision trees of 0.9, 4 decision trees of 0.7, and another decision tree with a weight of 0.5 before the prediction was deleted from the decision tree model set.

[0042] When a decision tree model Tree m After being deleted, a set of data is randomly selected from the data set Q and put into the new data set Qnew (Q new is a similar new dataset generated by the same method as the previous datasets Q1 and Q2. new The amount of data in the new Duplicate data is allowed in the data set. new Train and get a new decision tree model denoted as Tree m , Tree m Add it to the original decision tree model set and initialize the weight of the decision tree model to 1.

Claims

1. A method for determining the bonding of a slab that can be self-learned online, characterized in that The steps include: S1. Collect the historical data of the thermocouple temperature of the copper plate of the crystallizer when the billet is bonded and when it is not bonded: The thermocouple data of the billet bonding includes the characteristic vector X composed of the temperature data of the first row, the second row, and the third row of thermocouples corresponding to the bonding position in the last 60 seconds after the bonding occurs. N The thermocouple data of the ingot without bonding includes the characteristic vector X composed of the temperature data of the first row, the second row, and the third row of thermocouples in a normal thermocouple column without bonding for the last 60 seconds. Z , X N With X Z The format is [T 1_1 , T 1_2 ,……T 1_60 , T 2_1 , T 2_2 ,……T 2_60 , T 3_1 , T 3_2 ,……T 3_60 ], mark the bonding data as (X N , 1), unbonded data is marked as (X Z , 0), each group of bonded data and each group of unbonded data are merged into a data set Q, and the total number of data groups in the data set Q is recorded as n; S2. Randomly sample the data: repeatedly randomly select a set of data from the data set Q and put it into the new data set Q 新 Until the new data set Q 新 The amount of data in the new dataset Q reaches n. 新 Duplicate data is allowed in Repeatedly randomly select data from data set Q and put it into the new data set Q 新 The process is repeated 100 times, and 100 data sets are obtained and recorded as {Q1, Q2, Q3...Q 100 }; S3, training data set: use decision tree model to train 100 data sets {Q1, Q2, Q3...Q 100 } respectively, and the set of 100 decision tree models is recorded as {Tree1, Tree2, Tree3...Tree 100 }; S4. Label weights: record the weights of the 100 decision tree models as {Weight1, Weight2, Weight3...Weight 100 }, and initialize the weights of 100 decision tree models to 1; S5. Real-time output of weighted prediction results: 100 decision tree models are arranged online, and weighted prediction results are output in real time, which are: bonding occurs or no bonding occurs; S6. Data marking: After the weighted prediction result is that bonding occurs, observe whether the actual billet bonding occurs. If the billet bonding actually occurs, the data is marked as (X N , 1) If the billet does not actually stick, the data is marked as (X Z , 0), add the labeled data to the data set Q; S7. Modify the weight of each decision tree model according to the actual ingot bonding observation results.

2. The method for determining the bonding of a casting blank capable of online self-learning according to claim 1, characterized in that: In S2, bagging method is used to randomly sample data.

3. The method for determining the bonding of a casting blank capable of online self-learning according to claim 1, characterized in that: The decision tree model in S3 uses the C4.5 algorithm.

4. The method for determining the bonding of a casting blank capable of online self-learning according to claim 1, characterized in that: In S5, the method of outputting the weighted prediction result is to calculate the sum of the weights of the decision tree models whose judgment results are that bonding occurs. 粘结 The sum of the weights of the decision tree model whose result is no bonding 未粘结 , when Weight 粘结 ≥Weight 未粘结 When , the weighted prediction result is that bonding occurs; when Weight 粘结 <Weight 未粘结 When , the weighted prediction result is that no bonding occurs.

5. The method for determining the bonding of a casting slab capable of online self-learning according to claim 1, characterized in that: In S7, "correcting the weights of each decision tree model according to the actual ingot sticking observation results" means: If the prediction result of the decision tree model is consistent with the actual observation result of the bonding of the ingot, its weight is increased by 0.05, and its weight is not greater than the maximum weight threshold; If the prediction result of the decision tree model does not conform to the actual observation result of the bonding of the ingot, its weight is reduced by 0.

1. When the weight of the decision tree model is less than the minimum weight threshold, the decision tree model is deleted from the decision tree model set; When a decision tree model is deleted, a set of data is randomly selected from the data set Q and put into the new data set Q new In, to the new data set Q new The amount of data in the new dataset Q reaches n. new Duplicate data is allowed in Use decision tree model to analyze new data set Q new Train and get a new decision tree model denoted as Tree m , the new decision tree model Tree m Add it to the original decision tree model set and initialize the weight of the decision tree model to 1.

6. The method for determining the bonding of a casting slab capable of online self-learning according to claim 5, characterized in that: The maximum weight threshold is 1.2, and the minimum weight threshold is 0.5.