Intelligent hot rolling process optimization method based on artificial labels

Abstract

The invention relates to an intelligent hot rolling process optimization method based on an artificial label, which comprises the following steps: performing labeling influence factor analysis, specifying a labeling judgment standard, affirming a labeling responsible person, performing labeling process standard specification, performing labeling data acquisition, performing data clearness and data pre-processing, constructing a machine learning scheme, and performing data training. Inputting a current equipment working condition label into the HMI, and optimizing process solution and application; real-time online rolling force prediction is carried out by utilizing the trained model, the rolling force prediction precision is improved, uncontrollable factors of equipment are fully considered, the uncontrollable factors of the equipment are introduced into a machine learning algorithm in a manual labeling form, the real situation is better met, the decision result is more accurate, and the method is suitable for large-scale popularization and application. And when the training data is different from the latest state, a better training result can be obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of hot-rolled strip steel, and in particular to an intelligent hot-rolling process optimization method based on artificial labels. Background Art

[0002] Hot-rolled strip production is an indispensable key process in the entire steel production process. From the perspective of the production process, hot-rolled strip production is a key link that connects the preceding and following stages. On the technical level, hot-rolled strip production involves multiple complex physical and chemical processes such as high-temperature heating, rolling deformation, and cooling control. The efficiency and quality of hot-rolled strip production also directly affect the economic benefits and market competitiveness of steel companies.

[0003] The core element of hot rolling production is the rolling mill, which provides the core capability of hot rolled strip deformation: rolling force. Accurate prediction of rolling force and optimization of rolling process parameters are essential to ensure the quality and performance of the strip. Traditional rolling force calculation methods are usually based on theoretical models, but due to the complexity and uncertainty of the hot rolling process, these models often have large errors. In addition, with the continuous improvement of market requirements for strip quality and performance, the traditional process parameter adjustment method based on experience and trial and error has been difficult to meet production needs.

[0004] Therefore, many researchers have proposed a variety of intelligent algorithms for predicting and calculating rolling force, such as the XGBoost algorithm used in CN110773580B and the random forest algorithm used in CN 111790762B. There are also many similar patents. These patents will perform data cleaning before modeling to obtain as much useful data as possible, and perform correlation analysis to obtain the influence weights of different influencing factors as much as possible. However, these patents do not distinguish the authenticity of the data when performing data analysis, especially the survival data is a hot rolling mill system composed of tens of thousands of spare parts and replacement parts. The production status of the system is changing all the time. For example, there is a resonance problem in the F1 frame in a certain data collection stage. The rolling force of the frame is limited during production, but these data are not marked as abnormal. Then these data will cause pollution to the entire model training, and the model trained based on these data will not be able to guide the normal production process. Summary of the invention

[0005] The present invention provides an intelligent hot rolling process optimization method based on artificial labels. In the historical data preparation stage, labeling influencing factor analysis is added, labeling judgment standards are formulated, labeling responsible persons are identified, and labeling process standard specification work is carried out. Then, a multi-objective intelligent optimization algorithm is used to build a model and train data. Finally, the trained model is used to perform real-time online rolling force prediction, thereby improving the rolling force prediction accuracy, taking full account of the uncontrollable factors of the equipment, being more in line with the actual situation, and making the decision results more accurate. When the training data is different from the latest state, better training results can also be obtained.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] An intelligent hot rolling process optimization method based on artificial labels comprises the following steps:

[0008] S1. Label determination: Analyze the factors affecting labeling, formulate labeling judgment standards, identify the person responsible for labeling, and carry out labeling process standard specifications;

[0009] S2. Data collection: Collect hot-rolled strip production data for no less than 3 months and combine it with manual labels to build a basic data platform;

[0010] S3, data pre-processing: perform data pre-processing on the collected production data;

[0011] S4. Build a machine learning solution: Select a machine learning solution with label scalability to classify the data and determine the training data set and validation data set;

[0012] S5. Data training: Train the data of process problems based on machine learning algorithms to obtain process models, and use validation data sets to verify the model results;

[0013] S6. Determine the label of the current equipment working condition: label the current equipment status based on the label label determined in S1;

[0014] S7, process solution when all equipment is currently operating normally: Use the machine learning model trained in S5 to solve the equipment condition labels of S6 and the production data to be produced to obtain the optimal process solution under the current equipment status.

[0015] Furthermore, the step S1 includes the following steps:

[0016] S11. Analyze various production failures that may occur in hot rolling mills, including roll eccentricity, roll surface defects, roll bearing failures, hydraulic failures, temperature detection failures, uneven heating of heating furnaces, roller deviation, plate shape defects and automatic control system failures;

[0017] S12. According to the magnitude of the factors affecting production during the production process, identify each fault and determine whether it affects the production process;

[0018] S13, introducing the purpose of data collection to the on-site operators, and making a preliminary quantitative determination based on the impact of the data collection on the rolling force;

[0019] S14. Make the process problems that need to be labeled into test papers and standardize the operators;

[0020] S15. Provide standard and regulatory training to operators to ensure data accuracy during the data collection phase.

[0021] Furthermore, step S2 includes the following steps:

[0022] S21, the operator determines the equipment status in the data collection stage;

[0023] S22. After collecting data every day, the data for that day is reviewed manually to determine the accuracy of the data, so as to ensure that the manually labeled data is consistent with the actual production situation.

[0024] Furthermore, step S3 includes the following steps:

[0025] S31. Initial attribute screening: Based on prior knowledge and data missing conditions, select attributes for subsequent modeling from the original data.

[0026] S32, abnormal value processing: when a positive attribute has a non-positive value abnormality, all non-positive values ​​of this type of attribute are replaced with null values;

[0027] S33. Negative value processing: For attributes with both positive and negative values, convert all attributes of this type into absolute values.

[0028] Furthermore, the step S4 adopts the random forest theory to construct a model, including the following steps:

[0029] S41, randomly sample training subsets to build decision trees:

[0030] Randomly extract multiple sample subsets D from the original training data set D i (i=1,2,……,n, where n is the number of decision trees) to construct a decision tree. For the sample subset D i , randomly select features from this subset for node splitting, there are m features in total, k features are randomly selected each time Used to evaluate the splitting criteria; the splitting criteria are used to determine the best splitting point. When splitting a node, the feature and splitting point that minimizes the sum of the Gini index of the child nodes after the split are selected for splitting;

[0031] S42, prediction by integrated decision trees: After building multiple decision trees, random forest integrates the prediction results of these decision trees to make the final prediction. For classification problems, the voting method is used, and for regression problems, the averaging method is used.

[0032] S44. Evaluate feature importance:

[0033] The average impurity reduction method is used to measure feature importance. For each decision tree, the impurity reduction brought by each feature when the node is split is calculated, and then the average is taken for all decision trees.

[0034] Furthermore, the classification problem in step S42 is:

[0035] For a new sample x, each decision tree will give a category prediction result, and the final prediction category is the category with the most votes. That is, if there are n decision trees, for category j, if one tree predicts category j, then the probability that sample x belongs to category j is expressed as

[0036] Furthermore, the regression problem in step S42 is:

[0037] For a new sample x, each decision tree gives a prediction value, and the final prediction value is the average of the prediction values ​​of all decision trees. That is, if there are n decision trees, the prediction value of each tree for the sample is y i (i=1,2,3,……,n,), then the final prediction value is

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1) Use the trained model to perform real-time online rolling force prediction to improve the rolling force prediction accuracy. When optimizing the hot rolling intelligent process, the uncontrollable factors of the equipment are fully considered. Through manual labeling, the uncontrollable factors of the equipment are introduced into the machine learning algorithm, which is more in line with the actual situation and the decision results are more accurate.

[0040] 2) When applying the model, the uncontrollable factors of the equipment are fully considered. Even if the training data is different from the latest status, good training results can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a method flow chart of an embodiment of the present invention.

[0042] Figure 2 It is a schematic diagram of data clarification-outlier removal according to an embodiment of the present invention.

[0043] Figure 3It is a schematic diagram of data pre-processing-data equalization according to an embodiment of the present invention.

[0044] Figure 4 Schematic diagram of random forest importance index analysis according to an embodiment of the present invention.

[0045] Figure 5 It is a schematic diagram of the effect of C40 mean optimization described in an embodiment of the present invention and the corresponding increase in the proportion of cold-rolled products less than 8um. DETAILED DESCRIPTION

[0046] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0047] See Figure 1 , is a flow chart of the method of the present invention. The present invention provides an intelligent hot rolling process optimization method based on artificial labels, comprising the following steps:

[0048] S1. Label determination: Analyze the factors affecting labeling, formulate labeling judgment standards, identify the responsible persons for labeling, and carry out labeling process standard specifications; analyze various production failures that may be encountered by the hot rolling mill, determine the corresponding label categories of different failures, and introduce the purpose of data collection to the operators of the four shifts A, B, C, and D on site, and make a preliminary quantitative judgment based on the impact on the rolling force. The specific steps include the following:

[0049] S11. Analyze various production failures that may occur in hot rolling mills, such as roll eccentricity, roll surface defects, roll bearing failure, hydraulic failure, temperature detection failure, uneven heating of the heating furnace, roller deviation, plate shape defects, and automatic control system failure;

[0050] S12. According to the magnitude of the factors affecting production during the production process, each fault is identified to determine whether it affects the production process, such as labeling the good and bad roll shifting functions of F1-F7 hot rolling mills and labeling the good and bad production stability of rolling mills;

[0051] S13. Introduce the purpose of data collection to the operators of the four shifts A, B, C and D on site, and make a preliminary quantitative judgment based on the impact of the data on the rolling force;

[0052] S14. Make the process problems that need to be labeled into test papers, and standardize the operators of the four shifts A, B, C, and D;

[0053] S15. Commend the operators of the four shifts A, B, C and D, and strive for data accuracy during the data collection stage.

[0054] S2. Data collection: Collect hot-rolled strip production data for no less than 3 months, including basic process data of heating furnace, rough rolling, finishing rolling, layer cooling, curling, etc., and combine it with manual labels to build a basic data platform, which includes the following steps:

[0055] S21. The operators of the four shifts A, B, C and D determine the equipment status during the data collection phase. For example, the A shift determines that the roller shifting function of the F4 frame is poor, and to ensure normal production, adjusts its roller shifting capacity to other units. For another example, the D shift determines that the resonance state of the F2 frame is poor, and transfers the load distribution to the front and rear frames;

[0056] S22. After collecting data every day, the data must be manually reviewed to determine the accuracy of the data to ensure that the manually labeled data is consistent with the actual production situation.

[0057] S3, data preprocessing: perform data preprocessing on the collected production data, including data cleaning, data smoothing and data standardization. Here, the manual labeling retains both the original data and the labeled data; it includes the following steps:

[0058] S31. Initial screening of attributes: There are a large number of empty or all-0 attributes in the original data. Based on prior knowledge and data missing conditions, and relying on the expert database, useless attributes are deleted from 2,270 attributes, and 144 attributes for subsequent modeling are screened out;

[0059] S32. Abnormal value processing: There are some attributes that must be positive values. If there are abnormal non-positive values, all non-positive values ​​of this type of attributes will be replaced with null values.

[0060] S33. Negative value processing: There are some attributes that have both positive and negative values. Convert all attributes of this type into absolute values.

[0061] S4. Build a machine learning solution: According to the goal of process optimization, select a suitable machine learning solution. The primary basis for selecting the solution is to have good label scalability, that is, when data other than labels appears, it can be expanded to a certain extent, and the data can be classified to determine the training data set and the verification data set; including the following steps:

[0062] S41, randomly sample training subsets to build decision trees:

[0063] Random forests are constructed by randomly sampling multiple sample subsets D from the original training dataset D. i (i=1,2,……,n, where n is the number of decision trees) to construct a decision tree. The construction process of each decision tree is as follows:

[0064] For the sample subset D i, randomly select a part of the features from the subset for node splitting. Assuming there are m features in total, k features are randomly selected each time to evaluate the splitting criteria.

[0065] Use some splitting criteria, including information gain and Gini index, to determine the best splitting point. For example, for the Gini index, the calculation formula is: Where C is the number of categories, P i is the probability that the sample belongs to the i-th category. When a node is split, the feature and split point that minimize the sum of the Gini index of the child nodes after the split are selected for splitting;

[0066] S42, integrated decision tree for prediction:

[0067] After building multiple decision trees, random forests make final predictions by integrating the prediction results of these decision trees. For classification problems, the voting method is used; for regression problems, the averaging method is used;

[0068] Classification problem: For a new sample x, each decision tree will give a category prediction result, and the final prediction category is the category with the most votes. That is, if there are n decision trees, for category j, if a tree predicts category j, then the probability that sample x belongs to category j can be expressed as

[0069] Regression problem: For a new sample x, each decision tree gives a prediction value, and the final prediction value is the average of all decision tree prediction values. That is, if there are n decision trees, the prediction value of each tree for the sample is y. i (i=1,2,3,……,n,), then the final prediction value is

[0070] S5. Data training: Train the data of process problems based on machine learning algorithms to obtain accurate process models, and use verification data sets to verify the model results.

[0071] S6. Based on the label annotation determined in S1, label the status of the current device.

[0072] S7, process solution when all equipment is currently operating normally: Use the machine learning model trained in S5 to solve the equipment condition labels of S6 and the production data to be produced to obtain the optimal process solution under the current equipment status.

[0073] The following examples are implemented on the premise of the technical solution of the present invention, and provide detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to the following examples. The methods used in the following examples are conventional methods unless otherwise specified.

[0074] [Example]

[0075] When a steel plant produces hot-rolled silicon steel convexity control, in order to reduce the actual control value of silicon steel convexity C40, and thus increase the proportion of cold-rolled finished products C15≤8um; an intelligent hot rolling process optimization method based on artificial labels specifically includes the following steps:

[0076] S1. Various factors affecting the convexity of hot-rolled silicon steel are sorted out to determine important influencing factors such as rolling force, load distribution, and proportional convexity distribution. Among them, there are uncontrollable factors such as roll eccentricity, roll surface defects, roll bearing failure, hydraulic failure, temperature detection failure, uneven heating of the heating furnace, roller deviation, plate shape defects, and automatic control system failure. These factors cannot be entered into the machine learning system in a digitized way, and they need to be manually labeled to introduce their influencing factors into the machine learning system.

[0077] S2. After quantifying factors such as roll shifting, rolling force, and motor power, the corresponding label format is formulated and the data collection stage is entered. This stage needs to take into account both the accuracy of the data and the accuracy of manual judgment. When a certain shift A determines that the roll shifting function of the F4 frame is abnormal, in order to ensure normal production, its roll shifting capacity is adjusted to other frames. At this time, the roll shifting capacity of the F4 frame will be labeled as poor;

[0078] After collecting data every day, the data for that day must be manually reviewed to determine its accuracy to ensure that the manually labeled data is consistent with the actual production situation.

[0079] S3, pre-process the collected production data, such as Figure 2 As shown in , outliers of certain values ​​in the material are removed; Figure 3 The data shown are equalized.

[0080] S4. Construct a machine learning solution. Here we choose the random forest theory: Random forest randomly extracts multiple sample subsets D from the original training data set D. i (i=1,2,……,n, where n is the number of decision trees) to construct a decision tree. The construction process of each decision tree is as follows:

[0081] For the sample subset D i , randomly select a part of the features from the subset for node splitting. Assuming there are m features in total, k features are randomly selected each time to evaluate the splitting criteria.

[0082] Use some splitting criteria, including information gain and Gini index, to determine the best splitting point. For the Gini index, the calculation formula is: Where C is the number of categories, Pi is the probability that the sample belongs to the i-th category. When a node is split, the feature and split point that minimize the sum of the Gini index of the child nodes after the split are selected for splitting;

[0083] After building multiple decision trees, random forests make final predictions by integrating the prediction results of these decision trees. For classification problems, the voting method is used; for regression problems, the averaging method is used;

[0084] Classification problem: For a new sample x, each decision tree will give a category prediction result, and the final prediction category is the category with the most votes. That is, if there are n decision trees, for category j, if a tree predicts category j, then the probability that sample x belongs to category j can be expressed as

[0085] Regression problem: For a new sample x, each decision tree gives a prediction value, and the final prediction value is the average of all decision tree prediction values. That is, if there are n decision trees, the prediction value of each tree for the sample is: y i (i=1,2,3,……,n,), then the final prediction value is

[0086] S5. Use random forest theory to train the data and obtain the corresponding training model. Figure 4 The importance index analysis conclusions are shown.

[0087] S6. The operator first determines the data label and inputs it into the HMI interactive interface, and the model automatically reads the data from the interactive interface.

[0088] S7. When the model is ready to guide on-site production, the optimization goal is to minimize the convexity of each strip. The model trained in step S5 is applied to the site. The operator first determines the data label and enters it into the HMI interactive interface. The model automatically reads the data from the interactive interface and performs optimization. The optimization result is as follows: Figure 5 As shown in the figure, the C40 control target of hot-rolled silicon steel gradually decreases, the smaller the better, and the corresponding proportion of cold-rolled finished products less than 8um gradually increases.

Claims

1. An intelligent hot rolling process optimization method based on artificial labels, characterized in that: The steps include: S1. Label determination: Analyze the factors affecting labeling, formulate labeling judgment standards, identify the person responsible for labeling, and carry out labeling process standard specifications; S2. Data collection: Collect hot-rolled strip production data for no less than 3 months and combine it with manual labels to build a basic data platform; S3, data pre-processing: perform data pre-processing on the collected production data; S4. Build a machine learning solution: Select a machine learning solution with label scalability to classify the data and determine the training data set and validation data set; S5. Data training: Train the data of process problems based on machine learning algorithms to obtain process models, and use validation data sets to verify the model results; S6. Determine the label of the current equipment working condition: label the current equipment status based on the label label determined in S1; S7, process solution when all equipment is currently operating normally: Use the machine learning model trained in S5 to solve the equipment condition labels of S6 and the production data to be produced to obtain the optimal process solution under the current equipment status.

2. The intelligent hot rolling process optimization method based on artificial labels according to claim 1 is characterized in that: The step S1 comprises the following steps: S11. Analyze various production failures that may occur in hot rolling mills, including roll eccentricity, roll surface defects, roll bearing failures, hydraulic failures, temperature detection failures, uneven heating of heating furnaces, roller deviation, plate shape defects and automatic control system failures; S12. According to the magnitude of the factors affecting production during the production process, identify each fault and determine whether it affects the production process; S13, introducing the purpose of data collection to the on-site operators, and making a preliminary quantitative determination based on the impact of the data collection on the rolling force; S14. Make the process problems that need to be labeled into test papers and standardize the operators; S15. Provide standard and regulatory training to operators to ensure data accuracy during the data collection phase.

3. The intelligent hot rolling process optimization method based on artificial labels according to claim 1 is characterized in that: The step S2 comprises the following steps: S21, the operator determines the equipment status in the data collection stage; S22. After collecting data every day, the data for that day is reviewed manually to determine the accuracy of the data and ensure that the manually labeled data is consistent with the actual production situation.

4. The intelligent hot rolling process optimization method based on artificial labels according to claim 1 is characterized in that: The step S3 comprises the following steps: S31. Initial attribute screening: Based on prior knowledge and data missing conditions, attributes for subsequent modeling are screened from the original data; S32, abnormal value processing: when a positive attribute has a non-positive value abnormality, all non-positive values ​​of this type of attribute are replaced with null values; S33. Negative value processing: For attributes with both positive and negative values, convert all attributes of this type into absolute values.

5. The intelligent hot rolling process optimization method based on artificial labels according to claim 1 is characterized in that: The step S4 adopts the random forest theory to build a model, including the following steps: S41, randomly sample training subsets to build decision trees: Randomly extract multiple sample subsets D from the original training data set D i (i=1,2,……,n, where n is the number of decision trees) to construct a decision tree. For the sample subset D i , randomly select features from this subset for node splitting, there are m features in total, k features are randomly selected each time Used to evaluate the splitting criteria; the splitting criteria are used to determine the best splitting point. When splitting a node, the feature and splitting point that minimizes the sum of the Gini index of the child nodes after the split are selected for splitting; S42, prediction by integrated decision trees: After building multiple decision trees, random forest integrates the prediction results of these decision trees to make the final prediction. For classification problems, the voting method is used, and for regression problems, the averaging method is used. S44. Evaluate feature importance: The average impurity reduction method is used to measure feature importance. For each decision tree, the impurity reduction brought by each feature when the node is split is calculated, and then the average is taken for all decision trees.

6. The intelligent hot rolling process optimization method based on artificial labels according to claim 5 is characterized in that: The classification problem in step S42 is: For a new sample x, each decision tree will give a category prediction result, and the final prediction category is the category with the most votes. That is, if there are n decision trees, for category j, if one tree predicts category j, then the probability that sample x belongs to category j is expressed as 7. The intelligent hot rolling process optimization method based on artificial labels according to claim 5 is characterized in that: The regression problem in step S42 is: For a new sample x, each decision tree gives a prediction value, and the final prediction value is the average of the prediction values ​​of all decision trees. That is, if there are n decision trees, the prediction value of each tree for the sample is y i (i=1,2,3,……,n,), then the final prediction value is

Images