Single-stand 20-roller mill ASU presetting method based on data driving
By applying deep neural networks on the twenty-roll rolling mill to establish multi-layer preset models, the problems of plate shape setting stability and reliability of the rolling mill are solved, and the stability of the rolling process and the improvement of plate shape quality are achieved.
Patent Information
- Application Number
- CN202510088419.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-21
- Publication Date
- 2025-05-23
AI Technical Summary
The twenty-fold rolling mill has stability and reliability problems in plate shape setting, resulting in plate shape defects and frequent strip breaks. The existing empirical regression model is difficult to improve the setting accuracy.
Using a data driving method based on deep neural networks, a multi-layer preset model is established, and the adjustment ratio of a single frame 20 roll mill ASU is accurately set through the combination of different modeling data layers.
It improves the stability and plate shape quality of the rolling process, increases the yield and economic benefits, and overcomes the limitations of traditional methods in complex plate shape control.
Smart Images
Figure CN120030885A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of metallurgical engineering and artificial intelligence, and relates to a data-driven single-stand 20-roller mill ASU presetting method. Background Art
[0002] At present, the shape setting of the twenty-high mill mainly relies on on-site manual experience, which has poor stability and reliability, and is prone to quality problems such as shape defects and frequent strip breakage. Especially in the initial rolling stage of each rolling pass, the rolling process is extremely unstable and the shape quality fluctuates greatly. The accuracy and stability of shape setting are not only related to the shape control of the head of the strip, but also affect the feedback control of the shape. By constructing a shape setting model to accurately set the shape, it can not only effectively improve the stability of the entire rolling process, but also improve the shape quality of the strip, thereby ensuring a high yield rate and increasing economic benefits.
[0003] The twenty-high Sendzimir mill has a complex structure, a large number of rolls, and a variety of plate shape adjustment methods, which results in many factors that control the accuracy of its plate shape and makes it very complex, making the plate shape control extremely nonlinear, strongly coupled, and multivariable. The plate shape setting empirical regression model currently used has some irreparable defects, and has little room for improving the accuracy of plate shape setting. Artificial neural networks have good self-learning capabilities and can self-adjust weights and biases through training data to optimize model performance without manually defining complex rules. At the same time, neural networks can achieve complex nonlinear mapping through activation functions, which makes them perform well in dealing with complex nonlinear problems.
[0004] Therefore, using neural networks for the shape setting of the twenty-high mill to improve the shape quality of the strip is an issue that needs to be urgently addressed in the industry. Summary of the invention
[0005] In view of this, the object of the present invention is to provide a data-driven single-stand 20-roll mill adjustable side support unit (ASU) pre-setting method, based on different modeling data levels, using deep neural networks, respectively, for each modeling data level to establish a corresponding single-stand 20-roll mill ASU pre-setting method, so as to achieve the precise setting of the ASU adjustment ratio of different rolling passes when the 20-roll mill rolls strips of different specifications.
[0006] In order to achieve the above object, the present invention provides the following technical solutions:
[0007] A data-driven single-stand 20-high mill ASU presetting method, comprising:
[0008] Establish several different modeling data layers according to the steel type, strip width, strip thickness and strip rolling passes;
[0009] A single-stand 20-roller mill ASU preset model based on a deep neural network is constructed for each modeling data layer, wherein the input variables of the preset model are set to process parameters, equipment parameters and rolled product parameters that affect the single-stand 20-roller cold-rolled strip shape, and the output variables of the preset model are set to the adjustment ratio of the ASU at different positions on the segmented support roller;
[0010] Collect and extract relevant actual production data from the single-stand 20-roller cold rolling site to form the original modeling data of each modeling data layer, and then obtain the final modeling data of each modeling data layer through data preprocessing;
[0011] The final modeling data in each modeling data layer is randomly divided into training data set and test data set according to a certain ratio;
[0012] The training data set of each modeling data layer is used to train the pre-set model corresponding to each modeling data layer, and the optimal hyperparameters of the pre-set model are determined by using the Bayesian method;
[0013] The test data set of each modeling data layer is used to test the preset models corresponding to each modeling data layer, and the performance of each preset model is evaluated by the evaluation index;
[0014] During on-site production of a single-stand 20-roller rolling mill, the corresponding modeling data level is searched according to the production process parameters of the strip steel, and the preset model corresponding to the modeling data level is called to calculate the initial adjustment ratio of the ASU.
[0015] Furthermore, the establishment of several different modeling data layers according to the steel type, strip width, strip thickness and strip rolling passes includes:
[0016] The steel grades are divided into different layers according to the steel grade names; the strip widths are divided into different layers according to the strip width ranges; the strip thicknesses are divided into different layers according to the strip thickness ranges; the rolling passes are divided into different layers according to the strip rolling passes;
[0017] Different steel grade layers, strip width layers, strip thickness layers and rolling pass layers are combined to establish several different modeling data layers; each modeling data layer includes a steel grade layer, a strip width layer, a strip thickness layer and a rolling pass layer.
[0018] Furthermore, the input variables include rolling force, left side eccentricity adjustment ratio of the roll system, right side eccentricity adjustment ratio of the roll system, rolling line adjustment ratio, previous intermediate roll shifting position, next intermediate roll shifting position, inlet tension, outlet tension, inlet strip thickness, outlet strip thickness, inlet strip temperature and outlet strip temperature.
[0019] Furthermore, after collecting and extracting relevant actual production data from the single-stand 20-roller cold rolling site, the actual rolling data of each coil of plate recorded in a time series manner is discretized according to a certain time interval Δt, and the discretized data of input and output variables used to construct the single-stand 20-roller rolling mill ASU pre-set model are extracted;
[0020] According to the steel grade layer, strip width layer, strip thickness layer and rolling pass layer to which each group of input and output variable data belongs, the input and output variable data are classified into the corresponding modeling data layer to form the original modeling data of each modeling data layer;
[0021] Eliminate the data samples belonging to the non-stable uniform rolling stage in each modeling data layer; Eliminate the data samples with missing values in each modeling data layer, and then determine and eliminate the data samples with abnormal data points in the input and output variables based on the Laida criterion;
[0022] The values of the input and output variables in each modeling data layer are normalized to the range of 0 to 1 to obtain the final modeling data of each modeling data layer.
[0023] Furthermore, the using of the training data set of each modeling data level to train the preset model corresponding to each modeling data level includes:
[0024] Obtain a training data set of any modeling data layer, and divide the training data set into two mutually exclusive parts, one part as training data of a preset model, and the other part as verification data of the preset model;
[0025] Determine the hyperparameters of the deep neural network, and traverse the deep neural network with the determined hyperparameters on the training data l times to complete l model training;
[0026] After each traversal, the prediction accuracy index ε of the preset model is obtained through the verification data.
[0027] Furthermore, the hyperparameters of the deep neural network are optimized by Bayesian optimization method during the training process of the preset model, including:
[0028] Based on the Bayesian optimization method, in the last k traversals of the loop traversal process, the average of the prediction accuracy index ε obtained after the k traversals is used as the validation loss, and the hyperparameters of the deep neural network are optimized within the value range of the hyperparameters;
[0029] The execution of l loop traversals and a Bayesian optimization after the loop traversal is regarded as a round. This is repeated for n rounds to obtain the validation loss under n different hyperparameter combinations. The neural network hyperparameter value corresponding to the minimum validation loss is determined as the best optimization result of the hyperparameter.
[0030] The beneficial effects of the present invention are as follows: the present invention overcomes the limitations of traditional theoretical methods and constructs a data-driven high-precision single-frame 20-roller mill ASU pre-setting model for the complex 20-roller plate shape problem, which can accurately set the ASU adjustment ratio of different specifications of strip steel in different rolling passes, which is of great significance to improving the stability of the rolling process and improving the plate shape.
[0031] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:
[0033] Figure 1 A schematic flow chart of a data-driven single-stand 20-roller mill ASU presetting method provided by an embodiment of the present invention;
[0034] Figure 2 Detailed process for building a pre-set model of a single-stand 20-high rolling mill ASU;
[0035] Figure 3 is the prediction result of the M1 model on the test set;
[0036] Figure 4 is the prediction result of the M2 model on the test set. DETAILED DESCRIPTION
[0037] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0038] In this embodiment, the relevant actual rolling data of a 1400mm single-stand 20-roller cold-rolled silicon steel production line is collected and extracted to establish a data-driven single-stand 20-roller rolling mill ASU pre-set model. The 1400mm single-stand 20-roller cold-rolled silicon steel production line is equipped with a ZR22 20-roller Sendzimir mill, with 7 ASU adjustment positions (ASU1-ASU7) on the segmented support rolls, and mainly performs oriented silicon steel rolling.
[0039] like Figure 1 As shown, based on the above collected data, the specific steps of the data-driven single-stand 20-high rolling mill ASU pre-setting method provided in this embodiment include:
[0040] Step 1: Establish the modeling data layer for constructing the preset model of the 20-high rolling mill ASU;
[0041] 1) According to the different steel grade names, the steel grade layers are divided;
[0042] Specifically, the steel grade layers are divided according to the steel grade name of the oriented silicon steel, as shown in Table 1.
[0043] Table 1 Classification of steel grades
[0044] Name Description Steel-1 CGO Oriented Silicon Steel Steel-2 HiB Oriented Silicon Steel
[0045] 2) According to the different width ranges of the strip steel, the strip steel width is divided into different layers. The results of the strip steel width division are shown in Table 2.
[0046] Table 2 Strip width layer classification
[0047] Name Width Range / mm Name Width Range / mm Width-1 850≤w＜900 Width-5 1050≤w＜1100 Width-2 900≤w＜950 Width-6 1100≤w＜1150 Width-3 950≤w＜1000 Width-7 1150≤w＜1200 Width-4 1000≤w＜1050 Width-8 1200≤w＜1250
[0048] 3) According to the different thickness ranges of the strip steel, the strip steel thickness is divided into different layers. The results of the strip steel thickness division are shown in Table 3.
[0049] Table 3 Strip steel thickness layer classification
[0050] Name Thickness Range / mm Name Thickness Range / mm Thick-1 0.18≤h＜0.20 Thick-4 0.24≤h＜0.26 Thick-2 0.20≤h＜0.22 Thick-5 0.26≤h＜0.28 Thick-3 0.22≤h＜0.24 Thick-6 0.28≤h＜0.30
[0051] 4) According to the different rolling passes of the strip steel, the rolling pass levels are divided into different levels. The results of the rolling pass level division are shown in Table 4.
[0052] Table 1 Rules for dividing rolling pass levels
[0053] Name Description Name Description Pass-1 The 1st Rolling Pass Pass-4 The 4th Rolling Pass Pass-2 The 2nd Rolling Pass Pass-5 The 5th Rolling Pass Pass-3 The 3rd Rolling Pass Pass-6 The 6th Rolling Pass
[0054] 5) By combining the steel grade layer, strip width layer, strip thickness layer and rolling pass layer, a modeling data layer for constructing the 20-high mill ASU preset model is established, each modeling data layer including a steel grade layer, a strip width layer, a strip thickness layer and a rolling pass layer.
[0055] According to the specific contents of the steel type layer, strip width layer, strip thickness layer and rolling pass layer shown in Tables 1 to 4, there are 576 types of modeling data layers. The following Table 5 shows some of the 576 types of modeling data layers. The remaining modeling data layers can refer to the modeling data layer format shown in Table 5, which can be obtained by combining the steel type layer, strip width layer, strip thickness layer and rolling pass layer.
[0056] Table 5: Some modeling data layers
[0057]
[0058]
[0059] Step 2: The process parameters, equipment parameters and workpiece parameters that affect the shape of the single-stand 20-roller cold-rolled strip are used as input variables of the ASU pre-setting model of the single-stand 20-roller mill, and the adjustment ratio of the ASU at different positions on the segmented support roll is used as the output variable of the ASU pre-setting model of the single-stand 20-roller mill;
[0060] The input variables include: (1) rolling force, (2) roller system left eccentricity adjustment ratio, (3) roller system right eccentricity adjustment ratio, (4) rolling line adjustment ratio, (4) previous intermediate roller shifting position, (6) next intermediate roller shifting position, (7) inlet tension, (8) outlet tension, (9) inlet strip thickness, (10) outlet strip thickness, (11) inlet strip temperature, (12) outlet strip temperature;
[0061] The output variable is the ASU adjustment ratio of 7 different positions ASU1 to ASU7 on the segmented support roller.
[0062] Step 3: According to the constructed modeling data layers and the input and output variables of the model, relevant actual production data are collected and extracted from the single-stand 20-roller cold rolling site to form the original modeling data of each modeling data layer, and the final modeling data of each modeling data layer is obtained through data preprocessing;
[0063] 1) According to a certain time interval Δt, the actual rolling data of each coil of plate recorded in a time series manner is discretized to extract the discretized data of the input and output variables used to construct the preset model of the 20-high rolling mill ASU. In this embodiment, Δt=0.5s.
[0064] 2) According to the steel grade layer, strip width layer, strip thickness layer and rolling pass layer to which each group of extracted input and output variable data belongs, it is divided into the corresponding modeling data layer to form the original modeling data of each modeling data layer; Table 6 shows the original modeling data of the modeling data layer in this embodiment.
[0065] Table 6 Modeling data layers that form the original modeling data
[0066]
[0067]
[0068] 3) Eliminate the data samples belonging to the non-stable uniform rolling stage in each modeling data layer;
[0069] Specifically, for the data shown in Table 6, the data samples belonging to the non-steady uniform rolling stage in the modeling data layers data-167 and data-261 are eliminated.
[0070] 4) Eliminate data samples with missing values in each modeling data layer, and then determine and eliminate data samples with abnormal data points in the input and output variables based on the Laida criterion, where the Laida criterion discriminant is:
[0071]
[0072] in, σ V are the maximum value, mean value and variance of variable V respectively.
[0073] Based on formula (1), the abnormal value of variable V is judged, that is, if the maximum value of variable V If equation (1) is satisfied, is the outlier value of variable V; otherwise, there is no outlier value in variable V, and we continue to judge whether there are outliers in other variables.
[0074] In the specific implementation, the data samples containing missing values in the modeling data layers data-167 and data-261 shown in Table 6 are eliminated, and then based on the Laida criterion, the data samples containing abnormal data points in the input and output variables in the modeling data layers data-167 and data-261 are respectively determined and eliminated.
[0075] 5) Based on the [0, 1] data normalization method, the values of the input and output variables in each modeling data layer are normalized to the interval of 0 to 1. The [0, 1] data normalization formula is:
[0076]
[0077] in, is the normalized result of the variable V of the i-th data sample; is the value of the variable V of the ith data sample; max(x V ) is the maximum value of the variable V; min(x V ) is the minimum value of the variable V; N is the number of data samples in each modeling data layer; Q is the total number of input and output variables.
[0078] In this embodiment, the N values in the modeling data layers data-167 and data-261 shown in Table 6 are 15371 and 5388 respectively, and Q=19.
[0079] Step 4: Randomly divide the final modeling data in each modeling data layer into training data set and test data set according to a certain ratio;
[0080] In the specific implementation, the final modeling data in the modeling data layers data-167 and data-261 shown in Table 6 are randomly divided into a training data set and a test data set in a ratio of 17:3.
[0081] Step 5: Based on the deep neural network, a corresponding data-driven single-stand 20-high mill ASU preset model is established for each modeling data layer. The preset model of each modeling data layer is trained through the training data set of the modeling data layer, and the Bayesian method is used to determine the optimal hyperparameters of the model. The specific process is as follows: Figure 2 As shown;
[0082] In this embodiment, based on the training data sets of the modeling data layers data-167 and data-261 shown in Table 6, a deep neural network is used to establish corresponding data-driven single-stand 20-roller mill ASU preset models M1 and M2 respectively, and the Bayesian method is used to determine the optimal hyperparameters of the model.
[0083] 1) Read the training data in a modeling data layer;
[0084] You can first read the training data in the data layer data-167, and after the M1 model is built, read the training data in the data layer data-261 to build the M2 model.
[0085] 2) Define the value range of the deep neural network hyperparameters and initialize the deep neural network hyperparameters within the hyperparameter value range;
[0086] Among them, the hyperparameters that need to be optimized for deep neural networks include the learning rate R learn , the ratio of neurons discarded during each training process R drop 、Number of hidden layers layers , the number of neurons in the hidden layer N neurons , the amount of data Batch_Size for each batch input to the neural network training, the activation function type Activation, the gradient descent optimization algorithm type Optimizer, and the value ranges of each hyperparameter are shown in Table 7.
[0087] Table 7. Limit intervals for deep neural network hyperparameter optimization
[0088] Serial Number Hyperparameter Symbol Optimization Interval Type 1 Learning Rate <![CDATA[R learn ]]> <![CDATA[[10 -5 ,10 -2 ]]]> Real Number 2 Dropout Ratio <![CDATA[R drop ]]> <![CDATA[[10 -6 ,10 -3 ]]]> Real Number 3 Number of Hidden Layers <![CDATA[N layers ]]> [2,20] Integer 4 Number of Neurons in Hidden Layers <![CDATA[N neurons > [15,250] Integer 5 Batch Size Batch_size [20,100] Integer 6 Activation Function Activation ["ReLu","Tanh","Sigmoid"] Strings 7 Optimization Algorithm Optermizer ["Adam","SGD","RMSprop"] Strings
[0089] 3) Perform deep neural network model training based on the determined neural network hyperparameters and training data;
[0090] ① Divide the training data set into two mutually exclusive parts, one for training data and the other for verification data;
[0091] Specifically, 85% of the data can be used for training and 15% of the data can be used for validation;
[0092] ② Based on the determined deep neural network hyperparameters, execute l Epoch cycles, that is, traverse l times on the divided training data and complete l model training. After each model training, the prediction accuracy index ε on the validation data will be obtained. The prediction accuracy index ε is described by the mean of the mean square error MSE of all ASU prediction results:
[0093]
[0094] Among them, MSE j is the mean square error of the prediction result of the jth ASU adjustment ratio; t is the number of ASUs; is the predicted value of the adjustment ratio of the jth ASU; y ji is the target value of the j-th ASU adjustment ratio; m is the number of verification data samples.
[0095] In the specific implementation, l=100; the total number m of strip steel samples in the verification data in the modeling data layers data-167 and data-261 shown in Table 6 are 1960 and 687 respectively; t=7.
[0096] 4) Based on the Bayesian optimization method, the mean of the prediction accuracy index ε on the validation data obtained by performing the last k model trainings (executing the last k Epoch cycles) in 3) is To verify the loss, the deep neural network hyperparameters are optimized within the range of hyperparameter values, and the cyclic training in 3) is repeated until the hyperparameter optimization process based on the Bayesian method is completed; where k = 10.
[0097] 5) By repeating steps 3) and 4) n times, that is, executing the hyperparameter optimization process based on the Bayesian method n times, the validation losses under n different hyperparameter combinations are obtained, and the neural network hyperparameter value corresponding to the minimum validation loss is determined as the optimal optimization result of the hyperparameter.
[0098] In the specific implementation, n=10. The hyperparameter optimization results of the M1 and M2 models are shown in Tables 8 and 9, respectively. After the M1 and M2 models repeat the hyperparameter optimization process based on the Bayesian method for the 7th and 10th times, the optimal hyperparameters are determined, and the validation loss corresponding to the optimal hyperparameters is All are minimum values.
[0099] Table 8 Optimization results of M1 model hyperparameters
[0100]
[0101] Table 9 Optimization results of M2 model hyperparameters
[0102]
[0103]
[0104] 6) Based on the training data and the optimal neural network hyperparameters, a single-stand 20-roller mill ASU preset model driven by the corresponding modeling data layer is established.
[0105] For each modeling data level, a corresponding data-driven single-stand 20-roller mill ASU preset model is established.
[0106] In this embodiment, based on the modeling data layers data-167 and data-261, corresponding data-driven single-stand 20-roller mill ASU preset models M1 and M2 are established respectively.
[0107] Step 6: Based on the test data set of each modeling data layer, the performance of each single-stand 20-roll mill ASU pre-set model established using the deep neural network is evaluated by evaluation indicators. The evaluation indicators include: the mean of the mean square error MSE of all ASU prediction results, the mean of the mean absolute error (MAE), and the mean of the coefficient of determination (expressed as R2).
[0108] The mean expression of the mean square error MSE of all ASU prediction results is:
[0109]
[0110] The mean expression of the mean absolute error MAE of all ASU prediction results is:
[0111]
[0112] The coefficient of determination R for all ASU prediction results 2 The mean expression of is:
[0113]
[0114] Where, t is the number of ASUs; is the average value of the j-th ASU adjustment ratio target value; is the predicted value of the adjustment ratio of the jth ASU; y ji is the target value of the j-th ASU adjustment ratio; M is the number of test samples.
[0115] In the specific implementation, t=7; the number of test samples M in the modeling data layers data-167 and data-261 are 2306 and 808 respectively.
[0116] Figure 3 and Figure 4 They are the prediction results of model M1 and model M2 on the corresponding test data sets. Table 10 shows the indicators MMSE, MMAE and MR 2 The calculation result of .
[0117] Table 10 Index calculation results
[0118]
[0119]
[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.
Claims
1. A data-driven single-stand 20-high mill ASU presetting method, characterized in that: The method includes: Establish several different modeling data layers according to the steel type, strip width, strip thickness and strip rolling passes; Constructing a single-stand 20-roller rolling mill ASU preset model based on a deep neural network for each of the modeling data layers, wherein the input variables of the preset model are set to process parameters, equipment parameters and rolled product parameters that affect the single-stand 20-roller cold-rolled strip shape, and the output variables of the preset model are set to the adjustment ratios of the ASUs at different positions on the segmented support rolls; Collect and extract relevant actual production data from the single-stand 20-roller cold rolling site to form the original modeling data of each modeling data layer, and then obtain the final modeling data of each modeling data layer through data preprocessing; The final modeling data in each modeling data layer is randomly divided into training data set and test data set according to a certain ratio; The training data set of each modeling data layer is used to train the pre-set model corresponding to each modeling data layer, and the optimal hyperparameters of the pre-set model are determined by using the Bayesian method; The test data set of each modeling data layer is used to test the preset models corresponding to each modeling data layer, and the performance of each preset model is evaluated by the evaluation index; During on-site production of a single-stand 20-roller rolling mill, the corresponding modeling data level is searched according to the production process parameters of the strip steel, and the preset model corresponding to the modeling data level is called to calculate the initial adjustment ratio of the ASU.
2. The method according to claim 1, characterized in that According to the different steel types, strip widths, strip thicknesses and strip rolling passes, several different modeling data layers are established, including: The steel grades are divided into different layers according to the steel grade names; the strip widths are divided into different layers according to the strip width ranges; the strip thicknesses are divided into different layers according to the strip thickness ranges; the rolling passes are divided into different layers according to the strip rolling passes; Different steel grade layers, strip width layers, strip thickness layers and rolling pass layers are combined to establish a number of different modeling data layers; each of the modeling data layers includes a steel grade layer, a strip width layer, a strip thickness layer and a rolling pass layer.
3. The method according to claim 1, characterized in that The input variables include rolling force, left side eccentricity adjustment ratio of the roll system, right side eccentricity adjustment ratio of the roll system, rolling line adjustment ratio, previous intermediate roll shifting position, next intermediate roll shifting position, inlet tension, outlet tension, inlet strip thickness, outlet strip thickness, inlet strip temperature and outlet strip temperature.
4. The method according to claim 1, characterized in that After collecting and extracting relevant actual production data from the single-stand 20-roller cold rolling site, the actual rolling data of each coil of plate recorded in a time series manner is discretized at a certain time interval Δt, and the discretized data of input and output variables used to construct the single-stand 20-roller rolling mill ASU pre-set model are extracted; According to the steel grade layer, strip width layer, strip thickness layer and rolling pass layer to which each group of input and output variable data belongs, the input and output variable data are classified into the corresponding modeling data layer to form the original modeling data of each modeling data layer; Eliminate the data samples belonging to the non-stable uniform rolling stage in each modeling data layer; Eliminate the data samples with missing values in each modeling data layer, and then determine and eliminate the data samples with abnormal data points in the input and output variables based on the Laida criterion; The values of the input and output variables in each modeling data layer are normalized to the range of 0 to 1 to obtain the final modeling data of each modeling data layer.
5. The method according to claim 1, characterized in that The training of the preset models corresponding to each modeling data level using the training data set of each modeling data level includes: Obtain a training data set of any modeling data layer, and divide the training data set into two mutually exclusive parts, one part as training data of a preset model, and the other part as verification data of the preset model; Determine the hyperparameters of the deep neural network, and traverse the deep neural network with the determined hyperparameters on the training data l times to complete l model training; After each traversal, the prediction accuracy index ε of the preset model is obtained through the verification data.
6. The method according to claim 5, characterized in that During the training process of the pre-set model, the hyperparameters of the deep neural network are optimized through the Bayesian optimization method, including: Based on the Bayesian optimization method, in the last k traversals of the loop traversal process, the average of the prediction accuracy index ε obtained after the k traversals is used as the validation loss, and the hyperparameters of the deep neural network are optimized within the value range of the hyperparameters; The execution of l loop traversals and a Bayesian optimization after the loop traversal is regarded as a round. This is repeated for n rounds to obtain the validation loss under n different hyperparameter combinations. The neural network hyperparameter value corresponding to the minimum validation loss is determined as the best optimization result of the hyperparameter.