A method for correcting the first pass roll gap zero point of a heavy plate finishing mill

Abstract

The application discloses a method for correcting the first-pass roll gap zero point of a heavy plate finishing mill, which comprises the following steps: S1, variable selection, selecting variables related to the correction of the first-pass roll gap zero point of the heavy plate finishing mill from historical data; S2, performing data cleaning and feature extraction on the variables selected in the step S1; S3, using a random configuration network algorithm to establish a model for correcting the first-pass roll gap zero point of the heavy plate finishing mill; and S4, using the model for correcting the first-pass roll gap zero point of the heavy plate finishing mill established in the step S3 to set the first-pass roll gap zero point correction value. The application is based on a large number of effective historical data related to the setting of the roll gap zero point correction value, and uses the random configuration network algorithm to correct the first-pass roll gap zero point of the heavy plate finishing mill after the gauge change in the finishing process of the heavy plate finishing mill, so that the thickness accuracy of the slab is improved.

Description

Technical Field

[0001] This invention belongs to the field of roll gap correction technology, and relates to a method for zero-point correction of the first roll gap in a thick plate finishing mill. Background Technology

[0002] Thick plates are steel plates with a thickness of 20 to 60 mm in plate and strip steel products. Their production process mainly includes several stages such as slab selection, slab heating, rolling, post-rolling accelerated cooling, hot straightening, cooling bed cooling, shearing, and cold straightening.

[0003] Thickness accuracy is a crucial indicator of the quality of thick plate products. Throughout the entire thick plate production process, finish rolling has the most direct and significant impact on the thickness accuracy of the finished product. Thick plate production is characterized by small batches, multiple production runs, and frequent specification changes. To ensure the thickness accuracy of thick plate products, after a specification change occurs in the continuously rolled slab, the zero-point value of the roll gap in the first finish rolling pass needs to be adjusted to ensure stable subsequent rolling production.

[0004] However, in the current thick plate finishing rolling process, the zero-point correction value of the roll gap in the first finishing pass after specification change is mainly set by on-site operators. This manual setting method relies on the operator's experience and level of attention during work. On the one hand, technologies such as industrial big data and artificial intelligence have been widely applied to various aspects of steel production, such as hot metal composition prediction, equipment fault diagnosis, and rolling parameter optimization; on the other hand, the product order types of thick plate production lines have become relatively stable, and a large amount of effective information and knowledge about the setting of roll gap zero-point correction values ​​has been accumulated. Therefore, extracting good roll gap zero-point correction data, analyzing and studying it using big data and artificial intelligence methods, and automatically adjusting the optimal roll gap zero-point correction value after specification change is of great significance for improving the thickness accuracy of rolled slabs. Summary of the Invention

[0005] In view of the above-mentioned defects in the prior art, the purpose of this invention is to provide a method for zero-point correction of the first roll gap in a thick plate finishing mill. Based on a large amount of valid historical data on the setting of roll gap zero-point correction values, and using a random configuration network algorithm, the zero point of the first roll gap in the finishing mill after a change in specifications occurs during the finishing process of the thick plate mill is corrected, so as to improve the thickness accuracy of the slab.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention provides a method for zero-point correction of the first roll gap in a thick plate finishing mill, comprising the following steps:

[0008] S1, Variable selection: Select variables related to the zero-point correction of the first pass roll gap of the heavy plate finishing mill from historical data.

[0009] S2, perform data cleaning and feature extraction on the variables selected in step S1;

[0010] S3. Use the randomized network algorithm to establish a zero-point correction model for the first roll gap of the thick plate finishing mill.

[0011] S4. The zero-point correction value of the first roll gap of the finishing mill is set using the zero-point correction model of the first roll gap established in step S3.

[0012] Preferably, in step S1, the variables related to the zero-point correction of the first roll gap of the thick plate finishing mill include the steel grade, target thickness, and zero-point correction value of the first roll gap of the finishing mill before the specification change, as well as the steel grade, target width, and target thickness of the slab after the specification change.

[0013] Preferably, the process in step S2 is as follows:

[0014] S21, divide the variables selected in step S1 into categorical variables and continuous variables;

[0015] S22, Represent the steel type in the classification variables using one-hot encoding;

[0016] S23, normalize the continuous variable, and the normalization formula is as follows:

[0017]

[0018] In the formula, For the i-th normalized data in a continuous variable;

[0019] x i Let i be the i-th original data in a certain continuous variable;

[0020] x min The minimum value of the original data in a continuous variable;

[0021] x max It represents the maximum value of the original data in a continuous variable.

[0022] Preferably, the process in step S3 is as follows:

[0023] S31, take the zero-point correction value of the first roll gap of the finishing mill after the slab is changed in size as the output variable y, and take the remaining variable after the processing of step S2 as the input variable x. The input variable x and the output variable y form the original sample set D.

[0024] S32, the original sample set D is divided into training set D1 and validation set D2 in a 4:1 ratio;

[0025] S33, the training set D1 is trained using the random configuration network algorithm, and the trained model is verified using the validation set D2 to establish the zero-point correction model for the first pass roll gap of the thick plate finishing mill.

[0026] Preferably, the specific process of step S33 is as follows:

[0027] S331, Network Initialization: Number of hidden layer nodes L, Number of random parameter configurations T max e0 represents the preset allowable error, and the range for generating random parameters. Parameter r;

[0028] S332, in range Internally randomly generated parameter w l and b l ;

[0029] S333, select nodes that satisfy the following formula;

[0030]

[0031] In the formula, e L-1 This represents the residual of the hidden layer of the SCN network after the (L-1)th addition of a node;

[0032] G L This represents the value of the Lth hidden node in the SCN network;

[0033] μ L Produced by parameter r, μ L = (1-r) / (L+1).

[0034] S334, select the node with ζ>0 in step S333 as the candidate node, and select the node corresponding to the maximum value of ζ as the new node;

[0035] S335, calculate the output weight β of the current SCN model using the following formula;

[0036]

[0037] In the formula, f L-1 (x n ;β) represents the objective function value after adding a node for the (L-1)th time;

[0038] yes The matrix representation of .

[0039] S336, Calculate the training error and repeat steps S332 to S335 until the preset termination condition is met;

[0040] S337, The trained model is validated using the validation set D2. If the termination condition preset in step S336 is met, the training is successful; otherwise, steps S332 to S336 are repeated.

[0041] S338, obtain the SCN output weight β and the SCN model with L-1 (L=1,2,3,…) nodes, which is the zero-point correction model for the first pass roll gap of the finishing mill.

[0042] Preferably, in step S336, the preset termination condition is: e L Less than a pre-defined error criterion, where e L This represents the residual of the hidden layer of the SCN network after the Lth addition of a node;

[0043] Preferably, in step S338, the SCN model with L-1 (L = 1, 2, 3, ...) nodes is as follows:

[0044]

[0045] In the formula, β l This represents the weight parameters between the Lth hidden node and the output layer, and is the output weight.

[0046] g is the activation function.

[0047] The zero-point correction method for the first roll gap of a thick plate finishing mill provided by this invention has the following beneficial effects:

[0048] 1. The zero-point correction method for the first roll gap of the thick plate finishing mill of the present invention can replace the operator in performing zero-point correction of the first roll gap after changing specifications. It can make timely adjustments according to the changes in the on-site process conditions and the changes in the working status of the equipment, thereby ensuring the thickness accuracy of the slab.

[0049] 2. The zero-point correction method for the first pass roll gap of the thick plate finishing mill of the present invention can improve the level of intelligence of roll gap zero-point correction during finishing rolling, reduce the labor intensity of operators and the defect rate. Attached Figure Description

[0050] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0051] Figure 1 This is a schematic flowchart of the zero-point correction method for the first pass roll gap in a thick plate finishing mill according to the present invention.

[0052] Figure 2 This is an illustration of the effect of the zero-point correction method for the first pass roll gap of the thick plate finishing mill of the present invention. Detailed Implementation

[0053] To better understand the above-mentioned technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with embodiments.

[0054] Combination Figure 1 As shown, the method for zero-point correction of the first roll gap in a thick plate finishing mill of the present invention includes the following steps:

[0055] S1, Variable selection: Select variables related to the zero-point correction of the first pass roll gap of the heavy plate finishing mill from historical data.

[0056] Specifically, for continuously rolled slabs, a change in specifications is considered to have occurred, i.e., a change in specifications, when any of the following situations occur:

[0057] (1) The steel grade of the slab has changed;

[0058] (2) The difference in target width between the slabs is greater than 0.5m;

[0059] (3) The difference in target thickness of the slab is greater than 2 mm.

[0060] Therefore, when selecting variables, we select thick plate production data including multiple steel grades from historical data of a certain period. Based on process knowledge, we select variables related to the zero-point correction of the first roll gap of the thick plate finishing mill. Among them, the variables related to the zero-point correction of the first roll gap of the thick plate finishing mill include the steel grade, target thickness, and zero-point correction value of the first roll gap of the slab before the specification change, as well as the steel grade, target width, and target thickness of the slab after the specification change.

[0061] S2, perform data cleaning and feature extraction on the variables selected in step S1;

[0062] Specifically, the data cleaning and feature extraction processes are as follows:

[0063] S21, divide the variables selected in step S1 into categorical variables and continuous variables;

[0064] S22, represent the steel type in the categorical variable using one-hot encoding;

[0065] S23, normalize the continuous variable, and the normalization formula is as follows:

[0066]

[0067] In the formula, For the i-th normalized data in a continuous variable;

[0068] x i Let i be the i-th original data in a certain continuous variable;

[0069] xmin The minimum value of the original data in a continuous variable;

[0070] x max It represents the maximum value of the original data in a continuous variable.

[0071] S3. Use the randomized network algorithm to establish a zero-point correction model for the first roll gap of the thick plate finishing mill.

[0072] The specific procedures in step S3 are as follows:

[0073] S31, take the zero-point correction value of the first roll gap of the finishing mill after the slab is changed in size as the output variable y, and take the remaining variable after the processing of step S2 as the input variable x. The input variable x and the output variable y form the original sample set D.

[0074] S32, the original sample set D is divided into training set D1 and validation set D2 in a 4:1 ratio;

[0075] S33, the training set D1 is trained using the random configuration network algorithm, and the trained model is verified using the validation set D2 to establish the zero-point correction model for the first pass roll gap of the thick plate finishing mill.

[0076] In specific implementation, the process of establishing the zero-point correction model for the first pass roll gap of the heavy plate finishing mill is as follows:

[0077] S331, Network Initialization: Number of hidden layer nodes L, Number of random parameter configurations T max e0 represents the preset allowable error, and the range for generating random parameters. Parameter r;

[0078] S332, in range Internally randomly generated parameter w l and b l ;

[0079] S333, select nodes that satisfy the following formula;

[0080]

[0081] In the formula, e L-1 This represents the residual of the hidden layer of the SCN network after the (L-1)th addition of a node;

[0082] G L This represents the value of the Lth hidden node in the SCN network;

[0083] μ L Produced by parameter r, μ L = (1-r) / (L+1).

[0084] S334, select the node with ζ>0 in step S333 as the candidate node, and select the node corresponding to the maximum value of ζ as the new node;

[0085] S335, calculate the output weight β of the current SCN model using the following formula;

[0086]

[0087] In the formula, f L-1 (x n ;β) represents the objective function value after adding a node for the (L-1)th time;

[0088] yes The matrix representation of .

[0089] S336, Calculate the training error and repeat steps S332 to S335 until a preset termination condition is met; wherein the preset termination condition is: e L Less than a pre-defined error criterion, where e L This represents the residual of the hidden layer of the SCN network after the Lth addition of a node.

[0090] S337, The trained model is validated using the validation set D2. If the termination condition preset in step S336 is met, the training is successful; otherwise, steps S332 to S336 are repeated.

[0091] S338, obtain the SCN output weight β and the SCN model with L-1 (L=1,2,3,…) nodes, which is the zero-point correction model for the first pass roll gap of the finishing mill.

[0092] The SCN model with L-1 (L = 1, 2, 3, ...) nodes is as follows:

[0093]

[0094] In the formula, β l This represents the weight parameters between the Lth hidden node and the output layer, and is the output weight.

[0095] g is the activation function (i.e., the Gaussian radial basis function);

[0096] S4. The zero-point correction value of the first roll gap of the finishing mill is set using the zero-point correction model of the first roll gap established in step S3.

[0097] The following section provides a further description of the zero-point correction method for the first pass roll gap of a thick plate finishing mill, using specific examples.

[0098] Example 1

[0099] The zero-point correction method for the first pass roll gap of a heavy plate mill was validated using historical production data from a heavy plate mill. Multiple heavy plates were samples from April and May 2019, representing various steel grades. 1000 plates were selected as the initial sample points. The training, validation, and test sets were divided in an 8:1:1 ratio.

[0100] Step 1: Variable Selection. Based on process knowledge, select variables related to the zero-point correction of the first pass roll gap in the heavy plate finishing mill at the sample points. These variables include the steel grade, target thickness, and zero-point correction value of the first pass roll gap for the slab before specification change, as well as the steel grade and target width of the slab after specification change. Finally, select 10 variables.

[0101] Step 2: Perform data cleaning and feature extraction on the variables selected in Step 1.

[0102] S21: Divide the variables selected in step 1 into categorical variables and continuous variables. Categorical variables include the steel type of the slab before specification change, and continuous variables include the target thickness and target width.

[0103] S22: Represent the categorical variables obtained in S21 using one-hot encoding.

[0104] S23: Calculate the minimum and maximum values ​​of the continuous variables obtained in S21, and normalize the continuous variables.

[0105] Step 3: Use the randomized network algorithm to establish a zero-point correction model for the first roll gap of the thick plate finishing mill based on the variables processed in Step 2.

[0106] S31: Take the zero-point correction value of the first roll gap after the slab is changed in size as the output variable y, and take the remaining 9 variables after the processing in step 2 as the input variable x. The input variable x and the output variable y form the original sample set D.

[0107] S32: Divide the data into three sets: 800 sample points as the training set D1, 100 sample points as the validation set D2, and 100 sample points as the test set D3.

[0108] S33: The training set D1 is trained using a randomized network algorithm, and the trained model is validated using the validation set D2 to establish a zero-point correction model for the first pass roll gap of a thick plate finishing mill.

[0109] In specific implementation, the process of establishing the zero-point correction model for the first pass roll gap of the heavy plate finishing mill is as follows:

[0110] S331: Network initialization: Number of hidden layer nodes L is 8, number of random parameter configurations T max The value is 30, the preset allowable error e0 is 0.05, and the random parameter generation range is... The range is [-5,5], j = 1,...,9, and the parameter r is 0.1.

[0111] S332: Randomly generate parameter w within the range [-5,5], j = 1,...,9. l and b l .

[0112] S333: Select nodes that satisfy the following formula;

[0113]

[0114] In the formula, e L-1 This represents the residual of the hidden layer of the SCN network after the (L-1)th addition of a node;

[0115] G L This represents the value of the Lth hidden node in the SCN network;

[0116] μ L Produced by parameter r, μ L = (1-r) / (L+1).

[0117] S334: Select the node from step S333 as a candidate node, and select the node corresponding to the maximum value as a new node.

[0118] S335: Calculate the output weights of the current model using the following formula;

[0119]

[0120] In the formula, f L-1 (x n ;β) represents the objective function value after adding a node for the (L-1)th time;

[0121] yes The matrix representation of .

[0122] S336: Calculate the training error and repeat steps S332 to S335 until a preset termination condition is met; wherein the preset termination condition is: e L Less than a pre-given error criterion e0, where e L This represents the residual of the hidden layer of the SCN network after the Lth addition of a node.

[0123] S337: The trained model is validated using the validation set D2. If the termination condition preset in step S336 is met, the training is successful; otherwise, steps S332 to S336 are repeated.

[0124] S338: Obtain the output weights and a model with 9 nodes, namely the zero-point correction model for the first pass roll gap of the finishing mill.

[0125]

[0126] In the formula, β l This represents the weight parameters between the Lth hidden node and the output layer, and is the output weight.

[0127] g is a Gaussian radial basis function.

[0128] Step 4: Use the zero-point correction model for the first roll gap of the thick plate finishing mill established in Step 3 to set the zero-point correction value for the first roll gap of the finishing mill.

[0129] Using the above method, the zero-point correction model for the first pass roll gap of the finishing mill was tested using 100 sample points from test set D3. The model was compared with the corresponding manually set values ​​and plotted. Figure 2 The renderings in the image; by Figure 2 It can be seen that the zero-point correction method for the first roll gap of the thick plate finishing mill of the present invention can effectively provide the zero-point correction value for the first roll gap of the finishing mill.

[0130] Example 2

[0131] The zero-point correction method for the first pass roll gap of a heavy plate mill was validated using historical production data from a certain heavy plate mill. Multiple heavy plates were samples from three months, from May 2021 to July 2019, including heavy plates of various steel grades. 2000 plates were extracted as the original sample points. The training set, validation set, and test set were divided in an 8:1:1 ratio.

[0132] Step 1: Variable Selection. Based on process knowledge, select variables related to the zero-point correction of the first roll gap in the heavy plate finishing mill at the sample points. These variables include the steel grade, target thickness, and zero-point correction value of the first roll gap in the slab before specification change, as well as the steel grade, target width, and initial rolling temperature of the slab after specification change. Finally, 14 variables are selected.

[0133] Step 2: Perform data cleaning and feature extraction on the variables selected in Step 1.

[0134] S21: Divide the variables selected in step 1 into categorical variables and continuous variables. Categorical variables include the steel grade of the slab before specification change, the heating furnace number, etc., and continuous variables include the target thickness, target width, etc.

[0135] S22: Represent the categorical variables obtained in S21 using one-hot encoding.

[0136] S23: Calculate the minimum and maximum values ​​of the continuous variables obtained in S21, and normalize the continuous variables.

[0137] Step 3: Use the randomized network algorithm to establish a zero-point correction model for the first roll gap of the thick plate finishing mill based on the variables processed in Step 2.

[0138] S31: Take the zero-point correction value of the first roll gap after the slab is changed in size as the output variable y, and take the remaining 13 variables after the processing in step 2 as the input variable x. The input variable x and the output variable y form the original sample set D.

[0139] S32: Divide the dataset into three sets: 1600 sample points as the training set D1, 200 sample points as the validation set D2, and 200 sample points as the test set D3.

[0140] S33: The training set D1 is trained using a randomized network algorithm, and the trained model is validated using the validation set D2 to establish a zero-point correction model for the first pass roll gap of a thick plate finishing mill.

[0141] In specific implementation, the process of establishing the zero-point correction model for the first pass roll gap of the heavy plate finishing mill is as follows:

[0142] S331: Network initialization: The number of hidden layer nodes L is 14, and the number of random parameter configurations T is... max The value is 80, the preset allowable error e0 is 0.05, and the random parameter generation range is... The range is [-5,5], j = 1,...,14, and the parameter r is 0.1.

[0143] S332: Randomly generate parameter w within the range [-5,5], j = 1,...,14. l and b l .

[0144] S333: Select nodes that satisfy the following formula;

[0145]

[0146] In the formula, e L-1 This represents the residual of the hidden layer of the SCN network after the (L-1)th addition of a node;

[0147] G L This represents the value of the Lth hidden node in the SCN network;

[0148] μ L Produced by parameter r, μL = (1-r) / (L+1).

[0149] S334: Select the node from step S333 as a candidate node, and select the node corresponding to the maximum value as a new node.

[0150] S335: Calculate the output weight β of the current model using the following formula;

[0151]

[0152] In the formula, f L-1 (x n ;β) represents the objective function value after adding a node for the (L-1)th time;

[0153] yes The matrix representation of .

[0154] S336: Calculate the training error and repeat steps S332 to S335 until a preset termination condition is met; wherein the preset termination condition is: e L Less than a pre-given error criterion e0, where e L This represents the residual of the hidden layer of the SCN network after the Lth addition of a node.

[0155] S337: The trained model is validated using the validation set D2. If the termination condition preset in step S336 is met, the training is successful; otherwise, steps S332 to S336 are repeated.

[0156] S338: Obtain the output weights and a model with 12 nodes, namely the zero-point correction model for the first pass roll gap of the finishing mill.

[0157]

[0158] In the formula, β l Let represent the weight parameters between the Lth hidden node and the output layer, and be the output weights. g is the Gaussian radial basis function.

[0159] Step 4: Use the zero-point correction model for the first roll gap of the thick plate finishing mill established in Step 3 to set the zero-point correction value for the first roll gap of the finishing mill.

[0160] Using the above method, the zero-point correction model for the first pass roll gap of the finishing mill obtained above is used to predict the 200 sample points of test set D3. With the corresponding manually set value y i If the mean squared error (MSE) is calculated to be 0.85, which is less than the preset MSE standard of 1.00, then the method is deemed effective.

[0161] Example 3

[0162] The zero-point correction method for the first pass roll gap of a heavy plate mill was validated using historical production data from a certain heavy plate mill. Multiple heavy plates were samples from six months, from October 2021 to March 2022, including heavy plates of various steel grades. 4000 plates were extracted as the original sample points. The training set, validation set, and test set were divided in an 8:1:1 ratio.

[0163] Step 1: Variable Selection. Based on process knowledge, select variables related to the zero-point correction of the first pass roll gap in the heavy plate finishing mill at the sample points. These variables include the steel grade, target thickness, and zero-point correction value of the first pass roll gap for the slab before specification change, as well as the steel grade, target width, and initial rolling temperature of the slab after specification change. Finally, 16 variables are selected.

[0164] Step 2: Perform data cleaning and feature extraction on the variables selected in Step 1.

[0165] S21: Divide the variables selected in step 1 into categorical variables and continuous variables. Categorical variables include the steel grade of the slab before specification change, the heating furnace number, etc., and continuous variables include the target thickness, target width, etc.

[0166] S22: Represent the categorical variables obtained in S21 using one-hot encoding.

[0167] S23: Calculate the minimum and maximum values ​​of the continuous variables obtained in S21, and normalize the continuous variables.

[0168] Step 3: Use the randomized network algorithm to establish a zero-point correction model for the first roll gap of the thick plate finishing mill based on the variables processed in Step 2.

[0169] S31: Take the zero-point correction value of the first roll gap after the slab is changed in size as the output variable y, and take the remaining 14 variables after the processing in step 2 as the input variable x. The input variable x and the output variable y form the original sample set D.

[0170] S32: Divide the dataset into three sets: 3200 sample points for training, 400 sample points for validation, and 400 sample points for test.

[0171] S33: The training set D1 is trained using a randomized network algorithm, and the trained model is validated using the validation set D2 to establish a zero-point correction model for the first pass roll gap of a thick plate finishing mill.

[0172] In specific implementation, the process of establishing the zero-point correction model for the first pass roll gap of the heavy plate finishing mill is as follows:

[0173] S331: Network initialization: The number of hidden layer nodes L is 17, and the number of random parameter configurations T is... max The value is 100, the preset allowable error e0 is 0.05, and the random parameter generation range is... The range is [-5,5], j = 1,...,17, and the parameter r is 0.1.

[0174] S332: Randomly generate parameter w within the range [-5,5], j = 1,...,17. l and b l .

[0175] S333: Select nodes that satisfy the following formula;

[0176]

[0177] In the formula, e L-1 This represents the residual of the hidden layer of the SCN network after the (L-1)th addition of a node;

[0178] G L This represents the value of the Lth hidden node in the SCN network;

[0179] μ L Produced by parameter r, μ L = (1-r) / (L+1).

[0180] S334: Select the node from step S333 as a candidate node, and select the node corresponding to the maximum value as a new node.

[0181] S335: Calculate the output weights of the current model using the following formula;

[0182]

[0183] In the formula, f L-1 (x n ;β) represents the objective function value after adding a node for the (L-1)th time;

[0184] yes The matrix representation of .

[0185] S336: Calculate the training error and repeat steps S332 to S335 until a preset termination condition is met; wherein the preset termination condition is: e L Less than a pre-given error criterion e0, where e L This represents the residual of the hidden layer of the SCN network after the Lth addition of a node.

[0186] S337: The trained model is validated using the validation set D2. If the termination condition preset in step S336 is met, the training is successful; otherwise, steps S332 to S336 are repeated.

[0187] S338: Obtain the output weights and a model with 16 nodes, namely the zero-point correction model for the first pass roll gap of the finishing mill.

[0188]

[0189] In the formula, β l This represents the weight parameters between the Lth hidden node and the output layer, and is the output weight.

[0190] g is a Gaussian radial basis function.

[0191] Step 4: Use the zero-point correction model for the first roll gap of the thick plate finishing mill established in Step 3 to set the zero-point correction value for the first roll gap of the finishing mill.

[0192] Using the above method, the zero-point correction model for the first pass roll gap of the finishing mill obtained above is used to predict the 400 sample points of test set D3. With the corresponding manually set value y i The mean squared error (MSE) was calculated to be 1.01, which is close to the pre-set MSE standard of 1.00, indicating that the method is effective. However, it is noted that the method still needs to be improved for correcting production sample points with a longer time span.

[0193] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the essential spirit of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for zero-point correction of the first roll gap in a thick plate finishing mill, characterized in that, Includes the following steps: S1, Variable Selection: Select variables related to the zero-point correction of the first pass roll gap in the heavy plate finishing mill from historical data. The variables related to the zero-point correction of the first roll gap of the thick plate finishing mill include the steel grade, target thickness, and zero-point correction value of the first roll gap of the slab before the specification change, as well as the steel grade, target width, and target thickness of the slab after the specification change. S2, perform data cleaning and feature extraction on the variables selected in step S1, as follows: S21, divide the variables selected in step S1 into categorical variables and continuous variables; S22, Represent the steel type in the classification variables using one-hot encoding; S23, normalize the continuous variable, and the normalization formula is as follows: ； In the formula, For the i-th normalized data in a continuous variable; Let i be the i-th original data in a certain continuous variable; The minimum value of the original data in a continuous variable; The maximum value of the original data in a continuous variable; S3. Using a randomized network algorithm, a zero-point correction model for the first roll gap of a thick plate finishing mill is established based on the variables processed in step S2. The process is as follows: S31, take the zero-point correction value of the first roll gap of the finishing mill after the slab is changed in size as the output variable y, and take the remaining variable after the processing of step S2 as the input variable x. The input variable x and the output variable y form the original sample set D. S32, the original sample set D is divided into training set D1 and validation set D2 in a 4:1 ratio; S33, the training set D1 is trained using the random configuration network algorithm, and the trained model is verified using the validation set D2 to establish the zero-point correction model for the first pass roll gap of the thick plate finishing mill. S4. The zero-point correction value of the first roll gap of the finishing mill is set using the zero-point correction model of the first roll gap established in step S3.

2. The method for zero-point correction of the first roll gap in a thick plate finishing mill according to claim 1, characterized in that, The specific process of step S33 is as follows: S331, Network Initialization: Number of Hidden Layer Nodes Number of times random parameter configuration , The range of random parameter generation is set to a preset allowable error. , ,parameter ; S332, in range , Internally generated parameters and ; S333, select nodes that satisfy the following formula: ； In the formula, This indicates that the hidden layer of the SCN network is at the 1st... The residual after adding nodes; Indicates the first in the SCN network The value of each hidden node; From parameters produce, ; S334, Select step S333 Select the node as the candidate node. The node corresponding to the maximum value is added as a new node; S335, calculate the current using the following formula Model output weights ; ； In the formula, Indicates the first The objective function value after adding nodes; Matrix representation; S336, Calculate the training error and repeat steps S332 to S335 until the preset termination condition is met; S337. The trained model is validated using the validation set D2. If the termination condition preset in step S336 is met, the training is successful; otherwise, steps S332 to S336 are repeated. S338, obtained Output weights and have Each node The model is the zero-point correction model for the first pass roll gap of the finishing mill.

3. The method for zero-point correction of the first roll gap in a thick plate finishing mill according to claim 2, characterized in that, In step S336, the preset termination condition is: Less than the pre-defined error criterion, where This indicates that the hidden layer of the SCN network is at the 1st... The residual after adding nodes.

4. The method for zero-point correction of the first roll gap in a thick plate finishing mill according to claim 2, characterized in that, In step 338, the one with Each node The model is as follows: ； In the formula, Indicates the first The weight parameters between each hidden node and the output layer are the output weights. This is the activation function.

Images