A method for predicting the efficiency coefficient of flatness control in cold rolling flatness control

Abstract

The present invention proposes a method for predicting the plate shape control efficiency coefficient in cold rolling plate shape control, which belongs to the field of metallurgical rolling technology. The method first calculates the plate shape control efficiency coefficient matrix of different preset working points through offline simulation, uses production data to improve the optimal weight factor between the preset working point and the actual working point, and then uses a data-driven method to establish an optimal weight factor prediction model for the actual working point, thereby accurately obtaining the plate shape control efficiency coefficient under any rolling condition. The optimal weight factor prediction method for the actual working point not only has high accuracy, but also has good stability and fast convergence speed, providing a new method for optimizing the plate shape control efficiency coefficient.

Description

Technical Field

[0001] The invention belongs to the technical field of metallurgical rolling and relates to a method for predicting a plate shape regulation efficiency coefficient in cold-rolled plate shape control. Background Art

[0002] The shape control efficiency coefficient can be regarded as a quantitative description of the influence of the shape adjustment mechanism on the strip shape. Only by obtaining an accurate shape control efficiency coefficient can we accurately calculate the optimal adjustment amount of the shape adjustment mechanism such as bending roll, lateral displacement, and tilting required to eliminate the shape deviation. Therefore, the accuracy of the shape control efficiency coefficient directly affects the accuracy of the shape closed-loop control. How to obtain a high-precision shape control efficiency coefficient online is a fundamental issue in shape control.

[0003] Many researchers at home and abroad have used influence function method, finite element analysis and other methods to analyze the adjustment characteristics of plate shape adjustment mechanism and its influence on plate shape, providing a theoretical basis for the calculation of plate shape control efficiency coefficient. Since the factors that affect the calculation accuracy of plate shape control efficiency coefficient include strip width, strip thickness, rolling force, pressure reduction, etc., the plate shape control efficiency coefficient obtained by traditional methods such as single constant parameter mechanism model or offline simulation model has poor accuracy, lacks quantitative analysis of the transition process between two steady states, and cannot meet the plate shape control accuracy requirements under all working conditions. In order to solve this problem, some scholars have proposed a model-free adaptive optimization based entirely on rolling process data to improve the accuracy of plate shape control efficiency coefficient. However, due to the complex on-site environment and the existence of noise in the collected data, the adaptation without initial model often deviates from the correct optimization direction and fails to achieve the purpose of improving accuracy. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a method for predicting the plate shape control efficiency coefficient in cold-rolled plate shape control, and adopt a hybrid modeling method that combines simulation modeling and data-driven modeling into a series structure to establish an optimal weight factor prediction model for the actual working condition point. First, a three-dimensional finite element simulation model of the strip cold rolling process is established to obtain the plate shape control efficiency coefficient matrix of the preset working condition point, and the production data is used to improve the optimal weight factor between the preset working condition point and the actual working condition point. Then, a data-driven method is used to establish the optimal weight factor prediction model for the actual working condition point, and then the plate shape control efficiency coefficient under any rolling condition is accurately obtained.

[0005] The present invention provides a method for predicting a plate shape control efficiency coefficient in cold rolling plate shape control, comprising:

[0006] Step 1: Set rolling force, strip width and strip thickness as parameters of the working point, and select multiple preset working points in the simulation experiment;

[0007] Step 2: Obtain the plate shape control efficiency coefficient matrix of each preset operating point through offline simulation, and export the measured rolling data from the rolling mill;

[0008] Step 3: Use the mean square error to detect and eliminate abnormal values ​​of the plate shape change in the measured rolling data;

[0009] Step 4: Use Lagrange interpolation method to fill the missing data of the plate shape change after removing the outliers;

[0010] Step 5: Obtain the actual operating points from the measured rolling data in sequence, and calculate the optimal weight factor array between the actual operating point and the preset operating point by using the DBO algorithm;

[0011] Step 6: The parameters of the actual operating point and the optimal weight factor array constitute a data set and divide it into a training set and a test set;

[0012] Step 7: Establish a BP neural network prediction model, train the BP neural network prediction model through the training set, and obtain the optimal number of hidden layer nodes;

[0013] Step 8: Use the DBO algorithm to optimize the initial weights and thresholds of the BP neural network prediction model;

[0014] Step 9: Use the trained BP neural network prediction model to predict the plate shape control efficiency coefficient.

[0015] The method for predicting the efficiency coefficient of plate shape regulation in cold rolling plate shape control of the present invention has the following beneficial effects:

[0016] (1) First, the plate shape control efficiency coefficient matrix under different working conditions is calculated through offline simulation, and then the value range of the optimal weight factor is limited by the Euclidean distance method to prevent deviation from the correct optimization direction during the online adaptive process.

[0017] (2) The weight factor between the actual operating point and the preset operating point is determined by using the three-dimensional characteristic operating condition space representation technology, which significantly improves the calculation accuracy compared with the traditional two-dimensional operating condition division method.

[0018] (3) The optimal weight factor prediction model of the actual working point is established by data-driven method, and the optimal weight factor is found by using the actual working point data, so as to accurately obtain the plate shape control efficiency coefficient matrix under any rolling condition. The optimal weight factor prediction method of the actual working point has not only high accuracy, but also good stability and fast convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of a method for predicting the efficiency coefficient of plate shape regulation in cold-rolled plate shape control of the present invention;

[0020] Figure 2a is the measured plate shape change at a certain actual working point before the mean square error detection of the present invention;

[0021] Figure 2b is the plate shape change at a certain actual working point after Lagrangian interpolation filling of the present invention;

[0022] Figure 3 It is a schematic diagram of the positions of the actual operating point and the 8 related preset operating points in the present invention;

[0023] Figure 4 This is a flow chart of the present invention using the DBO algorithm to optimize the BP neural network. DETAILED DESCRIPTION

[0024] like Figure 1 As shown, a method for predicting the efficiency coefficient of plate shape control in cold rolling plate shape control of the present invention comprises the following steps:

[0025] Step 1: Set rolling force, strip width and strip thickness as the parameters of the working point, and select multiple preset working points in the simulation experiment, specifically:

[0026] Step 1.1: Determine the upper and lower limits of the rolling force, strip width, and strip thickness parameter values ​​during actual rolling, and use them as the upper and lower limits of the parameter values ​​of the preset operating point to ensure that the actual operating point parameter values ​​during actual rolling fall within the preset operating point parameter value range.

[0027] Step 1.2: Divide the parameter value intervals of rolling force, strip width, and strip thickness at the preset working point into V parts, divide each parameter into V+1 parameter values, and arrange and combine these three parameter values ​​to obtain (V+1) 3 nodes, which serve as preset operating points in the offline simulation experiment.

[0028] Step 2: Obtain the plate shape control efficiency coefficient matrix of each preset working point through offline simulation, and export the measured rolling data from the rolling mill, specifically:

[0029] Step 2.1: Use the parameter values ​​of each preset operating point to conduct an offline simulation experiment, and give different plate shape adjustment amounts to form a plate shape adjustment amount matrix ΔU = [Δu1 Δu2 … Δu j ], Δu j is the adjustment amount of the j-th plate shape adjustment mechanism, and the plate shape change matrix in the plate width direction is obtained. ΔY = [Δy1 Δy2 … Δy i ], Δy i It is the change of the plate shape at the i-th measuring point in the width direction of the strip.

[0030] In specific implementation, the plate shape adjustment mechanism includes working roll bending, intermediate roll bending, roll lateral shifting and roll tilting, etc.

[0031] Step 2.2: Calculate the plate shape control efficiency coefficient matrix for each preset working point through the plate shape change matrix in the plate width direction and the plate shape adjustment matrix. The calculation formula is:

[0032]

[0033] Where, Eff is the plate shape control efficiency coefficient matrix at the preset working point, eff ij is the plate shape control efficiency coefficient of the j-th plate shape adjustment mechanism at the i-th measuring point in the plate width direction.

[0034] Step 2.3: After obtaining the plate shape control efficiency coefficient matrix of several preset operating points and the parameter values ​​of the preset operating points, make a table and then save it in the form of a text file.

[0035] Step 2.4: Obtain the actual rolling data from the rolling mill, including the rolling force, strip width, strip thickness, adjustment amount of the plate shape adjustment mechanism, and the plate shape change at each measuring point in the strip width direction. Organize these data and make a table, and save them in the measured rolling data Excel file.

[0036] Step 3: Use the mean square error to detect and eliminate abnormal values ​​of the plate shape change in the measured rolling data, specifically:

[0037] The mean square error is used to detect abnormal values ​​in each group of plate shape changes under actual working conditions. Data between [μ-3σ, μ+3σ] are considered valid data, and data outside of these ranges are considered abnormal data. The calculation formulas for μ and σ are as follows:

[0038]

[0039] Where: Δy oi is the measured plate shape change of the ith measuring point in the plate width direction at the actual working point O, and M is the number of measuring points in the strip width direction.

[0040] Figure 2a It is the measured plate shape change before mean square error detection under the parameters of rolling force 255KN, strip width 1275mm and strip thickness 0.5mm.

[0041] Step 4: Use the Lagrange interpolation method to fill the missing data of the plate shape change after removing the outliers, specifically:

[0042] Step 4.1: Select a set of plate shape changes from the plate shape change data set after removing outliers, and select L measurement points that have not been removed from this set of plate shape changes: (C1, △y o1)(C2,△y o2 )…(C L , △y oL ).

[0043] Step 4.2: Based on the given L measurement points, construct the Lagrange basic polynomial corresponding to each measurement point:

[0044]

[0045] Among them, l i (C) is the Lagrange basic polynomial of the i-th measurement point, C r Indicates the serial number of the measurement point to be eliminated, C i Indicates the serial number of the i-th measurement point in the plate width direction among the L measurement points that have not been eliminated, C m Indicates the serial number of the mth measurement point in the plate width direction among the L measurement points that have not been eliminated.

[0046] Step 4.3: Use the Lagrange basic polynomial l corresponding to each measurement point obtained in step 4.2 i (C) Calculate the Lagrange interpolation polynomial L(C) corresponding to the known L measurement points, which is the plate shape change Δy of the rth measurement point in the plate width direction that was eliminated or , to complete the filling of missing characteristic parameters, the Lagrange interpolation polynomial L(C) is:

[0047]

[0048] Figure 2b It is the change of plate shape after Lagrangian interpolation filling under the parameters of rolling force 255KN, strip width 1275mm and strip thickness 0.5mm.

[0049] Step 5: Obtain the actual operating points from the measured rolling data in sequence, and calculate the optimal weight factor array between the actual operating point and the preset operating point through the DBO algorithm, specifically:

[0050] Step 5.1: Figure 3 As shown in the figure, let the strip width be the X axis, the strip thickness be the Y axis, and the rolling force be the Z axis. The three parameters form a three-dimensional space structure. Each of the three coordinate axes has V+1 parameter values. These parameter values ​​divide the three-dimensional space structure into V 3 The actual working point O is obtained from the measured rolling data in sequence, the small cube where the actual working point O is located is determined, and the 8 vertices of the small cube, i.e., the 8 preset working points related to the actual working point O, are determined.

[0051] Step 5.2: Calculate the weight factor γ between the actual operating point O and the eight related preset operating points by using the Euclidean distance method. 1、 γ 2、 γ 3、 γ 4、 γ 5、 γ 6、 γ 7、 γ8, The calculation formula of the weight factor calculated by the Euclidean distance method is as follows:

[0052]

[0053]

[0054] Wherein: x0, y0, z0 are the strip width value, strip thickness value and rolling force value of the actual operating point O respectively; x1, x2, x3, x4, x5, x6, x7, x8 are the strip width values ​​of 8 preset operating points respectively, y1, y2, y3, y4, y5, y6, y7, y8 are the strip thickness values ​​of 8 preset operating points respectively, z1, z2, z3, z4, z5, z6, z7, z8 are the rolling force values ​​of 8 preset operating points respectively.

[0055] Step 5.3: The weight factor calculated by the Euclidean distance method is not the optimal weight factor, but it represents the similarity between the preset operating point and the actual operating point to a certain extent. The optimal weight factor value range is set based on the weight factor calculated by the Euclidean distance method to prevent the model from deviating from the correct direction when searching for the optimal weight factor. The optimal weight factor value range is as follows:

[0056]

[0057] Where: γ′ k is the optimal weight factor between the actual operating point and the kth preset operating point, γ k is the weight factor between the actual operating point and the kth preset operating point calculated by the Euclidean distance method, k = 1, 2, 3, 4, 5, 6, 7, 8.

[0058] Step 5.4: Use the DBO algorithm to obtain the optimal weight factor array. The mathematical model of the DBO algorithm is as follows:

[0059] Create R = 90 arrays, each array consists of H numbers. The maximum number of iterations of the DBO algorithm in this implementation is T = 100. This step is to find the optimal weight factor array, so in this step H = 8, each number represents a weight factor, and the values ​​of the 8 weight factors are randomly generated within the range of weight factor values ​​in step 5.3. Use the evaluation function to evaluate the quality of the array. The lower the evaluation function value, the better the array. The evaluation function calculation method is as follows:

[0060]

[0061] In the formula: J q is the evaluation function, and g i is the weight coefficient of each measurement point in the plate width direction. The weight coefficient of the measurement points within 60% of the plate width center is 0.8, and the weight coefficient of the measurement points on the remaining two sides is 1; Δu oj is the adjustment amount of the j-th measured plate shape adjustment mechanism at the actual working condition point O, and N is the number of plate shape adjustment mechanisms; Eff oq is the plate shape control efficacy coefficient matrix calculated with the q-th weight factor array created in the DBO algorithm as the weight factor at the actual working condition point O. The calculation formula is:

[0062]

[0063] In the formula: Eff s is the plate shape control efficacy coefficient matrix of the s-th preset working condition point related to the actual working condition point O calculated in step 2, and X qs is the s-th number of the q-th weight factor array created, where s = 1, 2, 3, 4, 5, 6, 7, 8.

[0064] Step 5.5: Divide the R created weight factor arrays into 4 parts according to 6:6:7:11, calculate the evaluation functions of all weight factor arrays, sort the four parts of weight factor arrays respectively from low to high according to the evaluation function, and at the same time sort all weight factor arrays from low to high according to the evaluation function. The array with the highest evaluation function among all weight factor arrays is used as X worst , and the array with the lowest evaluation function is used as X best . The array with the smallest evaluation function in the second part of the weight factor arrays is X gbest , and the array with the smallest evaluation function in the third part of the weight factor arrays is X lbest . Each part of the weight factor arrays is updated through different iteration methods to find the weight factor array with the highest adaptability.

[0065] (1) There are two iteration update formulas for the first part of the weight factor arrays, which are respectively:

[0066]

[0067] In the formula: t represents the current iteration number, represents the n-th array after the t-th iteration, 0 < c ≤ 0.2 is a constant value, B is a constant value between 0 and 1, and α is a natural coefficient assigned -1 or 1; represents the array with the worst fitness after the t-th iteration, θ is randomly generated in [0, π], and when θ = 0, θ = π / 2, θ = π, the array will not be updated in this round.

[0068] Each array in the first part will take a random number between 0 and 1 before iterative update. If the random number is less than 0.9, it will be iteratively updated according to the first formula, otherwise it will be iteratively updated according to the second formula.

[0069] (2) The iterative update formula of the second part weight factor array is:

[0070]

[0071] Where: B1, B2 are 1×8 independent random variables, Lb * ,Ub * is the upper and lower bounds of the second part of the weight factor array, and the calculation formula is:

[0072]

[0073] Where: is the array with the smallest evaluation function in the second part of the weight factor array after the tth iteration, Lb is the lower limit array of the optimal weight factor value range in step 5.3, Ub is the upper limit array of the optimal weight factor value range in step 5.3, R = (1-t) / T.

[0074] (3) The iterative update formula of the weight factor array in the third part is:

[0075]

[0076] Where: C1 is a random number that obeys a normal distribution, C2 is a 1×8 random vector between (0,1); Lb l ,Ub l is the upper and lower bounds of the third part of the array, and the calculation formula is:

[0077]

[0078] Where: It is the array with the smallest evaluation function in the third part of the weight factor array after the tth iteration.

[0079] (4) The fourth part of the array iteration update formula is:

[0080]

[0081] Where G is a 1×D random vector between (0,1), and S is a constant.

[0082] Each time the array is iterated and updated, the evaluation function is calculated once, and the new X worst Arrays and The evaluation function of the array is compared, and the one with the higher evaluation function becomes the new X worst ; X best Arrays and The evaluation function of the array is compared, and the one with the lower evaluation function becomes the new X best ; X gbest Arrays and The evaluation function of the array is compared, and the one with the lower evaluation function becomes the new X gbest ; X lbest Arrays and The evaluation function of the array is compared, and the one with the lower evaluation function becomes the new X lbest .

[0083] Step 5.6: Determine whether the number of iterations t is greater than the maximum number of iterations T. If not, proceed to step 5.5. If it is, stop iterating and obtain the array X with the minimum evaluation function after T generations. best , and use this set of arrays as the optimal weight factor array of the actual operating point O.

[0084] Step 6: The parameters of the actual operating point and the optimal weight factor array constitute a data set and divide it into a training set and a test set, specifically:

[0085] After finding the optimal weight factor arrays for all actual operating points in the measured production data, the rolling force, strip width, strip thickness and optimal weight factor arrays of these actual operating points are tabulated as the data set for BP neural network training. The data order in the data set is shuffled, the first 80% is used as the training set, and the last 20% is used as the test set.

[0086] Step 7: Establish a BP neural network prediction model. Train the BP neural network prediction model with the training set to obtain the optimal number of hidden layer nodes, specifically:

[0087] Step 7.1: Determine the BP neural network structure, where the input nodes are 3, which are rolling force, strip width and strip thickness respectively; the output nodes are 8, which are the optimal weight factors between the actual working point and the 8 related preset working points. The range of the number of hidden layer nodes of the BP neural network is determined according to the empirical formula:

[0088]

[0089] Where: β∈[1,10], h is the number of hidden layer nodes of BP neural network, a is the number of input nodes, and b is the number of output nodes.

[0090] The calculated range of the number of nodes in the hidden layer of the BP neural network is h = [5,13].

[0091] Step 7.2: Input the training set, set the initial weights and thresholds of the BP neural network, including the connection weight between the input layer and the hidden layer is w ef , the threshold between the input layer and the hidden layer is b f ; The connection weight between the hidden layer and the output layer is w fg , the threshold between the hidden layer and the output layer is b g .

[0092] Neural network input P e The calculation formula of output Q is:

[0093]

[0094] Where: P e is the e-th input variable of the neural network, F is the Sigmoid activation function, and f represents the f-th hidden node.

[0095] Step 7.3: Calculate the predicted output value Q of the BP neural network with different numbers of hidden layer nodes dg and target value The mean square error is calculated as:

[0096]

[0097] In the formula, MSE is the total mean square error of the BP neural network output, E is the number of samples, Q is dg is the predicted output value of the d-th sample data of the g-th output, is the target value of the dth sample data of the gth output.

[0098] Step 7.4: Take the number of hidden layer nodes when the mean square error value is the smallest as the optimal number of hidden layer nodes, and take the number of hidden layer nodes as 10.

[0099] The experimental results are shown in Table 1. When the number of hidden layer nodes is 10, the minimum mean square error is 0.0038310, so the BP neural network is set to a 3-10-8 structure.

[0100] Table 1 Mean square error of training set under different hidden layer nodes

[0101]

[0102] Step 8: Use the DBO algorithm to optimize the initial weights and thresholds of the BP neural network prediction model to improve the prediction accuracy and convergence speed of the BP neural network. Figure 4 As shown, specifically:

[0103] Step 8.1: Create R = 90 arrays, each of which consists of the initial weights and thresholds w of the BP neural network. ef 、bf 、w fg 、b g The elements of the array are arranged, that is, the array consists of v = 121 numbers, let the uth array in the R array be a u ,Right now:

[0104] a u =[a u1 a u2 … a uv ]=[w ef w fg bf g ];

[0105] in

[0106] [a u1 a u2 … a u30 ]=[w ef ];

[0107] [a u31 a u32 … a u110 ]=[w fg ];

[0108] [a u111 a u112 a u113 ]=[b f ];

[0109] [a u114 a u115 … a u121 ]=[b g ];

[0110] Step 8.2: Use fitness to represent the prediction ability of the BP neural network under different initial weights and thresholds, and calculate the predicted output value Q of the BP neural network under different initial weights and thresholds. dg and target value The root mean square error is used as the fitness value of all arrays. The better the prediction ability of the neural network, the smaller the fitness value.

[0111] Step 8.3: Use the DBO algorithm to iteratively update the arrays, find the array with the smallest fitness, and determine the optimal initial weights and thresholds of the BP neural network; perform T = 100 rounds of iterative updates on the R arrays according to step 5.5, and select the array a with the smallest fitness after 100 iterative updates. best , as the optimal initial weight and threshold.

[0112] Step 8.4: Use the test set to verify the prediction accuracy of the trained BP neural network prediction model.

[0113] Step 9: Use the trained BP neural network prediction model to predict the plate shape control efficiency coefficient.

[0114] The above description is only a preferred embodiment of the present invention and is not intended to limit the concept of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the efficiency coefficient of plate shape control in cold rolling plate shape control, characterized in that: include: Step 1: Set rolling force, strip width and strip thickness as parameters of the working point, and select multiple preset working points in the simulation experiment; Step 2: Obtain the plate shape control efficiency coefficient matrix of each preset operating point through offline simulation, and export the measured rolling data from the rolling mill; Step 3: Use the mean square error to detect and eliminate abnormal values ​​of the plate shape change in the measured rolling data; Step 4: Use Lagrange interpolation method to fill the missing data of the plate shape change after removing the outliers; Step 5: Obtain the actual operating points from the measured rolling data in sequence, and calculate the optimal weight factor array between the actual operating point and the preset operating point by using the DBO algorithm; Step 6: The parameters of the actual operating point and the optimal weight factor array constitute a data set and divide it into a training set and a test set; Step 7: Establish a BP neural network prediction model, train the BP neural network prediction model through the training set, and obtain the optimal number of hidden layer nodes; Step 8: Use the DBO algorithm to optimize the initial weights and thresholds of the BP neural network prediction model; Step 9: Use the trained BP neural network prediction model to predict the plate shape control efficiency coefficient.

2. The method for predicting the efficiency coefficient of shape control in cold rolling shape control according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1.1: Determine the upper and lower limits of the rolling force, strip width, and strip thickness parameter values ​​during actual rolling, and use them as the upper and lower limits of the parameter values ​​of the preset working point, so as to ensure that the actual working point parameter values ​​during actual rolling fall within the preset working point parameter value range; Step 1.2: Divide the parameter value intervals of rolling force, strip width, and strip thickness at the preset working point into V parts, divide each parameter into V+1 parameter values, and arrange and combine these three parameter values ​​to obtain (V+1) 3 nodes, which serve as preset operating points in the offline simulation experiment.

3. The method for predicting the efficiency coefficient of shape control in cold rolling shape control according to claim 1, characterized in that: The step 2 is specifically as follows: Step 2.1: Use the parameter values ​​of each preset operating point to conduct an offline simulation experiment, and give different plate shape adjustment amounts to form a plate shape adjustment amount matrix ΔU = [Δu1 Δu2 … Δu j ], Δu j is the adjustment amount of the j-th plate shape adjustment mechanism, and the plate shape change matrix in the plate width direction is obtained. ΔY = [Δy1 Δy2 … Δy i ], Δy i is the change in shape of the i-th measuring point in the width direction of the steel strip; Step 2.2: Calculate the plate shape control efficiency coefficient matrix for each preset working point through the plate shape change matrix in the plate width direction and the plate shape adjustment matrix. The calculation formula is: Where, Eff is the plate shape control efficiency coefficient matrix at the preset working point, eff ij is the plate shape control efficiency coefficient of the jth plate shape adjustment mechanism at the i-th measuring point in the plate width direction; Step 2.3: After obtaining the plate shape control efficiency coefficient matrix of several preset working points and the parameter values ​​of the preset working points, make a table and then save it in the form of a text file; Step 2.4: Obtain the actual rolling data from the rolling mill, including the rolling force, strip width, strip thickness, adjustment amount of the plate shape adjustment mechanism, and the plate shape change at each measuring point in the strip width direction. Organize these data and make a table, and save them in the measured rolling data Excel file.

4. The method for predicting the efficiency coefficient of shape control in cold rolling shape control according to claim 3, characterized in that: The step 3 is specifically as follows: The mean square error is used to detect abnormal values ​​in each group of plate shape changes under actual working conditions. Data between [μ-3σ, μ+3σ] are considered valid data, and data outside of these ranges are considered abnormal data. The calculation formulas for μ and σ are as follows: Where: Δy oi is the measured plate shape change of the ith measuring point in the plate width direction at the actual working point O, and M is the number of measuring points in the strip width direction.

5. The method for predicting the efficiency coefficient of shape control in cold rolling shape control according to claim 4, characterized in that: The step 4 is specifically as follows: Step 4.1: Select a set of plate shape changes from the plate shape change data set after removing outliers, and select L measurement points that have not been removed from this set of plate shape changes: (C1, Δy o1 )(C2,Δy o2 )…(C L , Δy oL ); Step 4.2: Based on the given L measurement points, construct the Lagrange basic polynomial corresponding to each measurement point: Among them, l i (C) is the Lagrange basic polynomial of the i-th measurement point, C r Indicates the serial number of the measurement point to be eliminated; C i Indicates the sequence number of the i-th measurement point in the plate width direction among the L measurement points that have not been eliminated, C m Indicates the sequence number of the mth measurement point in the plate width direction among the L measurement points that have not been eliminated; Step 4.3: Use the Lagrange basic polynomial l corresponding to each measurement point obtained in step 4.2 i (C) Calculate the Lagrange interpolation polynomial L(C) corresponding to the known L measurement points, which is the plate shape change Δy of the rth measurement point in the plate width direction that was eliminated or , to complete the filling of missing characteristic parameters, the Lagrange interpolation polynomial L(C) is:

6. The method for predicting the efficiency coefficient of shape control in cold rolling shape control according to claim 5, characterized in that: The step 5 is specifically as follows: Step 5.1: Let the strip width be the X axis, the strip thickness be the Y axis, and the rolling force be the Z axis. The three parameters form a three-dimensional space structure. Each of the three coordinate axes has V+1 parameter values. These parameter values ​​divide the three-dimensional space structure into V 3 Small cubes of the same size, each cube has 8 vertices, i.e. 8 preset working points; The actual working point O is obtained from the measured rolling data in sequence, a small cube where the actual working point O is located is determined, and 8 vertices of the small cube are determined, that is, 8 preset working points related to the actual working point O; Step 5.2: Calculate the weight factor γ between the actual operating point O and the eight related preset operating points by using the Euclidean distance method. 1、 γ 2、 γ 3、 γ 4、 γ 5、 γ 6、 γ 7、 γ 8, , the calculation formula of the weight factor calculated by the Euclidean distance method is as follows: Wherein: x0, y0, z0 are the strip width value, strip thickness value and rolling force value of the actual working point O respectively; x1, x2, x3, x4, x5, x6, x7, x8 are the strip width values ​​of 8 preset working points respectively, y1, y2, y3, y4, y5, y6, y7, y8 are the strip thickness values ​​of 8 preset working points respectively, z1, z2, z3, z4, z5, z6, z7, z8 are the rolling force values ​​of 8 preset working points respectively; Step 5.3: Set the optimal weight factor value range based on the weight factor calculated by the Euclidean distance method to prevent the model from deviating from the correct direction when searching for the optimal weight factor. The optimal weight factor value range is as follows: Where: γ′ k is the optimal weight factor between the actual operating point and the kth preset operating point, γ k is the weight factor between the actual operating point and the kth preset operating point calculated by the Euclidean distance method, k = 1, 2, 3, 4, 5, 6, 7, 8; Step 5.4: Use the DBO algorithm to obtain the optimal weight factor array. The mathematical model of the DBO algorithm is as follows: Create R = 90 weight factor arrays, each weight factor array consists of 8 weight factors. The maximum number of iterations of the DBO algorithm is T = 100. The 8 weight factors are randomly generated within the range of the optimal weight factor values ​​in step 5.

3. Use an evaluation function to evaluate the quality of each weight factor array. The lower the evaluation function value, the better the weight factor array. The evaluation function calculation method is as follows: Where: J q is the evaluation function, g i is the weight coefficient of each measuring point in the board width direction. The weight coefficient of the measuring points within 60% of the board width center is 0.8, and the weight coefficient of the measuring points on the remaining two sides is 1; Δu oj is the adjustment amount of the jth measured plate shape adjustment mechanism at the actual working point O, N is the number of plate shape adjustment mechanisms; Eff oq The plate shape control efficiency coefficient matrix is ​​calculated by taking the qth weight factor array created in the DBO algorithm as the weight factor for the actual operating point O. The calculation formula is: Where: Eff s is the plate shape control efficiency coefficient matrix of the sth preset operating point related to the actual operating point O calculated in step 2, X qs is the sth number of the qth weight factor array created, s=1,2,3,4,5,6,7,8; Step 5.5: Divide the created R weight factor arrays into 4 parts, calculate the evaluation function of all weight factor arrays, sort the four weight factor arrays from low to high according to the evaluation function, and sort all weight factor arrays from low to high according to the evaluation function. The array with the highest evaluation function among all weight factor arrays is taken as X. worst , the array with the lowest evaluation function is taken as X best ; The array with the smallest evaluation function in the second part of the weight factor array is X gbest , the array with the smallest evaluation function in the third part of the weight factor array is X lbest , each part of the weight factor array is updated through different iterations to find the weight factor array with the highest adaptability; (1) There are two iterative update formulas for the weight factor array in the first part, which are: where: t represents the current iteration number, represents the nth array after the t-th iteration, 0 < c ≤ 0.2 is a constant value, B is a constant value between 0 and 1, and α is a natural coefficient assigned -1 or 1; represents the array with the worst fitness after the t-th iteration, θ is randomly generated in [0, π], and when θ = 0, θ = π / 2, θ = π, the array will not be updated in this round; Each array in the first part will take a random number between 0 and 1 before iterative update. If the random number is less than 0.9, it will be iteratively updated according to the first formula, otherwise it will be iteratively updated according to the second formula; (2) The iterative update formula of the second part weight factor array is: Where: B1, B2 are 1×8 independent random variables, Lb * ,Ub * is the upper and lower bounds of the second part of the weight factor array, and the calculation formula is: Where: is the array with the smallest evaluation function in the second part of the weight factor array after the tth iteration, Lb is the lower limit array of the optimal weight factor value range in step 5.3, Ub is the upper limit array of the optimal weight factor value range in step 5.3, R = (1-t) / T; (3) The iterative update formula of the weight factor array in the third part is: Where: C1 is a random number that obeys a normal distribution, C2 is a 1×8 random vector between (0,1); Lb l ,Ub l is the upper and lower bounds of the third part of the array, and the calculation formula is: Where: It is the array with the smallest evaluation function in the third part of the weight factor array after the tth iteration; (4) The iterative update formula of the fourth part weight factor array is: Where: G is a 1×8 random vector between (0,1), S is a constant; Each time the array is iterated and updated, the evaluation function is calculated once, and the new X worst Arrays and The evaluation function of the array is compared, and the one with the higher evaluation function becomes the new X worst ; X best Arrays and The evaluation function of the array is compared, and the one with the lower evaluation function becomes the new X best ; X gbest Arrays and The evaluation function of the array is compared, and the one with the lower evaluation function becomes the new X gbest ; X lbest Arrays and The evaluation function of the array is compared, and the one with the lower evaluation function becomes the new X lbest ; Step 5.6: Determine whether the number of iterations t is greater than the maximum number of iterations T. If not, proceed to step 5.

5. If it is, stop iterating and obtain the array X with the minimum evaluation function after T generations. best , and use this set of arrays as the optimal weight factor array of the actual operating point O.

7. The method for predicting the efficiency coefficient of shape control in cold rolling shape control according to claim 6, characterized in that: The step 6 is specifically as follows: after finding the optimal weight factor arrays of all actual operating points in the measured production data, the rolling force, strip width, strip thickness and optimal weight factor arrays of these actual operating points are made into a table as a data set for BP neural network training, the order of the data in the data set is shuffled, the first 80% is used as a training set, and the last 20% is used as a test set.

8. The method for predicting the efficiency coefficient of shape control in cold rolling shape control according to claim 6, characterized in that: The step 7 is specifically as follows: Step 7.1: Determine the BP neural network structure, where the input nodes are 3, which are rolling force, strip width and strip thickness respectively; the output nodes are 8, which are the optimal weight factors between the actual working point and the 8 related preset working points. The range of the number of hidden layer nodes of the BP neural network is determined according to the empirical formula: Where: β∈[1,10], h is the number of hidden layer nodes of BP neural network, a is the number of input nodes, and b is the number of output nodes; The calculated range of the number of nodes in the hidden layer of the BP neural network is h = [5, 13]; Step 7.2: Input the training set, set the initial weights and thresholds of the BP neural network, including the connection weight between the input layer and the hidden layer is w ef , the threshold between the input layer and the hidden layer is b f ; The connection weight between the hidden layer and the output layer is w fg , the threshold between the hidden layer and the output layer is b g ; Neural network input P e The calculation formula of output Q is: Where: P e is the e-th input variable of the neural network, F is the Sigmoid activation function, and f represents the f-th hidden node; Step 7.3: Calculate the predicted output value Q of the BP neural network with different numbers of hidden layer nodes dg and target value The mean square error is calculated as: In the formula, MSE is the total mean square error of the BP neural network output, E is the number of samples, Q is dg is the predicted output value of the d-th sample data of the g-th output, is the target value of the d-th sample data of the g-th output; Step 7.4: Take the number of hidden layer nodes when the mean square error value is the smallest as the optimal number of hidden layer nodes, and take the number of hidden layer nodes as 10.

9. The method for predicting the efficiency coefficient of shape control in cold rolling shape control according to claim 8, characterized in that: The step 8 is: Step 8.1: Create R = 90 arrays, each of which consists of the initial weights and thresholds w of the BP neural network. ef 、b f 、w fg 、b g The elements of the array are arranged, that is, the array consists of v = 121 numbers, let the uth array in the R array be a u ,Right now: a u =[a u1 a u2 … a uv ]=[w ef w fg b f b g ]; in [a u1 a u2 … a u30 ]=[w ef ]; [a u31 a u32 …a u110 ]=[w fg ]; [a u111 a u112 a u113 ]=[b f ]; [a u114 a u115 … a u121 ]=[b g ]; Step 8.2: Use fitness to represent the prediction ability of the BP neural network under different initial weights and thresholds, and calculate the predicted output value Q of the BP neural network under different initial weights and thresholds. dg and target value The root mean square error is used as the fitness value of all arrays. The better the prediction ability of the neural network, the smaller the fitness value. Step 8.3: Use the DBO algorithm to iteratively update the arrays, find the array with the smallest fitness, and determine the optimal initial weights and thresholds of the BP neural network; perform T = 100 rounds of iterative updates on the R arrays according to step 5.5, and select the array a with the smallest fitness after 100 iterative updates. best , as the optimal initial weight and threshold; Step 8.4: Use the test set to verify the prediction accuracy of the trained BP neural network prediction model.

Images