A data and mechanism fusion hot-rolled strip crown prediction method

Abstract

The application provides a kind of data and mechanism fusion hot-rolled strip steel crown prediction method, comprising the following steps: collecting the production data of strip steel hot continuous rolling process, and establishing hot-rolled strip steel crown mechanism model;From the production data collected and the data calculated from the hot-rolled strip steel crown mechanism model, the characteristic data affecting the crown is selected;The collected production data is preprocessed, including data smoothing and data standardization;Based on the data set obtained after feature selection and data preprocessing, a convolutional neural network is established;Determine the number of neural network convolution layers, convolution kernel size, number of hidden layer neurons, activation function, optimizer, learning rate;Train the neural network model to predict the strip steel crown value.The application provides a kind of data and mechanism fusion hot-rolled strip steel crown prediction method, which uses the rolling data collected in industrial field to establish convolutional neural network to predict the crown.Solve the problem of poor control effect caused by complex model and inaccurate modeling in traditional control.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rolling process prediction, in particular to a hot-rolled strip crown prediction method based on data and mechanism fusion. BACKGROUND

[0002] Hot-rolled strip is an important industrial raw material in the production process of products such as automobiles, aerospace, instruments and daily life. With the rapid development of social economy and the continuous improvement of people's living standards, the demand for strip products is also increasing. Improving the product quality of strip production not only can avoid a large amount of energy and capital waste, but also can provide high-quality raw materials for downstream industries and ensure the quality of downstream products. The rolling control system is a complex dynamic system with great difficulty in control. Using the production data collected during rolling to build a model and making full use of the information in the data can achieve the purpose of accurately controlling the exit crown and improve the product quality, which is a scientific and application value method.

[0003] The crown of the strip in the rolling process is affected by many factors such as the state of the rolling mill and the state of the roll, and has characteristics such as nonlinearity, large time delay and strong coupling, and is a complex dynamic control system. Traditional control is mostly based on the established mechanism model of the strip. Because the model is too complex, a large number of approximations are used in the modeling process, and the accuracy of the established model is very low, and the parameters of the rolling mill are set by expert experience, and with the increasing demand of users for quality and the increasing product specifications, this setting method not only greatly increases the workload of experts but also is very inefficient. Taking a hot rolling field as an example, according to the regulation that the actual crown of the strip within ±10μm of the target crown is a qualified product, the strip produced by the above method has only 61.95% that meets the requirements. In order to meet the needs of customers for yield and quality, there is an urgent need for a hot-rolled plate crown prediction system based on data-driven. SUMMARY

[0004] In view of the shortcomings of the prior art, the present application provides a hot-rolled strip crown prediction method based on data and mechanism fusion, which can predict the exit crown in real time according to the collected data during rolling, and control the rolling mill according to the crown prediction value to improve the product quality.

[0005] The technical scheme adopted by the present application to solve the technical problems is: a hot-rolled strip crown prediction method based on data and mechanism fusion, comprising the following steps:

[0006] Step 1. Establish a mechanism model: collect production data of the hot continuous rolling process of the strip, and establish a hot-rolled strip crown mechanism model;

[0007] Step 2. Feature selection: select feature data that affects the crown from the collected production data and hot-rolled strip crown mechanism model calculation data;

[0008] Step 3. Data preprocessing: data preprocessing is performed on the collected production data, including data smoothing and data standardization;

[0009] Step 4. Building a neural network: a convolutional neural network (CNN) is established based on the data set obtained after step 2 feature selection and step 3 data preprocessing;

[0010] Step 5. Neural network hyperparameter determination: determine the number of convolutional layers of the neural network, the size of the convolution kernel, the number of hidden layer neurons, the activation function, the optimizer, and the learning rate;

[0011] Step 6. Convexity prediction: train the neural network model to predict the strip convexity value.

[0012] Further, the calculation formula of the strip convexity mechanism model established in step 1 is specifically:

[0013] Step 1.1. The mathematical formula of the hot rolled strip convexity mechanism model is as follows:

[0014] C h (i)=C W (i)×W(i)+C PC (i)×P(i)+C FC (i)×F(i)+C WC (i)×{C hR (i)+C hW (i)+C hT (i)}+C θC (i)×C hθ (i)+C ηC (i)×C h (i-1)

[0015] In the formula, C h (i) is the convexity of the strip at the mill outlet, W(i) is the strip width, P(i) is the rolling force, F(i) is the bending force, C hR (i) is the work roll convexity, C hW (i) is the work roll wear convexity, C hT (i) is the work roll thermal convexity, C hθ (i) is the equivalent convexity of roll shift, C W (i) is the correction coefficient of strip width, C PC (i) is the transverse stiffness correction coefficient, C FC (i) is the bending force correction coefficient, C WC (i) is the work roll convexity correction coefficient, C θC (i) is the roll shift position correction coefficient, C ηC (i) is the strip convexity genetic correction coefficient, and i is the number of rolling mills in the finishing mill train;

[0016] Step 1.2. The thermal crown of the work roll caused by thermal expansion is calculated using the following formulas according to different working conditions:

[0017] When the temperature gradient in the radial direction of the work roll is large during the initial stage of rolling, the calculation formula is:

[0018]

[0019] When a certain amount of strip steel has been rolled, and the work rolls are (near) thermal equilibrium, the calculation formula is:

[0020]

[0021] In the formula, β is the average temperature rise within the cross-section of the work roll, β is the linear expansion coefficient of the work roll material, v is Poisson's ratio, and T is the average temperature rise within the cross-section of the work roll. R T represents the current temperature of the work roll. R0 The initial temperature of the working roll;

[0022] in The calculation formula is:

[0023] In the formula, R is the radius of the working roll, and r is a variable along the radial direction of the working roll.

[0024] Furthermore, the production data in the feature data of step 2 includes: intermediate billet thickness, average thickness and average width during rolling, strip temperature at the entrance of the finishing mill (first mill), strip temperature at the exit of the finishing mill (last mill), rolling speed of each mill in the finishing mill, rolling force of each mill, bending force of each mill, amount of roll shifting of each mill, rolling mileage, and the roll crown of the last mill in the finishing mill. The calculation data based on the hot-rolled strip crown mechanism model is the roll crown of each mill in the finishing mill except for the last mill.

[0025] Furthermore, the smoothing method in step 3 adopts the five-point cubic smoothing method.

[0026] Furthermore, the specific implementation method of each module of the convolutional neural network in step 4 is as follows:

[0027] Step 4.1. One-dimensional convolution operation: Read data sequentially from the dataset processed in Step 3. The input vector is convolved with a one-dimensional convolution kernel Y(1*n1) to obtain the convolution result; where: The dimension of the data collected in step 2, feature selection. And take odd numbers;

[0028] Step 4.2. Pooling operation: Perform pooling operation on the result of the convolution operation;

[0029] Step 4.3. Flatten operation: The flatten operation is used to connect the first and last ends of the multi-channel one-dimensional vector to form a one-dimensional vector, which is then connected to the hidden layer of the neural network.

[0030] Step 4.4. Forward Propagation: Let the one-dimensional vector expanded by the flatten operation be X, the weights between the input layer and the hidden layer be w, the threshold be b1, and the activation function be f1; then the hidden layer output net1 is:

[0031]

[0032] Where i and j represent the input neuron and hidden layer neuron numbers, respectively;

[0033] Using the hidden layer output net1 as the input to the next layer, with weights v between the hidden layer and the output layer, a threshold of b2, and an activation function of f2, the output layer output net2 is:

[0034]

[0035] Where m and n represent the numbers of hidden layer neurons and output neurons, respectively;

[0036] Step 4.5. Backward differentiation: Based on the final output net2 obtained in step 4.4 and the expected output, the error term E can be obtained. Then, the gradient descent method is used to update the weights and thresholds.

[0037] Furthermore, the specific method for updating the weights and thresholds using gradient descent is as follows:

[0038] Calculate the partial derivatives of the error term E with respect to the hidden layer weights and thresholds using the chain rule.

[0039]

[0040]

[0041] Calculate the partial derivatives of the error term E with respect to the output layer weights and threshold using the chain rule.

[0042]

[0043]

[0044] Based on the partial derivatives of the weights and thresholds of each layer calculated above, the values ​​of the thresholds and weights are updated using the following formula:

[0045]

[0046]

[0047]

[0048]

[0049] wherein η represents a learning rate, k = 1, 2, …, n represents the number of updates or iterations, k = 1 represents the first update, and so on.

[0050] Further, the specific method of step 5 is:

[0051] Step 5.1. Determine the number of convolution layers and hidden layers, determine the approximate interval of the number of convolution layers and hidden layers according to the data characteristics and data distribution characteristics in the data set, and traverse the interval values two by two to select the optimal number of convolution layers and hidden layers.

[0052] Step 5.2. Determine the size of the convolution kernel in the convolution operation, and search 1-5 to determine the size of the convolution kernel.

[0053] Step 5.3. Determine the number of neurons in the neural network hidden layer, first calculate the empirical value of the number of neurons using the empirical formula, which is as follows:

[0054] n = log2T

[0055] wherein T is the number of input features, and n is the number of hidden layer neurons; based on the calculated empirical value and the number of neurons after the convolution layer is unfolded, an interval is determined, and the number of neurons in the interval is traversed to determine the optimal number of neurons n.

[0056] Step 5.4. Determine the activation function between the fully connected layers, test these combinations respectively by combining the commonly used activation functions Sigmoid, tanh and Relu two by two, and determine the optimal activation function.

[0057] Step 5.5. Determine the optimizer of the neural network, and search for several commonly used optimizers in neural network training: SGD, MBGD, AdaGrad and Adam to determine the optimal optimizer.

[0058] Step 5.6. Determine the learning rate of the neural network, and search 0.001-0.1 to determine the optimal learning rate lr.

[0059] Further, the specific implementation method of step 6 is: based on the on-site data processed in step 3 and the hyperparameters of the convolutional neural network determined in step 5, train the convolutional neural network, and the trained convolutional neural network is the convexity prediction model. The new collected rolling data is input into the convexity prediction model after data preprocessing to obtain the convexity prediction value of the strip steel.

[0060] Compared with the prior art, the present application has the beneficial effects that:

[0061] The present application provides a data and mechanism fusion hot-rolled strip steel crown prediction method, which uses the rolling data collected in the industrial field to establish a convolutional neural network for crown prediction. The present application provides a method that can fully utilize the data collected in the production process, establish a high-precision prediction model, and realize crown prediction. The present application provides a method that can fully utilize the data collected in the production process, establish a high-precision prediction model, and realize crown prediction. The present application provides a method that can fully utilize the data collected in the production process, establish a high-precision prediction model, and realize crown prediction. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 A data-driven hot-rolled plate crown prediction system schematic diagram is provided in the present application.

[0063] Figure 2 A five-point cubic data smoothing before and after effect comparison chart is provided.

[0064] Figure 3 A convolutional neural network in different situations convolution operation process schematic diagram is provided.

[0065] Figure 4 A convolutional neural network structure schematic diagram used in the data and mechanism fusion hot-rolled strip steel crown prediction method provided in the present application is provided.

[0066] Figure 5 A crown prediction value and actual measurement value scatter plot is provided. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below with examples.

[0068] In this embodiment, a hot continuous rolling production line composed of six HCW rolling mills is used, the roll diameter is 630mm-700mm, the width of the rolled strip is 700mm-2130mm, and the thickness is 1.2mm-25.4mm.

[0069] A data and mechanism fusion hot-rolled strip steel crown prediction method, as shown in Figure 1 includes the following steps:

[0070] Step 1. Collect the production data of a hot continuous rolling mill, a total of 4000 steel coils; according to the collected data related to the crown, establish a hot-rolled strip steel crown mechanism model, the specific calculation method is as follows:

[0071] Step 1.1. Mathematical formula of hot rolling strip crown mechanism model is as follows:

[0072] C h (i)=C W (i)×W(i)+C PC (i)×P(i)+C FC (i)×F(i)+C WC (i)×{C hR (i)+C hW (i)+C hT (i)}+C θC (i)×C hθ (i)+C ηC (i)×C h (i-1)

[0073] In the formula, C h (i) is the crown of strip at mill exit, W(i) is the strip width, P(i) is the rolling force, F(i) is the bending force, C hR (i) is the work roll crown, C hW (i) is the work roll wear crown, C hT (i) is the work roll thermal crown, C hθ (i) is the equivalent crown of roll shift, C W (i) is the correction coefficient of strip width, C PC (i) is the correction coefficient of lateral stiffness, C FC (i) is the correction coefficient of bending force, C WC (i) is the correction coefficient of work roll crown, C θC (i) is the correction coefficient of roll shift position, C ηC (i) is the genetic correction coefficient of strip crown, i is the number of mills in the finishing mill train;

[0074] Step 1.2. Work roll thermal crown caused by thermal expansion of work roll, the calculation formula is respectively as follows according to different working conditions:

[0075] When the temperature gradient in the radial direction of work roll is large at the beginning of rolling, the calculation formula is as follows:

[0076]

[0077] When the work roll approaches to thermal equilibrium after rolling a certain amount of strip, the calculation formula is as follows:

[0078]

[0079] In the formula, is the average temperature rise in the cross section of work roll, β is the linear expansion coefficient of work roll material, v is the Poisson's ratio, T R is the current temperature of work roll, TR0 is the initial temperature of the work roll.

[0080] wherein The calculation formula is:

[0081] In the formula, R is the radius of the work roll, and r is a variable along the radial direction of the work roll.

[0082] Step 1.3. In the hot-rolled strip crown mechanism model, the remaining variables except the strip width, rolling force, bending force and work roll thermal crown are regarded as fixed values, the strip crown value at the exit of the finishing mill is calculated, and the value is taken as the input data of the convolutional neural network.

[0083] Step 2. From the measured data and calculated data obtained in step 1, 36 features that have the greatest impact on the crown are selected for research from a total of 122 features, including variable values collected by sensors on the production line: intermediate blank thickness, average thickness and average width during rolling, strip temperature at the entrance of the finishing mill unit, strip temperature at the exit of the finishing mill unit, rolling speed of the finishing mill unit (F1-F6), rolling force of F1-F6, bending force of F1-F6, roll shifting amount of F1-F6, rolling kilometers, strip crown after the finishing mill F6, and variable values calculated by the crown mechanism model: crown after rolling of the finishing mill (F1-F5).

[0084] Step 3. The five-point cubic data smoothing technique is used to smooth the data collected on site, and the effects before and after smoothing are shown in Figure 2 , and data standardization processing is performed;

[0085] The specific method is:

[0086] Step 3.1. The data collected on site has the characteristics of large data volume, large noise and high complexity. In order to remove high-frequency noise in the data and improve the smoothness of the data, five-point cubic smoothing is used for pre-processing of the crown data after rolling of the finishing mill F6, and the calculation formula is as follows:

[0087]

[0088] wherein Y i refers to the i-th crown data, is the smoothed value of Y i .

[0089] Step 3.2. Many kinds of sensors are arranged on the rolling production line to collect data during rolling, and the collected data has different dimensions. In order to ensure error convergence in the network training process, the data in the data set is standardized, and the calculation formula is as follows:

[0090]

[0091] wherein x i represents x i the standardized value, x i is the initial value of the i-th x feature data, μ is the mean of all data of the feature, and σ is the standard deviation of all data of the feature.

[0092] Step 4. A convolutional neural network (CNN) is established based on the data after feature selection and data cleaning, and the specific implementation method of each module of the CNN is as follows:

[0093] Step 4.1. One-dimensional convolution operation:

[0094] The data set used in the present application contains 5656 pieces of data of 36 dimensions, which are collected in the field and calculated by a mechanism model. One piece of data X (1*36) is taken from the data set as an input vector and a one-dimensional convolution kernel Y (1*n1) is used for convolution operation (wherein n1<36 and is usually an odd number); the convolution operation process is as follows: the first n1 elements of the input vector X are multiplied with the n1 elements of the convolution kernel Y at the corresponding positions, and then the results are added to obtain the first element of the output vector Z; the convolution kernel Y is shifted to the right by one step, that is, the first element and the n1th element of the convolution kernel Y are respectively multiplied with the second element and the n1+1th element of the input vector X at the corresponding positions, and then the results are added to obtain the second element of the output vector Z; the input vector is traversed in sequence, and the convolution output Z of the input vector X and the convolution kernel Y is obtained; the above is the case of single input channel and single output channel; similarly, one piece of data X (1*36) is taken from the data set as an input vector and a plurality of one-dimensional convolution kernels are used for convolution operation to obtain a plurality of convolution outputs, which is the case of single input channel and multiple output channels; further, the plurality of convolution outputs obtained are used for convolution operation with a plurality of convolution kernels to obtain a convolution result, which is the case of multiple input channels and multiple output channels. The convolution operation schematic diagrams of the three cases are shown in Figure 3 .

[0095] Step 4.2. Pooling operation:

[0096] The main features can be extracted and the operation complexity can be reduced, and the pooling operation is generally used after the convolution layer to pool the convolution result, which is divided into two categories of maximum pooling and mean pooling. The following takes (1*2) maximum pooling as an example to describe the pooling operation, and the maximum value of the first two elements of the output vector Z calculated in step 4.1 is taken as the first element of the pooling output vector; then, the pooling output vector is shifted to the right by two steps, that is, the maximum value of the 3rd and 4th elements of the output vector Z is taken as the second element of the pooling output vector; the vector Z is traversed in sequence with a step of 2 to obtain the pooling output; the pooling operation does not change the number of channels of the data;

[0097] Step 4.3. flatten operation:

[0098] After the convolution and pooling operations are performed on the input data, the output vector usually has more than one channel. In order to effectively connect the hidden layer of the neural network, the multi-channel output vector needs to be compressed and expanded. The flatten operation can connect the first and last one-dimensional vectors of the multi-channel output vector to form a one-dimensional vector, which is convenient for connection with the hidden layer.

[0099] Step 4.4. Forward propagation: Let the one-dimensional vector expanded by the flatten operation be X, the weight between the input layer and the hidden layer be w, the threshold value be b1, and the activation function be f1. Then the output net1 of the hidden layer is:

[0100]

[0101] where, where i and j represent the input neuron and hidden layer neuron numbers, respectively;

[0102] Take the output net1 of the hidden layer as the input of the next layer, the weight between the hidden layer and the output layer be v, the threshold value be b2, and the activation function be f2. Then the output net2 of the output layer is:

[0103]

[0104] where, where m and n represent the hidden layer neuron and output neuron numbers, respectively;

[0105] Step 4.5. Backward derivation: According to the final output net2 obtained in step 4.4 and the expected output, the error term E can be obtained. Then the gradient descent method is used to update the weights and thresholds.

[0106] According to the chain rule, the partial derivative of the error term E with respect to the hidden layer weight and threshold value is calculated

[0107]

[0108]

[0109] According to the chain rule, the partial derivative of the error term E with respect to the output layer weight and threshold value is calculated

[0110]

[0111]

[0112] According to the partial derivatives of the weights and thresholds of each layer obtained above, the values of the thresholds and weights are updated, and the update formula is as follows:

[0113]

[0114]

[0115]

[0116]

[0117] wherein η represents a learning rate, k = 1, 2,..., n represents the number of updates or iterations, k = 1 represents the first update, and so on.

[0118] Step 5. The established convolutional neural network has a structure as shown in Figure 4 The number of convolutional layers, the size of the convolution kernel, the number of hidden layer neurons, the activation function, the optimizer, the learning rate and other hyperparameters used in the application are determined as follows:

[0119] Step 5.1. Determine the number of convolutional layers and hidden layers. According to the data distribution characteristics and data features, the approximate interval of the number of layers is determined, and then the results are tested and compared in turn to finally determine the optimal combination as the number of convolutional layers and hidden layers both being 2.

[0120] Step 5.2. Determine the size of the convolution kernel. Increase from 1 to 5 with an increment of 1, and do 5 tests for each value to find the optimal number of convolution kernels as 3.

[0121] Step 5.3. Determine the number of hidden layer neurons. First, calculate the number of hidden layer neurons according to the empirical formula, and then determine the number of hidden layer neurons as 300 in combination with the actual number of neurons after the convolution layer is unfolded.

[0122] The formula is as follows:

[0123] n = log2T

[0124] wherein T is the number of input features, and n is the number of hidden layer neurons. Based on the calculated empirical value and the number of neurons after the convolution layer is unfolded, an interval is determined, and the optimal number of neurons n is determined by traversing the number of neurons in the interval.

[0125] Step 5.4. Determine the activation function between the fully connected layers. For the three commonly used Sigmoid, tanh and Relu functions, nine different combinations are determined by combining them in pairs, and their effects are tested respectively to select the optimal activation function combination as Relu, Relu.

[0126] Step 5.5. Determine the optimizer in the convolutional neural network. Use the four optimizers of SGD, MBGD, AdaGrad and Adam in turn for neural network training, and finally determine the optimal optimizer as Adam.

[0127] Step 5.6. Determine the learning rate of the neural network, starting from 0.001, each time the learning rate is increased to twice the original, and the range of 0.001-0.1 is traversed to determine the optimal learning rate of 0.002.

[0128] Step 6. The newly collected data is preprocessed, and then a convexity prediction model is established using a convolutional neural network, and the final performance of the model is tested using on-site rolling data. Based on the on-site data processed in step 3 and the hyperparameters of the convolutional neural network determined in step 5, the convolutional neural network is trained, and the trained convolutional neural network is the convexity prediction model. The newly collected rolling data is preprocessed and input into the convexity prediction model (input is 35-dimensional data excluding the convexity after the finishing mill F6) to obtain the predicted value of the convexity of the strip after the finishing mill F6.

[0129] For industrial production, the actual convexity value is considered to be a qualified product within the interval of 10 μm above and below the set value. The existing method combining neural network and genetic algorithm has 94.42% of the predicted values meeting this interval. The prediction model based on the convolutional neural network proposed in the present application has 96.29% of the predicted values meeting the requirements, as shown in Table 2. Figure 5

[0130] The above technical solutions set forth the technical idea of the present application, and cannot limit the protection scope of the present application. Any modification and modification of the above technical solutions according to the technical essence of the present application, without departing from the content of the technical solutions of the present application, belongs to the protection scope of the technical solutions of the present application.​

Claims

1. A method for predicting the crown of hot-rolled strip steel by fusing data and mechanisms, characterized in that, Includes the following steps: Step 1. Establish a mechanism model: Collect production data of the hot strip rolling process and establish a mechanism model of hot-rolled strip crown; Step 2. Feature selection: Select the feature data that affects the crown from the collected production data and the calculation data of the hot-rolled strip crown mechanism model; Step 3. Data Preprocessing: Perform data preprocessing on the collected production data, including data smoothing and data standardization; Step 4. Construct a neural network: Build a convolutional neural network based on the dataset obtained after feature selection in Step 2 and data preprocessing in Step 3; Step 5. Determine the hyperparameters of the neural network: Determine the number of convolutional layers, kernel size, number of hidden layer neurons, activation function, optimizer, and learning rate; Step 6. Convexity Prediction: Train a neural network model to predict the convexity value of the strip; Step 1, based on the collected production data, establishes a strip crown mechanism model, specifically as follows: Step 1.

1. The mathematical formula for the crown mechanism model of hot-rolled strip is as follows: In the formula, The crown of the strip at the mill exit. For strip width, For rolling force, For the bending roller force, For the crown of the work roll, For the wear crown of the work roll, For the thermal crown of the work roll, For roller displacement equivalent convexity, This is a correction factor for the strip width. This is the lateral stiffness correction factor. This is the correction factor for the bending roller force. This is the correction factor for the crown of the work roll. This is the roller shift position correction factor. Genetic correction factor for strip crown. This refers to the number of rolling mills in the finishing mill unit; Step 1.

2. The thermal crown of the work roll caused by thermal expansion is calculated using the following formulas according to different working conditions: When the temperature gradient in the radial direction of the work roll is large during the initial stage of rolling, the calculation formula is: When a certain amount of strip steel has been rolled and the work rolls reach thermal equilibrium, the calculation formula is as follows: In the formula, The average temperature rise within the cross-section of the work roll. The coefficient of linear expansion of the work roll material is... Poisson's ratio, This is the current temperature of the work roll. The initial temperature of the working roll; in The calculation formula is: In the formula, The radius of the work roll, This is a variable along the radial direction of the work roll.

2. The method for predicting the crown of hot-rolled strip steel by fusing data and mechanisms according to claim 1, characterized in that, The production data in step 2 includes: intermediate billet thickness, average thickness and average width during rolling, strip temperature at the entrance of the finishing mill, strip temperature at the exit of the finishing mill, rolling speed of each mill in the finishing mill, rolling force of each mill, bending force of each mill, amount of roll shifting of each mill, rolling mileage, and the roll crown of the last mill in the finishing mill. The calculation data based on the hot-rolled strip crown mechanism model is the roll crown of each mill in the finishing mill except for the last mill.

3. The method for predicting the crown of hot-rolled strip steel by fusing data and mechanisms according to claim 1, characterized in that, The smoothing method used in step 3 is the five-point cubic smoothing method.

4. The method for predicting the crown of hot-rolled strip steel by fusing data and mechanisms according to claim 1, characterized in that, The specific implementation method of each module of the convolutional neural network in step 4 is as follows: Step 4.

1. One-dimensional convolution operation: Read data sequentially from the dataset processed in Step 3. (1*) () as input vector and one-dimensional convolution kernel (1*) Perform a convolution operation to obtain the convolution result; where: The dimension of the data collected in step 2, feature selection. < And take odd numbers; Step 4.

2. Pooling operation: Perform pooling operation on the result of the convolution operation; Step 4.

3. Flatten operation: The flatten operation is used to connect the first and last ends of the multi-channel one-dimensional vector to form a one-dimensional vector, which is then connected to the hidden layer of the neural network. Step 4.

4. Forward Propagation: Let the one-dimensional vector expanded by the flatten operation be... The weights between the input layer and the hidden layer are The threshold is The activation function is Then the hidden layer output for: Where i and j represent the input neuron and hidden layer neuron numbers, respectively; Output the hidden layer As the input to the next layer, the weights between the hidden layer and the output layer are: The threshold is The activation function is Then the output layer outputs... for: Where m and n represent the numbers of hidden layer neurons and output neurons, respectively; Step 4.

5. Inverse Differentiation: Based on the final output obtained in Step 4.4 The error term can be obtained by comparing it with the expected output. Then, the gradient descent method is used to update the weights and thresholds.

5. The method for predicting the crown of hot-rolled strip steel by fusing data and mechanisms according to claim 1, characterized in that, The specific method for updating weights and thresholds using gradient descent is as follows: Calculate the error term using the chain rule. Partial derivatives with respect to hidden layer weights and thresholds , : Calculate the error term using the chain rule. Partial derivatives with respect to output layer weights and thresholds , : Based on the partial derivatives of the weights and thresholds of each layer calculated above, the values ​​of the thresholds and weights are updated using the following formula: in Indicates the learning rate. Indicates the number of updates or iterations. This indicates the first update, and so on.

6. The method for predicting the crown of hot-rolled strip steel by fusing data and mechanisms according to claim 1, characterized in that, The specific method for step 5 is as follows: Step 5.

1. Determine the number of convolutional layers and hidden layers. Based on the characteristics of the data in the dataset and the distribution of the data, determine the approximate range of the number of convolutional layers and hidden layers. Then, iterate through the range of values ​​and combine them in pairs to select the optimal number of convolutional layers and hidden layers. Step 5.

2. Determine the kernel size in the convolution operation by traversing and searching from 1 to 5 to determine the kernel size; Step 5.

3. Determine the number of neurons in the hidden layer of the neural network. First, calculate the empirical value of the number of neurons using an empirical formula, as follows: Where T is the number of input features and n is the number of hidden layer neurons; Step 5.

4. Determine the activation functions between fully connected layers. For commonly used activation functions such as Sigmoid, tanh, and ReLU, test each combination in pairs to determine the optimal activation function. Step 5.

5. Determine the optimizer for the neural network. Iterate through several commonly used optimizers in neural network training: SGD, MBGD, AdaGrad, and Adam, and find the optimal optimizer. Step 5.

6. Determine the learning rate of the neural network by iterating through the range of 0.001 to 0.1 to determine the optimal learning rate lr.

7. The method for predicting the crown of hot-rolled strip steel by fusing data and mechanisms according to claim 1, characterized in that, The specific implementation method of step 6 is as follows: based on the field data processed in step 3 and the hyperparameters of the convolutional neural network determined in step 5, the convolutional neural network is trained. The trained convolutional neural network is the convexity prediction model. The newly collected rolling data is preprocessed and then input into the convexity prediction model to obtain the convexity prediction value of the strip.

Images