Method for constructing standard pattern curve of strip shape of twenty-high rolling mill for silicon steel

Abstract

The invention relates to the technical field of cold-rolled silicon steel production strip shape control, in particular to a silicon steel twenty-high roll mill strip shape standard mode curve construction method which comprises the steps of collecting process data and conducting preprocessing, fitting a model through training data, calculating a model output value, and then conducting traversal search in the interval of + / -15% of the output value. Obtaining a hidden layer neuron node number n; traversing and searching 0.1-0.4, and determining the learning rate lr of the neural network of the hidden layer; combining the purelin transfer function, the logsig transfer function and the sigmod transfer function to generate three different combinations, and selecting the transfer functions of the double-hidden-layer network and the output layer; determining a training function of the hidden layer neural network by adopting a traversal method; optimizing the hidden layer neural network model by adopting a fast non-dominated sorting genetic algorithm, decoding a weight and a threshold value of a neural network, endowing the neural network model, and outputting the neural network model optimized by the non-dominated sorting genetic algorithm; and the standard pattern curve of the silicon steel twenty-high roll mill plate shape is constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of plate shape control technology of cold-rolled silicon steel production, and particularly relates to a method for constructing a plate shape standard mode curve of a silicon steel twenty-high rolling mill. BACKGROUND

[0002] With the development of high-grade silicon steel products in the direction of thinner and smaller iron loss, users have higher requirements for the dimensional accuracy and shape accuracy of the strip, therefore, the research and development of an intelligent plate shape system with fast response and high control accuracy has become a key task for improving the shape control level of the twenty-high rolling mill. Since the rolling deformation of the strip involves elastic and plastic deformation, internal structure and mechanical properties during the production process, it brings challenges in precision, stability and adaptability to data modeling. From the cold-rolled plate rolling process, the post-rolling plate shape curve is the residual stress deviation distribution curve of the post-rolling strip along the transverse direction (plate width direction). Since the plate shape control is composed of standard curves through pattern recognition, the control accuracy is largely dependent on the fitting accuracy of the standard curve.

[0003] Since the neural network has the characteristics of approximating any nonlinear mapping relationship, the present application constructs a standard mode curve using a BP neural network with double hidden layers. The literature discloses a study on high-precision wide strip cold rolling mill plate shape fuzzy neural control, which selects the Legendre polynomial method to construct a standard normalized equation to represent left waves, right waves, middle waves, double-sided waves, four-part waves and edge middle waves. The present application mainly considers two points: (1) based on the characteristics of BP neural network with arbitrary approximation of nonlinear mapping relationship and data-driven optimization, a neural network with double hidden layers is constructed to construct a plate shape standard mode curve; (2) in order to lay a foundation for accurate control of plate shape, the constructed standard mode curve is more detailed and diverse. SUMMARY

[0004] The present application provides a method for constructing a plate shape standard mode curve of a silicon steel twenty-high rolling mill, which is based on the characteristics of neural network with arbitrary approximation of nonlinear mapping relationship, and realizes more kinds and more accurate typical wave shape standard mode curves by constructing a neural network with double hidden layers, which can accurately calculate the standard mode curve expression of complex wave shape, and lays a foundation for accurate control of plate shape.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0006] A method for constructing a plate shape standard mode curve of a silicon steel twenty-high rolling mill, comprising the following steps:

[0007] Step 1: collecting process data of silicon steel production;

[0008] Step 2: preprocessing the collected data;

[0009] Step 3: Establishing a 4-layer BP neural network model, including: using training data to fit the model, performing model training or predicting task to calculate output value, then performing traversal search in the interval of output value ± 15%, obtaining the number of hidden layer neuron nodes n; performing traversal search of 0.1-0.4 to determine the learning rate lr of the hidden layer neural network; combining purelin, logsig and sigmod transfer functions to produce three different combinations and selecting the transfer functions of the double hidden layer network and the output layer; and using the traversal method to determine the training function of the hidden layer neural network;

[0010] Step 4: Optimizing the hidden layer neural network model by using the fast non-dominated sorting genetic algorithm;

[0011] Step 5: Decoding the weight and threshold values of the neural network, and assigning the neural network model, outputting the non-dominated sorting genetic algorithm optimized neural network model.

[0012] Further, the process data includes an actual plate shape curve, a silicon steel variety, a strip steel width, a strip steel thickness, a strip steel speed and an intermediate roll shape curve.

[0013] Further, the preprocessing of the collected data includes calculating the average value and the standard deviation of each process data, eliminating the abnormal values of the data, performing smoothing processing on the data after eliminating the abnormal values, and performing normalization processing on the data after the smoothing processing.

[0014] Further, step 3 specifically includes the following steps:

[0015] (1) Network structure: 4-layer BP neural network with 16 inputs and 10 outputs, the hidden layer is set to 10 nodes, the first hidden layer activation function is purelin and sigmoid function, the second hidden layer activation function is purelin and logsig function, and the output layer is purelin function;

[0016] (2) Initialization: setting all the weighted coefficients to the smallest random number, i.e. the function (rand(x) / 32767)*((W max )-(W min ))+(W min ), wherein W max =0.3, W min =-0.3, and setting all the bias initial values to 0; wherein rand(x) is a random function, W max is the maximum threshold value, and W min is the minimum threshold value;

[0017] (3) Providing a training set: giving an input vector X={x 1 ,x2 ,…,x 16} and the expected output vector T = {t 1 ,t 2 ,…,t 10}; x is the element of the input vector, and t is the element of the output vector;

[0018] (4) Calculate the actual output:

[0019]

[0020] in, w is the input to the p-th neuron in the i-th hidden layer. ij The weights from the input layer to the i-th hidden layer are... For input, θ i Let M be the bias of the i-th hidden layer, M be the number of neurons in the i-th hidden layer, i be the i-th layer, and j be the input layer. Let be the activation function of the i-th hidden layer. For the input of the output layer, w kj For output layer weights, For the hidden layer output, θ k The output layer bias is q, where q is the number of neurons in the output layer.

[0021] (5) Calculate the error between the expected value and the actual output:

[0022] in, To output the expected value, To output the actual values, L is the number of output values;

[0023] (6) Adjust the weight coefficients of the output layer and hidden layer:

[0024]

[0025] Among them, w ki (k+1) represents the (k+1)th adjustment weight of the output layer, w ki (k) represents the k-th adjustment weight of the output layer, and η is the step size. For the output layer gradient, w ij (k+1) represents the weight of the hidden layer after the (k+1)th adjustment, w ij (k) represents the k-th adjustment weight of the hidden layer. For hidden layer gradients;

[0026] (7) Return to step (3) until the error meets the requirement of 0.001.

[0027] Furthermore, step 4 specifically includes the following steps:

[0028] 1) Set the maximum number of generations G max ;

[0029] 2) Train the hidden layer neural network model according to the training function, then input the preprocessed production data into the trained hidden layer neural network model to make predictions and output the prediction results;

[0030] 3) Calculate the training error of the two hidden layer neural network model;

[0031] 4) Apply binary encoding to convert the weights and thresholds of the two hidden layer neural network into binary numbers;

[0032] 5) Randomly generate a population of n individuals, representing the set of all solutions;

[0033] 6) Perform a fast non-dominated sort on the individuals in the population. Determine whether the non-dominated sorts of two individuals in the population are the same. If they are the same, select the individual with the smaller crowding and continue to step 7). If they are different, select the individual with the smaller non-dominated sort level and continue to step 5.

[0034] 7) The selected individuals are grouped into a new population and genetic operations are performed;

[0035] 8) Determine if the current evolution number G is less than the maximum evolution number G. max If yes, return to step 2; otherwise, output the weights and thresholds of the optimized hidden neural network model and continue to step 5.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1) By applying a BP neural network to establish 12 standard pattern curves, steel strip defects can be classified into center waves, edge waves, edge-center waves, quarter waves, tight edge waves, and broken edge waves. Since neural networks can approximate arbitrary nonlinear mapping relationships, a BP neural network with two hidden layers is used to construct the standard pattern curves, which can more accurately express the curve characteristics.

[0038] 2) In view of the complex characteristics of the 20-roll mill shape control, the segmented pressing and intermediate roll adjustment characteristics of AS-U are better simulated. By constructing more detailed and diversified standard mode curves, a solid foundation is laid for precise shape control.

[0039] 3) Achieve high-precision control of the plate shape of high-grade silicon steel rolled by a 20-roll mill, reducing the plate shape deviation from 50i in raw materials to less than 10i in finished products, improving the quality and output of high-grade and oriented silicon steel products, and generating an annual benefit of approximately 600,000 yuan. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the construction of the standard pattern curve of the plate shape described in this invention.

[0041] Figure 2 This is a wave curve diagram in the plate-shaped pattern 1 constructed by the neural network described in this invention.

[0042] Figure 3 This is the edge wave curve diagram of the plate-shaped mode 2 constructed by the neural network described in this invention.

[0043] Figure 4 This is a wave curve diagram of the three sides of the plate-shaped pattern constructed by the neural network described in this invention.

[0044] Figure 5 This is a diagram of the four-wave curve of the plate-shaped pattern constructed by the neural network described in this invention.

[0045] Figure 6 This is a diagram of the broken edge wave curve of the plate shape mode 5 constructed by the neural network described in this invention.

[0046] Figure 7 This is the tight-edge curve diagram of the plate-shaped pattern 6 constructed by the neural network described in this invention.

[0047] Figure 8 This is a single-sided wave curve diagram of the plate-shaped pattern 7 constructed by the neural network described in this invention.

[0048] Figure 9 This is a diagram of the eight side wave curves of the plate-shaped pattern constructed by the neural network described in this invention.

[0049] Figure 10 This is a wave curve diagram in the plate-shaped pattern 9 constructed by the neural network described in this invention.

[0050] Figure 11 This is a wave curve diagram of the plate-shaped pattern 10 constructed by the neural network described in this invention.

[0051] Figure 12 This is a wave curve diagram in the plate-shaped pattern 11 constructed by the neural network described in this invention.

[0052] Figure 13 This is a wave curve diagram of the plate-shaped pattern 12 constructed by the neural network described in this invention. Detailed Implementation

[0053] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0054] See Figure 2 This is a schematic diagram illustrating the construction of the standard pattern curve for plate shape in this invention. This invention provides a method for constructing the standard pattern curve for plate shape in a 20-roll mill for silicon steel, which adds multiple typical wave-shaped standard curves and accurately calculates the standard pattern curve expression of composite waves. Specifically, it includes the following steps:

[0055] Step 1: Collect process data for silicon steel production. Collect production data for a hot-dip galvanizing process, including target coating thickness, coating type, strip width, strip thickness, and strip speed, totaling 200 steel coils. The data contains 22 features. Select the 5 features that have the greatest impact on air knife distance for further study.

[0056] Step 2: Preprocess the collected data, including: calculating the average and standard deviation of each process data, removing outliers, smoothing the data after removing outliers, and normalizing the smoothed data.

[0057] (1) Calculate the mean and standard deviation of each production data;

[0058] In this embodiment, the average value of each production data is calculated. and standard deviation S x See formulas (1) and (2):

[0059]

[0060] Where m is the number of data points, x i This is the i-th process data;

[0061] (2) Remove outliers from the data;

[0062] The method for removing outliers in production data is the Layda criterion, which removes outliers from production data that meet the formula (3).

[0063]

[0064] (3) Smooth the production data after removing outliers;

[0065] The production data after outlier removal was smoothed using the 5-point 3-fold smoothing method, as shown in formula (4).

[0066]

[0067] Among them, Y i For the i-th data after removing outliers, For Y i Smoothed values;

[0068] (4) Normalize the smoothed data using the minimum-maximum normalization method, as shown in formula (5):

[0069]

[0070] in, Let x be the normalized value of the i-th data point. i Let x be the initial value of the i-th data. min x is the minimum value in the dataset. max This represents the maximum value in the dataset.

[0071] Step 3: Establish a 4-layer BP neural network model, including: fitting the model using training data, performing model training or prediction tasks, calculating the model output value, and then performing a traversal search within the range of ±15% of the output value to obtain the number of hidden layer neurons n; traversing the search range from 0.1 to 0.4 to determine the learning rate lr of the hidden layer neural network; combining purelin, logsig, and sigmoid transfer functions to generate three different combinations and selecting the transfer functions of the two hidden layer network and the output layer; and using a traversal method to determine the training function of the hidden layer neural network.

[0072] Specifically, establishing a two-hidden-layer neural network model includes:

[0073] (1) Network structure: 4-layer BP neural network with 16 inputs and 10 outputs, with 10 hidden layers. The activation functions of the first hidden layer are purelin and sigmoid functions, the activation functions of the second hidden layer are purelin and logsig functions, and the output layer is purelin function.

[0074] (2) Initialization: Set all weighting coefficients to the smallest random number, i.e., the function (rand(x) / 32767)*((W max )-(W min ))+(W min ), where W max =0.3, W min = -0.3, set all bias initial values ​​to 0; where rand(x) is a random function, W max For the maximum threshold, W min Minimum threshold;

[0075] (3) Provide a training set: Provide the sequentially assigned input vector X = {x 1 ,x 2 ,…,x 16} and the expected output vector T = {t 1 ,t 2 ,…,t 10}; x is the element of the input vector, and t is the element of the output vector;

[0076] (4) Calculate the actual output:

[0077]

[0078] in, w is the input to the p-th neuron in the i-th hidden layer. ij The weights from the input layer to the i-th hidden layer are... For input, θ i Let M be the bias of the i-th hidden layer, M be the number of neurons in the i-th hidden layer, i be the i-th layer, and j be the input layer. Let be the activation function of the i-th hidden layer. For the input of the output layer, w kj For output layer weights, For the hidden layer output, θ k The output layer bias is q, where q is the number of neurons in the output layer.

[0079] (5) Calculate the error between the expected value and the actual output:

[0080] in, To output the expected value, To output the actual values, L is the number of output values;

[0081] (6) Adjust the weight coefficients of the output layer and hidden layer:

[0082]

[0083] Among them, w ki (k+1) represents the (k+1)th adjustment weight of the output layer, w ki (k) represents the k-th adjustment weight of the output layer, and η is the step size. For the output layer gradient, w ij (k+1) represents the weight of the hidden layer after the (k+1)th adjustment, w ij (k) represents the k-th adjustment weight of the hidden layer. For hidden layer gradients;

[0084] (7) Return to step (3) until the error meets the requirement of 0.001.

[0085] Step 4: Optimize the hidden layer neural network model using a fast non-dominated sorting genetic algorithm;

[0086] 1) Set the maximum number of generations G max =200;

[0087] 2) Train the hidden layer neural network model according to the training function, then input the preprocessed production data into the trained hidden layer neural network model to make predictions and output the prediction results;

[0088] 3) Calculate the training error of the two hidden layer neural network model;

[0089] 4) Use binary encoding to convert the weights and thresholds of a single hidden layer neural network into a string of numbers consisting of 0s and 1s;

[0090] 5) Randomly generate a population of n individuals to represent the set of all solutions; in this embodiment, we test the population size of 30, 40, 50, and 60, and iterate through the population in turn to determine the optimal population size as 50, i.e., let n = 50;

[0091] 6) Perform a fast non-dominated sort on the individuals in the population. The specific steps are as follows:

[0092] Step 4.6.1: For each individual i in the population, examine the set of individuals S it governs. i Find the number of individual solutions n i Find individuals with a value of 0 and store them in the current set F1;

[0093] Step 4.6.2: For each individual j in the current set F1, examine the set of individuals S it governs. j Find the number of individual solutions n k For individuals where -1 = 0, the individuals are stored in another set H1;

[0094] Step 4.6.3: Treat the current set F1 as the first-level non-dominated individual set, and assign the same non-dominated order i to the individuals in this set. rank Then, continue to perform the above hierarchical operation on set H1 and assign the corresponding non-dominated order until all individuals are hierarchically classified.

[0095] Step 4.6.4: Calculate two attributes for each individual in the population: non-dominated order i rank and crowding level i d ;

[0096] Step 4.6.5: Determine whether the non-dominated order of two individuals in the population is the same. If they are the same, select the individual with the smaller crowding and continue with step (7). If they are different, select the individual with the smaller non-dominated order level and continue with step (7).

[0097] 7) The selected individuals are grouped into a new population and genetic operations are performed;

[0098] 8) Determine if the current evolution number G is less than the maximum evolution number G. max If yes, return to step 2; otherwise, output the weights and thresholds of the optimized hidden neural network model and continue to step 5.

[0099] Step 5: Decode the weights and thresholds of the neural network and assign them to the neural network model, outputting a neural network model optimized by a non-dominated sorting genetic algorithm.

[0100] likeFigures 2 to 13 As shown, this ultimately results in 12 standard plate-shaped pattern curves.

[0101] The above embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the above embodiments. Unless otherwise specified, the methods used in the above embodiments are conventional methods.

Claims

1. A method for constructing a standard pattern curve for the shape of a silicon steel 20-roll mill, characterized in that, Includes the following steps: Step 1: Collect process data for silicon steel production; Step 2: Preprocess the collected data; Step 3: Establish a 4-layer BP neural network model, including: fitting the model using training data, performing model training or predicting tasks to calculate output values, and then performing a traversal search within the range of ±15% of the output values ​​to obtain the number of hidden layer neurons n; traversing the search range from 0.1 to 0.4 to determine the learning rate lr of the hidden layer neural network; combining purelin, logsig, and sigmoid transfer functions to generate three different combinations and selecting the transfer functions of the two hidden layer network and the output layer; and using a traversal method to determine the training function of the hidden layer neural network. Step 4: Optimize the hidden layer neural network model using a fast non-dominated sorting genetic algorithm; Step 5: Decode the weights and thresholds of the neural network and assign them to the neural network model, outputting a neural network model optimized by a non-dominated sorting genetic algorithm.

2. The method for constructing a standard pattern curve for silicon steel 20-roll mill according to claim 1, characterized in that, The process data includes the actual strip shape curve, silicon steel type, strip width, strip thickness, strip speed, and intermediate roll shape curve.

3. The method for constructing a standard pattern curve for silicon steel 20-roll rolling mill according to claim 1, characterized in that, The preprocessing of the collected data includes calculating the average value and standard deviation of each process data, removing outliers, smoothing the data after removing outliers, and normalizing the smoothed data.

4. The method for constructing a standard pattern curve for silicon steel 20-roll rolling mill according to claim 1, characterized in that, Step 3 specifically includes the following steps: (1) Network structure: 4-layer BP neural network with 16 inputs and 10 outputs, with 10 hidden layers. The activation functions of the first hidden layer are purelin and sigmoid functions, the activation functions of the second hidden layer are purelin and logsig functions, and the output layer is purelin function. (2) Initialization: Set all weighting coefficients to the smallest random number, i.e., the function (rand(x) / 32767)*((W max )-(W min ))+(W min ), where W max =0.3, W min = -0.3, set all bias initial values ​​to 0; where rand(x) is a random function, W max For the maximum threshold, W min Minimum threshold; (3) Provide a training set: Provide the sequentially assigned input vector X = {x 1 ,x 2 ,…,x 16 } and the expected output vector T = {t 1 ,t 2 ,…,t 10 }; x is the element of the input vector, and t is the element of the output vector; (4) Calculate the actual output: in, w is the input to the p-th neuron in the i-th hidden layer. ij The weights from the input layer to the i-th hidden layer are... For input, θ i Let M be the bias of the i-th hidden layer, M be the number of neurons in the i-th hidden layer, i be the i-th layer, and j be the input layer. Let be the activation function of the i-th hidden layer. For the input of the output layer, w kj For output layer weights, For the hidden layer output, θ k The output layer bias is q, where q is the number of neurons in the output layer. (5) Calculate the error between the expected value and the actual output: in, To output the expected value, To output the actual values, L is the number of output values; (6) Adjust the weight coefficients of the output layer and hidden layer: Among them, w ki (k+1) represents the (k+1)th adjustment weight of the output layer, w ki (k) represents the k-th adjustment weight of the output layer, and η is the step size. For the output layer gradient, w ij (k+1) represents the weight of the hidden layer after the (k+1)th adjustment, w ij (k) represents the k-th adjustment weight of the hidden layer. For hidden layer gradients; (7) Return to step (3) until the error meets the requirement of 0.

001.

5. The method for constructing a standard pattern curve for silicon steel 20-roll mill shape according to claim 1, characterized in that, Step 4 specifically includes the following steps: 1) Set the maximum number of generations G max ; 2) Train the hidden layer neural network model according to the training function, then input the preprocessed production data into the trained hidden layer neural network model to make predictions and output the prediction results; 3) Calculate the training error of the two hidden layer neural network model; 4) Apply binary encoding to convert the weights and thresholds of the two hidden layer neural network into binary numbers; 5) Randomly generate a population of n individuals, representing the set of all solutions; 6) Perform a fast non-dominated sort on the individuals in the population. Determine whether the non-dominated sorts of two individuals in the population are the same. If they are the same, select the individual with the smaller crowding and continue to step 7). If they are different, select the individual with the smaller non-dominated sort level and continue to step 5. 7) The selected individuals are grouped into a new population and genetic operations are performed; 8) Determine if the current evolution number G is less than the maximum evolution number G. max If yes, return to step 2; otherwise, output the weights and thresholds of the optimized hidden neural network model and continue to step 5.

Images