A strip full-length thickness prediction method based on a memory function network

Abstract

A strip full-length thickness prediction method based on a memory function network of the present invention includes: collecting real-time process rolling data and measured values of strip outlet thickness at equal time intervals to form an input feature data set; dividing the input feature data set into a training set and a test set according to a ratio and in chronological order and performing normalization processing; constructing a GLSTM network for strip thickness point prediction, and training to obtain the connection weights and biases of each layer of the GLSTM network for point prediction; using a particle swarm algorithm to optimize the connection weights of the output layer of the GLSTM network for point prediction to obtain the optimal connection weights W of the output layer opt ; based on the connection weights and biases of each layer except the output layer and W opt , constructing a GLSTM network for strip thickness interval prediction, determining the optimal parameters of the GLSTM network for interval prediction according to evaluation indexes; using the optimal GLSTM network for interval prediction to perform strip full-length thickness prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of rolling process control, and relates to a method for predicting the full-length thickness of strip steel based on a memory function network. Background Art

[0002] As an important rolled product, strip steel is widely used in many fields of the national economy, and the demand for strip steel is also increasing day by day. With the rapid development of modern social science and technology and production, the industry has higher and higher requirements for the quality of strip steel. The strip steel thickness has become an important standard for measuring the performance of strip steel. The strip steel thickness deviation may lead to serious process interruptions, and result in defects and unqualified products, ultimately leading to product rejection. It can be seen that the strip steel thickness is directly related to the quality of the entire strip steel product and the economic benefits of the enterprise. Therefore, the high-precision prediction of strip steel thickness is particularly important. However, the factors affecting the strip steel thickness are diverse, and each influencing factor has the characteristics of non-linearity and strong coupling. The establishment of the model is very complex, and it is difficult to complete the high-precision prediction of strip steel thickness based on traditional mathematical methods. In the context of the current application of industrial big data, relying on machine learning prediction algorithms to establish an accurate strip steel thickness prediction model provides a new method and idea for solving such problems.

[0003] Neural networks have strong non-linear processing capabilities, so they have been widely used in process control research. In recent years, some scholars have successively proposed to use neural networks for the prediction of strip steel thickness, such as the prediction of strip steel thickness based on BP neural networks, the prediction of strip steel thickness based on the OSELM algorithm, etc. At the same time, scholars have also used optimization algorithms such as whale algorithm and sparrow search algorithm to improve the accuracy of the prediction model. However, the current prediction research focuses on the prediction of the strip steel head thickness, and there is no research on the thickness prediction in the full-length direction of the entire coil of strip steel. However, this kind of point prediction of the strip steel head thickness itself has certain defects, that is, the prediction model only provides a single prediction value, lacking information related to accuracy. That is to say, point prediction cannot express the probability of correct prediction, and lacks the characterization and analysis of prediction errors. The mechanism analysis of the rolling process shows that the rolling process of strip steel has strong time continuity, and the strip steel continuously passes through the rolling mill along the length direction. Therefore, the full-length thickness of strip steel can be expressed as the strip steel outlet thickness at different times. And the strip steel outlet thickness at a certain moment is related to the rolling process and thickness at the previous moment, and shows strong time series characteristics. However, the existing neural network models have weak ability to capture the non-linear deep features of strip steel data and the time dependence between data, and the prediction accuracy is not high, and the interval prediction of strip steel thickness cannot be realized. 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 full-length thickness of strip steel based on a memory function network.

[0005] A strip full-length thickness prediction method based on a memory function network of the present invention includes:

[0006] Step 1: Collect real-time process rolling data and measured values of strip exit thickness at equal time intervals to form an input feature data set;

[0007] Step 2: Divide the input feature data set into a training set and a test set according to a ratio and in chronological order, and normalize the data in the training set and the test set;

[0008] Step 3: Construct a GLSTM network for strip thickness point prediction, and train the GLSTM network for point prediction through the training set to obtain the connection weights and biases of each layer of the GLSTM network for point prediction;

[0009] Step 4: Use the particle swarm optimization algorithm to optimize the connection weights of the output layer of the GLSTM network for point prediction in Step 3 to obtain the optimal connection weight W opt ;

[0010] Step 5: Based on the connection weights and biases of each layer except the output layer obtained in Step 3 and W opt obtained in Step 4, construct a GLSTM network for strip thickness interval prediction, and determine the optimal parameters of the GLSTM network for interval prediction according to the evaluation indexes PICP, PINAW and CWC to obtain the optimal GLSTM network for interval prediction;

[0011] Step 6: Use the optimal GLSTM network for interval prediction obtained in Step 5 to predict the strip full-length thickness, and calculate the PICP, PINAW and CWC evaluation indexes according to the upper and lower limits of the predicted strip thickness to complete the performance evaluation of the GLSTM network for interval prediction.

[0012] Further, the real-time process rolling data includes: rolling kilometers, inlet thickness, and real-time process rolling data of 34 important influencing factors such as rolling speed, rolling force, front and rear tension, bending roll force, and roll gap of each stand.

[0013] Further, in Step 2, the data in the training set and the test set are normalized according to the following formula:

[0014]

[0015] In the formula: x′ k is the normalized feature data, x k is the kth feature data, x min is the minimum value in this feature data, x max is the maximum value in this feature data.

[0016] Further, the GLSTM network for strip thickness point prediction in step 3 is specifically as follows:

[0017]

[0018] f T = 1 - i T

[0019]

[0020]

[0021] c T = f T × h T-1 + g T × i T

[0022] h T = o T × Relu(c T )

[0023] In the formula, x T is the input data at time T; is the weight related to the input x, is the weight of the input gate related to the input data x, is the weight of the output gate related to the input data x, is the weight of the new information vector related to the input data x; is the weight related to the hidden state h T-1 at the previous time step; b (i) is the bias of the input gate, b (o) is the bias of the output gate, b (g) is the bias of the new information vector; i T represents the input gate, f T represents the forget gate, o T represents the output gate, g T represents the new information vector, c T represents the cell state of the network at time T, h T represents the hidden layer state of the network at time T; Iσ represents the Isigmoid activation function, and both LRelu and Relu are activation functions.

[0024] Further, in step 4, the connection weights of the output layer of the GLSTM for point prediction in step 3 are optimized using the particle swarm algorithm, specifically as follows:

[0025] Step 4.1: The connection weight W oPerform replication to obtain an \(n\times2\) \(W\). ho And add a random matrix with the same dimension whose elements are in the range of \([-1.1, 0.1]\) to obtain \(W\). no ;

[0026] Step 4.2: Initialization of the population. Randomly initialize the velocity of each particle in the solution space of the particle, and use \(W\) to initialize the position of each particle. no Initialization;

[0027] Step 4.3: Use the CWC evaluation index as the fitness function to evaluate the fitness function value of each particle.

[0028] Step 4.4: For each particle, compare its current fitness value with the fitness value corresponding to its individual historical best position. If the current fitness value is better, use the current position to update the individual historical best position of the particle.

[0029] Step 4.5: For each particle, compare its current fitness value with the fitness value corresponding to the global best position. If the current fitness value is better, use the current position to update the historical optimal position of the particle swarm.

[0030] Step 4.6: Update the velocity and position of the particle according to the following formula:

[0031]

[0032]

[0033] Where, \(Pbest\) j represents the historical best position of the \(j\)-th particle, \(Gbest\) represents the historical best position of the entire population, \(c1\) and \(c2\) are learning factors, \(r1\) and \(r2\) are uniformly distributed random numbers in the range of \([0, 1]\), and \(t\) is the number of iteration steps; when the velocity iteration update is greater than the maximum value \(V\) of its limit range max , make it equal to \(V\) max ; Similarly, when it is less than the minimum value \(V\) of the limit range min , make it equal to \(V\) min ;

[0034] Step 4.7: If the maximum number of iterations is satisfied or the value of the fitness function CWC remains unchanged for \(M\) times during the iteration process, output the global best position, that is, the connection weight \(W\) of the best output layer. opt ; If the iteration termination condition is not satisfied, go to Step 4.3 to continue the iterative update.

[0035] Further, the optimal parameters of the GLSTM network for interval prediction in step 5 include: the number of hidden layer nodes, the combination of activation functions for each layer, the network learning rate, and the parameters to be determined by the particle swarm optimization algorithm. The parameters to be determined by the particle swarm optimization algorithm include: learning factors c1 and c2, and the population size.

[0036] Further, the definitions of the PICP, PINAW, and CWC evaluation metrics are as follows:

[0037] (1) The coverage probability PICP is an important feature of interval prediction, representing the probability that the target value lies within the interval. Its definition is as follows:

[0038]

[0039] where N represents the total number of samples, and ∈ k is a boolean decision variable used to indicate whether the k-th target has been covered by the interval. If the target is covered by the interval, then ∈ k = 1; otherwise, ∈ k = 0. Therefore, the mathematical formula for ∈ k is defined as follows:

[0040]

[0041] where y k represents the k-th target, i.e., the measured value of the strip outlet thickness; L k and U k represent the lower and upper bounds of the prediction interval generated for y k respectively. To ensure the accuracy of interval prediction and obtain an effective prediction interval, the coverage probability PICP is required to be not lower than the nominal confidence level 1-α of the prediction interval, where α represents the significance level; otherwise, the prediction interval is invalid and should be discarded.

[0042] (2) The quantitative measurement of the interval prediction width is defined as the normalized average width of the prediction interval PINAW, which is expressed by the following mathematical formula:

[0043]

[0044] where R g refers to the range of change of the target, i.e., the difference between the maximum and minimum values of the target.

[0045] (3) Define a comprehensive cost function based on interval prediction, called the standard of coverage width CWC, for optimizing the above two evaluation metrics. The specific definition is as follows:

[0046] CWC = PINAW(1 + γ(PICP)e -η(PICP-μ) )

[0047]

[0048] Among them, η and μ are two parameters used to control the optimization speed and accuracy. η is used as a penalty factor to penalize the prediction intervals with poor quality. If the PICP does not reach the pre-specified confidence level, the CWC will increase exponentially in this term. If the PICP reaches the pre-specified confidence level, the CWC will no longer focus on the PICP but only on the PINAW; μ corresponds to the confidence level related to the interval prediction and is set to 1 - α; γ(PICP) is a boolean variable.

[0049] A strip steel full-length thickness prediction method based on a memory function network of the present invention has at least the following beneficial effects:

[0050] (1) The present invention first proposes the concept of strip steel thickness prediction in the full-length direction, and according to the characteristic that the strip steel full-length thickness has a time series, proposes to use a network with a memory function to predict the strip steel thickness. This network can enhance the temporal correlation of strip steel data and improve the model prediction accuracy.

[0051] (2) The present invention uses a GLSTM network to predict the strip steel full-length thickness. Compared with the traditional memory network, the activation functions of each layer are improved, and the forgetting gate and the input gate are made complementary. The prediction accuracy of the GLSTM network model that improves the new information vector to the latest state of the data is improved, and the model is more stable.

[0052] (3) In order to improve the reliability of the GLSTM network for strip steel thickness interval prediction, the present invention proposes to complete the strip steel thickness prediction based on the point prediction combined with the interval prediction method and the particle swarm optimization algorithm, enhancing the optimization breadth and accuracy.

[0053] (4) The present invention proposes that the optimization iteration termination condition in the particle swarm algorithm is to reach the maximum number of iterations or the fitness function remains unchanged for M times. Because in the iteration process, there may be a situation where the optimal solution no longer changes from a certain iteration, and continuing to iterate for optimization is meaningless in this case. And the present invention proposes that the fitness function remains unchanged for M times, which is considered to find the optimal solution, which will effectively avoid unnecessary time consumption in the model training process.

[0054] (5) The strip steel full-length thickness prediction method based on the memory function network of the present invention can simultaneously achieve the point prediction and interval prediction of the strip steel thickness. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flowchart of a strip steel full-length thickness prediction method based on a memory function network of the present invention;

[0056] Figure 2 is an error comparison diagram of five network models;

[0057] Figure 3 It is a comparison chart of the absolute coefficients of five network models;

[0058] Figure 4 It is a comparison chart of the evaluation indexes of five network models;

[0059] Figure 5 It is the strip full-length thickness point prediction result of the GLSTM network for point prediction of the present invention;

[0060] Figure 6 It is the strip full-length thickness interval prediction result of the GLSTM network for interval prediction of the present invention. Detailed implementation mode

[0061] As shown in the figure, a strip full-length thickness prediction method based on a memory function network of the present invention includes:

[0062] Step 1: Collect the real-time process rolling data and the measured values of the strip outlet thickness at equal time intervals to form an input feature data set;

[0063] In specific implementation, the real-time process rolling data recorded over time is collected once at equal time intervals in the PDA file of a certain factory, including: the rolling kilometers, the inlet thickness, and the real-time process rolling data of 34 important influencing factors such as the rolling speed, rolling force, front and rear tensions, bending roll force, and roll gap of each stand.

[0064] Step 2: Divide the input feature data set into a training set and a test set according to a ratio and in chronological order. Since there are large differences in the order of magnitude between the data, in order to eliminate the influence of the dimension of each input feature data and reduce the error caused by the dimension, make each input feature data be in the same order of magnitude, and facilitate the subsequent processing of the data and accelerate the network learning speed, it is necessary to perform normalization processing on the data;

[0065] In specific implementation, the data of the training set and the test set are normalized according to the following formula:

[0066]

[0067] In the formula: x′ k is the feature data after normalization processing, x k is the kth feature data, x min is the minimum value in this feature data, x max is the maximum value in this feature data.

[0068] Step 3: Construct a GLSTM network for strip thickness point prediction. Since the interval prediction network of the present invention is based on point prediction, first train the GLSTM network for point prediction with the training set to obtain the connection weights and biases of each layer of the GLSTM network for point prediction, and store them separately.

[0069] The GLSTM network for strip thickness point prediction is specifically:

[0070]

[0071] f T = 1 - i T

[0072]

[0073]

[0074] c t = f T × h T-1 + g T × i T

[0075] h T = o T × Relu(c T )

[0076] In the formula, x T is the input data at time T; is the weight related to the input x, is the weight of the input gate related to the input data x, is the weight of the output gate related to the input data x, is the weight of the new information vector related to the input data x; is the weight related to the hidden state h T-1 at the previous time step; b (i) is the bias of the input gate, b (o) is the bias of the output gate, b (g) is the bias of the new information vector; i T represents the input gate, f T represents the forget gate, o T represents the output gate, g T represents the new information vector, c T represents the cell state of the network at time T, h T represents the hidden layer state of the network at time T; Iσ represents the Isigmoid activation function, and both LRelu and Relu are activation functions.

[0077] The GLSTM network for strip thickness point prediction in the present invention improves the forget gate based on the traditional LSTM network, making the forget gate form a complementary relationship with the input gate. At the same time, in order to enhance the input data features of the network and reduce the number of parameters, the new information vector g involved in the LSTM is T improved to the current data state. At the same time, the cell state c of the network is also T associated with the hidden state h at the previous moment, and it can be considered that the information from the past to time T is saved in it. T-1 Since the Sigmoid activation function has the inherent defect of gradient disappearance, the present invention combines the Relu-like function and proposes to use the Isigmoid function to alleviate this problem. The Tanh activation function is an S-shaped function and also has the problem of gradient disappearance (explosion), and it involves exponential calculations of e, which are computationally complex. While the Relu-like function effectively overcomes the problem of gradient disappearance (explosion), so this patent proposes to use the Relu-like function instead of the Tanh function to improve the accuracy of the prediction model. Compared with the standard recurrent neural network, the GLSTM network of the present invention has the characteristics of fewer parameters, simple calculation, strong parallel ability, and obvious original data features.

[0078] Step 4: Use the particle swarm optimization algorithm to optimize the connection weights of the output layer of the GLSTM network for point prediction in Step 3 to obtain the optimal connection weights W of the output layer

[0079] Specifically: opt

[0080] Step 4.1: Copy the n×1 connection weights W of the output layer obtained in Step 3 o to obtain W of n×2 ho , and add a random matrix with the same dimension whose elements are between [-1.1, 0.1] to obtain W no ;

[0081] Since the prediction interval is composed of an upper limit and a lower limit, this means that the network model for interval prediction requires two output nodes, while the neural network has only one output node in point prediction. Therefore, in order to obtain the upper and lower limits of the interval at the beginning of the iteration and accelerate convergence, the connection weights W of the output layer obtained in Step 3 o are processed, and finally W of n×2 is obtained no , and then the particle swarm optimization algorithm is used to optimize W no .

[0082] Step 4.2: Initialization of the population, randomly initialize the velocity of each particle in the solution space of the particle, and initialize the position of each particle using W no ;

[0083] ​Step 4.3: Use the CWC evaluation index as the fitness function to evaluate the fitness function value of each particle;

[0084] Step 4.4: For each particle, compare its current fitness value with the fitness value corresponding to its individual historical best position. If the current fitness value is better, use the current position to update the individual historical best position of the particle;

[0085] Step 4.5: For each particle, compare its current fitness value with the fitness value corresponding to the global best position. If the current fitness value is better, use the current position to update the historical optimal position of the particle swarm;

[0086] Step 4.6: Update the velocity and position of the particle according to the following formula:

[0087]

[0088]

[0089] where, Pbest j represents the historical best position of the j-th particle, Gbest represents the historical best position of the entire swarm, c1 and c2 are learning factors, r1 and r2 are uniformly distributed random numbers in the range of [0,1], and t is the number of iteration steps; when the velocity iteration update is greater than the maximum value V max of its limit range, make it equal to V max ; similarly, when it is less than the minimum value V min of the limit range, make it equal to V min .

[0090] Step 4.7: The termination condition of the entire optimization process is to reach the maximum number of iterations or the value of the fitness CWC remains unchanged for M times during the iteration process. When the termination condition is met, output the global best position, that is, the connection weight W opt of the best output layer; if the iteration termination condition is not met, go back to Step 4.3 to continue the iterative update.

[0091] Step 5: Based on the connection weights and biases of each layer except the output layer obtained in Step 3 and W opt obtained in Step 4, construct a GLSTM network for strip thickness interval prediction based on the test set data, and determine the best parameters of the GLSTM network for interval prediction according to the evaluation indexes PICP, PINAW and CWC to obtain the optimal GLSTM network for interval prediction;

[0092] In specific implementation, the optimal parameters of the GLSTM network for interval prediction include: the number of hidden layer nodes, the combination of activation functions for each layer, the network learning rate, and the parameters to be determined by the particle swarm optimization algorithm. The parameters to be determined by the particle swarm optimization algorithm include: learning factors c1 and c2, and the population size.

[0093] Step 6: Use the optimal GLSTM network for interval prediction obtained in Step 5 to predict the full-length thickness of the strip steel, and calculate the PICP, PINAW, and CWC evaluation indicators based on the upper and lower limits of the predicted strip steel thickness to complete the performance evaluation of the GLSTM network for interval prediction.

[0094] In specific implementation, the definitions of the PICP, PINAW, and CWC evaluation indicators are as follows:

[0095] (1) The coverage probability PICP is an important feature of interval prediction, indicating the probability that the target value is within the interval. Its definition is as follows:

[0096]

[0097] where, N represents the total number of samples, ∈ k is a Boolean decision variable used to indicate whether the kth target has been covered by the interval. If the target is covered by the interval, then ∈ k = 1; otherwise, ∈ k = 0. Therefore, the mathematical formula of ∈ k is defined as follows:

[0098]

[0099] where, y k represents the kth target, that is, the measured value of the strip steel exit thickness; L k and U k respectively represent the lower and upper bounds of the prediction interval generated for y k To ensure the accuracy of interval prediction and obtain an effective prediction interval, the coverage probability PICP is required to be not lower than the nominal confidence level 1 - α of the prediction interval, where α represents the significance level; otherwise, the prediction interval is invalid and should be discarded;

[0100] The coverage probability PICP is used as an evaluation index for interval prediction, and its value range is between 0 and 1. Obviously, the larger the probability of the prediction interval coverage, the better. It is easy to see that the ideal value of PICP is 1, which means that all targets are within the predicted interval. Although the coverage probability PICP is an important index for evaluating the quality of interval prediction, it is not the only index. This is because when the interval coverage rate is large enough, the interval width may also be very large, providing very little certainty information and unable to convey meaningful information, so the entire interval prediction is of little significance. Therefore, in addition to considering the coverage probability PICP, the width of the interval should also be considered.

[0101] (2) The quantitative measurement of the interval prediction width is defined as the predicted interval normalized average width PINAW, which is expressed by the following mathematical formula:

[0102]

[0103] where R g refers to the change range of the target, that is, the difference between the maximum value and the minimum value of the target.

[0104] In the ideal case, it is desired to include as many targets as possible within the narrowest possible interval range. It is easy to see that to achieve high-quality interval prediction, a higher interval coverage rate and a narrower interval width are required. From the above formulas for calculating the coverage probability PICP and the predicted interval normalized average width PINAW, it can be seen that these two requirements are contradictory. Therefore, a comprehensive cost function based on interval prediction, called the coverage width criterion CWC (CoverageWidth-based Criterion), needs to be defined to optimize the above two evaluation indexes.

[0105] (3) The coverage width criterion CWC is specifically defined as follows:

[0106] CWC = PINAW(1 + γ(PICP)e -η(PICP-μ) )

[0107]

[0108] where η and μ are two parameters used to control the optimization speed and accuracy. η is used as a penalty factor to penalize the prediction intervals with poor quality. If PICP does not reach the pre-specified confidence level, CWC will increase exponentially in this term. If PICP reaches the pre-specified confidence level, CWC will no longer focus on PICP but only on PINAW; μ corresponds to the confidence level related to interval prediction and is set to 1 - α; γ(PICP) is a Boolean variable.

[0109] In the process of training the network model with CWC as the fitness function, it can be seen from the above formula that CWC realizes single-objective optimization, helps to balance the interval prediction between accuracy and effectiveness, so as to achieve high-quality interval prediction.

[0110] Embodiment:

[0111] To prove the effectiveness of the strip full-length thickness prediction method based on the memory function network proposed by the present invention, data during the rolling process of a certain factory are collected at equal time intervals. The measured data and calculated data of 34 important influencing factors such as rolling speed, rolling force, front and back tensions, work roll bending force, roll gap, rolling kilometers, and inlet thickness, as well as the measured value of the strip outlet thickness at the previous moment, are selected as the input of the model, and the measured value of the strip outlet thickness is used as the output data. There are a total of 9 continuous strips of steel, 1528 groups of continuous data, among which the data of the first 8 strips of steel are used as the training set, a total of 1371 groups of data, and the data of the last strip of steel are used as the test set, a total of 157 groups of data. Experiments are respectively carried out on the traditional RNN, LSTM, GRU networks, and the GLSTM network of the present invention, and a comparative experiment is carried out using the artificial neural network (ANN). The mean absolute error, root mean square error, and mean absolute percentage error of the five different network models on the data set during point prediction are as follows Figure 2 shown, and the corresponding coefficient of determination is as Figure 3 shown.

[0112] It can be seen from Figure 2 that the mean absolute error, root mean square error, and mean absolute percentage error of the point prediction of the four memory networks are significantly smaller than those of the ANN network. The four memory networks achieve better point prediction effects on the data set. At the same time, the GLSTM network proposed by the present invention has smaller mean absolute error, root mean square error, and mean absolute percentage error than the other three traditional memory networks RNN, LSTM, and GRU, and the prediction performance is the best. The coefficient of determination is generally between 0 and 1. The closer it is to 1, the better the model fitting performance, and the closer it is to 0, the worse the model fitting performance. It can be seen from Figure 3 that compared with the ANN network, the point prediction performance of the four memory networks is better, and among them, the GLSTM network adopted by the present invention has the best prediction performance.

[0113] During interval prediction, the comparisons of PICP, PINAW, and CWC of the five different network models on the data set are as follows Figure 4 shown. It can be seen from Figure 4It can be seen that the comprehensive evaluation index of the four memory network interval predictions is significantly smaller than that of the ANN network. Among them, the CWC value of the GLSTM network adopted in the present invention is the smallest, indicating that the prediction interval performance obtained by the memory network of the present invention for predicting the strip full-length thickness is better. Due to the confidence level set at 0.95, it can be seen that among the five networks, only the PICP value of the GLSTM network adopted in the present invention is 0.951592, which meets the condition. At the same time, it can be seen that the PINAW value of the GLSTM network is the smallest, indicating that the prediction interval of the GLSTM network is the narrowest. It can be seen that the GLSTM network model proposed in the present invention has the best interval prediction effect.

[0114] During specific implementation, the training set is used to train the model of the GLSTM network of the present invention. The number of hidden layer nodes of the model is set to 40, the learning rate is set to 0.01, and the two control parameters η and μ in the interval prediction evaluation index CWC are respectively set to 30 and 0.95. The trained model is used for point prediction and interval prediction on the test set, and the results of the point prediction are as Figure 5 shown, and the interval prediction results are as Figure 6 shown. According to Figure 5 and Figure 6 it can be seen that whether it is point prediction or interval prediction, the GLSTM network proposed in the present invention can effectively predict the strip full-length thickness at future moments.

[0115] The above are only the preferred embodiments of the present invention and are not intended to limit the idea of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the full-length thickness of strip steel based on a memory function network, characterized in that, Including: Step 1: Collect the real-time process rolling data and the measured values of the strip exit thickness at equal time intervals to form an input feature data set. Step 2: Divide the input feature data set into a training set and a test set according to a ratio and in chronological order, and normalize the data in the training set and the test set. Step 3: Construct a GLSTM network for strip thickness point prediction, and train the GLSTM network for point prediction through the training set to obtain the connection weights and biases of each layer of the GLSTM network for point prediction. The GLSTM network for strip thickness point prediction is specifically: f T = 1 - i T C T = f T × h T-1 + g T × i T h T = o T × Relu(c T ) where, x T is the input data at time T; is the weight related to the input x, is the weight of the input gate related to the input data x, is the weight of the output gate related to the input data x, is the weight of the new information vector related to the input data x; is the weight related to the hidden state h T-1 at the previous time step; b (i) is the bias of the input gate, b (o) is the bias of the output gate, b (g) is the bias of the new information vector; i T represents the input gate, f T represents the forget gate, o T represents the output gate, g T represents the new information vector, c T represents the cell state of the network at time T, h T represents the hidden layer state of the network at time T; Iσ represents the Isigmoid activation function, and both LRelu and Relu are activation functions; Step 4: Optimize the connection weights of the output layer of the GLSTM network for point prediction in Step 3 using the particle swarm algorithm to obtain the optimal connection weights \(W\) of the output layer opt ; Step 5: Based on the connection weights and biases of each layer except the output layer obtained in Step 3 and W obtained in Step 4 opt , construct a GLSTM network for strip thickness interval prediction, and determine the optimal parameters of the GLSTM network for interval prediction according to the evaluation indexes PICP, PINAW and CWC to obtain the optimal GLSTM network for interval prediction; Step 6: Use the optimal GLSTM network for interval prediction obtained in Step 5 to predict the strip full-length thickness, and calculate the PICP, PINAW, and CWC evaluation indicators according to the upper and lower limits of the predicted strip thickness to complete the performance evaluation of the GLSTM network for interval prediction. PICP represents the probability that the target value is within the interval, PINAW represents the normalized average width of the prediction interval, and CWC represents the standard of the coverage width.

2. The method for predicting the full-length thickness of strip steel based on a memory function network according to claim 1, characterized in that, The real-time process rolling data includes: rolling kilometers, entrance thickness, and real-time process rolling data of multiple important influencing factors such as the rolling speed, rolling force, front and back tensions, bending roll force, and roll gap of each stand.

3. The method for predicting the full-length thickness of strip steel based on a memory function network according to claim 1, characterized in that, In Step 2, the data in the training set and the test set are normalized according to the following formula: Where: x' k is the feature data after normalization, x k is the k-th feature data, x min is the minimum value in the feature data, x max is the maximum value in the feature data.

4. The method for predicting the full-length thickness of strip steel based on a memory function network according to claim 1, characterized in that, In Step 4, the particle swarm algorithm is used to optimize the connection weights of the output layer of the GLSTM for point prediction in Step 3. Specifically: Step 4.1: Copy the connection weights W of the n×1 output layer obtained in Step 3 o to obtain an n×2 W ho , and add a random matrix of the same dimension with elements in the range [-1.1, 0.1] to obtain W no ; Step 4.2: Initialization of the population. Randomly initialize the velocity of each particle in the solution space of the particles, and use W to initialize the position of each particle no Initialization; Step 4.3: Use the CWC evaluation indicator as the fitness function to evaluate the fitness function value of each particle. Step 4.4: Compare the current fitness value of each particle with the fitness value corresponding to its individual historical best position. If the current fitness value is better, use the current position to update the individual historical best position of the particle. Step 4.5: Compare the current fitness value of each particle with the fitness value corresponding to the global best position. If the current fitness value is better, use the current position to update the historical optimal position of the particle swarm. Step 4.6: Update the velocity and position of the particle according to the following formula: Among them, Pbest j represents the historical best position of the j-th particle, Gbest represents the historical best position of the entire population, c1 and c2 are learning factors, r1 and r2 are uniformly distributed random numbers within the range of [0, 1], and t is the number of iteration steps; when the velocity iteration update is greater than the maximum value V of its limit range max it is set equal to V max ; similarly, when it is less than the minimum value V of the limit range min it is set equal to V min ; Step 4.7: If the maximum number of iterations is satisfied or the value of the fitness function CWC remains unchanged for M times during the iteration process, then output the global best position, that is, the connection weights W of the optimal output layer opt ; If the iteration termination condition is not satisfied, go to Step 4.3 to continue the iterative update.

5. The method for predicting the full-length thickness of strip steel based on a memory function network according to claim 4, characterized in that, The best parameters of the GLSTM network for interval prediction in Step 5 include: the number of hidden layer nodes, the combination of activation functions for each layer, the network learning rate, and the parameters that need to be determined by the particle swarm optimization algorithm. The parameters that need to be determined by the particle swarm optimization algorithm include: learning factors c1 and c2 and the population size.

6. The method for predicting the full-length thickness of strip steel based on a memory function network according to claim 1, characterized in that, The definitions of the PICP, PINAW, and CWC evaluation indicators are as follows: (1) The coverage probability PICP is an important feature of interval prediction, representing the probability that the target value is within the interval. Its definition is as follows: where, N represents the total number of samples, ∈ k is a Boolean decision variable used to indicate whether the k-th target has been covered by the interval. If the target is covered by the interval, then ∈ k = 1; otherwise k = 0. Therefore, ∈ k is defined by the following mathematical formula: Among them, y k represents the k-th target, that is, the measured value of the strip outlet thickness; L k and U k respectively represent the lower bound and the upper bound of the prediction interval generated for y k To ensure the accuracy of the interval prediction and obtain an effective prediction interval, the coverage probability PICP is required to be not lower than the nominal confidence level 1-α of the prediction interval, where α represents the significance level; otherwise, the prediction interval is invalid and should be discarded; (2) The quantitative measurement of the prediction interval width is defined as the normalized average width PINAW of the prediction interval, which is expressed by the following mathematical formula: wherein, R g refers to the variation range of the target, that is, the difference between the maximum value and the minimum value of the target; (3) Define a comprehensive cost function based on interval prediction, called the standard CWC of the coverage width, to optimize the above two evaluation indicators. The specific definition is as follows: CWC = PINAW(1 + γ(PICP)e -η(PICP-μ) ) Among them, η and μ are two parameters used to control the optimization speed and accuracy. η is used as a penalty factor to penalize the prediction intervals with poor quality. If the PICP does not reach the pre-specified confidence level, the CWC will exponentially increase in this term. If the PICP reaches the pre-specified confidence level, the CWC will no longer focus on the PICP but only on the PINAW; μ corresponds to the confidence level related to interval prediction and is set to 1 - α; γ(PICP) is a boolean variable.

Images