Method and device for predicting aerial fog cooling heat transfer coefficient of cast-rolled thin strip after rolling

Abstract

The invention provides a method and device for predicting an aerial fog cooling heat transfer coefficient of a cast-rolled thin strip after rolling, and the method comprises the steps: carrying out the normalization processing of heat transfer experiment data, and dividing the data into a training sample and a verification sample; and creating an initial three-layer feed-forward full-connection neural network according to preset factors and the number of samples. And training the network by using a particle swarm optimization algorithm and an extreme learning machine to obtain a target network. And verifying the target network based on the verification set sample, and if a verification error meets a preset standard, determining a final network. And according to the actual aerial fog cooling condition parameters and the final network, predicting an aerial fog cooling heat transfer coefficient, and setting an aerial fog cooling model. By using the global optimization characteristic of a particle swarm optimization algorithm and the rapid convergence learning characteristic of an extreme learning machine, supervised learning is performed on aerial fog cooling heat transfer experimental data through a three-layer feedforward full-connection network, so that aerial fog cooling heat transfer coefficients under different actual working conditions can be accurately predicted, regulated and controlled; the cast-rolled thin strip cooling effect is improved, and the product quality is ensured.

Description

Technical Field

[0001] The invention relates to the field of metallurgical automation and intelligent technology, and in particular to a method and device for predicting the heat transfer coefficient of gas mist cooling after rolling of a cast-rolled thin strip. Background Art

[0002] In the steel manufacturing industry, especially in the production process of cast and rolled thin strip, the selection and control of cooling technology has a vital impact on product quality and production efficiency. The mist cooling technology atomizes the water flow through a two-phase nozzle to form atomized droplets with small particle size and uniform distribution, which can provide a wider and more uniform cooling effect than traditional columnar laminar cooling. This technology is widely used in the fields of continuous casting secondary cooling water control, H-beam and wire post-rolling cooling, etc.

[0003] However, despite the significant application of mist cooling technology in these fields, there is still a lag problem in the post-rolling cooling of cast and rolled thin strips. This is because the heat transfer coefficient of mist cooling is affected by many factors, and the traditional method mainly relies on experimental data to deduce the coefficient. However, due to the fluctuation of actual process parameters, the direct application of experimental data is difficult to meet the needs of precise control. This makes it difficult to achieve a stable effect of mist cooling during the cooling process of cast and rolled thin strips, resulting in certain challenges in the precise control and optimization of post-rolling cooling.

[0004] Therefore, how to accurately predict and control the heat transfer coefficient of mist cooling under different actual working conditions has become a key technical problem to improve the cooling effect of cast and rolled thin strips, ensure product quality and increase the yield rate. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a method and device for predicting the heat transfer coefficient of aerosol cooling of cast-rolled thin strip after rolling, so as to solve the problem that the heat transfer coefficient of aerosol cooling of cast-rolled thin strip after rolling is difficult to meet the precise control and ensure the cooling effect.

[0006] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0007] The first aspect of the present invention discloses a method for predicting the heat transfer coefficient of aerosol cooling of a cast-rolled thin strip after rolling, the method comprising:

[0008] Dividing the heat transfer experiment result data into initial training set samples and initial verification set samples, and performing normalization processing on the initial training set samples and the initial verification set samples to obtain training set samples and verification set samples;

[0009] Creating an initial three-layer feedforward fully connected neural network according to the number of preset factors, the number of samples in the training set, and the number of all samples;

[0010] Based on the training samples in the training set samples, various parameters of the particle swarm optimization algorithm set in advance, the particle swarm optimization algorithm and the extreme learning machine, the initial three-layer feedforward fully connected neural network is trained to obtain a target three-layer feedforward fully connected neural network;

[0011] The target three-layer feedforward fully connected neural network is verified based on all verification samples in the verification set samples, and if the prediction relative errors of all verification samples are less than a preset relative error percentage value, a final three-layer feedforward fully connected neural network is determined;

[0012] According to the actual aerosol cooling condition parameters and the final three-layer feedforward fully connected neural network, the aerosol cooling heat transfer coefficient is predicted, and the aerosol cooling model is set based on the aerosol cooling heat transfer coefficient.

[0013] Preferably, the step of creating an initial three-layer feedforward fully connected neural network according to the number of preset factors, the number of samples in the training set and the number of all samples includes:

[0014] Set the number of input layer nodes according to the preset number of factors;

[0015] Set the number of output layer nodes to 1;

[0016] Obtain the number of samples in the training set and the number of all samples, and calculate the value range of the number of hidden layer nodes according to a rounding-down function;

[0017] The S-shaped growth curve function is determined as the activation function of the hidden layer nodes;

[0018] An initial three-layer feedforward fully connected neural network is created based on the value ranges of the number of input layer nodes, the number of output layer nodes, the number of hidden layer nodes, and the activation function of the hidden layer nodes.

[0019] Preferably, the training of the initial three-layer feedforward fully connected neural network based on the training samples in the training set samples, various parameters of the pre-set particle swarm optimization algorithm, the particle swarm optimization algorithm and the extreme learning machine to obtain the target three-layer feedforward fully connected neural network includes:

[0020] For any training sample in the training set samples, input the feature quantity of the training sample into the initial three-layer feedforward fully connected neural network, and perform network forward calculation on the activation output of each hidden layer node according to the extreme learning machine algorithm;

[0021] Perform network forward calculation based on the activation output of each hidden layer node to obtain the network output expression;

[0022] Calculating an error amount based on the network output expression and the label amount of the training sample;

[0023] Constructing a loss function according to the corresponding error amounts of the plurality of training samples;

[0024] Minimize the loss function based on the extreme learning machine algorithm to determine the connection weight vector between the hidden layer and the output layer;

[0025] Performing an iterative operation, according to the number of particles of the particle swarm optimization algorithm, for each particle, calculating the fitness of the individual particle according to the loss function and the connection weight vector between the hidden layer and the output layer;

[0026] According to the fitness of the individual particles, the optimal fitness and optimal position of the individual particles, the optimal fitness and optimal position of the particle group, and the optimal weight vector are adjusted until the current number of iterations reaches a preset maximum number of iterations, and the global optimal position and optimal weight vector of the particle group are determined;

[0027] All weights and biases of a three-layer feedforward fully connected neural network are determined according to the global optimal position and optimal weight vector of the particle swarm, and a target three-layer feedforward fully connected neural network is obtained.

[0028] Preferably, the step of adjusting the optimal fitness and optimal position of the individual particle, the optimal fitness and optimal position of the particle group, and the optimal weight vector according to the fitness of the individual particle comprises:

[0029] Determine whether the fitness of the individual particle is less than the current optimal fitness of the individual particle;

[0030] If it is less than, the fitness of the particle individual is taken as the latest optimal fitness of the particle individual;

[0031] Taking the current position of the particle as the latest optimal position of the particle;

[0032] Determine whether the fitness of the individual particle is less than the optimal fitness of the current particle group;

[0033] If it is less than, the fitness of the individual particle is taken as the optimal fitness of the latest particle group;

[0034] Using the current position of the particle as the optimal position of the latest particle group;

[0035] The weight vector of the particle is taken as the latest optimal weight vector.

[0036] Preferably, the target three-layer feedforward fully connected neural network is verified based on all verification samples in the verification set samples, and if the prediction relative errors of all verification samples are less than a preset relative error percentage value, then the final three-layer feedforward fully connected neural network is determined, including:

[0037] For any verification sample in the verification set samples, input the feature quantity of the verification sample into the target three-layer feedforward fully connected neural network to obtain the prediction value of the network;

[0038] Compare the predicted value with the label quantity in the validation sample to obtain an error;

[0039] The corresponding errors of all validation samples are counted to obtain the relative prediction error;

[0040] If the predicted relative error is less than the preset relative error percentage value, the structure and all weight and bias parameters of the target three-layer feedforward fully connected neural network are solidified according to the number of hidden layer nodes, the global optimal position of the particle swarm and the optimal weight vector determined in the verification process to obtain the final three-layer feedforward fully connected neural network.

[0041] Preferably, the aerosol cooling heat transfer coefficient is predicted according to the actual aerosol cooling condition parameters and the final three-layer feedforward fully connected neural network, and the aerosol cooling model is set based on the aerosol cooling heat transfer coefficient, including:

[0042] Obtain actual aerosol cooling condition parameters and perform normalization processing;

[0043] Inputting the normalized actual aerosol cooling condition parameters into the final three-layer feedforward fully connected neural network for prediction, and outputting the network prediction value;

[0044] The network prediction value is processed by using the anti-normalization method to obtain the predicted value of the aerosol cooling heat transfer coefficient;

[0045] The mist cooling model is set based on the predicted value of the mist cooling heat transfer coefficient.

[0046] The second aspect of the present invention discloses a device for predicting the heat transfer coefficient of gas mist cooling after rolling of a cast and rolled thin strip, the device comprising:

[0047] A division unit, used for dividing the heat transfer experiment result data into an initial training set sample and an initial verification set sample, and performing normalization processing on the initial training set sample and the initial verification set sample to obtain a training set sample and a verification set sample;

[0048] A creation unit, used for creating an initial three-layer feedforward fully connected neural network according to the number of preset factors, the number of samples in the training set and the number of all samples;

[0049] A training unit, configured to train the initial three-layer feedforward fully connected neural network based on the training samples in the training set samples, various parameters of the pre-set particle swarm optimization algorithm, the particle swarm optimization algorithm and the extreme learning machine to obtain a target three-layer feedforward fully connected neural network;

[0050] A verification unit, used to verify the target three-layer feedforward fully connected neural network based on all verification samples in the verification set samples, and if the prediction relative error of all verification samples is less than a preset relative error percentage value, then determine to obtain a final three-layer feedforward fully connected neural network;

[0051] The prediction unit is used to predict the aerosol cooling heat transfer coefficient according to the actual aerosol cooling condition parameters and the final three-layer feedforward fully connected neural network, and set the aerosol cooling model based on the aerosol cooling heat transfer coefficient.

[0052] Preferably, the creation unit includes:

[0053] A setting module, used to set the number of input layer nodes according to the number of preset factors;

[0054] A first determination module is used to determine the number of output layer nodes to be 1;

[0055] A first calculation module is used to obtain the number of samples in the training set and the number of all samples, and calculate the value range of the number of hidden layer nodes according to a rounding-down function;

[0056] A second determination module is used to determine the S-shaped growth curve function as the activation function of the hidden layer node;

[0057] A creation module is used to create an initial three-layer feedforward fully connected neural network based on the value range of the number of input layer nodes, the number of output layer nodes, the number of hidden layer nodes and the activation function of the hidden layer nodes.

[0058] Preferably, the training unit comprises:

[0059] A second calculation module is used for inputting the feature quantity of any training sample in the training set samples into the initial three-layer feedforward fully connected neural network, and performing network forward calculation of the activation output of each hidden layer node according to the extreme learning machine algorithm;

[0060] The third calculation module is used to perform network forward calculation according to the activation output of each hidden layer node to obtain a network output expression;

[0061] A fourth calculation module, used for calculating an error amount based on the network output expression and the label amount of the training sample;

[0062] A construction module, used to construct a loss function according to the corresponding error amounts of the plurality of training samples;

[0063] A third determination module is used to minimize the loss function based on an extreme learning machine algorithm to determine a connection weight vector between the hidden layer and the output layer;

[0064] A fifth calculation module is used to perform an iterative operation, and according to the number of particles of the particle swarm optimization algorithm, for each particle, calculate the fitness of the individual particle according to the loss function and the connection weight vector between the hidden layer and the output layer;

[0065] A fourth determination module is used to adjust the optimal fitness and optimal position of the individual particle, the optimal fitness and optimal position of the particle group, and the optimal weight vector according to the fitness of the individual particle, until the current number of iterations has reached a preset maximum number of iterations, and determine the global optimal position and optimal weight vector of the particle group;

[0066] The fifth determination module is used to determine all weights and biases of the three-layer feedforward fully connected neural network according to the global optimal position and optimal weight vector of the particle swarm to obtain a target three-layer feedforward fully connected neural network.

[0067] Preferably, the fourth determining module is specifically used to:

[0068] Determine whether the fitness of the individual particle is less than the current individual optimal fitness of the particle; if so, take the individual fitness of the particle as the latest individual optimal fitness of the particle; take the current position of the particle as the latest individual optimal position of the particle;

[0069] Determine whether the fitness of the individual particle is less than the optimal fitness of the current particle group; if so, take the fitness of the individual particle as the optimal fitness of the latest particle group; take the current position of the particle as the optimal position of the latest particle group; take the weight vector of the particle as the latest optimal weight vector.

[0070] Based on the above-mentioned embodiment of the present invention, a method and device for predicting the heat transfer coefficient of aerosol cooling of cast and rolled thin strip after rolling is provided, and the heat transfer experimental data is normalized and divided into training samples and verification samples. An initial three-layer feedforward fully connected neural network is created according to preset factors and sample quantity. The network is trained using a particle swarm optimization algorithm and an extreme learning machine to obtain a target network. The target network is verified based on the validation set samples, and if the verification error meets the preset standard, the final network is determined. According to the actual aerosol cooling condition parameters and the final network, the aerosol cooling heat transfer coefficient is predicted and used to set the aerosol cooling model. By utilizing the global optimization characteristics of the particle swarm optimization algorithm and the fast convergence learning characteristics of the extreme learning machine, the aerosol cooling heat transfer experimental data is supervised and learned through a three-layer feedforward fully connected network, which helps to accurately predict and regulate the aerosol cooling heat transfer coefficient under different actual working conditions, improve the cooling effect of the cast and rolled thin strip, and ensure product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0072] Figure 1 A flow chart of a method for predicting the heat transfer coefficient of aerosol cooling of a cast-rolled thin strip after rolling provided by an embodiment of the present invention;

[0073] Figure 2 A structural diagram of an initial three-layer feedforward fully connected neural network provided in an embodiment of the present invention;

[0074] Figure 3 Another flow chart of a method for predicting the heat transfer coefficient of aerosol cooling of a cast-rolled thin strip after rolling provided by an embodiment of the present invention;

[0075] Figure 4 A structural block diagram of a device for predicting the heat transfer coefficient of post-rolling mist cooling of a cast-rolled thin strip provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] In this application, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element.

[0078] As can be seen from the background technology, the traditional method mainly relies on experimental data to infer the heat transfer coefficient of gas mist cooling, but the direct application of experimental data is difficult to meet the needs of precise control, making it difficult to stably achieve the effect of gas mist cooling during the cooling process of cast and rolled thin strips.

[0079] Therefore, an embodiment of the present invention provides a method and device for predicting the heat transfer coefficient of aerosol cooling of cast and rolled thin strip after rolling, and normalizes the heat transfer experimental data and divides them into training samples and verification samples. An initial three-layer feedforward fully connected neural network is created according to preset factors and sample quantity. The network is trained using a particle swarm optimization algorithm and an extreme learning machine to obtain a target network. The target network is verified based on the validation set samples, and if the verification error meets the preset standard, the final network is determined. According to the actual aerosol cooling condition parameters and the final network, the aerosol cooling heat transfer coefficient is predicted and used to set the aerosol cooling model. By utilizing the global optimization characteristics of the particle swarm optimization algorithm and the fast convergence learning characteristics of the extreme learning machine, supervised learning of the aerosol cooling heat transfer experimental data is performed through a three-layer feedforward fully connected network, which helps to accurately predict and regulate the aerosol cooling heat transfer coefficient under different actual working conditions, improve the cooling effect of the cast and rolled thin strip, and ensure product quality.

[0080] See also Figure 1 , shows a flow chart of a method for predicting the heat transfer coefficient of aerosol cooling of a cast-rolled thin strip after rolling provided by an embodiment of the present invention. The method comprises:

[0081] Step S101: Divide the heat transfer experiment result data into initial training set samples and initial verification set samples, and perform normalization processing on the initial training set samples and the initial verification set samples to obtain training set samples and verification set samples.

[0082] In the specific implementation of step S101, the heat transfer experiment result data is obtained, and the heat transfer experiment result data is divided into initial training set samples and initial validation set samples. The initial training set samples and initial validation set samples are normalized by the Min-Max Normalization method shown in formula (1), thereby obtaining training set samples and validation set samples.

[0083] (1)

[0084] Among them, x is the sample original data; is the transformed data; x min is the minimum value of the original data obtained after sample statistics; x max It is the maximum value of the original data obtained after sample statistics.

[0085] It can be understood that the initial training set sample data and the initial validation set sample data are transformed into the range of [-1, +1] by formula (1).

[0086] In some specific embodiments, when the heat transfer experimental result data is divided into the initial training set samples and the initial validation set samples, the range of the characteristic water pressure and air pressure in the initial validation set samples should be smaller than that of the initial training set samples and closer to the actual working water pressure and air pressure of the aerosol cooling after the casting and rolling of the thin strip. This setting helps to ensure that the network can better adapt to the fluctuations of water pressure and air pressure in the actual aerosol cooling process.

[0087] In actual application, the Z-Score normalization method can also be used to normalize the initial training set samples and the initial validation set samples to obtain training set samples and validation set samples.

[0088] Step S102: Create an initial three-layer feedforward fully connected neural network according to the preset number of factors, the number of training set samples and the number of all samples.

[0089] In the specific implementation of step S102, an initial three-layer feedforward fully connected neural network is created, including an input layer, an output layer and a hidden layer, wherein the three layers are set according to the preset number of factors, the number of training set samples and the number of all samples.

[0090] Specifically, the process of creating the initial three-layer feedforward fully connected neural network is as follows (process A1 to process A5):

[0091] Process A1: Set the number of input layer nodes according to the number of preset factors.

[0092] It should be noted that the preset number of factors is the preset number of main factors affecting the heat transfer coefficient of aerosol cooling, which is specifically known based on heat transfer theory and aerosol cooling heat transfer experiments.

[0093] In other words, the number of input layer nodes N of the initial three-layer feedforward fully connected neural network is consistent with the number of main factors affecting the heat transfer coefficient of aerosol cooling.

[0094] Process A2: Set the number of output layer nodes to 1.

[0095] It can be understood that the output layer of the initial three-layer feedforward fully connected neural network only outputs the aerosol cooling heat transfer coefficient, so there is only one node, that is, M=1.

[0096] Process A3: Obtain the number of training set samples and the number of all samples, and calculate the value range of the number of hidden layer nodes based on the floor function.

[0097] Specifically, the value range of the number of hidden layer nodes L is calculated as shown in formula (2). Further, the number of hidden layer nodes L is initialized with the minimum value in the value range.

[0098] (2)

[0099] Among them, floor() is the rounding function; Q is the number of training set samples; k is the number of samples corresponding to each network structure parameter.

[0100] Process A4: Determine the S-shaped growth curve function as the activation function of the hidden layer nodes.

[0101] It can be understood that the S-shaped growth curve function is a Sigmoid function.

[0102] It should be noted that the activation function of the hidden layer node is determined according to the content shown in formula (3).

[0103] (3)

[0104] Among them, g(x) is the activation function of the hidden layer nodes.

[0105] Process A5: Create an initial three-layer feedforward fully connected neural network based on the value ranges of the number of input layer nodes, the number of output layer nodes, the number of hidden layer nodes, and the activation function of the hidden layer nodes.

[0106] It can be understood that, according to the value ranges of the number of input layer nodes, the number of output layer nodes, the number of hidden layer nodes and the activation function of the hidden layer nodes determined by processes A1 to A4, an initial three-layer feedforward fully connected neural network is created, in which the input layer nodes are fully connected to the hidden layer nodes, and the hidden layer nodes are also fully connected to the output layer nodes.

[0107] It should be noted that see Figure 2 , which shows the structure of the initial three-layer feedforward fully connected neural network provided by the embodiment of the present invention. i is the input value of the i-th node in the input layer. j is the output value of the activation function of the jth node in the hidden layer. l is the output value of the lth node in the output layer.

[0108] β L−1 ,β L ,… represents the connection weight from the hidden layer to the output layer, and represents the weight of the edge connecting the hidden layer and the output layer.

[0109] w ij is the connection weight from the input layer to the hidden layer, which represents the weight of the edge connecting the input layer and the hidden layer.

[0110] Step S103: Based on the training samples in the training set samples, various parameters of the pre-set particle swarm optimization algorithm, the particle swarm optimization algorithm and the extreme learning machine, the initial three-layer feedforward fully connected neural network is trained to obtain a target three-layer feedforward fully connected neural network.

[0111] It should be noted that the various parameters of the particle swarm optimization algorithm are preset, including the particle swarm size (number of particles) n, the maximum number of iterations MaxEpoch, the learning factors c1 and c2, and the inertia weight ω.

[0112] The position vector and velocity vector of each particle are determined by the connection weights between the input layer nodes and the hidden layer nodes of the initial three-layer feedforward fully connected neural network. ( ) and the bias of the hidden layer nodes ( ). And the position vector and velocity vector of each particle have the same dimension, both .

[0113] In practical applications, the connection weights in the position vector and the velocity vector are randomly and bias are initialized to random numbers ranging from 0 to 1. The optimal fitness of the particle group And the optimal fitness of each particle , , are initialized to a large number (that is, a very large number, such as 10 to the power of 10. In the first iteration, the fitness will be replaced by the calculated value, thereby achieving gradual optimization judgment).

[0114] In the specific implementation of step S103, based on the training samples in the training set samples and the various parameters of the pre-set particle swarm optimization algorithm, the initial three-layer feedforward fully connected neural network is trained using the particle swarm optimization algorithm (PSO) and the extreme learning machine algorithm (ELM), so as to obtain the target three-layer feedforward fully connected neural network.

[0115] The specific training process is as follows (process B1 to process B8):

[0116] It can be understood that the extreme learning machine algorithm trains the initial three-layer feedforward fully connected neural network, which can be divided into two main stages: random feature mapping and linear parameter solution.

[0117] The first stage is to randomly initialize the hidden layer parameters, select the nonlinear mapping activation function, and map the input data to a new feature space (called ELM feature space), as shown in process B1.

[0118] Process B1: For any training sample in the training set, the feature quantity of the training sample is input into the initial three-layer feedforward fully connected neural network, and the network forward calculates the activation output of each hidden layer node.

[0119] Taking the pth training sample in the training set as an example, the feature quantity of the pth training sample is Input the initial three-layer feedforward fully connected neural network. The connection weights of the input layer and hidden layer of the initial three-layer feedforward fully connected neural network and the bias of the hidden layer nodes are determined by the position vector of the sth particle ( )Decide.

[0120] Specifically, the network forward calculation is performed through formula (4) to obtain the activation output of each hidden layer node .

[0121] (4)

[0122] in, is the input value of the i-th node in the input layer; is the connection weight between the i-th node in the input layer and the j-th node in the hidden layer; is the bias of the jth node in the hidden layer; g() is the activation function; is the activation output of the jth node in the hidden layer.

[0123] It can be understood that the second stage is to minimize the loss function composed of the training error terms, solve the generalized inverse matrix, and obtain the connection weights between the hidden layer and the output layer, as shown in process B2 to process B5.

[0124] Process B2: Perform network forward calculation based on the activation output of each hidden layer node to obtain the network output expression.

[0125] Assume that the connection weight vector between the hidden layer and the output layer of the initial three-layer feedforward fully connected neural network is , according to the activation output of each hidden layer node , continue the network forward calculation and obtain the network output expression (as shown in formula (5)).

[0126] (5)

[0127] in, .

[0128] Understandably, is the connection weight between the jth node in the hidden layer and the kth node in the output layer. It is a row vector with a dimension of 1×M, which is the predicted value of the network.

[0129] Process B3: Based on the network output expression and the label quantity of the training samples, the error amount is calculated.

[0130] It is understandable that the network output expression and the number of labels for training samples Compare and get the error .

[0131] Process B4: Construct a loss function based on the corresponding error amounts of multiple training samples.

[0132] It can be understood that by inputting Q training samples one by one into the initial three-layer feedforward fully connected neural network, Q error vectors can be obtained after forward calculation.

[0133] Therefore, the loss function is constructed based on the corresponding error amounts of multiple training samples (as shown in formula (6)):

[0134] (6)

[0135] in, is a matrix of dimension Q×L, ; is a label matrix of dimension Q×M, .

[0136] Process B5: Minimize the loss function based on the extreme learning machine algorithm and determine the connection weight vector between the hidden layer and the output layer.

[0137] It should be noted that the loss function based on the extreme learning machine algorithm By minimizing, we can get the optimal solution ; The connection weight vector between the hidden layer and the output layer can be determined as .

[0138] in, For the matrix The Moore-Penrose generalized inverse matrix can be solved in a general way using the singular value decomposition (SVD) method.

[0139] It is understandable that although the extreme learning machine can show good performance in most cases, the initial parameters of the hidden layer (such as connection weights, biases, and the number of nodes) still have a great influence on the accuracy of the extreme learning machine, resulting in unstable learning performance. In order to solve this problem, the particle swarm optimization algorithm can be used to improve the extreme learning machine in order to improve its generalization ability and achieve higher prediction accuracy in the prediction task, as shown in process B6 to process B8.

[0140] Process B6: Perform iterative operations. According to the number of particles in the particle swarm optimization algorithm, for each particle, the fitness of the individual particle is calculated based on the loss function and the connection weight vector between the hidden layer and the output layer.

[0141] When implementing process B6, perform an iterative operation and convert the loss function Directly use it as the fitness function of the particle. Substitute, through , we can calculate the fitness of the sth particle individual ,Furthermore, according to the number of particles n of the particle swarm optimization algorithm, the fitness of n individual particles can be ,calculated.

[0142] Process B7: According to the fitness of the individual particles, the optimal fitness and optimal position of the individual particles, the optimal fitness and optimal position of the particle group, and the optimal weight vector are adjusted until the current number of iterations has reached the preset maximum number of iterations, and the global optimal position and optimal weight vector of the particle group are determined.

[0143] In the specific implementation process B7, the optimal fitness and optimal position of the individual particle, the optimal fitness and optimal position of the particle group, and the optimal weight vector are adjusted according to the fitness of the individual particle, and it is determined whether the number of iterations has reached the preset maximum number of iterations MaxEpoch; if reached, the global optimal position and optimal weight vector of the particle group are determined; if not reached, return to the execution process B1.

[0144] Specifically, the process of adjusting the optimal fitness and optimal position of individual particles, the optimal fitness and optimal position of the particle group, and the optimal weight vector according to the fitness of the individual particles is as follows (process C1 to process C4):

[0145] Process C1: Determine whether the fitness of the individual particle is less than the current optimal fitness of the individual particle; if it is, execute process C2; if it is not, execute process C3.

[0146] Process C2: If it is less than, the fitness of the individual particle is taken as the latest optimal fitness of the individual particle; the current position of the particle is taken as the latest optimal position of the individual particle, and then process C3 is executed.

[0147] Process C3: Determine whether the fitness of the individual particle is less than the optimal fitness of the current particle group; if so, execute process C4; if not, end.

[0148] Process C4: If it is less than , the fitness of the individual particle is taken as the optimal fitness of the latest particle group; the current position of the particle is taken as the optimal position of the latest particle group; the weight vector of the particle is taken as the latest optimal weight vector.

[0149] Specifically, take the sth particle as an example:

[0150] like ,but ,and ;

[0151] like ,but ,and , while keeping in mind .

[0152] in, is the fitness of individual particles; is the optimal fitness of the current individual particle; is the latest optimal position of individual particles; is the current position of the particle; is the optimal fitness of the current particle group; is the optimal position of the latest particle group; is the latest optimal weight vector; is the weight vector of the particle.

[0153] Furthermore, the particle is moved. Taking the sth particle as an example, the velocity vector of each particle can be updated one by one according to formula (7) and formula (8): and the position vector .

[0154] (7)

[0155] (8)

[0156] in, and is the learning factor; is the inertia weight, between 0 and 1; and A random number between 0 and 1.

[0157] It should be noted that the inertia weight It is not a fixed value, but decreases linearly with the increase of the number of iterations, that is:

[0158] .

[0159] in MaxEpoch is the maximum number of iterations, epoch is the current number of iterations, It can be between 0.75 and 0.95. It can be between 0.25 and 0.45, and can be selected according to the maximum number of iterations. A larger inertia weight is used in the early stage of iteration. It is conducive to global optimization and uses a smaller inertia weight in the later stages of iteration. It is conducive to local optimization.

[0160] Process B8: Determine all weights and biases of the three-layer feedforward fully connected neural network according to the global optimal position and optimal weight vector of the particle swarm, and obtain the target three-layer feedforward fully connected neural network.

[0161] It can be understood that the global optimal position of the particle swarm determined by the iterative process and optimal , all weights and biases of the three-layer feedforward fully connected neural network can be determined, thereby completing the training of the initial three-layer feedforward fully connected neural network and obtaining the target three-layer feedforward fully connected neural network.

[0162] Step S104: The target three-layer feedforward fully connected neural network is verified based on all verification samples in the verification set samples. If the prediction relative error of all verification samples is less than the preset relative error percentage value, the final three-layer feedforward fully connected neural network is determined.

[0163] In the specific implementation of step S104, all verification samples in the verification set samples are used to verify the target three-layer feedforward fully connected neural network. When the prediction relative error of all verification samples is less than the preset relative error percentage value, it is determined that the final three-layer feedforward fully connected neural network is obtained. When the prediction relative error of all verification samples is not less than the preset relative error percentage value, the hidden layer node L+1 is returned to step S102 to recreate the initial three-layer feedforward fully connected neural network, and then the parameter settings of the particle swarm optimization algorithm are kept unchanged, and step S103 is executed.

[0164] It should be noted that the target three-layer feedforward fully connected neural network is verified using all the verification samples in the verification set samples. When the prediction relative error of all verification samples is less than the preset relative error percentage value, the specific process of determining the final three-layer feedforward fully connected neural network is as follows (process D1 to process D4):

[0165] Process D1: For any validation sample in the validation set, the feature value of the validation sample is input into the target three-layer feedforward fully connected neural network to obtain the network's predicted value.

[0166] It can be understood that for any validation sample in the validation set, the feature value of the validation sample is input into the target three-layer feedforward fully connected neural network to obtain the network's predicted value .

[0167] Process D2: Compare the predicted value with the label quantity in the validation sample to obtain the error.

[0168] It should be noted that the predicted value of the network and the number of labels in the validation sample Compare and get the error .

[0169] Process D3: Count the corresponding errors of all validation samples to obtain the prediction relative error.

[0170] It can be understood that the corresponding errors of all verification samples are statistically calculated as shown in formula (9).

[0171] (9)

[0172] in, is the prediction relative error of all validation samples.

[0173] Then judge % (preset relative error percentage value, such as 3% to 10%) is established. If so, execute process D4.

[0174] Process D4: If the predicted relative error is less than the preset relative error percentage value, the structure and all weight and bias parameters of the target three-layer feedforward fully connected neural network are solidified according to the number of hidden layer nodes, the global optimal position of the particle swarm and the optimal weight vector determined in the verification process to obtain the final three-layer feedforward fully connected neural network.

[0175] It is understandable that if % (that is, the predicted relative error is less than the preset relative error percentage value), then according to the number of hidden layer nodes, the global optimal position of the particle swarm and the optimal value determined in the network verification process, the structure and all weights and bias parameters of the three-layer feedforward fully connected neural network can be solidified, thereby completing the network verification work and obtaining the final three-layer feedforward fully connected neural network.

[0176] Step S105: predicting the mist cooling heat transfer coefficient according to the actual mist cooling condition parameters and the final three-layer feedforward fully connected neural network, and setting the mist cooling model based on the mist cooling heat transfer coefficient.

[0177] In the specific implementation of step S105, the actual aerosol cooling condition parameters are first obtained and normalized; then the normalized actual aerosol cooling condition parameters are input into the final three-layer feedforward fully connected neural network for prediction, and the network prediction value is output. ; Use the anti-normalization method to predict the network value The mist cooling heat transfer coefficient prediction value y is obtained by processing; and the mist cooling model is set based on the mist cooling heat transfer coefficient prediction value y.

[0178] It is understandable that the network prediction value is The process of obtaining the predicted value y of the mist cooling heat transfer coefficient is shown in formula (10).

[0179] (10)

[0180] in, is the network prediction value; y is the predicted value of the mist cooling heat transfer coefficient; y min The minimum value obtained by counting the number of labels in all samples; y max It is the maximum value obtained by counting the number of labels in all samples.

[0181] In an embodiment of the present invention, by utilizing the global optimization characteristics of the particle swarm optimization algorithm and the fast convergence learning characteristics of the extreme learning machine, a simple three-layer feedforward fully connected network can be used to perform supervised learning on small sample data obtained from the mist cooling heat transfer experiment, and has good generalization ability, which helps to accurately predict and control the mist cooling heat transfer coefficient under different actual working conditions, improve the cooling effect of the cast and rolled thin strip, ensure product quality and increase the yield rate.

[0182] In order to more clearly explain the above embodiments of the present invention Figure 1 For an explanation of the contents in Figure 3 , shows another flow chart of a method for predicting the heat transfer coefficient of a cast-rolled thin strip after rolling provided by an embodiment of the present invention, comprising:

[0183] like Figure 3 As shown, the method is divided into three parts, namely network training (step S301 to step S314), network verification (step S315 to step S320) and network prediction (step S321 to step S324).

[0184] Step S301: Divide the aerosol cooling experiment result data into initial training set samples and initial validation set samples.

[0185] Step S302: normalize the initial training set samples and the initial validation set samples to obtain training set samples and validation set samples.

[0186] Step S303: Determine the network input and output, and set the number of hidden layer nodes L=L min .

[0187] That is, determine the network input and output of the initial three-layer feedforward fully connected neural network, and set the number of hidden layer nodes L=L min .

[0188] Step S304: Create the MistHTCNet network.

[0189] It can be understood that an initial three-layer feed-forward fully connected neural network is created, named MistHTCNet network.

[0190] Step S305: Setting PSO optimization algorithm parameters.

[0191] That is to say, set the various parameters of the particle swarm optimization algorithm.

[0192] Step S306: Initialize the position and velocity of the particle.

[0193] It can be understood that the position and velocity of the particle are initialized.

[0194] Step S307: Combined with the ELM algorithm, calculate the particle fitness and the MistHTCNet network weight vector β.

[0195] Step S308: Particle fitness < Is it established? If so, execute step S309; ​​if not, execute step S310.

[0196] in, is the optimal fitness of the current individual particle.

[0197] Step S309: Update the optimal position pbest of the individual particle.

[0198] It is understandable that when the particle fitness < When it is established, update the optimal position pbest of the individual particle.

[0199] Step S310: Particle fitness < Is it true? If so, execute step S311; if not, execute step S312.

[0200] in, is the optimal fitness of the current particle group.

[0201] It is understandable that when the particle fitness < When it is not true, or after updating the optimal position pbest of the individual particle, determine whether the particle fitness is < Is it true?

[0202] Step S311: Update the optimal position gbest and optimal weight vector β* of the particle swarm.

[0203] That is, when the particle fitness < When it is established, update the optimal position gbest and optimal weight vector β* of the particle group.

[0204] Step S312: Calculate the speed and new position of the particle.

[0205] It is understandable that when the particle fitness < When it is not true, or after updating the optimal position gbest of the particle group and the optimal weight vector β*, the speed and new position of the particle movement are calculated.

[0206] Step S313: whether the iterative calculation conditions are met; if the iterative calculation conditions are met, execute step SS307; if the iterative calculation conditions are not met, execute step S314.

[0207] Step S314: According to the global optimal position and weight vector β* of the particle swarm in the iterative process, all weights and biases of the MistHTCNet network are obtained.

[0208] It can be understood that when the iterative calculation conditions are not met, all weights and biases of the MistHTCNet network are obtained according to the global optimal position and weight vector β* of the particle swarm in the iterative process, and thus the target three-layer feedforward fully connected neural network is obtained.

[0209] Step S315: The validation set samples are input into the MistHTCNet network for prediction.

[0210] It can be understood that the specific process is to input the validation set samples one by one into the trained MistHTCNet network (i.e., the target three-layer feedforward fully connected neural network) for prediction.

[0211] Step S316: determine whether the prediction relative error < δ% is true; if so, execute step S319; if not, execute step S317.

[0212] Step S317: Whether the number of hidden layer nodes L>Lmax holds; if so, execute step S319; if not, execute step S318.

[0213] Step S318: L=L+1.

[0214] It can be understood that when the number of hidden layer nodes L>Lmax does not hold, L=L+1 is executed, and the process returns to step S304.

[0215] Step S319: According to the global optimal position and weight vector β* of the particle swarm in the verification process, all weights and biases of the MistHTCNet network are obtained.

[0216] It can be understood that when the prediction relative error <δ%, or the number of hidden layer nodes L>Lmax, all weights and biases of the MistHTCNet network are obtained according to the global optimal position and weight vector β* of the particle swarm in the verification process.

[0217] Step S320: Solidify the MistHTCNet network structure and parameters.

[0218] In other words, based on the number of hidden layer nodes, the global optimal position and optimality of the particle swarm determined during the network verification process, the structure and all weights and bias parameters of the three-layer feedforward fully connected neural network are solidified, thereby completing the network verification work and obtaining the final three-layer feedforward fully connected neural network.

[0219] Step S321: normalizing actual aerosol cooling condition parameters.

[0220] That is to say, the actual aerosol cooling condition parameters are obtained and normalized.

[0221] Step S322: MistHTCNet network prediction.

[0222] It can be understood that the normalized actual aerosol cooling condition parameters are input into the MistHTCNet network (i.e., the final three-layer feedforward fully connected neural network) for prediction, and the network prediction value is output.

[0223] Step S323: Perform inverse normalization to obtain a predicted value of the mist cooling heat transfer coefficient.

[0224] That is to say, the network prediction value is denormalized to obtain the predicted value of the aerosol cooling heat transfer coefficient.

[0225] Step S324: mist cooling model setting calculation.

[0226] It can be understood that the mist cooling model is set according to the predicted value of the mist cooling heat transfer coefficient.

[0227] It should be noted that the specific implementation principles of steps S301 to S324 are detailed in the embodiments of the present invention. Figure 1 The contents in will not be repeated here.

[0228] In an embodiment of the present invention, by utilizing the global optimization characteristics of the particle swarm optimization algorithm and the fast convergence learning characteristics of the extreme learning machine, a simple three-layer feedforward fully connected network can be used to perform supervised learning on small sample data obtained from the mist cooling heat transfer experiment, and has good generalization ability, which helps to accurately predict and control the mist cooling heat transfer coefficient under different actual working conditions, improve the cooling effect of the cast and rolled thin strip, ensure product quality and increase the yield rate.

[0229] In order to better understand the embodiments of the present invention Figure 1 The following is a detailed description of the contents in the article. The following is an example of 1172 heat transfer experimental results obtained by cooling a steel plate made of Q235B, with a thickness of 1.0~2.8mm, a speed of 1.167~3.75m / s, and a temperature of 1100℃ under multiple groups of aerosol nozzles with a water pressure of 0.1~0.6MPa and an air pressure of 0.1~0.4MPa. Figure 1 The prediction method shown predicts the value of the mist cooling heat transfer coefficient under given mist cooling conditions.

[0230] Step 1: The data format of the heat transfer experiment results is shown in Table 1. The samples with water pressure of 1, 2, 3, 4, 5, and 6 bar and air pressure of 1, 2, 3, and 4 bar are divided into training set samples, totaling 928 samples, and the samples with water pressure of 1.5, 2.5, 3.5, and 4.5 bar and air pressure of 1.5 and 2.5 bar are divided into validation set samples, totaling 244 samples; the maximum and minimum values ​​of each feature quantity and label quantity are counted, and all sample data can be transformed into the range according to the minimum-maximum normalization method.

[0231] Table 1 (Sample data structure example)

[0232]

[0233] Step 2: Create a three-layer feedforward fully connected neural network named MistHTCNet. Since the sample has 5 feature quantities, the number of input layer nodes N=5. Since the sample has only 1 label quantity, the number of output layer nodes M=1. The range of the number of hidden layer nodes L can be set according to the embodiment of the present invention. Figure 1 The formula (2) shown in the figure can be used to calculate, and k can be set to 10, so L min =2,L max =13, L is initialized according to the minimum value, that is, L = 2. The activation function of the hidden layer node adopts the Sigmoid function.

[0234] Step 3: Set the PSO particle swarm optimization algorithm parameters, particle number n=30, maximum number of iterations MaxEpoch=100, learning factor c1=2, learning factor c2=2, and the inertia weight decreases linearly according to the number of iterations, that is, , where epoch is the current iteration number, The value can be 0.8. It can be taken as 0.3. The position vector and velocity vector of each particle have the same dimension, which is determined by the connection weights between the input layer node and the hidden layer node of the MistHTCNet network. ( ) and the bias of the hidden layer nodes ( ), the dimensions of the position vector and the velocity vector are .

[0235] By randomly changing the connection weights between the position vector and the velocity vector and bias are initialized to random numbers ranging from 0 to 1. The optimal fitness of the particle group And the optimal fitness of each particle , , are initialized to a large number 10 8 .

[0236] Step 4: The normalized The feature quantities of the training set samples are input into the MistHTCNet network. The connection weights between the input layer and the hidden layer of the network and the bias of the hidden layer nodes are determined by the position vector of the sth particle ( )Decide.

[0237] Through the embodiments of the present invention Figure 1 The network forward calculation can obtain the activation output of each hidden layer node by using formula (4) shown in .

[0238] Assume that the connection weight vector between the hidden layer and the output layer of the MistHTCNet network is , then continue the network forward calculation and get the The network output expression of samples is: .

[0239] Compare it with the number of labels of the sample By comparing, we can get an error By inputting 928 training samples into the MistHTCNet network one by one, we can establish a loss function constructed by the training error term: .

[0240] in, is a matrix of dimension 928×2, , is a label vector with a dimension of 928, .

[0241] Using the extreme learning machine ELM method, the loss function By minimizing, we can solve the connection weight vector between the hidden layer and the output layer. . The loss function is directly used as the fitness function of the particle. Substitute, through , we can calculate the fitness of the sth particle individual .

[0242] Similarly, by executing this step 30 times, the fitness of 30 individual particles can be obtained.

[0243] Step 5: Find the optimal position of individual particles and the optimal position of the particle group, then move the particles, update the particle speed and position, and perform the following steps in sequence:

[0244] Step 5.1: Perform the following judgments and operations on the calculated fitness of the 30 individual particles (taking the sth particle as an example), and you can update the optimal fitness and optimal position of the individual particle and the optimal fitness and optimal position of the particle group:

[0245] like ,but ,and ;

[0246] like ,but ,and , while keeping in mind .

[0247] Step 5.2: Move particles. Taking the sth particle as an example, according to the embodiment of the present invention, Figure 1 Formulas (7) and (8) shown in the figure can update the velocity vector of each particle one by one. and the position vector .

[0248] Step 5.3: Determine whether the number of PSO iterations exceeds the limit. If the number of iterations does not exceed the maximum number of iterations MaxEpoch = 100, go to step 4 to continue iterative calculation; otherwise, the global optimal position of the particle swarm determined by the iterative process is and optimal , all weights and biases of the MistHTCNet network can be determined, thus completing the network training.

[0249] Step 6: Put the normalized validation set The feature value of each sample is input into the trained MistHTCNet network, and the prediction value of the network is calculated. , and compare it with the label quantity in the sample Compare and get the error .

[0250] Count the relative errors of predictions for all validation samples , set the relative error measure ,like %, go to step 7, otherwise, determine whether the number of hidden layer nodes L of the MistHTCNet network exceeds the limit. If L>13, also go to step 7, otherwise, the number of hidden layer nodes L increases by 1, that is, L=L+1, and then go to step 2, recreate the MistHTCNet network, keep the PSO optimization algorithm parameter settings unchanged, and repeat steps 3 to 6.

[0251] Step 7: According to the network verification process, the number of hidden layer nodes L=6 and the global optimal position of the particle swarm are determined and optimal , we can finally determine the structure of the MistHTCNet network, which includes 5 nodes in the input layer, 6 nodes in the hidden layer, and 1 node in the output layer. The first 30 items in the vector are used as the connection weights between the input layer and the hidden layer. The last 6 items in the vector are used as the bias of the hidden layer nodes. The 6 items in the vector serve as the connection weights between the hidden layer and the output layer, thus completing the solidification of the MistHTCNet network.

[0252] The following is a prediction of the heat transfer coefficient of a steel plate made of Q235B, with a thickness of 2.5mm, a speed of 1167mm / s, and a temperature of 600℃ when subjected to mist cooling at a nozzle water pressure of 3.2bar and an air pressure of 2.3bar. First, the given thickness, speed, temperature, water pressure, and air pressure are normalized, and then input into the solidified MistHTCNet network. After network forward calculation, the network prediction value is obtained, and then the predicted value of the mist cooling heat transfer coefficient is obtained by the inverse normalization method. 389W / m 2 .K.

[0253] When setting the calculation of the mist cooling model, the MistHTCNet network can be directly called according to the aforementioned mist cooling condition parameters to obtain the predicted value of the mist cooling heat transfer coefficient, participate in the temperature field calculation, and set the opening and closing state of the mist cooling nozzle.

[0254] In an embodiment of the present invention, by utilizing the global optimization characteristics of the particle swarm optimization algorithm and the fast convergence learning characteristics of the extreme learning machine, a simple three-layer feedforward fully connected network can be used to perform supervised learning on small sample data obtained from the mist cooling heat transfer experiment, and has good generalization ability, which helps to accurately predict and control the mist cooling heat transfer coefficient under different actual working conditions, improve the cooling effect of the cast and rolled thin strip, ensure product quality and increase the yield rate.

[0255] Corresponding to the method for predicting the heat transfer coefficient of the cast and rolled thin strip after rolling provided by the above embodiment of the present invention, see Figure 4 , showing a structural block diagram of a device for predicting the heat transfer coefficient of post-rolling mist cooling of a cast thin strip provided by an embodiment of the present invention, the device includes: a division unit 401, a creation unit 402, a training unit 403, a verification unit 404 and a prediction unit 405.

[0256] The division unit 401 is used to divide the heat transfer experiment result data into initial training set samples and initial verification set samples, and perform normalization processing on the initial training set samples and the initial verification set samples to obtain training set samples and verification set samples.

[0257] The creation unit 402 is used to create an initial three-layer feedforward fully connected neural network according to the preset number of factors, the number of training set samples and the number of all samples.

[0258] The training unit 403 is used to train the initial three-layer feedforward fully connected neural network based on the training samples in the training set samples, the various parameters of the pre-set particle swarm optimization algorithm, the particle swarm optimization algorithm and the extreme learning machine to obtain the target three-layer feedforward fully connected neural network.

[0259] The verification unit 404 is used to verify the target three-layer feedforward fully connected neural network based on all verification samples in the verification set samples. If the prediction relative error of all verification samples is less than a preset relative error percentage value, the final three-layer feedforward fully connected neural network is determined.

[0260] The prediction unit 405 is used to predict the aerosol cooling heat transfer coefficient according to the actual aerosol cooling condition parameters and the final three-layer feedforward fully connected neural network, and set the aerosol cooling model based on the aerosol cooling heat transfer coefficient.

[0261] The prediction unit 405 is specifically used to: obtain actual aerosol cooling condition parameters and perform normalization processing; input the normalized actual aerosol cooling condition parameters into the final three-layer feedforward fully connected neural network for prediction, and output the network prediction value; use the anti-normalization method to process the network prediction value to obtain the aerosol cooling heat transfer coefficient prediction value; set the aerosol cooling model based on the aerosol cooling heat transfer coefficient prediction value.

[0262] In an embodiment of the present invention, by utilizing the global optimization characteristics of the particle swarm optimization algorithm and the fast convergence learning characteristics of the extreme learning machine, a simple three-layer feedforward fully connected network can be used to perform supervised learning on small sample data obtained from the mist cooling heat transfer experiment, and has good generalization ability, which helps to accurately predict and control the mist cooling heat transfer coefficient under different actual working conditions, improve the cooling effect of the cast and rolled thin strip, ensure product quality and increase the yield rate.

[0263] Combination Figure 4 The content shown, the creation unit 402, includes: a setting module, a first determination module, a first calculation module, a second determination module and a creation module.

[0264] The setting module is used to set the number of input layer nodes according to the preset number of factors.

[0265] The first determination module is used to determine the number of output layer nodes to be 1.

[0266] The first calculation module is used to obtain the number of training set samples and the number of all samples, and calculate the value range of the number of hidden layer nodes according to the floor function.

[0267] The second determination module is used to determine the S-shaped growth curve function as the activation function of the hidden layer node.

[0268] Create a module for creating an initial three-layer feedforward fully connected neural network based on the range of values ​​of the number of input layer nodes, the number of output layer nodes, the number of hidden layer nodes, and the activation function of the hidden layer nodes.

[0269] Combination Figure 4 The content shown, training unit 403, includes: a second calculation module, a third calculation module, a fourth calculation module, a construction module, a third determination module, a fifth calculation module, a fourth determination module and a fifth determination module.

[0270] The second calculation module is used to input the feature quantity of any training sample in the training set samples into the initial three-layer feedforward fully connected neural network, and perform network forward calculation of the activation output of each hidden layer node according to the extreme learning machine algorithm.

[0271] The third calculation module is used to perform network forward calculation according to the activation output of each hidden layer node to obtain a network output expression.

[0272] The fourth calculation module is used to calculate the error amount based on the network output expression and the label amount of the training sample.

[0273] The construction module is used to construct a loss function according to the corresponding error amounts of multiple training samples.

[0274] The third determination module is used to minimize the loss function based on the extreme learning machine algorithm to determine the connection weight vector between the hidden layer and the output layer.

[0275] The fifth calculation module is used to perform iterative operations, and according to the number of particles of the particle swarm optimization algorithm, for each particle, calculates the fitness of the individual particle according to the loss function and the connection weight vector between the hidden layer and the output layer.

[0276] The fourth determination module is used to adjust the optimal fitness and optimal position of the individual particle, the optimal fitness and optimal position of the particle group, and the optimal weight vector according to the fitness of the individual particle, until the current number of iterations has reached a preset maximum number of iterations, and determine the global optimal position and optimal weight vector of the particle group.

[0277] The fourth determination module is specifically used to: determine whether the fitness of the individual particle is less than the current individual particle optimal fitness; if less than, take the individual particle fitness as the latest individual particle optimal fitness; take the current position of the particle as the latest individual particle optimal position.

[0278] Determine whether the fitness of the individual particle is less than the optimal fitness of the current particle group; if so, take the fitness of the individual particle as the optimal fitness of the latest particle group; take the current position of the particle as the optimal position of the latest particle group; take the weight vector of the particle as the latest optimal weight vector.

[0279] The fifth determination module is used to determine all weights and biases of the three-layer feedforward fully connected neural network according to the global optimal position and optimal weight vector of the particle swarm, and obtain the target three-layer feedforward fully connected neural network.

[0280] Combination Figure 4 The content shown, the verification unit 404, includes: an input module, a comparison module, a statistical module and a solidification module.

[0281] The input module is used to input the feature quantity of any verification sample in the verification set samples into the target three-layer feedforward fully connected neural network to obtain the predicted value of the network.

[0282] The comparison module is used to compare the predicted value with the label quantity in the validation sample to obtain the error.

[0283] The statistical module is used to count the corresponding errors of all validation samples to obtain the relative prediction error.

[0284] The solidification module is used to solidify the structure and all weight and bias parameters of the target three-layer feedforward fully connected neural network according to the number of hidden layer nodes, the global optimal position of the particle swarm and the optimal weight vector determined in the verification process if the predicted relative error is less than the preset relative error percentage value, so as to obtain the final three-layer feedforward fully connected neural network.

[0285] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without creative work.

[0286] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0287] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the heat transfer coefficient of gas mist cooling after casting and rolling thin strip, characterized in that: The method comprises: Dividing the heat transfer experiment result data into initial training set samples and initial verification set samples, and performing normalization processing on the initial training set samples and the initial verification set samples to obtain training set samples and verification set samples; Creating an initial three-layer feedforward fully connected neural network according to the number of preset factors, the number of samples in the training set, and the number of all samples; Based on the training samples in the training set samples, various parameters of the pre-set particle swarm optimization algorithm, the particle swarm optimization algorithm and the extreme learning machine, the initial three-layer feedforward fully connected neural network is trained to obtain a target three-layer feedforward fully connected neural network; The target three-layer feedforward fully connected neural network is verified based on all verification samples in the verification set samples, and if the prediction relative errors of all verification samples are less than a preset relative error percentage value, a final three-layer feedforward fully connected neural network is determined; According to the actual aerosol cooling condition parameters and the final three-layer feedforward fully connected neural network, the aerosol cooling heat transfer coefficient is predicted, and the aerosol cooling model is set based on the aerosol cooling heat transfer coefficient.

2. The method according to claim 1, characterized in that: The step of creating an initial three-layer feedforward fully connected neural network according to the number of preset factors, the number of samples in the training set, and the number of all samples includes: Set the number of input layer nodes according to the preset number of factors; Set the number of output layer nodes to 1; Obtain the number of samples in the training set and the number of all samples, and calculate the value range of the number of hidden layer nodes according to a rounding-down function; The S-shaped growth curve function is determined as the activation function of the hidden layer nodes; An initial three-layer feedforward fully connected neural network is created based on the value ranges of the number of input layer nodes, the number of output layer nodes, the number of hidden layer nodes, and the activation function of the hidden layer nodes.

3. The method according to claim 1, characterized in that The initial three-layer feedforward fully connected neural network is trained based on the training samples in the training set samples, various parameters of the pre-set particle swarm optimization algorithm, the particle swarm optimization algorithm and the extreme learning machine to obtain the target three-layer feedforward fully connected neural network, including: For any training sample in the training set, input the feature value of the training sample into the initial three-layer feedforward fully connected neural network, and perform network forward calculation on the activation output of each hidden layer node according to the extreme learning machine algorithm; Perform network forward calculation based on the activation output of each hidden layer node to obtain the network output expression; Calculating an error amount based on the network output expression and the label amount of the training sample; Constructing a loss function according to the corresponding error amounts of the plurality of training samples; Minimize the loss function based on the extreme learning machine algorithm to determine the connection weight vector between the hidden layer and the output layer; Performing an iterative operation, according to the number of particles of the particle swarm optimization algorithm, for each particle, calculating the fitness of the individual particle according to the loss function and the connection weight vector between the hidden layer and the output layer; According to the fitness of the individual particles, the optimal fitness and optimal position of the individual particles, the optimal fitness and optimal position of the particle group, and the optimal weight vector are adjusted until the current number of iterations reaches a preset maximum number of iterations, and the global optimal position and optimal weight vector of the particle group are determined; All weights and biases of a three-layer feedforward fully connected neural network are determined according to the global optimal position and optimal weight vector of the particle swarm, and a target three-layer feedforward fully connected neural network is obtained.

4. The method according to claim 3, characterized in that: The step of adjusting the optimal fitness and optimal position of the individual particle, the optimal fitness and optimal position of the particle group, and the optimal weight vector according to the fitness of the individual particle includes: Determine whether the fitness of the individual particle is less than the current optimal fitness of the individual particle; If it is less than, the fitness of the particle individual is taken as the latest optimal fitness of the particle individual; Taking the current position of the particle as the latest optimal position of the particle; Determine whether the fitness of the individual particle is less than the optimal fitness of the current particle group; If it is less than, the fitness of the individual particle is taken as the optimal fitness of the latest particle group; Using the current position of the particle as the optimal position of the latest particle group; The weight vector of the particle is taken as the latest optimal weight vector.

5. The method according to claim 1, characterized in that The target three-layer feedforward fully connected neural network is verified based on all verification samples in the verification set samples, and if the prediction relative errors of all verification samples are less than a preset relative error percentage value, a final three-layer feedforward fully connected neural network is determined, including: For any verification sample in the verification set samples, input the feature quantity of the verification sample into the target three-layer feedforward fully connected neural network to obtain the prediction value of the network; Compare the predicted value with the label quantity in the validation sample to obtain an error; The corresponding errors of all validation samples are counted to obtain the relative prediction error; If the predicted relative error is less than the preset relative error percentage value, the structure and all weight and bias parameters of the target three-layer feedforward fully connected neural network are solidified according to the number of hidden layer nodes, the global optimal position of the particle swarm and the optimal weight vector determined in the verification process to obtain the final three-layer feedforward fully connected neural network.

6. The method according to claim 1, characterized in that The method predicts the aerosol cooling heat transfer coefficient according to the actual aerosol cooling condition parameters and the final three-layer feedforward fully connected neural network, and sets the aerosol cooling model based on the aerosol cooling heat transfer coefficient, including: Obtain actual aerosol cooling condition parameters and perform normalization processing; Inputting the normalized actual aerosol cooling condition parameters into the final three-layer feedforward fully connected neural network for prediction, and outputting the network prediction value; The network prediction value is processed by using the anti-normalization method to obtain the predicted value of the aerosol cooling heat transfer coefficient; The mist cooling model is set based on the predicted value of the mist cooling heat transfer coefficient.

7. A device for predicting heat transfer coefficient of gas mist cooling after casting and rolling thin strip, characterized in that: The device comprises: A division unit, used for dividing the heat transfer experiment result data into an initial training set sample and an initial verification set sample, and performing normalization processing on the initial training set sample and the initial verification set sample to obtain a training set sample and a verification set sample; A creation unit, used for creating an initial three-layer feedforward fully connected neural network according to the number of preset factors, the number of samples in the training set and the number of all samples; A training unit, configured to train the initial three-layer feedforward fully connected neural network based on the training samples in the training set samples, various parameters of the pre-set particle swarm optimization algorithm, the particle swarm optimization algorithm and the extreme learning machine to obtain a target three-layer feedforward fully connected neural network; A verification unit, used to verify the target three-layer feedforward fully connected neural network based on all verification samples in the verification set samples, and if the prediction relative error of all verification samples is less than a preset relative error percentage value, then determine to obtain a final three-layer feedforward fully connected neural network; The prediction unit is used to predict the aerosol cooling heat transfer coefficient according to the actual aerosol cooling condition parameters and the final three-layer feedforward fully connected neural network, and set the aerosol cooling model based on the aerosol cooling heat transfer coefficient.

8. The device according to claim 7, characterized in that The creation unit comprises: A setting module, used to set the number of input layer nodes according to the number of preset factors; A first determination module is used to determine the number of output layer nodes to be 1; A first calculation module is used to obtain the number of samples in the training set and the number of all samples, and calculate the value range of the number of hidden layer nodes according to a rounding-down function; A second determination module is used to determine the S-shaped growth curve function as the activation function of the hidden layer node; A creation module is used to create an initial three-layer feedforward fully connected neural network based on the value range of the number of input layer nodes, the number of output layer nodes, the number of hidden layer nodes and the activation function of the hidden layer nodes.

9. The device according to claim 7, characterized in that The training unit comprises: A second calculation module is used for inputting the feature quantity of any training sample in the training set samples into the initial three-layer feedforward fully connected neural network, and performing network forward calculation of the activation output of each hidden layer node according to the extreme learning machine algorithm; The third calculation module is used to perform network forward calculation according to the activation output of each hidden layer node to obtain a network output expression; A fourth calculation module, used for calculating an error amount based on the network output expression and the label amount of the training sample; A construction module, used to construct a loss function according to the corresponding error amounts of the plurality of training samples; A third determination module is used to minimize the loss function based on an extreme learning machine algorithm to determine a connection weight vector between the hidden layer and the output layer; A fifth calculation module is used to perform an iterative operation, and according to the number of particles of the particle swarm optimization algorithm, for each particle, calculate the fitness of the individual particle according to the loss function and the connection weight vector between the hidden layer and the output layer; A fourth determination module is used to adjust the optimal fitness and optimal position of the individual particle, the optimal fitness and optimal position of the particle group, and the optimal weight vector according to the fitness of the individual particle, until the current number of iterations has reached a preset maximum number of iterations, and determine the global optimal position and optimal weight vector of the particle group; The fifth determination module is used to determine all weights and biases of the three-layer feedforward fully connected neural network according to the global optimal position and optimal weight vector of the particle swarm to obtain a target three-layer feedforward fully connected neural network.

10. The device according to claim 9, characterized in that The fourth determining module is specifically used for: Determine whether the fitness of the individual particle is less than the current individual optimal fitness of the particle; if so, take the individual fitness of the particle as the latest individual optimal fitness of the particle; take the current position of the particle as the latest individual optimal position of the particle; Determine whether the fitness of the individual particle is less than the optimal fitness of the current particle group; if so, take the fitness of the individual particle as the optimal fitness of the latest particle group; take the current position of the particle as the optimal position of the latest particle group; take the weight vector of the particle as the latest optimal weight vector.

Images