Multi-index prediction method for continuous annealing strip properties based on multi-objective evolutionary deep learning

Abstract

The present invention provides a multi-index prediction method for the performance of continuously annealed strip steel based on multi-objective evolutionary deep learning, and relates to the technical field of data processing in the cold rolling production control process of steel enterprises. The multi-objective evolutionary algorithm is applied to the hyperparameter selection of the deep learning model to further improve the accuracy of the online detection method; at the same time, the training set is obtained by random selection, and its search strategy in the hyperparameter space can quickly obtain the optimal hyperparameter combination of the deep learning model, and finally improve the generalization ability of the online detection model. The method proposed in the present invention overcomes the shortcomings of the traditional deep neural network model based on artificial experience and realizes the simultaneous prediction of multiple performance indicators, realizes the autonomous optimization construction of the deep learning model, and the multi-index prediction of the strip steel performance, which can help cold rolling production improve the level of product quality control.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing in a cold rolling production control process of a steel enterprise, and in particular to a multi-index prediction method for continuous annealing strip steel performance based on multi-objective evolutionary deep learning. Background Art

[0002] Continuous annealing is an important step in the cold rolling process of the steel industry, which can improve the physical properties and processing properties of metal materials, making them more suitable for various industrial applications. Since the continuous annealing production process involves multiple processing stages such as heating, heat preservation, rapid cooling, slow cooling, aging, quenching, and each processing stage contains multiple control process parameters with mutual coupling relationships, the process mechanism is complex, making it difficult to accurately judge the multiple performance indicators of the continuous annealing strip steel based on manual experience in the actual production process, which leads to large fluctuations in the strip steel quality indicators during the production process, affecting the continuous and stable operation of production.

[0003] Therefore, data-driven modeling methods represented by deep learning have become an important means of predictive modeling of performance indicators of continuous annealing strip steel, and are key technologies for improving the quality of continuous annealing strip steel products and the level of intelligence in production process control. For example: The invention patent "A method and system for predicting the performance of cold-rolled steel sheets based on deep learning" with application number CN116796627A proposes a method for predicting the performance of cold-rolled steel sheets based on deep learning, in which the parameters of the deep neural network are mainly set based on artificial experience. Although deep neural networks have many advantages in capturing implicit relationships between process control parameters and obtaining associated features, the architecture design of deep learning models and the setting of network parameters usually rely heavily on artificial experience, which makes the strip performance prediction model based on traditional deep learning prone to reduced generalization ability when the working conditions change, and it is difficult to meet the needs of actual production. In addition, most machine learning-based continuous annealing strip performance prediction models can usually only predict a single performance indicator (for example, the invention patent with patent number ZL201710159565.8 "A method for online detection of continuous annealing strip quality based on hybrid ensemble learning" and the invention patent with patent number ZL201410843307.8 "A method for online ensemble learning prediction of hardness of continuous annealing products"), and cannot use one model to predict multiple performance indicators at the same time. Summary of the invention

[0004] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and to provide a multi-index prediction method for continuous annealing strip performance based on multi-objective evolutionary deep learning, which overcomes the shortcomings of traditional deep neural network models based on artificial experience and realizes the simultaneous prediction of multiple performance indicators. A global search is performed in the hyperparameter space of the deep learning model through a multi-objective evolutionary algorithm to obtain the hyperparameter combination with the best performance, thereby realizing the autonomous optimization construction of the deep learning model and the multi-index prediction of strip performance.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A multi-index prediction method for continuous annealing strip performance based on multi-objective evolutionary deep learning, comprising the following steps:

[0007] Step 1: Build a sample set;

[0008] The output of a sample is multiple performance indicators of a continuous annealing strip, and the input is the process data of the strip during continuous annealing production; n samples under normal production conditions are collected from the historical data of the cold rolling continuous annealing unit to establish a sample set S sam ;

[0009] Step 2: Data preprocessing;

[0010] For the established sample data set S sam Normalize the n sample data in;

[0011] Step 3: Generate a multi-objective evolutionary deep learning model;

[0012] According to the n samples obtained in step 2, the network hyperparameters are trained using the multi-objective evolutionary deep learning method based on NSGA-Ⅱ;

[0013] Step 4: Select the optimal solution;

[0014] Based on the objective function MSE of each solution in the final population obtained in step 3 sum and model complexity, using the TOPSIS method to select and obtain a top-ranked solution as the optimal hyperparameter combination for the CNN architecture;

[0015] Step 5: Build a hybrid deep learning network model;

[0016] According to the values ​​of each hyperparameter in the final solution selected in step 4, a corresponding deep convolutional neural network is constructed, and then the network is trained on all data samples to obtain the corresponding deep convolutional neural network CNN;

[0017] Step 6: Based on the deep convolutional neural network CNN obtained in step 5, a Transformer network is superimposed on it to construct a hybrid deep learning network model CNN-Transformer, and then the network is trained on all data samples to obtain the final strip performance multi-index prediction model CNN-Transformer;

[0018] Step 7: Obtain the continuous annealing production process data of each strip in real time through the OPC interface, and use the CNN-Transformer model obtained in step 6 to predict the three key performance indicators of yield strength, tensile strength and elongation to achieve online accurate prediction of multiple indicators; when the deviation between the predicted value and the expected value of the performance indicator exceeds the given range, a timely alarm is issued to ensure the stability of the quality of the continuous annealing strip product.

[0019] Furthermore, the specific method of step 3 is:

[0020] Step 3.1: Randomly divide the preprocessed n samples into 0.8n and 0.2n samples to form the training set D and D respectively. train , validation set D validate ; If 0.8n and 0.2n are not integers, round them down;

[0021] Step 3.2: Set the maximum number of iterations L of the multi-objective evolutionary deep learning algorithm max , population size N pop , network hyperparameter set S para ={ x 1, x 2, x 3, x 4, x 5, x 6}, set the iteration counter l =0; where x 1. x 2. x 3. x 4. x 5. x 6 correspond to the six parameters of the convolutional neural network model: number of convolutional layers, number of output channels, convolution kernel size, number of hidden units, learning rate, and weight decay;

[0022] Step 3.3: Initialize the population; use all hyperparameters of the deep learning network model as decision variables and randomly generate N pop A real number vector with a length equal to the number of hyperparameters is used as the initial population to form the first l Generation population; where each vector x=( x 1, x 2, x 3,x 4, x 5, x 6) is a solution in the population, representing a hyperparameter combination scheme, based on which a corresponding deep learning network can be constructed;

[0023] Step 3.4: For each solution in the population, build a corresponding deep convolutional neural network according to its corresponding hyperparameter value, and then use it in the training data set D train The deep convolutional neural network is trained on the validation set D validate Calculate the prediction accuracy and complexity of the deep convolutional neural network model as the two optimization objective functions of the solution; the prediction accuracy is defined as the sum of the mean square errors (MSE) of multiple indicators sum ;

[0024] Step 3.5: Perform non-dominated sorting on the solutions in the current population according to the objective function and crowding distance, and select N from the current population using the binary tournament strategy. pop solutions as the parent solution set;

[0025] Step 3.6: Randomly select two solutions from the population, perform crossover and mutation operations in sequence to generate new offspring individuals; repeat N times pop times to get N pop Sub-generation solution;

[0026] Step 3.7: For each child solution, use the objective function calculation method in step 3.4 to obtain the objective function of the child solution;

[0027] Step 3.8: The newly generated N pop The sub-generation solutions are merged with the current population to form a population of size 2N pop population;

[0028] Step 3.9: Set the new population to an empty set, and then use the fast non-inferior sorting method of NSGA-II to sort the 2N pop The solutions are divided into different Pareto frontiers; secondly, starting from the best first Pareto frontier, solutions are selected from each frontier in turn to generate the next generation population;

[0029] Step 3.10: Set the number of iterations l = l +1; if l ≥L max , go to step 4; otherwise, return to step 3.6.

[0030] Furthermore, in step 3.2, the architecture of the deep learning network model needs to be designed accordingly, and the establishment of the model includes the following steps:

[0031] Step 3.2.1: Convolutional layer design; the network includes convolutional layers, each of which uses output channels and size is x The convolution kernel is 3=3, and the convolution layer introduces nonlinear transformation through the ReLU activation function;

[0032] Step 3.2.2: Optimizer and loss function selection: Adam optimizer and mean square error loss function MSELoss are used. The loss function measures the difference between the predicted value and the actual value by the square difference between the network output and the target.

[0033] Step 3.2.3: Fully connected layer design; the network includes 3 fully connected layers, each of which contains hidden units, each of which is followed by a ReLU activation function;

[0034] Step 3.2.4: Set weight decay; the weights of the convolutional layer and the linear layer are initialized using the Xavier initialization method. During training, the entire network continuously optimizes parameters to minimize the loss function through back propagation and gradient descent. The learning rate is set to , weight decay setting .

[0035] Furthermore, the specific method of the crossover and mutation operations in step 3.6 is:

[0036] Step 3.6.1: Perform SBX crossover on the selected individuals; first generate random numbers u ,if u <0.9, then perform crossover, otherwise, directly copy the parent individuals to the next generation;

[0037] Step 3.6.2: Perform mutation operation on the newly generated individuals; first generate random numbers u ',if u ′<0.5, then perform a mutation operation and select the parameters of an individual from the current population x i , according to the formula Get the individual parameters after mutation, where x i,new is the individual parameter after mutation, x i,old are the selected individual parameters, x i,upper and x i,lower The decision variables are x i The upper and lower bounds of , i=1,2,…6;δ represents the amount of variation; otherwise, the individual is kept unchanged.

[0038] Furthermore, the specific process of step 3.9 is as follows:

[0039] If all the solutions contained in the current Pareto front are added to the new population, the number of solutions in the new population is still less than N. pop , then all the solutions in the Pareto front are added to the new population and moved to the next Pareto front; otherwise, the most crowded solutions are deleted from the current Pareto front based on the crowding distance until the remaining solutions are added to the new population so that the size of the new population can reach N pop , and finally add all the remaining solutions to the new population to form a new population for the next iteration.

[0040] Furthermore, the specific method of step 4 is:

[0041] Step 4.1: Construct a decision matrix;

[0042] Build contains v The decision matrix X with 2 rows and 2 columns, MSE sum and Complexity as different columns, where the decision matrix is ;

[0043] Step 4.2: Standardize the decision matrix;

[0044] The decision matrix is ​​standardized to ensure that all indicators are in the same numerical range. The standardized matrix is ​​recorded as , where x ij represents the elements of the original decision matrix;

[0045] Step 4.3: Determine the weights;

[0046] Determine the weight for each indicator to indicate its relative importance in the comprehensive evaluation, where w1=0.6 and w2=0.4 are the weights of the two indicators, which correspond to the characteristic MSE of the two columns of the decision matrix X. sum and complexity;

[0047] Step 4.4: Construct a weighted normalized decision matrix;

[0048] Multiply the standardized decision matrix by the weights to get the weighted normalized decision matrix , represents the elements of the standardized decision matrix;

[0049] Step 4.5: Calculate and determine the positive ideal solution and negative ideal solution ;

[0050] Step 4.6: Calculate the distance;

[0051] Calculate the distance of each solution to the positive ideal solution The distance to the negative ideal solution ;

[0052] Step 4.7: Calculate the comprehensive score;

[0053] Calculate the closeness of each solution The index value corresponding to each solution, ;

[0054] Step 4.8: Select the best solution;

[0055] According to the comprehensive score, the C i Sort them and select the top-ranked solution from high to low as the optimal hyperparameter setting for the convolutional network model.

[0056] Furthermore, the specific method of step 5 is:

[0057] Step 5.1: Obtain the optimal individual; use the optimal individual selected in step 4 as the object of decoding, where the optimal individual includes a set of optimal hyperparameter settings of the convolutional neural network CNN;

[0058] Step 5.2: Decode different types of hyperparameters respectively to ensure that the decoded hyperparameters meet the defined parameter range. If they exceed the defined range, truncate them and find the closest real value within the decision variable value range.

[0059] Step 5.3: Apply the decoded hyperparameters to the architecture and training process of the convolutional neural network model CNN.

[0060] The beneficial effects of adopting the above technical solution are: the multi-index prediction method for the performance of continuous annealing strip steel based on multi-objective evolutionary deep learning provided by the present invention applies the multi-objective evolutionary algorithm to the hyperparameter selection of the deep learning model, further improving the accuracy of the online detection method; at the same time, the training set is obtained by random selection, and its search strategy in the hyperparameter space can quickly obtain the optimal hyperparameter combination of the deep learning model, and finally improve the generalization ability of the online detection model. After actual production data testing, the method proposed by the present invention can realize the prediction of multiple performance indicators of continuous annealing strip steel, thereby helping cold rolling production to improve the level of product quality control. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a cold rolling production process and data collection based on data drive provided by an embodiment of the present invention;

[0062] Figure 2 A flowchart of a multi-objective evolutionary deep learning model algorithm provided by an embodiment of the present invention;

[0063] Figure 3 A diagram of a network architecture for predicting multiple indices of continuous strip steel performance provided by an embodiment of the present invention;

[0064] Figure 4 A comparison chart of tensile strength prediction provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0066] For a strip steel of a certain tempering degree, this embodiment proposes a multi-index prediction method for continuous annealing strip steel performance based on multi-objective evolutionary deep learning, as described below.

[0067] Step 1: Build a sample set.

[0068] The output of a sample is multiple performance indicators (yield strength, tensile strength, elongation, etc.) of a continuous annealing strip, and the input is the process data of the strip during continuous annealing production. n = 15201 samples under normal production conditions are collected from the historical data of the cold rolling continuous annealing unit to establish a sample set S sam In this embodiment, the data collection items of the cold rolling production process are shown in Table 1. Figure 1 shown.

[0069] Table 1 Collection item information of cold rolling production process data

[0070] Collection item name unit Thickness of continuous withdrawal plan mm Width of continuous withdrawal plan mm Rp0.2 N / mm2 Elongation % Upper surface roughness μm Lower surface roughness μm Furnace speed m / min JPF ℃ … … Oven temperature ℃ Average coiling temperature ℃ Average finishing temperature ℃ Average final rolling temperature ℃ Continuous pickling speed m / s Rolling elongation % Elongation of flat mill % Leveling machine entrance tension kN

[0071] Step 2: Data preprocessing. sam The n=15201 sample data in are normalized and scaled to the range of [0,1] to ensure the regularity of the data.

[0072] Step 3: Generation of multi-objective evolutionary deep learning model. Based on the 15,201 samples obtained in step 2, the network hyperparameters are trained using the multi-objective evolutionary deep learning method based on NSGA-Ⅱ. The flowchart is as follows: Figure 2 As shown, the main process includes:

[0073] Step 3.1: Randomly divide 0.8n=12160 and 0.2n=3040 samples from the preprocessed n=15201 samples to form the training set D train and validation set Dvalidate .

[0074] Step 3.2: Set the maximum number of iterations L of the multi-objective evolutionary deep learning algorithm max =100, population size N pop =100, crossover probability p c =0.9, mutation probability p m =0.5, SBX crossover distribution index η c =5, network hyperparameter set S para ={ x 1, x 2, x 3, x 4, x 5, x 6}, set the iteration counter l =0. x 1. x 2. x 3. x 4. x 5. x 6 correspond to the various parameters of the convolutional neural network model: number of convolutional layers, number of output channels, convolution kernel size, number of hidden units, learning rate, and weight decay.

[0075] In the deep learning model, the model architecture needs to be designed accordingly. The establishment of the model includes the following steps:

[0076] Step 3.2.1: Convolutional layer design: The network contains convolutional layers, each of which uses output channels and size is x The convolution kernel is 3=3, and the convolution layer introduces nonlinear transformation through the ReLU activation function.

[0077] Step 3.2.2: Optimizer and loss function selection: Adopt Adam optimizer and mean square error loss function MSELoss. The loss function measures the difference between the predicted value and the actual value by the square difference between the network output and the target.

[0078] Step 3.2.3: Fully connected layer design: The network contains three fully connected layers, each of which contains There are hidden units, and each hidden unit is followed by a ReLU activation function to improve the nonlinear ability of the model.

[0079] Step 3.2.4: Set weight decay: The weights of the convolutional layer and the linear layer are initialized using the Xavier initialization method. During training, the entire network continuously optimizes parameters to minimize the loss function through back propagation and gradient descent. The learning rate is set , weight decay setting .

[0080] Step 3.3: Initialize the population: Take all the hyperparameters of the deep learning model as decision variables, randomly generate 100 real number vectors with a length equal to the number of hyperparameters as the initial population, and form the first l Generation population. Among them, each vector x=( x 1, x 2, x 3, x 4, x 5, x 6) is a solution in the population, representing a hyperparameter combination scheme. A corresponding deep learning network can be constructed based on this individual.

[0081] Step 3.4: For each solution in the population, build a corresponding deep convolutional neural network according to its corresponding hyperparameter value, and then use it in the training data set D train The network is trained on the validation set D validate Calculate the prediction accuracy of the network model (defined as the sum of the mean square errors of multiple indicators ) and complexity as the two optimization objective functions of the solution. 0.2n represents the number of samples in the validation set, j={1,2,3} represents the indicators of yield strength, tensile strength and elongation, is the actual value of the jth indicator of the ith sample, is the predicted value of the jth indicator of the ith sample.

[0082] Step 3.5: Perform non-dominated sorting on the solutions in the current population according to the objective function and crowding distance, and select 100 solutions from the current population as the parent solution set using the binary tournament strategy.

[0083] Step 3.6: Randomly select two solutions from the population, perform crossover and mutation operations in sequence to generate new offspring individuals. Repeat 100 times to obtain 100 offspring solutions.

[0084] Step 3.6.1: Perform SBX crossover on the selected individuals: First generate random numbers u ,if u <0.9, a crossover is performed, otherwise, the parent individuals are directly copied to the next generation.

[0085] Step 3.6.2: Perform mutation operation on the newly generated individuals: First generate random numbers u ',if u ′<0.5, then perform a mutation operation and select the parameters of an individual from the current population x i , according to the formula Get the individual parameters after mutation. Otherwise, keep the individual unchanged. x i,new is the individual parameter after mutation, x i,old are the selected individual parameters, x i,upper and x i,lower The decision variables are x i (i=1,2,…6) upper and lower bounds, δ Indicates the amount of variation, set δ =0.2.

[0086] Step 3.7: For each child solution, use the objective function calculation method in step 3.4 to obtain the objective function of the solution.

[0087] Step 3.8: Merge the newly generated 100 offspring solutions with the current population to form a population of size 200.

[0088] Step 3.9: Set the new population to an empty set, and then use the fast non-inferior sorting method of NSGA-II to divide the 200 solutions into different Pareto fronts for the merged population; secondly, starting from the best first Pareto front, select solutions from each front in turn to generate the next generation population. The main process includes: if all the solutions contained in the current Pareto front are added to the new population, and the number of solutions in the new population is still less than 100, then all the solutions in the Pareto front are added to the new population and transferred to the next Pareto front; otherwise, based on the crowding distance, the most crowded solutions are deleted from the current Pareto front in turn until the remaining solutions are added to the new population so that the size of the new population reaches 100, and finally all the remaining solutions are added to the new population to form a new population for the next iteration.

[0089] Step 3.10: Set the number of iterations l = l +1; if l ≥L max , go to step 4; otherwise, return to step 3.6.1.

[0090] Step 4: Select the optimal solution: based on the objective function MSE of each solution in the final population sumand model complexity, and the TOPSIS method is used to select and obtain a top-ranked solution as the optimal hyperparameter combination of the CNN architecture.

[0091] Step 4.1: Construct a decision matrix: Construct a decision matrix X with 100 rows and 2 columns. sum and Complexity as different columns, where the decision matrix is .

[0092] Step 4.2: Standardize the decision matrix: Standardize the decision matrix to ensure that all indicators are in the same numerical range. The standardized matrix is ​​recorded as , where x ij represents the elements of the original decision matrix, v This means that each column has 100 elements.

[0093] Step 4.3: Determine the weights: For each indicator (i.e., the feature MSE of the two columns of the decision matrix X sum and complexity) to determine the weights, indicating their relative importance in the comprehensive evaluation, where w1=0.6 and w2=0.4.

[0094] Step 4.4: Construct a weighted normalized decision matrix: Multiply the standardized decision matrix by the weights to obtain a weighted normalized decision matrix , Represents the elements of the standardized decision matrix.

[0095] Step 4.5: Determine the positive ideal solution and the negative ideal solution: The positive ideal solution is calculated as , the negative ideal solution is .

[0096] Step 4.6: Calculate the distance: Calculate the distance of each solution to the positive ideal solution , the distance to the negative ideal solution .

[0097] Step 4.7: Calculate the overall score: Calculate how close each solution is The index value corresponding to each solution.

[0098] Step 4.8: Select the best solution: C of all solutions according to the comprehensive score i Sort them and select the top-ranked solution from high to low as the optimal hyperparameter setting for the convolutional network model.

[0099] Step 5: Build a hybrid deep learning network model. According to the values ​​of each hyperparameter in the final solution selected in step 4, build a corresponding deep convolutional neural network, and then train the network on all data samples to obtain the corresponding deep convolutional neural network.

[0100] Step 5.1: Obtain the optimal individual: Take the optimal individual selected in step 4 as the decoding object, where the optimal individual contains a set of optimal hyperparameter settings of the convolutional neural network CNN.

[0101] Step 5.2: Decode different types of hyperparameters separately to ensure that the decoded hyperparameters meet the defined parameter range. If they exceed the defined range, truncate them and find the closest real value within the decision variable value range.

[0102] Step 5.3: Apply the decoded hyperparameters to the architecture and training process of the Convolutional Neural Network (CNN) model.

[0103] Step 6: Based on the obtained deep convolutional neural network CNN, a Transformer network is superimposed on it to build a hybrid deep learning network model CNN-Transformer. The network architecture is shown in the figure below: Figure 3 As shown, the network is then trained on all data samples to obtain the final strip performance multi-index prediction model CNN-Transformer, in which the output dimension of the convolutional network CNN is used as the input of the Transformer network, the number of multi-head attention mechanisms is set to 4, and the repeated decoder unit in the Transformer network is set to 1.

[0104] Step 7: Application of CNN-Transformer learning model: In actual production, the continuous annealing production process data of each strip is obtained in real time through the OPC interface, and the constructed CNN-Transformer model is used to predict the three key performance indicators of yield strength, tensile strength and elongation, so as to achieve online accurate prediction of multiple indicators; when the deviation between the predicted value and the expected value of the performance indicator exceeds the given range, a timely alarm is issued to ensure the stability of the quality of the continuous annealing strip product.

[0105] Table 2 shows the comparison results of hyperparameter tuning and empirical value setting of 12160 training samples by the multi-index prediction method of continuous strip steel performance based on multi-objective evolutionary deep learning proposed in this embodiment. Table 3 shows the experimental results of test samples (samples that the model has not been trained on). In the experiment, the test set samples were independently run 30 times for statistical analysis. As can be seen from Table 3, the hit rates of the root mean square error RMSE and the mean average error MAE in the multi-index prediction task of the method based on multi-objective evolutionary deep learning proposed in this embodiment are at a low level, which well reflects the closeness between the model prediction value and the actual value. It also shows that the method of the present invention has good stability after parameter tuning and adopts a hybrid deep learning model for multi-index prediction tasks, and the prediction accuracy is highly consistent with the actual situation. Figure 4 A graphical illustration comparing the predicted value and the true value of the tensile strength index is given.

[0106] Table 2 Comparison of evolutionary algorithm tuning results and empirical value setting results

[0107] parameter Tuning Results Experience Number of convolutional layers 4 5 Number of output channels 192 128 Convolution kernel size 3 4 Number of hidden units 141 50 Learning Rate 0.06039172533384231 0.001 Weight decay 1.3798747705105333E-06 0.001

[0108] Table 3 Evaluation results of mechanical property prediction indicators

[0109] Mechanical properties Evolved MAE RMSE after evolution MAE before evolution RMSE before evolution Yield Strength 2.308 4.785 2.405 4.862 tensile strength 1.908 3.712 2.902 4.213 Elongation 0.761 1.386 0.775 1.406

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A multi-index prediction method for continuous annealing strip performance based on multi-objective evolutionary deep learning, characterized by: The steps include: Step 1: Build a sample set; The output of a sample is multiple performance indicators of a continuous annealing strip, and the input is the process data of the strip during continuous annealing production; n samples under normal production conditions are collected from the historical data of the cold rolling continuous annealing unit to establish a sample set S sam ; Step 2: Data preprocessing; For the established sample data set S sam Normalize the n sample data in; Step 3: Generate a multi-objective evolutionary deep learning model; According to the n samples obtained in step 2, the network hyperparameters are trained using the multi-objective evolutionary deep learning method based on NSGA-Ⅱ; Step 4: Select the optimal solution; Based on the objective function MSE of each solution in the final population obtained in step 3 sum and model complexity, using the TOPSIS method to select and obtain a top-ranked solution as the optimal hyperparameter combination for the CNN architecture; Step 5: Build a hybrid deep learning network model; According to the values ​​of each hyperparameter in the final solution selected in step 4, a corresponding deep convolutional neural network is constructed, and then the network is trained on all data samples to obtain the corresponding deep convolutional neural network CNN; Step 6: Based on the deep convolutional neural network CNN obtained in step 5, a Transformer network is superimposed on it to construct a hybrid deep learning network model CNN-Transformer, and then the network is trained on all data samples to obtain the final strip performance multi-index prediction model CNN-Transformer; Step 7: Obtain the continuous annealing production process data of each strip in real time through the OPC interface, and use the CNN-Transformer model obtained in step 6 to predict the three key performance indicators of yield strength, tensile strength and elongation to achieve online accurate prediction of multiple indicators; when the deviation between the predicted value and the expected value of the performance indicator exceeds the given range, a timely alarm is issued to ensure the stability of the quality of the continuous annealing strip product.

2. The method for predicting the performance of continuous annealing strip steel based on multi-objective evolutionary deep learning according to claim 1 is characterized in that: The specific method of step 3 is: Step 3.1: Randomly divide the preprocessed n samples into 0.8n and 0.2n samples to form the training set D and D respectively. train , validation set D validate ; If 0.8n and 0.2n are not integers, round them down; Step 3.2: Set the maximum number of iterations L of the multi-objective evolutionary deep learning algorithm max , population size N pop , network hyperparameter set S para ={ x 1, x 2, x 3, x 4, x 5, x 6}, set the iteration counter l =0; where x 1. x 2. x 3. x 4. x 5. x 6 correspond to the six parameters of the convolutional neural network model: number of convolutional layers, number of output channels, convolution kernel size, number of hidden units, learning rate, and weight decay; Step 3.3: Initialize the population; use all hyperparameters of the deep learning network model as decision variables and randomly generate N pop A real number vector with a length equal to the number of hyperparameters is used as the initial population to form the first l Generation population; where each vector x=( x 1, x 2, x 3, x 4, x 5, x 6) is a solution in the population, representing a hyperparameter combination scheme based on the current evolution l The hyperparameter combination of each solution in the generation population can construct a corresponding deep learning network, 0≤ l ≤L max ; Step 3.4: For each solution in the population, build a corresponding deep convolutional neural network according to its corresponding hyperparameter value, and then use it in the training data set D train The deep convolutional neural network is trained on the validation set D validate Calculate the prediction accuracy and complexity of the deep convolutional neural network model as the two optimization objective functions of the solution; the prediction accuracy is defined as the sum of the mean square errors (MSE) of multiple indicators sum ; Step 3.5: Perform non-dominated sorting on the solutions in the current population according to the objective function and crowding distance, and select N from the current population using the binary tournament strategy. pop solutions as the parent solution set; Step 3.6: Randomly select two solutions from the population, perform crossover and mutation operations in sequence to generate new offspring individuals; repeat N times pop times to get N pop Sub-generation solution; Step 3.7: For each child solution, use the objective function calculation method in step 3.4 to obtain the objective function of the child solution; Step 3.8: The newly generated N pop The sub-generation solutions are merged with the current population to form a population of size 2N pop population; Step 3.9: Set the new population to an empty set, and then use the fast non-inferior sorting method of NSGA-II to sort the 2N pop The solutions are divided into different Pareto frontiers; secondly, starting from the best first Pareto frontier, solutions are selected from each frontier in turn to generate the next generation population; Step 3.10: Set the number of iterations l = l +1; if l ≥L max , go to step 4; otherwise, return to step 3.

6.

3. The method for predicting the performance of continuous annealing strip steel based on multi-objective evolutionary deep learning according to claim 2 is characterized in that: In step 3.2, the architecture of the deep learning network model needs to be designed accordingly. The establishment of the model includes the following steps: Step 3.2.1: Convolutional layer design; the network includes convolutional layers, each of which uses output channels and size is x The convolution kernel is 3=3, and the convolution layer introduces nonlinear transformation through the ReLU activation function; Step 3.2.2: Optimizer and loss function selection: Adam optimizer and mean square error loss function MSELoss are used. The loss function measures the difference between the predicted value and the actual value by the square difference between the network output and the target. Step 3.2.3: Fully connected layer design; the network includes 3 fully connected layers, each of which contains hidden units, each of which is followed by a ReLU activation function; Step 3.2.4: Set weight decay; the weights of the convolutional layer and the linear layer are initialized using the Xavier initialization method. During training, the entire network continuously optimizes parameters to minimize the loss function through back propagation and gradient descent. The learning rate is set to , weight decay setting .

4. The method for predicting the performance of continuous annealing strip steel based on multi-objective evolutionary deep learning according to claim 2 is characterized in that: The specific method of the crossover and mutation operation in step 3.6 is: Step 3.6.1: Perform SBX crossover on the selected individuals; first generate random numbers u ,if u <0.9, then perform crossover, otherwise, directly copy the parent individuals to the next generation; Step 3.6.2: Perform mutation operation on the newly generated individuals; first generate random numbers u ',if u ′<0.5, then perform a mutation operation and select the parameters of an individual from the current population x i , according to the formula Get the individual parameters after mutation, where x i,new is the individual parameter after mutation, x i,old are the selected individual parameters, x i,upper and x i,lower The decision variables are x i The upper and lower bounds of , i=1,2,…6; δ represents the amount of variation; otherwise, the individual is kept unchanged.

5. The method for predicting the performance of continuous annealing strip steel based on multi-objective evolutionary deep learning according to claim 2 is characterized in that: The specific process of step 3.9 is as follows: If all the solutions contained in the current Pareto front are added to the new population, the number of solutions in the new population is still less than N. pop , then all the solutions in the Pareto front are added to the new population and moved to the next Pareto front; otherwise, the most crowded solutions are deleted from the current Pareto front based on the crowding distance until the remaining solutions are added to the new population so that the size of the new population can reach N pop , and finally add all the remaining solutions to the new population to form a new population for the next iteration.

6. The method for predicting the performance of continuous annealing strip steel based on multi-objective evolutionary deep learning according to claim 2 is characterized in that: The specific method of step 4 is: Step 4.1: Construct a decision matrix; Build contains v The decision matrix X with 2 rows and 2 columns, MSE sum and Complexity as different columns, where the decision matrix is ; Step 4.2: Standardize the decision matrix; The decision matrix is ​​standardized to ensure that all indicators are in the same numerical range. The standardized matrix is ​​recorded as , where x ij represents the elements of the original decision matrix; Step 4.3: Determine the weights; Determine the weight for each indicator to indicate its relative importance in the comprehensive evaluation, where w1=0.6 and w2=0.4 are the weights of the two indicators, which correspond to the characteristic MSE of the two columns of the decision matrix X. sum and complexity; Step 4.4: Construct a weighted normalized decision matrix; Multiply the standardized decision matrix by the weights to get the weighted normalized decision matrix , represents the elements of the standardized decision matrix; Step 4.5: Calculate and determine the positive ideal solution and negative ideal solution ; Step 4.6: Calculate the distance; Calculate the distance of each solution to the positive ideal solution The distance to the negative ideal solution ; Step 4.7: Calculate the comprehensive score; Calculate the closeness of each solution The index value corresponding to each solution, ; Step 4.8: Select the best solution; According to the comprehensive score, the C i Sort them and select the top-ranked solution from high to low as the optimal hyperparameter setting for the convolutional network model.

7. The method for predicting the performance of continuous annealing strip steel based on multi-objective evolutionary deep learning according to claim 6 is characterized in that: The specific method of step 5 is: Step 5.1: Obtain the optimal individual; use the optimal individual selected in step 4 as the object of decoding, where the optimal individual includes a set of optimal hyperparameter settings of the convolutional neural network CNN; Step 5.2: Decode different types of hyperparameters respectively to ensure that the decoded hyperparameters meet the defined parameter range. If they exceed the defined range, truncate them and find the closest real value within the decision variable value range. Step 5.3: Apply the decoded hyperparameters to the architecture and training process of the convolutional neural network model CNN.

Images