A Multi-Indicator Quality Prediction Method for Strip Steel Based on Evolutionary Learning

Abstract

This invention provides a multi-index quality prediction method for strip steel based on evolutionary learning, belonging to the field of automatic control technology. This invention establishes a prediction model database by collecting historical data from actual continuous annealing production processes. Relevant information from the continuous annealing process is used as the historical data set, and the carbon equivalent is calculated based on two representations of carbon equivalent. The data in the established database is preprocessed to obtain a processed standard training dataset. Then, a multi-index quality prediction model for the continuous annealing process is established, combining a two-stage model of a deep sparse autoencoder network and an extreme gradient boosting algorithm with a multi-objective optimization algorithm. Finally, the Knee point strategy is used to select the optimal result from the Pareto optimal solution set of the multi-objective optimization based on the preferences of the actual production process, as the parameters of the multi-index quality prediction model. The multi-index quality prediction method of this invention can be effectively applied to actual production processes, providing operators with a basis for timely understanding of strip steel quality.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology, and in particular to a method for predicting the quality of strip steel based on evolutionary learning. Background Technology

[0002] Continuous annealing, as a downstream process in cold rolling, plays a crucial role in the quality of finished steel. The continuous annealing process involves gradually heating the rolled strip to the required process temperature, maintaining it at that temperature for a specific time, and then cooling it to eliminate internal stress, improve its plasticity, and achieve standard mechanical properties to meet quality requirements. However, in actual production, mechanical property testing lags far behind modern standards, relying on offline performance testing of randomly selected products. This method has significant time delays, failing to promptly detect substandard products and causing losses. Mechanical property prediction models are mainly divided into mechanistic models and data models. Mechanistic models are mostly established based on experimental research. Traditional physical metallurgical models can be used for performance calculations, but they also have significant limitations. They require extensive professional knowledge and related experiments to determine parameters, and the complex on-site environment often prevents the models from achieving satisfactory accuracy. Therefore, establishing a data-driven quality performance prediction model for finished steel can, to some extent, overcome the problem of high data quality requirements for metallurgical mechanism models. It can replace traditional product quality performance testing methods and has practical significance for improving product quality. Applying data-driven performance prediction models to actual production can increase production rate and product yield, reduce production costs, and provide guidance for the research and development of new steel grades.

[0003] Studies have found that the quality of strip steel depends on the characteristics of the raw materials (chemical composition and process parameters). Tensile strength, yield strength, and elongation are three key parameters for evaluating the mechanical properties of finished steel. Recent research has revealed the correlation between chemical composition, process parameters, and mechanical properties, with quantitative research gradually emerging based on qualitative studies. Currently, existing technologies can measure the necessary chemical composition and process parameters during continuous annealing. Operators control the chemical composition of the strip steel and the continuous annealing process parameters according to the mechanical property requirements of different steel grades to obtain strip steel products that meet the requirements.

[0004] While existing performance prediction models can predict the mechanical properties of strip steel, these traditional quantitative research methods are relatively singular and have limitations in predicting multi-index quality problems in continuous annealing processes, failing to effectively analyze the relatively complex multi-factor composite relationships during continuous annealing. Chinese patent CN112149272A proposes a mechanical property prediction model for cold-rolled steel strip based on multiple linear regression analysis. This method establishes regression equations between the mechanical properties of steel strip and related process parameters based on professional process knowledge and historical data, calculating mechanical properties under different process conditions to predict the mechanical properties of cold-rolled annealed steel strip. Chinese patent CN107609647A uses an improved neural network model to establish a mapping relationship between the alloy composition, process parameters, and mechanical properties of rolls, achieving prediction of the mechanical properties of roll alloys. Chinese patent CN110569566A discloses a method for predicting the mechanical properties of strip steel, employing multiple performance prediction sub-models to predict the samples under test, achieving an accurate evaluation of the yield strength prediction results. However, relying solely on yield strength as a performance index cannot comprehensively evaluate the quality of strip steel. Chinese patent CN111241750A uses an improved neural network model to establish the mapping relationship between the composition, process parameters, and mechanical properties of cold-rolled steel sheets, and combines this model with a genetic algorithm for further improvement. Although this solves the problem of model parameter optimization, traditional network models are prone to underfitting or overfitting when dealing with high-dimensional features and noise. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting the quality of strip steel based on evolutionary learning and multiple indicators.

[0006] A method for predicting the quality of strip steel based on evolutionary learning includes the following steps:

[0007] Step 1: Establish a dataset for a multi-index quality prediction model of the continuous annealing process;

[0008] Step 1.1: Determine 20 chemical composition parameters of the strip steel product, including sulfur content (S), copper content (Cu), nickel content (Ni), chromium content (Cr), molybdenum content (Mo), vanadium content (V), niobium content (Nb), total aluminum content (Al), acid-soluble aluminum content, titanium content (Ti), boron content (B), tin content (Sn), arsenic content (As), zirconium content (Zr), calcium content (Ca), lead content (Pb), antimony content (Sb), nitrogen content (N), oxygen content (O), and tungsten content (W).

[0009] Step 1.2: Based on the measured chemical composition parameters of the strip steel, calculate the carbon equivalent using the following formula:

[0010]

[0011]

[0012] Where Ceq is carbon equivalent #1, Pcm is carbon equivalent #2; C is carbon content, Si is silicon content, and Mn is manganese content;

[0013] Step 1.3: Obtain five process parameters and strip information data during the process from the cold-rolled coil entering the uncoiler, through the continuous annealing production process, to the coiler, including furnace temperature, current temperature, annealing temperature, thickness, and width;

[0014] Step 1.4: Measure three mechanical properties of the finished steel obtained after continuous annealing in the actual production environment, including: yield strength, tensile strength, and elongation.

[0015] Step 1.5: Map the composition parameters, carbon equivalent, process parameters, strip information data, and mechanical property indicators of the continuous annealing process obtained in Steps 1.1 to 1.4;

[0016] Step 1.6: Repeat steps 1.1 to 1.5 to obtain a set of sample data that meets the set number, and construct the training dataset for the multi-index quality prediction model. The data obtained in steps 1.1 to 1.3 are used as the input data of the multi-index quality prediction model, and the mechanical performance index data measured in step 1.4 are used as the output data of the multi-index quality prediction model.

[0017] Step 2: Perform data preprocessing on the multi-index quality prediction model dataset for the continuous annealing process constructed in Step 1 to obtain standard training and test datasets;

[0018] Step 2.1: Read in the multi-index quality prediction model dataset for the continuous annealing process constructed in Step 1, and randomly shuffle the multi-index quality prediction model dataset for the entire continuous annealing process.

[0019] Step 2.2: Visualize the dataset of the multi-index quality prediction model for the continuous annealing process after random shuffling. After data analysis, select the data items with strong correlation as the final input sample dataset for the multi-index quality prediction model of the continuous annealing process.

[0020] Step 2.3: Standardize the data obtained in Step 2.2 using data preprocessing techniques;

[0021] The standardization process is as follows: First, calculate the median M and interquartile range (IQR) of the data obtained in step 2.2. Then, for a certain value d of the data obtained in step 2.2... i The data values ​​after data preprocessing and standardization are:

[0022]

[0023] Step 2.4: Store the processed standard multi-index quality prediction model dataset for the continuous annealing process. Divide the stored dataset into two parts at a ratio of 17:3 to obtain the training dataset and test set for the multi-index quality prediction model of the continuous annealing process.

[0024] Step 3: Construct a multi-index quality prediction model for the continuous annealing process based on evolutionary learning;

[0025] Step 3.1: Establish a deep sparse autoencoder network to learn the deep semantic information of the continuous annealing dataset and the implicit information of the interaction between data, and obtain a high-dimensional mapping representation of the inherent features of the input data;

[0026] Step 3.1.1: Establish an autoencoder, which consists of an encoder and a decoder. The encoder comprises an input layer and hidden layers, while the decoder consists of multiple hidden layers and a reconstruction layer. First, the encoder encodes the input data of the training dataset for the multi-index quality prediction model of the standard continuous annealing process obtained in Step 2. Then, the decoder minimizes the output results of the continuous annealing process parameter data and chemical composition data. and input data x i The squared error is used to reconstruct the original information;

[0027] The encoder is defined as follows:

[0028] h = f(x) i )=σ(Wx i +b) (4)

[0029] Where, x i ∈X represents each data sample in the input sample dataset X, i = 1, 2, ..., N, where N is the number of data samples, σ(·) represents the activation function, W represents the weight matrix inside the encoder, and b represents the bias vector of the encoder;

[0030] The decoder is defined as follows:

[0031]

[0032] in, Indicates the input sample x i The corresponding output is where W′ is the weight matrix inside the decoder, and b′ represents the bias vector of the decoder. The weight parameters inside the network are trained by minimizing the squared error, which represents the difference between the input and output data, and is defined as follows:

[0033]

[0034] in, It is the output of the decoder in the autoencoder, ||·||2 represents the 2-norm of the autoencoder, used to measure x. i and Reconstruction error between them.

[0035] Step 3.1.2: Introduce sparsity constraints based on autoencoders to construct a sparse autoencoder network; by introducing sparsity constraints, the automatic feature selection and dimensionality reduction process is completed. The cost function of the sparse autoencoder network is defined as follows:

[0036]

[0037] Where β is the penalty factor, used to control the sparse weights of the sparse terms, η is the reconstruction error weight, used to control the weights of the reconstruction error, and D... KL It is a sparsity metric, specifically using KL divergence to measure the difference between the expected sparse distribution and the actual sparse distribution in the hidden layer, achieved by minimizing D. KL To constrain the sparsity of sparse autoencoders, the KL divergence is defined as follows:

[0038]

[0039]

[0040] Where m represents the total number of input patterns, and ρ is the sparsity parameter. Represents the sparse distribution of the network. It is the j-th node of the hidden layer of the sparse autoencoder. For the i-th input x i The output of .

[0041] Step 3.1.3: Stack the hidden layers of the sparse autoencoder network in sequence to construct a deep sparse autoencoder network. Reconstruct the input data by minimizing the reconstruction error. Use the output features of the last hidden layer of the network as the input data for the multi-index quality prediction model of the continuous annealing process.

[0042] Step 3.2: Construct the structure of the extreme gradient boosting algorithm model as a multi-index quality prediction model for the continuous annealing process. Then, use the high-dimensional feature representation of the last hidden layer of the deep sparse autoencoder network in Step 3.1 as the input of the extreme gradient boosting algorithm to predict the mechanical performance index values ​​of the strip steel.

[0043] Assume the input data is x i The boosting strategy for the extreme gradient boosting algorithm model is defined as follows:

[0044]

[0045]

[0046] Where n is the total number of input data. f is the output of the t-th iteration. k (x i Let be the predicted value of the k-th decision tree for input data i. Then, the cost function of the extreme gradient boosting algorithm is defined as:

[0047]

[0048] In the formula, n is the number of input data, Ω(f i ) represents the model complexity of the i-th decision tree, and t represents the total number of decisions;

[0049] Step 3.3: Construct an improved non-dominated sorting genetic algorithm II with Knee point strategy, perform the optimal solution search task in the solution space, optimize the two-stage model structure and hyperparameters of deep sparse autoencoder network and extreme gradient boosting algorithm, and generate a strip steel index quality prediction model that meets the accuracy requirements.

[0050] The non-dominated sorting genetic algorithm II specifically involves: stratifying individuals based on their superiority and defining different individuals in each stratum; then employing an elite selection strategy to select and iterate individuals with different advantages, thereby generating new individuals; preserving superior individuals from the parent generation to the offspring; and using improved crossover and variational operators to obtain a subpopulation, thereby improving the convergence speed of the non-dominated sorting genetic algorithm II and preserving population diversity.

[0051] Step 3.3.1: Parent population P t A new offspring population Q is obtained using a genetic algorithm. t This includes selection, crossover, and mutation processes;

[0052] Step 3.3.2: Parent population P t and the newly generated offspring population Q t Forming the t-th generation population G t The population size is 2N;

[0053] Step 3.3.3: For population G t Perform a non-dominated sort, generating a series of non-dominated solution sets Z1, Z2, ..., Z based on the dominance level. m Then, they are sequentially placed into the next generation population P. t+1 middle;

[0054] Step 3.3.4: If the population size of the non-dominated solution set is less than or equal to N, then add it to population P. t+1 If the population size of the non-dominated solution set is greater than N, then the crowding degree of the individuals exceeding the non-dominated solution set is calculated, and then they are placed into P in descending order of size. t+1In this process, until the number of individuals reaches N, the remaining solutions are eliminated.

[0055] Step 3.3.5: For the new population P t+1 By performing selection, crossover, and mutation operations, a new subpopulation Q is obtained. t+1 Then G t+1 The population then undergoes a new round of iteration.

[0056] Step 4: Initialize the starting parameters of the multi-objective optimization algorithm, and set the search range for the structural parameters and hyperparameters of the deep sparse autoencoder network and the extreme gradient boosting algorithm;

[0057] Step 4.1: Initialize the structural parameters and hyperparameters of the deep sparse autoencoder network to the default settings, including the number of network layers, number of nodes, sparsity, beta value, and learning rate;

[0058] Step 4.2: Initialize the network structure parameters and hyperparameters of the extreme gradient boosting algorithm to the default settings, including tree depth, number of evaluators, and learning rate;

[0059] Step 4.3: Set the hyperparameters and structural parameters involved in the multi-index quality prediction model of deep sparse autoencoder network and extreme gradient boosting algorithm;

[0060] The specific parameter configurations involved include: the number of layers N in the deep sparse autoencoder network. layer Number of nodes N node sparse distribution ρ, penalty factor β, learning rate lr DSAE The tree depth N of the extreme gradient boosting algorithm dept h Number of evaluators N estimator Learning rate lr XGBoost ;

[0061] Step 5: Execute the multi-index quality prediction model for the continuous annealing process based on evolutionary learning constructed in Step 3. Since the multi-index quality prediction model for the continuous annealing process is divided into two parts, each part of the multi-index quality prediction model for the continuous annealing process is optimized according to the marginal optimization method. First, the deep sparse autoencoder network is optimized, including the model structure and hyperparameters. Then, the structure and hyperparameters of the extreme gradient boosting algorithm are optimized. The Knee point strategy is used to select the parameter values ​​when the multi-index quality prediction model for the continuous annealing process has the best performance.

[0062] Step 5.1: Initialize the algorithm parameters of the non-dominated sorting genetic algorithm II, and define the maximum population size and maximum number of iterations for the multi-index quality prediction model during continuous annealing;

[0063] Step 5.2: Input the process parameters and chemical composition data of the continuous annealing process, train the deep sparse autoencoder network model, optimize the network structure and hyperparameters of the deep sparse autoencoder network, and select the optimal performance model parameters using the Knee point strategy;

[0064] When optimizing deep sparse autoencoders, considering both the network's time complexity and performance, the optimization objective for deep sparse autoencoders is set as follows:

[0065]

[0066]

[0067] Where F(.) is the objective function of the non-dominated sorting genetic algorithm II, L DSAE (.) represents the reconstruction error value of the deep sparse autoencoder network; θ represents the model parameters to be optimized. In the first stage, θ is the number of networks, and the structural parameters include the nodes of the i-th layer. Number of layers N layer and hyperparameters.

[0068] Step 5.2.1: Calculate the fitness function value for each population with the goal of minimizing the reconstruction error of the deep sparse autoencoder network;

[0069] Step 5.2.2: Save the current calculation result, increment the population size by one. When the population size p1 equals the maximum population size P1, the population size p1 = 0, and increment the iteration count g1 by one.

[0070] Step 5.2.3: Determine if the termination condition is met. Repeat steps 5.2.1 to 5.2.3 until the iteration count g1 equals the maximum iteration count G1;

[0071] Step 5.2.4: Use the Knee-point strategy to find the optimal solution in the Pareto optimal solution set of the deep sparse autoencoder network.

[0072] Step 5.2.5: Set the structural parameters and hyperparameters of the deep sparse autoencoder network corresponding to the optimal result as the optimal parameters of the multi-index quality prediction model in the continuous annealing process;

[0073] Step 5.3: Optimize the network structure and hyperparameters of the extreme gradient boosting algorithm prediction model; First, reset the optimal parameter results from Step 5.2 onto the deep sparse autoencoder network, using the same input process parameters and chemical composition data for continuous annealing as in Step 5.2. Then, retrain the multi-index quality prediction model for the continuous annealing process to optimize the extreme gradient boosting algorithm, using the Knee-point strategy to select the optimal performance model parameters; redesign the hyperparameter optimization objective of the extreme gradient boosting algorithm network as follows:

[0074] F(θ)={0.2(L YS +L TS ), 0.8L EL} (15)

[0075] Among them, L YS L TS L EL These are the root mean square errors between the predicted and actual measured values ​​of the three mechanical properties—yield strength, tensile strength, and elongation—during the continuous annealing process.

[0076] Step 5.3.1: Calculate the fitness function value for each population with the goal of minimizing the root mean square error between the predicted and actual measured values ​​of the three mechanical properties (yield strength, tensile strength, and elongation) during the continuous annealing process.

[0077] Step 5.3.2: Save the current calculation result, increment the population size by one, and when the population size p2 equals the maximum population size P2, the population size p2 = 0, and increment the iteration count g2 by one;

[0078] Step 5.3.3: Determine if the termination condition is met. Repeat steps 5.3.1 to 5.3.3 until the iteration count g2 equals the maximum iteration count G2;

[0079] Step 5.3.4: Use the Knee point strategy to determine the optimal solution from the non-dominated solution set of the overall framework optimization results;

[0080] Step 5.3.5: Set the network structure and hyperparameters of the extreme gradient boosting algorithm corresponding to the Knee solution to the optimal parameters of the extreme gradient boosting algorithm;

[0081] Step 6: Complete the multi-objective optimization process, and use the optimized hyperparameters and structural parameters as the hyperparameters and structural parameters of the multi-index quality prediction model in the continuous annealing process. Then, use the test dataset obtained in Step 2 to perform quality prediction.

[0082] The beneficial effects of adopting the above technical solution are as follows:

[0083] This invention provides a method for predicting the quality of multiple indicators of strip steel based on evolutionary learning. It proposes a two-stage model structure based on evolutionary learning to solve the problem of predicting the quality of multiple indicators of strip steel in the actual continuous annealing production process. In this invention, a deep sparse autoencoder network is designed to extract high-dimensional features of the input data, namely process parameters and composition parameters. An ensemble learning method based on the extreme gradient boosting algorithm is used to establish a performance indicator prediction model to predict three mechanical properties of strip steel, improving the model's stability. Finally, an improved non-dominated sorting genetic algorithm II is used to optimize the two-stage model structure and hyperparameters of the deep sparse autoencoder network and the extreme gradient boosting algorithm. The Knee point strategy is used to select the most realistic preference result from the Pareto optimal solution set, obtaining the optimal multi-indicator quality prediction model, enhancing the model's adaptability and generalization ability. Verified by actual production data, the multi-index quality prediction model based on evolutionary learning can significantly improve the accuracy and robustness of performance prediction results for continuously annealed products. This enables researchers to promptly grasp the quality of finished steel products, compensate for the shortcomings of traditional mechanical property testing methods, and thus help continuous annealing production lines improve product quality and reduce production costs. It also has guiding significance for the research and development of new steel grades. Attached Figure Description

[0084] Figure 1 This is a schematic diagram of the specific cold rolling continuous annealing process involved in this invention;

[0085] Figure 2 This is a framework diagram of a multi-index quality prediction model based on evolutionary learning, which is a specific embodiment of the present invention.

[0086] Figure 3 A feature visualization heatmap of the input data for the continuous annealing process of the multi-index quality prediction model in a specific embodiment of the present invention.

[0087] Figure 4 This is a schematic diagram of the population elite selection strategy in the evolutionary learning algorithm of a specific embodiment of the present invention.

[0088] Figure 5 This is a schematic diagram illustrating the simulated binary crossover process in the evolutionary learning algorithm of a specific embodiment of the present invention;

[0089] Figure 6 The optimization results of the evolutionary learning algorithm for different networks are shown in the specific embodiments of the present invention.

[0090] Figure (a) shows the optimization results of the evolutionary learning algorithm on the deep sparse autoencoder network, and Figure (b) shows the optimization results of the evolutionary learning algorithm on the extreme gradient boosting algorithm.

[0091] Figure 7This is a quantile-quantile plot showing the comparison between the predicted and actual values ​​of the three performance indicators obtained in this invention.

[0092] Figure (a) shows the yield strength, Figure (b) shows the tensile strength, and Figure (c) shows the elongation. Detailed Implementation

[0093] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0094] The multi-index quality prediction method based on evolutionary learning in this embodiment is designed for the continuous annealing process of cold-rolled strip steel. Continuous annealing is a crucial process in the cold rolling mill of steel enterprises, playing a vital role in the quality of strip steel. Its process flow is shown in the attached figure. Figure 1 As shown. The entire process mainly includes preheating, heating, homogenization, air cooling, roll cooling, aging, and final cooling. Its main purpose is to eliminate the hardening during cold rolling, soften the strip by recrystallization, and obtain high-quality finished steel. After uncoiling and welding, the cold-rolled strip enters the pickling tank to clean the oxides on the surface of the strip. Then, after the preheating-heating-homogenization process, the microstructure in the strip changes. Then, it enters the cooling zone, where the internal structure of the strip is altered through air cooling and roll cooling. Next, the strip undergoes an aging stage to maintain it in a certain low-temperature environment, thereby eliminating the internal stress and obtaining steel with better mechanical properties. Finally, in the final cooling zone, the temperature of the strip is gradually reduced to room temperature using air cooling or water cooling methods. Then, it is coiled into coils by a coiler and stored in the warehouse.

[0095] This implementation addresses the limitations of mechanistic models and general regression models in predicting multi-index quality in continuous annealing processes due to the complexity of the inherent correlations among process parameters. It provides a method for predicting multi-index quality of strip steel based on evolutionary learning. Figure 2 As shown, it includes the following steps:

[0096] Step 1: Establish a dataset for a multi-index quality prediction model of the continuous annealing process;

[0097] Step 1.1: Determine 20 chemical composition parameters of the strip steel product, including sulfur content (S), copper content (Cu), nickel content (Ni), chromium content (Cr), molybdenum content (Mo), vanadium content (V), niobium content (Nb), total aluminum content (Al), acid-soluble aluminum content, titanium content (Ti), boron content (B), tin content (Sn), arsenic content (As), zirconium content (Zr), calcium content (Ca), lead content (Pb), antimony content (Sb), nitrogen content (N), oxygen content (O), and tungsten content (W).

[0098] Step 1.2: Based on the measured chemical composition parameters of the strip steel, calculate the carbon equivalent using the following formula:

[0099]

[0100]

[0101] Where Ceq is carbon equivalent #1, Pcm is carbon equivalent #2; C is carbon content, Si is silicon content, and Mn is manganese content;

[0102] Step 1.3: Obtain five process parameters and strip information data during the process from the cold-rolled coil entering the uncoiler, through the continuous annealing production process, to the coiler, including furnace temperature, current temperature, annealing temperature, thickness, and width;

[0103] Step 1.4: Measure three mechanical properties of the finished steel obtained after continuous annealing in the actual production environment, including: yield strength, tensile strength, and elongation.

[0104] Step 1.5: Map the composition parameters, carbon equivalent, process parameters, strip information data, and mechanical property indicators of the continuous annealing process obtained in Steps 1.1 to 1.4;

[0105] The mapping rules in this embodiment are shown in Tables 1 and 2:

[0106] Table 1 Input features of the continuous annealing dataset

[0107]

[0108]

[0109] Table 2 Output characteristics of the continuous annealing dataset

[0110] Serial Number Output features 1 Yield strength (YS) 2 Tensile strength (TS) 3 Elongation (EL)

[0111] Step 1.6: Repeat steps 1.1 to 1.5 to obtain a set of sample data that meets the set number, and construct the training dataset for the multi-index quality prediction model. The 27 data obtained in steps 1.1 to 1.3 are used as the input data for the multi-index quality prediction model, and the mechanical performance index data measured in step 1.4 are used as the output data for the multi-index quality prediction model.

[0112] In this embodiment, the method proposed by the present invention requires the collection of approximately 60,000 data samples.

[0113] Step 2: Perform data preprocessing on the multi-index quality prediction model dataset for the continuous annealing process constructed in Step 1 to obtain standard training and test datasets;

[0114] Step 2.1: Read in the multi-index quality prediction model dataset for the continuous annealing process constructed in Step 1, and randomly shuffle the multi-index quality prediction model dataset for the entire continuous annealing process.

[0115] This step is to prevent the model learning effect from being poor due to random factors other than data, and to prevent the model training process from having a "bias" that reduces the model training accuracy.

[0116] Step 2.2: Visualize the dataset of the multi-index quality prediction model for the randomized continuous annealing process. After data analysis, select strongly correlated data items as the final input sample dataset for the multi-index quality prediction model of the continuous annealing process; see attached. Figure 3 The image shown is a feature visualization heatmap of the input data for the continuous annealing process;

[0117] Step 2.2.1: Perform visualization operations on the shuffled continuous annealing dataset, observe the significant features of the dataset under macroscopic conditions, and analyze the data distribution information and feature correlation.

[0118] Step 2.2.2: Based on the results of the visualization data analysis, remove missing data items, outlier data items, and data items with poor correlation to avoid poor-quality data samples affecting the model performance. Then select data items with strong correlation as the input dataset for the model.

[0119] Step 2.3: Use data preprocessing techniques to standardize the data obtained in Step 2.2 to eliminate the influence of different units between different data items;

[0120] The standardization process is as follows: First, calculate the median M and interquartile range (IQR) of the data obtained in step 2.2. Then, for a certain value d of the data obtained in step 2.2... i The data values ​​after data preprocessing and standardization are:

[0121]

[0122] Step 2.4: Store the processed standard multi-index quality prediction model dataset for the continuous annealing process. Divide the stored dataset into two parts at a ratio of 17:3 to obtain the training dataset and test set for the multi-index quality prediction model of the continuous annealing process.

[0123] Step 3: Construct a multi-index quality prediction model for the continuous annealing process based on evolutionary learning. First, a deep sparse autoencoder network is used to extract the latent features of the original data, enabling automatic feature selection, thereby better representing the input samples, reducing noise, and significantly reducing computational cost. Second, an extreme gradient boosting algorithm is used as the multi-index quality prediction model for the continuous annealing process, outputting predicted values ​​of the strip's mechanical properties. Finally, an improved non-dominated sorting genetic algorithm II is used to optimize the two-stage model layer by layer to obtain the Pareto optimal solution set. The Knee point strategy is then used to select the optimal network structure and hyperparameters, thus obtaining the best multi-index quality prediction model.

[0124] Step 3.1: Establish a deep sparse autoencoder network to learn the deep semantic information of the continuous annealing dataset and the implicit information of the interaction between data, and obtain a high-dimensional mapping representation of the inherent features of the input data;

[0125] Step 3.1.1: Establish an autoencoder, which consists of an encoder and a decoder. The encoder comprises an input layer and hidden layers, while the decoder consists of multiple hidden layers and a reconstruction layer. First, the encoder encodes the input data of the training dataset for the multi-index quality prediction model of the standard continuous annealing process obtained in Step 2. Then, the decoder minimizes the output results of the continuous annealing process parameter data and chemical composition data. and input data x i The squared error is used to reconstruct the original information;

[0126] The encoder is defined as follows:

[0127] h = f(x) i )=σ(Wx i +b) (19)

[0128] Where, x i ∈X represents each data sample in the input sample dataset X, i = 1, 2, ..., N, where N is the number of data samples, σ(·) represents the activation function, W represents the weight matrix inside the encoder, and b represents the bias vector of the encoder;

[0129] The decoder is defined as follows:

[0130]

[0131] in, Indicates the input sample x i The corresponding output is where W′ is the weight matrix inside the decoder, and b′ represents the bias vector of the decoder. The weight parameters inside the network are trained by minimizing the squared error, which represents the difference between the input and output data, and is defined as follows:

[0132]

[0133] in, It is the output of the decoder in the autoencoder, ||·||2 represents the 2-norm of the autoencoder, used to measure x. i and Reconstruction error between them.

[0134] Step 3.1.2: Introduce sparsity constraints based on autoencoders to construct a sparse autoencoder network. Compared with traditional autoencoders, sparse autoencoders allow more hidden layer elements than input layer elements. By introducing sparsity constraints, automatic feature selection and dimensionality reduction are achieved, and more representative general features of the data can be extracted to a certain extent. The cost function of the sparse autoencoder network is defined as follows:

[0135]

[0136] Where β is the penalty factor, used to control the sparse weights of the sparse terms, η is the reconstruction error weight, used to control the weights of the reconstruction error, and D... KL It is a sparsity metric, specifically using KL divergence to measure the difference between the expected sparse distribution and the actual sparse distribution in the hidden layer, achieved by minimizing D. KL To constrain the sparsity of sparse autoencoders, the KL divergence is defined as follows:

[0137]

[0138]

[0139] Where m represents the total number of input patterns, and ρ is the sparsity parameter. Represents the sparse distribution of the network. It is the j-th node of the hidden layer of the sparse autoencoder. For the i-th input x i The output of .

[0140] Step 3.1.3: Stack the hidden layers of the sparse autoencoder network in sequence to construct a deep sparse autoencoder network. Reconstruct the input data by minimizing the reconstruction error. Use the output features of the last hidden layer of the network as the input data for the multi-index quality prediction model of the continuous annealing process.

[0141] Step 3.2: Construct the structure of the extreme gradient boosting algorithm model as a multi-index quality prediction model for the continuous annealing process. Then, use the high-dimensional feature representation of the last hidden layer of the deep sparse autoencoder network in Step 3.1 as the input of the extreme gradient boosting algorithm to predict the mechanical performance index values ​​of the strip steel.

[0142] Assume the input data is x iThe boosting strategy for the extreme gradient boosting algorithm model is defined as follows:

[0143]

[0144]

[0145] Where n is the total number of input data. f is the output of the t-th iteration. k (x i Let be the predicted value of the k-th decision tree for input data i. Then, the cost function of the extreme gradient boosting algorithm is defined as:

[0146]

[0147] In the formula, n is the number of input data, Ω(f i ) represents the model complexity of the i-th decision tree, and t represents the total number of decisions;

[0148] Step 3.3: Construct an improved non-dominated sorting genetic algorithm II with Knee point strategy, perform the optimal solution search task in the solution space, optimize the two-stage model structure and hyperparameters of deep sparse autoencoder network and extreme gradient boosting algorithm, and generate a strip steel index quality prediction model that meets the accuracy requirements.

[0149] During neural network training, it is necessary to continuously adjust the network's structural parameters and hyperparameters to achieve the optimal performance of the network training model. Commonly used hyperparameter tuning methods are mainly divided into two types: manual empirical tuning and algorithm-adaptive tuning. This invention employs an improved non-dominated sorting genetic algorithm II with a Knee point strategy to optimize the structural parameters and hyperparameters of the two-stage model in the continuous annealing process: the deep sparse autoencoder network for feature extraction and the extreme gradient boosting algorithm for multi-index quality prediction. This results in a multi-index quality prediction model for the continuous annealing process that meets production requirements.

[0150] The non-dominated sorting genetic algorithm II specifically works as follows: Individuals are stratified based on their superiority, and different individuals are defined in each stratum. Then, an elite selection strategy is employed to select and iterate individuals with different advantages, generating new individuals. Superior individuals from the parent generation to the offspring are preserved, and improved crossover and variational operators are used to obtain a subpopulation, thereby improving the convergence speed of the non-dominated sorting genetic algorithm II and preserving population diversity. (Appendix) Figure 4 For G t Generations to G t+1 A diagram illustrating the elite selection strategy of a generational population.

[0151] Step 3.3.1: Parent population P t A new offspring population Q is obtained using a genetic algorithm.t This includes selection, crossover, and mutation processes, with appended... Figure 5 A schematic diagram simulating the binary crossover process;

[0152] Step 3.3.2: Parent population P t and the newly generated offspring population Q t Forming the t-th generation population G t The population size is 2N;

[0153] Step 3.3.3: For population G t Perform a non-dominated sort, generating a series of non-dominated solution sets Z1, Z2, ..., Z based on the dominance level. m Then, they are sequentially placed into the next generation population P. t+1 middle;

[0154] Step 3.3.4: If the population size of the non-dominated solution set is less than or equal to N, then add it to population P. t+1 If the population size of the non-dominated solution set is greater than N, then the crowding degree of the individuals exceeding the non-dominated solution set is calculated, and then they are placed into P in descending order of size. t+1 In this process, until the number of individuals reaches N, the remaining solutions are eliminated.

[0155] Step 3.3.5: For the new population P t+1 By performing selection, crossover, and mutation operations, a new subpopulation Q is obtained. t+1 Then G t+1 The population then undergoes a new round of iteration.

[0156] Step 4: Initialize the starting parameters of the multi-objective optimization algorithm, and set the search range for the structural parameters and hyperparameters of the deep sparse autoencoder network and the extreme gradient boosting algorithm;

[0157] Step 4.1: Initialize the structural parameters and hyperparameters of the deep sparse autoencoder network to the default settings, including the number of network layers, number of nodes, sparsity, beta value, and learning rate;

[0158] Step 4.2: Initialize the network structure parameters and hyperparameters of the extreme gradient boosting algorithm to the default settings, including tree depth, number of evaluators, and learning rate;

[0159] Step 4.3: Set the hyperparameters and structural parameters involved in the multi-index quality prediction model of deep sparse autoencoder network and extreme gradient boosting algorithm;

[0160] The specific parameter configurations involved include: the number of layers N in the deep sparse autoencoder network. layer Number of nodes N node sparse distribution ρ, penalty factor β, learning rate lr DSAEThe tree depth N of the extreme gradient boosting algorithm dept h Number of evaluators N estimator Learning rate lr XGBoost The setting range is shown in Table 3;

[0161] Table 3. Parameter setting ranges for deep sparse autoencoders and extreme gradient boosting algorithms.

[0162]

[0163] Step 5: Execute the multi-index quality prediction model for the continuous annealing process based on evolutionary learning constructed in Step 3. Since the multi-index quality prediction model for the continuous annealing process is divided into two parts, each part of the multi-index quality prediction model for the continuous annealing process is optimized according to the marginal optimization method. First, the deep sparse autoencoder network is optimized, including the model structure and hyperparameters. Then, the structure and hyperparameters of the extreme gradient boosting algorithm are optimized. The Knee point strategy is used to select the parameter values ​​when the multi-index quality prediction model for the continuous annealing process has the best performance.

[0164] Step 5.1: Initialize the algorithm parameters of the non-dominated sorting genetic algorithm II, and define the maximum population size and maximum number of iterations for the multi-index quality prediction model during continuous annealing;

[0165] In this embodiment, the maximum population size P1 and P2 are set to 30, the maximum number of iterations G1 and G2 are set to 15, and the initial population size p1 = 0, p2 = 0, and the initial number of generations g1 = 0, g2 = 0 are set.

[0166] Step 5.2: Input the process parameters and chemical composition data of the continuous annealing process, train the deep sparse autoencoder network model, optimize the network structure and hyperparameters of the deep sparse autoencoder network, and select the optimal performance model parameters using the Knee point strategy;

[0167] When optimizing deep sparse autoencoders, considering both the network's time complexity and performance, the optimization objective for deep sparse autoencoders is set as follows:

[0168]

[0169]

[0170] Where F(.) is the objective function of the non-dominated sorting genetic algorithm II, L DSAE (.) represents the reconstruction error value of the deep sparse autoencoder network; θ represents the model parameters to be optimized. In the first stage, θ is the number of networks, and the structural parameters include the nodes of the i-th layer. Number of layers N layer and hyperparameters.

[0171] Step 5.2.1: Calculate the fitness function value for each population with the goal of minimizing the reconstruction error of the deep sparse autoencoder network;

[0172] Step 5.2.2: Save the current calculation result, increment the population size by one. When the population size p1 equals the maximum population size P1, the population size p1 = 0, and increment the iteration count g1 by one.

[0173] Step 5.2.3: Determine if the termination condition is met. Repeat steps 5.2.1 to 5.2.3 until the iteration count g1 equals the maximum iteration count G1;

[0174] Step 5.2.4: The Knee-point strategy is used to find the optimal solution in the Pareto optimal solution set of the deep sparse autoencoder network. Since the optimization objectives of the deep sparse autoencoder network are the total number of nodes and the reconstruction error of the sample data, representing computational complexity and model accuracy respectively, and prediction accuracy is more important than computational complexity during continuous annealing, the Knee-point strategy is selected. Figure 6 The Knee solution in (a) is the optimal solution;

[0175] Step 5.2.5: Set the structural parameters and hyperparameters of the deep sparse autoencoder network corresponding to the optimal result as the optimal parameters of the multi-index quality prediction model in the continuous annealing process;

[0176] Step 5.3: Optimize the network structure and hyperparameters of the extreme gradient boosting algorithm prediction model; First, reset the optimal parameter results from Step 5.2 onto the deep sparse autoencoder network, using the same input process parameters and chemical composition data for continuous annealing as in Step 5.2. Then, retrain the multi-index quality prediction model for the continuous annealing process to optimize the extreme gradient boosting algorithm, using the Knee-point strategy to select the optimal performance model parameters; redesign the hyperparameter optimization objective of the extreme gradient boosting algorithm network as follows:

[0177] F(θ)={0.2(L YS +L TS ), 0.8L EL} (30)

[0178] Among them, L YS L TS L EL These are the root mean square errors between the predicted and actual measured values ​​of the three mechanical properties—yield strength, tensile strength, and elongation—during the continuous annealing process.

[0179] Step 5.3.1: Calculate the fitness function value for each population with the goal of minimizing the root mean square error between the predicted and actual measured values ​​of the three mechanical properties (yield strength, tensile strength, and elongation) during the continuous annealing process.

[0180] Step 5.3.2: Save the current calculation result, increment the population size by one, and when the population size p2 equals the maximum population size P2, the population size p2 = 0, and increment the iteration count g2 by one;

[0181] Step 5.3.3: Determine if the termination condition is met. Repeat steps 5.3.1 to 5.3.3 until the iteration count g2 equals the maximum iteration count G2;

[0182] Step 5.3.4: The optimal solution is determined from the non-dominated solution set of the overall framework optimization results using the Knee point strategy. Based on the mechanical properties of the strip steel, in this optimization process, the solution with lower losses in tensile strength TS and yield strength YS is selected as the Knee solution, as shown in the attached figure. Figure 6 As shown in (b).

[0183] Step 5.3.5: Set the network structure and hyperparameters of the extreme gradient boosting algorithm corresponding to the Knee solution to the optimal parameters of the extreme gradient boosting algorithm;

[0184] Step 6: Complete the multi-objective optimization process, and use the optimized hyperparameters and structural parameters as the hyperparameters and structural parameters of the multi-index quality prediction model in the continuous annealing process. Then, use the test dataset obtained in Step 2 to perform quality prediction.

[0185] Based on actual production data from a steel plant, the multi-index quality prediction model optimized in this embodiment was subjected to five-fold cross-validation to obtain statistical results. Table 4 shows the statistical results of the root mean square error, mean absolute error, and coefficient of determination for the three mechanical properties of strip steel—yield strength, tensile strength, and elongation—calculated using the specific implementation of this invention. For the root mean square error, Table 4 shows that the statistical results for the multi-index prediction of tensile strength, yield strength, and elongation are 22.6322, 17.5539, and 2.7772, respectively; the mean absolute error is 12.3772, 9.2225, and 1.6491, respectively; and the calculated coefficients of determination for tensile strength, yield strength, and elongation are 0.9655, 0.9792, and 0.9064, respectively. To further demonstrate the effectiveness of the multi-index quality prediction model based on evolutionary learning, this invention analyzed the average distribution between predicted and true values, using quantile-quantile plots to test model performance, as shown in the attached figure. Figure 7 As shown. From the appendix Figure 7As can be seen, the predicted values ​​are close to the observed values, and the overall prediction results tend to be a straight line with a slope of 1. The results indicate that the prediction error of this invention approximately follows a Gaussian distribution. Therefore, applying the multi-index quality prediction model of this invention in actual production processes can enable researchers to promptly grasp the quality of strip steel products, compensate for the shortcomings of traditional mechanical property testing methods, thereby helping continuous annealing production lines improve product quality and reduce costs, and also providing guidance for the research and development of new steel grades.

[0186] Table 4 Statistical Results of Mechanical Property Prediction for Continuous Annealing Process

[0187]

[0188]

[0189] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A strip steel multi-index quality prediction method based on evolutionary learning, characterized in that, The method comprises the following steps: Step 1: establishing a multi-index quality prediction model data set of a continuous annealing process; Step 2: performing data preprocessing on the multi-index quality prediction model data set constructed in step 1 to obtain a standard training data set and a test set; Step 3: constructing a multi-index quality prediction model in the continuous annealing process based on evolutionary learning, specifically including a deep sparse auto-encoding network and an extreme gradient boosting algorithm; Step 4: initializing starting parameters of a multi-objective optimization algorithm, and setting a search range for structure parameters and hyperparameters of the deep sparse auto-encoding network and the extreme gradient boosting algorithm; Step 5: executing the multi-index quality prediction model in the continuous annealing process based on evolutionary learning constructed in step 3, optimizing the structure parameters and hyperparameters of the deep sparse auto-encoding network and the extreme gradient boosting algorithm, and selecting parameter values when the multi-index quality prediction model in the continuous annealing process has the best performance by using a Knee point strategy; The step 5 specifically comprises the following steps: Step 5.1: initializing algorithm parameters of a non-dominated sorting genetic algorithm II, and defining a maximum population size and a maximum iteration number of the multi-index quality prediction model in the continuous annealing process; Step 5.2: inputting process parameters and chemical composition data of the continuous annealing process, training a deep sparse auto-encoding network model, optimizing network structure and hyperparameters of the deep sparse auto-encoding network, and selecting optimal model parameters by using a Knee point strategy; When optimizing the deep sparse auto-encoding network, the time complexity and network performance of the network are considered, and therefore the optimization target of the deep sparse auto-encoding network is set as: （16）； （17）； where F(.) is the objective function of the non-dominated sorting genetic algorithm II, L DSAE (.) is the reconstruction error value of the deep sparse auto-encoder network; θ is the model parameter to be optimized, in the first stage, θ is the number of networks, the structure parameter contains the number of nodes in the i-th layer , the number of layers N layer and the hyperparameters; Step 5.2.1: calculating fitness function values of each population by taking the minimum reconstruction error of the deep sparse auto-encoding network as a target; Step 5.2.2: Save the current calculation result, increment the population size by one, and when the population size... Equal to the maximum population size At that time, population size Number of iterations Add one; Step 5.2.3: judging whether a termination condition is met; repeating steps 5.2.1 to 5.2.3 until the iteration number is equal to a maximum iteration number ; Step 5.2.4: finding an optimal solution in a Pareto optimal solution set of the deep sparse auto-encoding network by using a Knee point strategy, Step 5.2.5: setting structure parameters and hyperparameters of the deep sparse auto-encoding network corresponding to the optimal result as optimal parameters of the multi-index quality prediction model in the continuous annealing process; Step 5.3: optimizing network structure and hyperparameters of the extreme gradient boosting algorithm prediction model; first, resetting the optimal parameter result optimized in step 5.2 to the deep sparse auto-encoding network, inputting process parameters and chemical composition data of the continuous annealing process as in step 5.2, then retraining the multi-index quality prediction model in the continuous annealing process, optimizing the extreme gradient boosting algorithm, and selecting optimal model parameters by using a Knee point strategy; the hyperparameter optimization target of the extreme gradient boosting algorithm network is redesigned as: （18）； wherein, L YS , L TS , L EL are the root mean square error results of the predicted values and the actual test values of the three mechanical property indexes of yield strength, tensile strength and elongation of the continuous annealing process, respectively; Step 5.3.1: calculating fitness function values of each population by taking the minimum root mean square error of predicted values and actual detection values of three mechanical performance indexes of the continuous annealing process as a target; Step 5.3.2: Save the current calculation result, increment the population size by one, and when the population size... Equal to the maximum population size At that time, population size Number of iterations Add one; Step 5.3.3: judging whether a termination condition is met; repeating steps 5.3.1 to 5.3.3 until the iteration number is equal to the maximum iteration number ; Step 5.3.4: determining an optimal solution from a non-dominated solution set of the optimization result of the overall framework by using a Knee point strategy; Step 5.3.5: setting network structure and hyperparameters of the extreme gradient boosting algorithm corresponding to the Knee solution as optimal parameters of the extreme gradient boosting algorithm; Step 6: After completing the multi-objective optimization process, the optimized hyperparameters and structure parameters are obtained, which are used as the hyperparameters and structure parameters of the multi-index quality prediction model in the continuous annealing process. Then, the test set obtained in step 2 is used for quality prediction.

2. The strip steel multi-index quality prediction method based on evolutionary learning according to claim 1, characterized in that, The step 1 specifically comprises the following steps: Step 1.1: 20 chemical composition parameters of the strip steel product are measured, including sulfur content S, copper content Cu, nickel content Ni, chromium content Cr, molybdenum content Mo, vanadium content V, niobium content Nb, total aluminum content Al, acid-soluble aluminum content, titanium content Ti, boron content B, tin content Sn, arsenic content As, zirconium content Zr, calcium content Ca, lead content Pb, antimony content Sb, nitrogen content N, oxygen content O, tungsten content W; Step 1.2: According to the measured data of the chemical composition parameters of the strip steel, the carbon equivalent is calculated using the following formula: （1）； （2）； Where Ceq is the carbon equivalent #1, Pcm is the carbon equivalent #2; C is the carbon content, Si is the silicon content, Mn is the manganese content; Step 1.3: 5 process parameters and strip steel information data are obtained during the process from entering the uncoiler to the coiler after the continuous annealing production process, including furnace temperature, existing temperature, annealing temperature, thickness, and width; Step 1.4: Three mechanical property indexes of the finished steel after the continuous annealing process in the actual production environment are measured, including yield strength, tensile strength, and elongation; Step 1.5: The composition parameters, carbon equivalent, process parameters and strip steel information data, and mechanical property indexes obtained in steps 1.1-1.4 are mapped; Step 1.6: Repeat steps 1.1 to 1.5 to obtain a sample data set that meets the set number of samples, and construct a training data set for the multi-index quality prediction model, wherein the data obtained in steps 1.1-1.3 are used as input data for the multi-index quality prediction model, and the mechanical property index data measured in step 1.4 are used as output data for the multi-index quality prediction model.

3. The strip steel multi-index quality prediction method based on evolutionary learning according to claim 1, characterized in that, The step 2 specifically comprises the following steps: Step 2.1: Read in the multi-index quality prediction model data set of the continuous annealing process constructed in step 1, and sequentially randomly shuffle the multi-index quality prediction model data set of the entire continuous annealing process; Step 2.2: Perform a visualization operation on the shuffled multi-index quality prediction model data set of the continuous annealing process, and select strongly correlated data items as the final input sample data set for the multi-index quality prediction model of the continuous annealing process after data analysis; Step 2.3: Standardize the data obtained in step 2.2 using data preprocessing techniques; The process of the standardization is as follows: firstly, the median value M and the interquartile range IQR of the data obtained in step 2.2 are calculated, then for a value d of the data obtained in step 2.2 i The data value after the data preprocessing and standardization is as follows: （5）； Step 2.4: Store the processed standard multi-index quality prediction model data set of the continuous annealing process, and divide the stored data set into a training data set and a test set in a ratio of 17:3 for the multi-index quality prediction model of the continuous annealing process.

4. The strip steel multi-index quality prediction method based on evolutionary learning according to claim 1, characterized in that, The step 3 specifically comprises the following steps: Step 3.1: Establish a deep sparse auto-encoding network to learn the deep semantic information and the implicit information of the interaction between data in the continuous annealing data set, and obtain a high-dimensional mapping representation of the intrinsic features of the input data; Step 3.1.1: Establishing the autoencoder, which is composed of an encoder and a decoder, wherein the encoder is composed of an input layer and a hidden layer, and the decoder is composed of multiple hidden layers and a reconstruction layer; first, the encoder encodes the input data of the training data set of the multi-index quality prediction model of the standard continuous annealing process obtained in step 2; then, the decoder reconstructs the original information by minimizing the square error of the output results of the continuous annealing process parameter data and the chemical composition data and the input data x i ​ Wherein the encoder is defined as follows: （6）； wherein, denotes each data sample in the input sample dataset X, N is the number of data samples, denotes an activation function, denotes a weight matrix inside the encoder, b denotes a bias vector of the encoder; The decoder is defined as follows: （7）； wherein, represents the input sample x i corresponding output result, W' is the weight matrix inside the decoder, and b' represents the bias vector of the decoder; the weight parameters inside the network are trained by minimizing the square error, and the square error represents the difference between the input data and the output data, which is defined as follows: （8）； wherein, is the output of the decoder in the autoencoder, represents the 2-norm of the autoencoder, used to measure the reconstruction error between x i and ; Step 3.1.2: Introduce a sparse self-encoding network based on the sparse constraint of the self-encoder; through the introduction of the sparse constraint, the automatic selection and dimension reduction process of the features are completed, and the cost function of the sparse self-encoding network is defined as follows: （9）； wherein, is a penalty factor to control the sparsity weight of the sparse term, η is a reconstruction error weight to control the weight of the reconstruction error, D KL is a sparsity measure term, i.e., the KL divergence is used to measure the difference between the expected sparse distribution in the hidden layer and the actual sparse distribution, by minimizing D KL to limit the sparsity of the sparse auto-encoding network, the formula of the KL divergence is defined as follows: （11）； （12）； where m denotes the total number of input modes, p is a sparsity parameter, representing the sparse distribution of the network, is the jth node of the sparse autoencoder hidden layer for the ith input x i the output; Step 3.1.3: Stack the hidden layers of the sparse self-encoding network in order to build a deep sparse self-encoding network, and through minimizing the reconstruction error to reconstruct the input data, the output features of the last hidden layer of the network at this time are used as the input data of the multi-index quality prediction model of the simulated annealing process; Step 3.2: Build the structure of the extreme gradient boosting algorithm model as the multi-index quality prediction model of the simulated annealing process, and then use the high-dimensional feature representation of the last hidden layer of the deep sparse self-encoding network in step 3.1 as the input of the extreme gradient boosting algorithm to predict the mechanical property index value of the strip steel; Assume the input data is x i The boosting strategy of the extreme gradient boosting algorithm model is defined as follows: （13）； （14）； where n is the total number of input data, is the output of the tth iteration, is the predicted value of the kth decision tree for the input data i; then, the cost function of the extreme gradient boosting algorithm is defined as: （15）； where n is the number of input data, Ω(f i ) represents the model complexity of the i-th decision tree, and t represents the total number of decisions. Step 3.3: Build an improved non-dominated sorting genetic algorithm II with a Knee point strategy to perform the optimal solution search task in the solution space to optimize the two-stage model structure and hyperparameters of the deep sparse self-encoding network and the extreme gradient boosting algorithm, and generate a strip steel index quality prediction model meeting the accuracy requirement.

5. The strip steel multi-index quality prediction method based on evolutionary learning according to claim 4, characterized in that, The non-dominated sorting genetic algorithm II in step 3.3 is as follows: according to the superiority of the individual, the individual is stratified, and different individuals in each layer are defined, then the elite selection strategy is performed to select and iterate the individuals with different advantages and generate new individuals; the superior individuals from the parent generation to the offspring are retained, and the improved crossover operator and variation operator are used to obtain the sub-population to improve the convergence speed of the non-dominated sorting genetic algorithm II and retain the population diversity, which specifically includes the following steps: Step 3.3.1 : Parent population Obtaining a new offspring population by a genetic algorithm including selection, crossover and mutation processes; Step 3.3.2: Parent population and a newly created offspring population composing the tthgeneration population of size 2N; Step 3.3.3: Selection of the population Non-dominated solutions are ranked and a set of non-dominated solutions is produced in order of dominance rank and placed in the next generation population in turn; and Step 3.3.4: Mutation of the population Step 3.3.4: If the population size of the non-dominated solution set is less than or equal to N, put it into the population If the population size of the non-dominated solution set is greater than N, calculate the crowding distance of the individuals in the non-dominated solution set that exceeds N, and then put them into the population in descending order until the number of individuals reaches N, and eliminate the remaining solutions ; Step 3.3.5: Selection on new population Selection, crossover and mutation operations are performed to obtain a new sub-population Then The new population is subjected to a new iteration process.

6. The strip steel multi-index quality prediction method based on evolutionary learning according to claim 1, characterized in that, The step 4 specifically includes the following steps: Step 4.1: Initialize the structure parameters and hyperparameters of the deep sparse self-encoding network to the default settings, including the number of network layers, the number of nodes, the sparsity, the beta value and the learning rate; Step 4.2: Initialize the network structure parameters and hyperparameters of the extreme gradient boosting algorithm to the default settings, including the tree depth, the number of evaluators and the learning rate; Step 4.3: Set the hyperparameters and structure parameters involved in the multi-index quality prediction model of the deep sparse self-encoding network and the extreme gradient boosting algorithm; The parameters specifically involved include: the number of layers N of the deep sparse auto-encoding network layer , the number of nodes N node , the sparse distribution , the penalty factor , the learning rate lr DSAE , the tree depth N of the extreme gradient boosting algorithm depth , the number of evaluators N estimator , the learning rate lr XGBoost .

Images