Neural network-based different steel pickling time prediction model and construction method

Abstract

This invention provides a neural network-based model and construction method for predicting pickling time for different steel grades, belonging to the field of steel rolling process optimization. The method first obtains historical production data of steel grades with different Si contents under different pickling processes to construct a sample set. From the sample set, Si content, furnace exit temperature, furnace time, final rolling temperature, layer cooling rate, coiling temperature, and iron oxide scale thickness are extracted, and an empirical formula for iron oxide scale thickness is established. A pickling time prediction model is constructed based on an LM-BP neural network. The input layer includes two input branches: the first branch is for iron oxide scale thickness, and the empirical formula for iron oxide scale thickness is introduced; the second branch is for the pickling process, setting input variables such as strip thickness, acid temperature, and acid concentration. The output is the pickling time. The model is trained, validated, and tested using the sample set to obtain pickling time prediction models for steel grades with different Si contents. This invention improves the accuracy and universality of pickling time prediction.

Description

Technical Field

[0001] This invention belongs to the field of steel rolling process optimization, specifically involving a prediction model and construction method for pickling time of different steel grades based on neural networks. Background Technology

[0002] Pickling is a crucial step in steel production, primarily used to remove the iron oxide scale that forms on the surface of hot-rolled strip steel during high-temperature heating and rolling. The quality of pickling directly determines the surface quality of the strip steel, playing a decisive role in the surface quality and performance of subsequent processes such as cold rolling, galvanizing, and coating. Iron oxide scale is typically composed of composite oxides such as FeO, Fe3O4, and Fe2O3, and its thickness, structure, and adhesion are influenced by various factors including steel composition and heating regime, requiring different pickling process parameters and corresponding pickling times. Precisely controlling the pickling time to avoid under-pickling or over-pickling is a critical technical issue for improving product surface quality and production efficiency.

[0003] Traditional methods for determining pickling process parameters primarily rely on empirical parameters. These parameters, such as acid temperature, acid concentration, and pickling time, are determined based on the type and thickness of the strip and the operational experience of the pickling line. However, this method has drawbacks. Different steel grades exhibit significant differences in pickling reaction rates under similar conditions, especially for steels with complex compositions, such as high-strength steel, duplex steel, and TRIP steel. Variations in the content of elements like Si, Mn, and Cr in the steel can significantly alter the phase composition and density of the iron oxide scale, making the pickling kinetics even more complex. Relying solely on empirical control often leads to uneven pickling, residual iron oxide scale, or corrosion of the substrate. In severe cases, this can cause significant differences in strip surface roughness, pitting, and other defects, thereby affecting the surface quality of the product and the stability of subsequent processes.

[0004] In recent years, with the widespread application of high-strength steel and multiphase steel in automobiles, the Si content in steel has generally increased to 0.5 wt.% or even above 1.7 wt.%. During heating, Si oxidation products easily form SiO2 or Fe2SiO4 layers at the interface between the steel matrix and the iron oxide scale. SiO2 forms a dense protective layer, while Fe2SiO4 pins the iron oxide scale, increasing the difficulty of pickling. Simultaneously, the dense silicon oxide layer also hinders the contact between the acid solution and the iron matrix, thus significantly slowing down the pickling reaction rate. Furthermore, there are complex nonlinear coupling relationships between factors such as strip thickness, iron oxide scale thickness, pickling temperature, and concentration. This makes it difficult for traditional empirical methods to accurately reflect the dissolution behavior of steel grades with different Si contents during the pickling process.

[0005] Currently, artificial intelligence (AI) is frequently used to solve complex engineering problems. Machine learning, based on existing historical data, can accurately obtain the complex interrelationships between multiple parameters and quickly predict the required output. AI methods in machine learning can discover the mapping from high-throughput input data to outputs used for decision-making. Based on this, to address the aforementioned issues, some scholars have developed a strip steel pickling prediction system based on a neural network model. This system includes an AI model to predict the pickling rate of strip steel. The model takes strip steel thickness, strip steel temperature, iron oxide scale thickness, acid temperature, acid concentration, and Fe ion concentration in the acid as input parameters, and pickling rate as the output. However, traditional artificial neural networks often require a large amount of prior data as training samples, and their internal "black box" structure cannot intuitively reflect the physical relationships between multiple sets of parameters, making them unsuitable for simultaneously predicting pickling processes for different steel grades. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a prediction model and construction method for pickling time of different steel grades based on neural networks. By constructing an empirical formula relating process parameters to the thickness of iron oxide scale and incorporating this empirical formula into the prediction model, the accuracy and universality of pickling time prediction are improved.

[0007] To achieve the above objectives, the technical solutions adopted in the embodiments of the present invention are as follows:

[0008] In a first aspect, embodiments of the present invention provide a method for constructing a prediction model for pickling time of different steel grades based on a neural network, the method comprising the following steps:

[0009] Step S1: Obtain historical production data of steel grades with different Si contents under different pickling processes, and construct a sample set including Si content, furnace exit temperature, furnace time, final rolling temperature, layer cooling rate, coiling temperature, strip thickness, acid solution temperature, acid solution concentration, iron oxide scale thickness and pickling time.

[0010] Step S2: Extract Si content, furnace exit temperature, furnace time, final rolling temperature, layer cooling rate, coiling temperature and iron oxide scale thickness from the sample set, establish an empirical formula for iron oxide scale thickness based on hot rolling process parameters, and use polynomial regression to calculate the coefficients of the empirical formula.

[0011] Step S3: Construct a preliminary pickling time prediction model based on the LM-BP neural network. The model includes an input layer, a hidden layer, and an output layer, and the number of nodes in each layer is unknown. The input layer includes two input branches: the first branch is the iron oxide scale thickness, and the second branch is the pickling process.

[0012] Step S4: The empirical formula for iron oxide scale thickness is introduced into the first input branch of the preliminary pickling time prediction model. The second input branch is set with input variables of strip thickness, acid temperature, and acid concentration, and the output is pickling time. Based on all input and output variables, the number of nodes in the input layer, hidden layer, and output layer is determined, the basic structure of the artificial neural network is determined, and the pickling time prediction model is obtained.

[0013] Step S5: After normalizing the data in the sample set, divide it into a training set, a validation set, and a test set. Train, validate, and test the pickling time prediction model to obtain pickling time prediction models for steel grades with different Si contents.

[0014] As a preferred embodiment of the present invention, the empirical formula for the thickness of the constructed iron oxide scale is as follows:

[0015] (1)

[0016] In equation (1), This indicates the Si content in the steel; Indicates the furnace exit temperature of the hot-rolled product; Indicates the time spent in the furnace; Indicates the final rolling temperature; Indicates the laminar cooling rate; Indicates the winding temperature; Represents a constant. This represents the coefficients of each item.

[0017] As a preferred embodiment of the present invention, a constant term and the values ​​of each coefficient are obtained by multinomial regression calculation.

[0018] As a preferred embodiment of the present invention, the Si content... With iron oxide scale thickness Satisfies the following physical laws: When When the content is ≤0.5wt.%, the thickness of the iron oxide scale should be controlled between 5 and 15 μm; when the content is <0.5wt.%, the thickness of the iron oxide scale should be controlled between 5 and 15 μm. When ≤1.0wt.%, the iron oxide scale thickness should be controlled between 10 and 25 μm; when 1.0wt.% < When the content is ≤1.7wt.%, the thickness of the iron oxide scale can reach 25~45μm.

[0019] As a preferred embodiment of the present invention, hot rolling parameters With iron oxide scale thickness It satisfies the following physical laws: furnace exit temperature For every 100°C increase, the thickness of the iron oxide scale increases. The thickness of the iron oxide scale increases by 5~10μm for every 30min increase in furnace time; the thickness of the iron oxide scale increases by 4~6μm for every 100℃ increase in final rolling temperature; the thickness of the iron oxide scale decreases by 5~8μm for every 50℃ / s increase in laminar cooling rate; and the thickness of the iron oxide scale increases by 2~4μm for every 50℃ increase in coiling temperature.

[0020] As a preferred embodiment of the present invention, the scope of the acquired historical production data is as follows:

[0021] The Si content of different steel grades is 0wt.%~2wt.%, the hot rolling furnace exit temperature is 1100℃~1250℃, the furnace time is 30min~120min, the final rolling temperature is 850℃~950℃, the layer cooling rate is 50℃ / s~300℃ / s, the coiling temperature is 450℃~650℃, the strip thickness is 1mm~5mm, the acid concentration is 5%~20%, the acid temperature is 70℃~90℃, the iron oxide scale thickness is 5μm~30μm, and the pickling time is 5s~60s.

[0022] In a preferred embodiment of the present invention, the acid solution includes, but is not limited to, HCl, H2SO4 and mixed acid solutions.

[0023] In a preferred embodiment of the present invention, the number of neurons in the input layer of the model is set to 9, the number of neurons in the output layer is set to 1, and the number of neurons in the hidden layer ranges from 2 to 12.

[0024] In a preferred embodiment of the present invention, during the model training process, the objective function E is set to minimize the sum of squared prediction errors.

[0025] Secondly, embodiments of the present invention also provide a prediction model for pickling time of different steel grades based on a neural network, wherein the model is constructed using the construction method described above.

[0026] The solutions of the embodiments of the present invention have the following beneficial effects:

[0027] The pickling time prediction model and construction method for different steel grades based on neural networks provided in this invention starts from the hot rolling process and the alloy composition of the steel grade, quantitatively correlates the hot rolling process parameters, Si content, and pickling behavior, and reflects the pickling time of the strip steel from the source of hot rolling. It expresses the iron oxide scale thickness in the form of hot rolling process parameters, avoiding the need for microscopic measurement of the iron oxide scale thickness of the steel grade, reducing the difficulty of model use, and expanding the model's applicability. Using the output of empirical formulas as input to the artificial neural network enhances the generalization ability of the prediction model and improves the model training efficiency and stability. This invention improves the accuracy and universality of pickling time prediction, increases the production efficiency of pickling lines, and can be used to guide the setting of acid temperature, concentration, and pickling time in actual production.

[0028] Of course, implementing any product or method of the present invention does not necessarily require achieving all of the advantages described above at the same time. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of the method for constructing a prediction model for pickling time of different steel grades based on neural networks, as described in an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the pickling time prediction model for different steel grades described in the embodiments of the present invention;

[0032] Figure 3 The prediction results are those of the pickling time prediction models for different steel grades described in the embodiments of the present invention.

[0033] Figure 4 This is an analysis chart showing the prediction results of the empirical formula for iron oxide scale thickness in the embodiments of the present invention;

[0034] Figure 5 This is the verification result of the generalization ability of the pickling time prediction model for different steel grades described in the embodiments of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. It should be noted that, without conflict, the embodiments and features in the embodiments of the present invention can also be combined with each other.

[0036] It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of the embodiments of the present invention, the terms "first," "second," "third," "fourth," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. In addition, sometimes a subscript such as W1 may be written in a non-subscript form such as W1, and their meanings are consistent unless the distinction is emphasized.

[0037] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0038] This invention provides a neural network-based model and method for predicting pickling time for different steel grades. By constructing a unified model, it predicts the pickling time of steel grades with different Si contents under different pickling processes. The method first collects hot rolling process parameters, strip oxide scale thickness, pickling parameters, and strip oxide scale cleaning time for steel grades with different Si contents from field data and literature. Second, it establishes empirical formulas for hot rolling parameters and strip oxide scale thickness, using the output of these formulas as input to the artificial neural network model. Then, it normalizes the strip oxide scale thickness, strip thickness, and pickling parameters, using this as the training data for the prediction model. Finally, it optimizes the BP neural network model using the LM algorithm, and uses the optimized BP neural network to establish a prediction model for pickling time. This invention incorporates empirical formulas as prior conditions into the artificial neural network, reducing the model's data dependency and enhancing its generalization ability. It achieves nonlinear coupling between steel composition and hot rolling and pickling process parameters, resulting in a high-precision intelligent prediction model for pickling time of steel strips with different Si contents, significantly improving the accuracy of process control and production efficiency.

[0039] like Figure 1 As shown, the method for constructing the pickling time prediction model for different steel grades based on neural networks includes the following steps:

[0040] Step S1: Obtain historical production data of steel grades with different Si contents under different pickling processes, and construct a sample set including Si content, furnace exit temperature, furnace time, final rolling temperature, layer cooling rate, coiling temperature, strip thickness, acid solution temperature, acid solution concentration, iron oxide scale thickness and pickling time.

[0041] In this step, the historical production data can come from the production site or from literature data, etc.

[0042] In one executable embodiment, the sample set obtained in this step includes input data: the Si content of different steel grades ( The content is 0wt.%~2wt.%, and the furnace exit temperature of hot rolling is ( The temperature range is 1100℃~1250℃, and the furnace time is ( The rolling time is 30-120 minutes, and the final rolling temperature is ( 850℃~950℃, layer cooling rate ( The winding speed is 50℃ / s to 300℃ / s. The temperature ranges from 450℃ to 650℃, the strip thickness (H) is 1mm to 5mm, the acid concentration (C) is 5% to 20%, and the acid temperature (T) is 70℃ to 90℃; Output data: iron oxide scale thickness ( The pickling time is 5μm~30μm. The time is 5s to 60s. The acid solution includes, but is not limited to, HCl, H2SO4 and mixed acid solutions.

[0043] Step S2: Extract Si content, furnace exit temperature, furnace time, final rolling temperature, layer cooling rate, coiling temperature and iron oxide scale thickness from the sample set, establish an empirical formula for iron oxide scale thickness based on hot rolling process parameters, and use polynomial regression to calculate the coefficients of the empirical formula.

[0044] In this step, the empirical formula for iron oxide scale thickness is input to hot rolling process parameters such as Si content, furnace exit temperature, furnace time, final rolling temperature, layer cooling rate, and coiling temperature, and outputs the predicted iron oxide scale thickness. Iron oxide scale thickness, as a key prerequisite for the subsequent pickling reaction rate, is directly related to the hot rolling process parameters. Higher furnace exit temperatures accelerate the oxidation reaction rate on the steel surface, increasing Fe content. 2+ Fe 3+ Enhanced diffusion leads to a significant increase in the thickness and density of the iron oxide scale. Longer furnace time results in more complete scale growth and a thicker scale. Higher final rolling temperatures cause scale compaction and regeneration, increasing density but hindering subsequent pickling. Lower final rolling temperatures, however, easily lead to cracks and pores, increasing acid permeability. The laminar cooling rate directly controls the oxidation time of the strip after final rolling; a faster cooling rate results in a shorter oxidation time and a thinner scale. Coiling temperature affects scale regeneration and phase transformation; high coiling temperatures result in composite scale with strong adhesion, while low coiling temperatures produce a thinner and looser oxide layer. Therefore, the slab process parameters directly reflect the thickness of the iron oxide scale on the surface of hot-rolled products. By constructing an empirical formula for scale thickness, hot rolling process parameters are combined with pickling kinetics, ultimately enabling the prediction of pickling time.

[0045] The empirical formula for the thickness of the iron oxide scale constructed in this step is as follows:

[0046] (1)

[0047] In equation (1), This indicates the Si content in the steel; Indicates the furnace exit temperature of the hot-rolled product; Indicates the time spent in the furnace; Indicates the final rolling temperature; Indicates the laminar cooling rate; Indicates the winding temperature; Represents a constant. The constant term and the values ​​of each coefficient are obtained by multinomial regression calculation.

[0048] In the empirical formula for the thickness of the iron oxide scale, the Si content... With iron oxide scale thickness Satisfies the following physical laws: When When the content is ≤0.5wt.%, the thickness of the iron oxide scale should be controlled between 5 and 15 μm; when the content is <0.5wt.%, the thickness of the iron oxide scale should be controlled between 5 and 15 μm. When ≤1.0wt.%, the iron oxide scale thickness should be controlled between 10 and 25 μm; when 1.0wt.% < When the content is ≤1.7wt.%, the thickness of the iron oxide scale can reach 25~45μm.

[0049] Hot rolling parameters With iron oxide scale thickness It satisfies the following physical laws: furnace exit temperature For every 100°C increase, the thickness of the iron oxide scale increases. The thickness of the iron oxide scale increases by 5~10μm for every 30min increase in furnace time; the thickness of the iron oxide scale increases by 4~6μm for every 100℃ increase in final rolling temperature; the thickness of the iron oxide scale decreases by 5~8μm for every 50℃ / s increase in laminar cooling rate; and the thickness of the iron oxide scale increases by 2~4μm for every 50℃ increase in coiling temperature.

[0050] By incorporating the above empirical formula for iron oxide scale thickness into the model, factors related to the variation of iron oxide scale thickness are introduced into the model, thereby improving the model's stability.

[0051] Step S3: Construct a preliminary pickling time prediction model based on the LM-BP neural network. The model includes an input layer, a hidden layer, and an output layer, and the number of nodes in each layer is unknown. The input layer includes two input branches: the first branch is the thickness of the iron oxide scale, and the second branch is the pickling process.

[0052] like Figure 2As shown, the LM-BP neural network adopts a three-layer framework. The three-layer neural network structure has been mathematically proven to be able to approximate any objective function with arbitrary precision, basically satisfying the relationship between input and output in the field of materials processing. In this embodiment, the three-layer neural network constructing the pickling time prediction model consists of an input layer, a hidden layer, and an output layer. The input layer is affected by various factors influencing pickling time, including the thickness of the strip oxide scale, the strip thickness, the acid concentration, and the acid temperature; therefore, the number of neurons in the input layer is set to 9. The thickness of the strip oxide scale is input by introducing an empirical formula for oxide scale thickness. The inputs of the formula include Si content, furnace exit temperature, furnace time, final rolling temperature, layer cooling rate, and coiling temperature. The basic mathematical model of the neurons is:

[0053] (2)

[0054] In equation (2), For input signal, For weight values, For the threshold, For activation function, This is the output signal.

[0055] The output layer corresponds to the acid washing time and has a neuron count of 1. A linear function is used in the output layer to maintain the continuity of the predicted values.

[0056] (3)

[0057] In equation (3), This represents the output layer function.

[0058] The hidden layer fits the complex nonlinear relationship between multiple parameters through a nonlinear activation function; preferably, the sigmoid function is used as the activation function; and the number of neurons in the hidden layer is calculated to be in the range of 2 to 12 based on empirical formulas. To simplify the model and ensure its computational power, the number of hidden layers is determined to be 10, and the empirical formula is as follows:

[0059] (4)

[0060] In equation (4), denoted as the number of neurons in the hidden layer, m as the number of neurons in the input layer, n as the number of neurons in the output layer, and a as a constant (usually 0 ≤ a ≤ 10).

[0061] Step S4: The empirical formula for iron oxide scale thickness is introduced into the first input branch of the preliminary pickling time prediction model. The second input branch is set with input variables of strip thickness, acid temperature, and acid concentration, and the output is pickling time. Based on all input and output variables, the number of nodes in the input layer, hidden layer, and output layer is determined, the basic structure of the artificial neural network is determined, and the pickling time prediction model is obtained.

[0062] In this step, the empirical formula for iron oxide scale thickness with determined coefficients is introduced into the preliminary pickling time prediction model through the first input branch.

[0063] Step S5: After normalizing the data in the sample set, divide it into a training set, a validation set, and a test set. Train, validate, and test the pickling time prediction model to obtain pickling time prediction models for steel grades with different Si contents.

[0064] In this step, the training set, validation set, and test set account for 60%, 30%, and 10% of the total dataset, respectively. Normalization achieves the following objectives: (1) Large range differences exist between different input data, leading to slow convergence of the neural network model and long training time; (2) Input data with a larger range may play a greater role in the model's classification process, while the role of input data with a smaller range may be weakened or even disappear. The Min-Max normalization method is used during normalization. The Min-Max normalization formula is:

[0065] (5)

[0066] In equation (5), For the normalized data, This is the original data. and These are the minimum and maximum values ​​of the same set of samples, respectively.

[0067] The process of training, validating, and testing a model specifically includes:

[0068] The training and validation sets are input into the pickling time prediction model. Variables in the empirical formula for oxide layer thickness are input through the first input branch, and pickling process-related data are input through the second input branch. The initial weights and thresholds in the neural network structure are determined based on the number of nodes in the input, hidden, and output layers. Based on the prediction error between the neural network's predicted and actual values, the weights and thresholds of each node are continuously adjusted using the Levenverg-Marquardt (LM) algorithm. After training for a preset number of rounds, the optimal weights and thresholds are determined, thereby improving the prediction accuracy and generalization ability of the neural network.

[0069] In this step, the LM algorithm formula is:

[0070] (6)

[0071] In equation (6), For Jacobian matrices, For the residual vector, To update the step size, This represents the variable that minimizes the loss function.

[0072] The formula for updating the weights is:

[0073] (7)

[0074] In equation (7), For Jacobian matrices, For the residual vector, The damping factor, It is the identity matrix. The algorithm automatically reduces this as the error decreases. The value is increased to make the update direction closer to the Gauss-Newton method, thus accelerating convergence; as the error increases, the value is increased. This value makes the update direction closer to gradient descent, improving the algorithm's stability.

[0075] The update formula for the threshold (bias value) is:

[0076] (8)

[0077] In equation (8), It is the threshold in the k-th iteration. It's the learning rate. The gradient of the bias.

[0078] During training, let the objective function E be the sum of squared prediction errors, as shown in the formula:

[0079] (9)

[0080] In equation (9), These are measured values. is the predicted value, and N is the total number of samples.

[0081] The prediction error thresholds are set as follows: Mean Absolute Value (MAE) ≤ 5%, Root Mean Square Error (RMSE) ≤ 5%, Regression Coefficient (R²) ≤ 5%. 2 )≥0.98.

[0082] After training and validation, the pickling time prediction model is tested using a test set to finally obtain pickling time prediction models for steels with different Si contents.

[0083] Based on the above model construction method, this embodiment of the invention also provides a neural network-based prediction model for pickling time of different steel grades. The model is constructed using the above method. The input layer of the model includes a first input branch and a second input branch, wherein the first input branch is used to input the constructed empirical formula for iron oxide scale thickness, and the second input branch is used to input the pickling process.

[0084] The pickling time prediction model and construction method based on neural networks for different steel grades described in this embodiment of the invention were applied to the construction process of a pickling time prediction model for a certain pickling production line. Based on the production history of this production line, historical data were obtained, including: Si content (0wt.%~0.2wt.%); hot rolling furnace exit temperature (1100℃~1250℃); furnace time (30min~200min); final rolling temperature (850℃~950℃); layer cooling rate (50℃ / s~300℃ / s); coiling temperature (450℃~650℃); strip thickness (1mm~5mm); acid concentration (5%~20%); and acid temperature (70℃~90℃). HCl was used as the acid.

[0085] The empirical formula for iron oxide scale thickness is in polynomial form. Based on the hot-rolling parameters and iron oxide scale thickness data, the material constants of the empirical formula are nonlinearly fitted to establish an empirical formula for strip iron oxide scale thickness. The output of this empirical formula is then used as the input parameter for an artificial neural network, resulting in the following empirical formula:

[0086]

[0087] To avoid training bias in neural networks caused by differences in the dimensions of different input parameters, the Min-Max normalization method is used to normalize the input parameters.

[0088] A three-layer backpropagation artificial neural network is used, consisting of an input layer, a hidden layer, and an output layer. The input layer contains 4 neurons, the output layer contains 1 neuron, and the number of neurons in the hidden layer is determined to be 10 based on an empirical formula. The number of neurons in the hidden layer directly affects the model's learning time and learning ability. If there are too few neurons, the model's learning time is shorter, but its learning and information processing capabilities are poorer; if there are too many neurons, although the model's learning and information processing capabilities are significantly improved, the model structure becomes more complex, and the learning time also increases. Considering all factors, the constant a is determined to be 8, and the number of neurons in the hidden layer is 10.

[0089] The initial weights and thresholds of the network are randomly generated between [-1, 1].

[0090] The LM algorithm was used to optimize the neural network model. The LM algorithm combines the advantages of gradient descent and Gauss-Newton methods, exhibiting fast convergence and strong nonlinear fitting capabilities. During training, the prediction error of pickling time was used as the objective function, and the convergence of the objective function or the error threshold was set to 5%.

[0091] After training, the validation set is input into the model, and the mean absolute error and fitting coefficients are calculated. When R² ≥ 0.98 and MAE ≤ 5%, the model is considered to have good prediction accuracy.

[0092] The constructed pickling time prediction model was applied to the pickling production line. Pickling times for different steel grades were statistically analyzed. The Si content (wt.%), hot-rolled furnace exit temperature (°C), furnace time (min), final rolling temperature (°C), layer cooling rate (°C / s), coiling temperature (°C), strip thickness (mm), acid concentration (%), and acid temperature (°C) of each steel grade were used as input parameters to the neural network. The oxide scale removal time for different steel grades was used as the output parameter. The data are listed in Table 1. Prediction results were plotted based on Table 1. Figure 3 As shown.

[0093] Table 1 shows the measurement data of pickling time for different Si contents.

[0094]

[0095]

[0096] Based on the data in Table 1, the predictions were validated using empirical formulas and an optimized LM-BP neural network. The predicted results were compared with the actual values, as shown in Tables 2 and 3. Analysis graphs were then plotted based on the data from Tables 2 and 3. Figure 4 As shown.

[0097] Table 2 Prediction Results Using Empirical Formulas

[0098]

[0099]

[0100] Table 3 Prediction results of LM-BP neural network

[0101]

[0102]

[0103] like Figure 4 As shown, through comparative analysis of actual and predicted values, it can be found that the prediction error (MAE) of the artificial neural network model is ≤5%, and the fitting coefficient (R²) is ≤5%. 2The result ≥0.99 indicates that the model can accurately predict the pickling and cleaning time for different Si contents under different hot rolling process conditions.

[0104] To verify the model's generalization ability, three sets of data outside the dataset were tested to observe the model's predictive ability on data outside the dataset. The results are shown in Tables 4 and 5.

[0105] Table 4 Additional Data

[0106]

[0107] Table 5. Generalization Ability Verification Results

[0108]

[0109] From Table 4, Table 5 and Figure 5 As can be seen, the artificial neural network can also predict the pickling time of iron oxide scale well for data outside the dataset. Therefore, it can be proven that the model has good generalization ability.

[0110] Therefore, the model constructed by the neural network-based pickling time prediction model construction method for different steel grades provided in this embodiment of the invention has high prediction accuracy and generalization ability, and can well reflect the nonlinear relationship between hot rolling process parameters and pickling parameters, and has high adaptability and universality.

[0111] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed, and is not intended to limit the scope of the claimed invention, but only to illustrate preferred embodiments of the invention. Those skilled in the art should understand that the scope of the invention is not limited to the 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 inventive concept. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A method for constructing a neural network-based pickling time prediction model for different steel grades, characterized in that, The method includes the following steps: Step S1: Obtain historical production data of steel grades with different Si contents under different pickling processes, and construct a sample set including Si content, furnace exit temperature, furnace time, final rolling temperature, layer cooling rate, coiling temperature, strip thickness, acid solution temperature, acid solution concentration, iron oxide scale thickness and pickling time. Step S2: Extract Si content, furnace exit temperature, furnace time, final rolling temperature, layer cooling rate, coiling temperature and iron oxide scale thickness from the sample set, establish an empirical formula for iron oxide scale thickness based on hot rolling process parameters, and use polynomial regression to calculate the coefficients of the empirical formula. Step S3: Construct a preliminary pickling time prediction model based on the LM-BP neural network. The model includes an input layer, a hidden layer, and an output layer, and the number of nodes in each layer is unknown. The input layer includes two input branches: the first branch is the iron oxide scale thickness, and the second branch is the pickling process. Step S4: The empirical formula for iron oxide scale thickness is introduced into the first input branch of the preliminary pickling time prediction model. The second input branch is set with input variables of strip thickness, acid temperature, and acid concentration, and the output is pickling time. Based on all input and output variables, the number of nodes in the input layer, hidden layer, and output layer is determined, the basic structure of the artificial neural network is determined, and the pickling time prediction model is obtained. Step S5: After normalizing the data in the sample set, divide it into a training set, a validation set, and a test set. Train, validate, and test the pickling time prediction model to obtain pickling time prediction models for steel grades with different Si contents.

2. The method according to claim 1, characterized in that, The empirical formula for the thickness of the iron oxide scale is as follows: (1) In equation (1), This indicates the Si content in the steel; Indicates the furnace exit temperature of the hot-rolled product; Indicates the time spent in the furnace; Indicates the final rolling temperature; Indicates the laminar cooling rate; Indicates the winding temperature; Represents a constant. This represents the coefficients of each item.

3. The method according to claim 2, characterized in that, The constant term and coefficient values ​​are obtained by using polynomial regression.

4. The method according to claim 2, characterized in that, Si content With iron oxide scale thickness Satisfies the following physical laws: When When the content is ≤0.5wt.%, the thickness of the iron oxide scale should be controlled between 5 and 15 μm; When 0.5wt.% < When ≤1.0wt.%, the iron oxide scale thickness should be controlled between 10 and 25 μm; when 1.0wt.% < When the content is ≤1.7wt.%, the thickness of the iron oxide scale can reach 25~45μm.

5. The method according to claim 2, characterized in that, Hot rolling parameters With iron oxide scale thickness It satisfies the following physical laws: furnace exit temperature For every 100°C increase, the thickness of the iron oxide scale increases. The thickness of the iron oxide scale increases by 5~10μm for every 30min increase in furnace time; the thickness of the iron oxide scale increases by 4~6μm for every 100℃ increase in final rolling temperature; the thickness of the iron oxide scale decreases by 5~8μm for every 50℃ / s increase in laminar cooling rate; and the thickness of the iron oxide scale increases by 2~4μm for every 50℃ increase in coiling temperature.

6. The method according to claim 1, characterized in that, The scope of the historical production data obtained is as follows: The Si content of different steel grades is 0wt.%~2wt.%, the hot rolling furnace exit temperature is 1100℃~1250℃, the furnace time is 30min~120min, the final rolling temperature is 850℃~950℃, the layer cooling rate is 50℃ / s~300℃ / s, the coiling temperature is 450℃~650℃, the strip thickness is 1mm~5mm, the acid concentration is 5%~20%, the acid temperature is 70℃~90℃, the iron oxide scale thickness is 5μm~30μm, and the pickling time is 5s~60s.

7. The method according to claim 1, characterized in that, The acid solution includes, but is not limited to, HCl, H2SO4, and mixed acid solutions.

8. The method according to claim 1, characterized in that, The model has 9 neurons in the input layer, 1 neuron in the output layer, and 2 to 12 neurons in the hidden layer.

9. The method according to claim 1, characterized in that, During model training, the objective function E is set to minimize the sum of squared prediction errors.

10. A neural network-based model for predicting pickling time for different steel grades, characterized in that, The model is constructed using the construction method described in any one of claims 1-9.

Images