A hot galvanizing iron strip surface roughness optimization control method based on hierarchical strategy

Abstract

The present application relates to a kind of based on hierarchical strategy's hot galvanizing iron strip surface roughness optimization control method, belong to metal plastic processing technical field, a kind of based on hierarchical strategy's hot galvanizing iron strip surface roughness optimization control method, comprising the following steps: S1: data acquisition;S2: target surface roughness index setting;S3: surface roughness prediction model construction based on LSTM;S4: first stage optimization: preliminary PSO search of flattening process elongation rate;S5: judgment and stage switching;S6: second stage optimization: multi-parameter joint PSO search;S7: output final optimization parameter;S8: closed-loop control and feedback mechanism, the present application realizes the hierarchical optimization control and cross-process collaborative optimization of hot galvanizing iron strip surface roughness;Break through single-process adjustment limit, and then accurately control the surface roughness of hot galvanizing iron strip.

Description

Technical Field

[0001] This invention belongs to the field of metal plastic processing technology, specifically relating to a method for optimizing and controlling the surface roughness of hot-dip galvanized iron strip based on a graded strategy. Background Technology

[0002] Hot-dip galvanized steel strip is widely used in the construction, automotive, and home appliance industries due to its excellent corrosion resistance, good processability and weldability. Its surface quality, especially surface roughness, has a crucial impact on subsequent coating adhesion, stamping performance, aesthetic appearance, and frictional properties with molds. Ideal surface roughness can effectively improve coating adhesion, reduce friction during the forming process, minimize mold wear, and give products a better appearance.

[0003] Currently, controlling surface roughness in the production process of hot-dip galvanized steel strip still faces many challenges. Numerous factors influence surface roughness and are interconnected, primarily including the substrate conditions of the steel strip, leveling process parameters, and hot-dip galvanizing process parameters (especially alloying). Traditional roughness control methods typically rely on operator experience, trial and error, or adjustments to single parameters. These methods are inefficient, struggle to achieve precise and stable roughness control, leading to product quality fluctuations and increased production costs and resource consumption.

[0004] For example, the elongation rate in the leveling process directly affects the surface morphology of the strip, but its adjustment range is limited, and excessive adjustment may affect the mechanical properties of the strip. Similarly, parameters such as alloying temperature, alloying rate, and iron content in the coating during the hot-dip galvanizing process significantly influence the microstructure, grain size, and surface morphology of the coating. The complex nonlinear relationships between these parameters make it difficult to meet increasingly stringent surface quality requirements with single or simple adjustment strategies.

[0005] Particle Swarm Optimization (PSO), as a swarm intelligence optimization algorithm, performs well in solving multidimensional continuous function optimization problems due to its advantages such as simple implementation, fast convergence speed, and few parameters. However, its application in the complex multi-stage, multi-parameter collaborative optimization of surface roughness of hot-dip galvanized steel strip, combined with the parameter search range constraints of specific stages, has not yet been fully utilized and researched.

[0006] Therefore, the industry urgently needs a new method that can systematically and intelligently optimize the production parameters of hot-dip galvanized steel strip, combined with the efficient optimization algorithm PSO, and taking into account the actual constraints of parameter adjustment, so as to achieve precise and efficient control of surface roughness. Summary of the Invention

[0007] The purpose of this invention is to provide a simple and rationally designed method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy in order to solve the above-mentioned problems.

[0008] The present invention achieves the above objectives through the following technical solutions: A method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy includes the following steps: S1: Data Acquisition; S2: Setting the target surface roughness index; S3: Construction of a surface roughness prediction model based on LSTM; S4: First stage optimization: Preliminary PSO search for the elongation rate of the leveling process; S5: Judgment and Phase Switching; S6: Second-stage optimization: Multi-parameter joint PSO search; S7: Output the final optimized parameters; S8: Closed-loop control and feedback mechanism.

[0009] As a further optimization of the present invention, the hot-dip galvanizing process includes: alloying temperature (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%) The leveling process includes: strip entry thickness (mm), elongation (%), Yield strength (MPa), rolling kilometers of the work rolls (km); Surface roughness: The surface roughness of hot-dip galvanized iron strip is measured using a roughness measuring instrument. In order to avoid uneven roughness in the width direction of the strip, three points are measured in the width direction: the working side, the operating side, and the middle. The average value is then taken as the roughness value Ra (µm) of the current strip.

[0010] As a further optimization of the present invention, the target surface roughness index can be a specific value (e.g., Ra=1.2µm) or an allowable range (e.g., 1.0µm≤Ra≤1.5µm).

[0011] As a further optimization of the present invention, the construction of the LSTM-based surface roughness prediction model specifically includes: in order to provide an accurate objective function evaluation basis for the PSO optimization algorithm, a deep learning prediction model is constructed to predict surface roughness values ​​under different combinations of process parameters in real time; the specific steps are as follows: S31: Data Preprocessing and Feature Sequence Construction Abnormal data cleaning: The collected raw process data is processed to remove abnormal process points such as belt breakage and deceleration stop sections in the production process, so as to ensure that the model training set reflects the normal production process. Normalization: Due to the huge difference in dimensions of the collected process parameters (such as alloying temperature and strip thickness), the Min-Max normalization method is used to scale all the original data to the (0,1) interval to accelerate the convergence of the neural network. Time series sliding window construction: Considering that the wear of the work rolls (rolling kilometers) in the leveling process is a continuous degradation process, this model does not only take single-point data, but also adopts the sliding window technique; the parameters of the current time t and the previous n historical volumes are selected to form the feature sequence. ; Input vector definition: The feature vector x at each time step includes: alloying temperature. (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%), strip entry thickness (mm), elongation (%), Yield strength (MPa), rolling kilometers of the work rolls (km); where the feature matrix X1 consists of the following parameters: ; S32: Dataset Partitioning and Temporal Feature Isolation To ensure the model has good generalization ability and maintains prediction accuracy on unseen production data, the preprocessed feature sequences are subjected to strict set partitioning: Time-sequence partitioning: Considering the continuity of the hot-dip galvanizing production process and the evolutionary characteristics of equipment conditions (such as roll wear), random sampling is not used. Instead, the sample library is strictly partitioned according to the chronological order of strip steel production. The processed sample library is divided into training set, validation set, and test set, with a ratio of [ratio missing]. ; Sample isolation: Ensure that the timestamp of the training set is earlier than that of the validation set, and the timestamp of the validation set is earlier than that of the test set, thereby simulating the real-world application scenario of the model "learning from history and predicting the future" in actual production. S33: Model Adaptive Hyperparameter Optimization To prevent underfitting or overfitting of the model due to manually set parameters, a Bayesian optimization algorithm is introduced to automatically optimize the core parameters of the LSTM network; the optimization objective is to minimize the mean squared error on the validation set. The optimization space includes the number of hidden layer neurons (32-256), the initial learning rate (0.0001-0.01), the time step n (3-15), and the Dropout ratio (0.1-0.5). Through multiple iterations, the optimal combination of hyperparameters that maximizes model performance is determined. ; S34: LSTM Network Architecture Design Input layer: receiving dimension is The tensor, where n is the optimal time step determined through optimization; LSTM Hidden Layer: The first LSTM layer captures the long-range dependence of the original process parameters in the time dimension (such as the evolution trend of the roll surface morphology on the rolling kilometer), and the second layer further extracts deep nonlinear features. Fully connected layer: 1-2 fully connected network layers are connected after the LSTM layer, and the ReLU activation function is used to map the extracted spatiotemporal features to the nonlinear roughness prediction space; Output layer: The output is a scalar value, i.e., the predicted surface roughness. ; S35: Model Training and Evaluation Metrics Loss function: The mean squared error (MSE) is used as the loss function. Performance evaluation metrics: Mean absolute percentage error (MAPE) and coefficient of determination (R²) are introduced to evaluate the model; The model is considered to have the accuracy to guide PSO optimization only when R2 reaches a preset threshold (e.g., above 0.92) and MAPE is below 5%. The training strategy uses the Adam optimizer and sets an early stopping mechanism, stopping training when the validation set loss no longer decreases within 10 consecutive epochs. S36: Model Denormalization and Calling Mechanism Inverse normalization: the predicted values ​​output by the model. It needs to be restored to a roughness value with actual physical meaning through the inverse operation of normalization. (µm) to align with target metrics Perform a comparison; Online optimization call: The trained model is fixed as a function object. In the subsequent first-stage optimization: the initial PSO search stage for the elongation rate of the leveling process and the second-stage optimization: the multi-parameter joint PSO search stage, whenever the particle updates its position (generating a new parameter combination), the algorithm automatically inputs the candidate parameter combination (such as elongation rate) into the model. The model quickly calculates and predicts the roughness through forward propagation and feeds it back to the objective function Fitness.

[0012] As a further optimization of the present invention, the first stage optimization: preliminary PSO search for the elongation rate of the leveling process specifically includes: The focus is on the elongation (%), which has the most significant impact on the surface morphology during the leveling process. Other leveling process parameters (such as entry thickness, yield strength, and rolling mileage) can be regarded as fixed input values ​​at this stage. The elongation value must be within the safe range allowed by mechanical properties to avoid strip breakage or other unacceptable defects. Search range setting: Set a preliminary, relatively broad search range [Emin1, Emax1] for the elongation rate (%); the setting of this range should be based on production experience, historical data or preliminary model prediction, and should be able to cover the elongation rate range corresponding to most possible high-quality roughness. Objective function definition: Establish the objective function Fitness1 to quantify the deviation between the surface roughness and the target value; the optimization objective is to minimize this deviation. in, It is the predicted value obtained by inputting the elongation to be optimized and other currently fixed process parameters into the LSTM prediction model; PSO Search and Output: Initialize the PSO population, with each particle representing a candidate value for elongation; within the range [Emin1, Emax1], perform iterative updates of the PSO algorithm (velocity and position updates) until convergence; this stage outputs one or a set of preliminary optimal elongation values. .

[0013] As a further optimization of the present invention, the judgment and stage switching specifically include: The optimal elongation value The surface roughness value is obtained by applying it to the LSTM prediction model. Conditional judgment: With target roughness index Compare; Condition met: If the target requirement has been met (within the allowable error range), the optimization is considered successful; the output is used as the final optimization parameters, and the process ends. Output The process ends when the final optimization parameters are used. Condition not met: If If the target requirement is still not met, it indicates that relying solely on extension rate optimization is insufficient, and a more refined joint optimization stage is needed.

[0014] As a further optimization of the present invention, the second stage optimization: multi-parameter joint PSO search specifically includes: Define a joint optimization variable set X2 containing five key parameters, including: elongation. (%), alloying temperature (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%) Refined Search Range: To improve efficiency and accuracy, the search range is intelligently limited based on the results of the first stage and the current production status. Specifically, the search range for the elongation rate is limited to... in Typically 10%. And the parameters in the hot-dip galvanizing process... , , , The search range is limited to 10% above or below its current measurement value or the best value of the previous period. Objective function definition: The objective function Fitness2 evaluates the roughness deviation caused by the combination of multiple parameters; in, It is the joint prediction value obtained by inputting a five-dimensional vector, including elongation and hot-dip galvanizing process parameters, into the LSTM prediction model; PSO Search and Output: Initialize the PSO population, where each particle is a five-dimensional vector containing five parameters; within the defined joint search space, perform iterative updates of the PSO algorithm until convergence; the output of this stage is the optimal combination of the five parameters after joint optimization, denoted as . .

[0015] As a further optimization of the present invention, the final output optimization parameters specifically include: The optimal combination of the five parameters obtained from the second stage optimization. (i.e., final elongation, alloying temperature, alloying speed, and iron content on the upper and lower surfaces) are output as the final optimized control parameters.

[0016] As a further optimization of the present invention, the closed-loop control and feedback mechanism specifically includes: The optimal parameter combination is transmitted to the underlying control system (Level 2 and PLC) via the industrial network, thereby driving the actuators to perform actions, including adjusting the rolling force or tension of the leveling mill to lock in the optimal elongation, adjusting the gas flow rate and fan frequency of the alloying furnace to match the temperature and speed, and fine-tuning the power distribution of the induction heater to control the iron content; subsequently, the actual surface roughness is collected using an online measuring instrument at the leveling exit. And calculate its residual with respect to the target. If the residual exceeds the preset threshold (e.g., 0.05µm), a short-term compensation mechanism is triggered to fine-tune the parameters. At the same time, the system stores the measured data of each volume into the historical database. After accumulating to the preset increment (20 volumes), step S3 is automatically called to perform incremental training and hyperparameter re-optimization of the LSTM prediction model, thereby realizing closed-loop control and feedback of the model.

[0017] The beneficial effects of this invention are as follows: 1. This invention proposes an intelligent control method for the surface roughness of hot-dip galvanized iron strip. Combining LSTM deep learning and PSO swarm intelligence optimization algorithm, it achieves hierarchical optimization control and cross-process collaborative optimization of the surface roughness of hot-dip galvanized iron strip. It breaks through the limitations of single-process adjustment, thereby accurately controlling the surface roughness of hot-dip galvanized iron strip. Through incremental learning and real-time error correction, it can dynamically adapt to extreme working conditions, improve the stability and control accuracy of long-term operation, and thus significantly improve product quality and production efficiency, effectively solving the problems faced by hot-dip galvanized iron strip in surface roughness control technology. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall process for optimizing the surface roughness of hot-dip galvanized iron strip based on a graded strategy. Figure 2 This is a schematic diagram of the cross-sectional structure of hot-dip galvanized iron strip; Figure 3 This is a flowchart of the construction process for a surface roughness prediction model based on LSTM. Figure 4 This is a schematic diagram of the first stage of PSO optimization; Figure 5 This is a schematic diagram of the second stage of PSO optimization. Detailed Implementation

[0019] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0020] Example 1: Below, according to Figure 1The overall flowchart illustrates in detail the surface roughness optimization and control method for hot-dip galvanized steel strip based on a hierarchical strategy in this invention.

[0021] S1: Data Acquisition. Collect key process parameters during the hot-dip galvanized steel strip production process. The data precisely includes: The hot-dip galvanizing process includes: alloying temperature (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%).

[0022] The leveling process includes: strip entry thickness (mm), elongation (%), Yield strength (MPa), rolling kilometers of the work rolls (km).

[0023] Surface roughness: The surface roughness of hot-dip galvanized iron strip is measured using a roughness measuring instrument. In order to avoid uneven roughness in the width direction of the strip, three points are measured in the width direction: the working side, the operating side, and the middle. The average value is then taken as the roughness value Ra (µm) of the current strip.

[0024] According to step S1, in the production task of hot-dip galvanized iron strip for a certain automobile outer panel, the steel grade is DC54D+ZF (ultra-deep drawing low-carbon steel). Its surface roughness is formed through two key stages: hot-dip galvanizing and leveling rolling. First, during the alloying process, the zinc-iron alloy coating itself determines the initial roughness of the strip (5.0-10.0µm). At this time, the substrate roughness does not affect this value but does affect the bonding strength. Subsequently, in the leveling stage, the surface condition of the work rolls and the leveling process parameters readjust the roughness to the target range (0.5-2.0µm) and finally determine it, which will not be changed in subsequent processes. Figure 2 As shown. The data acquisition system on its production line collects data every 3 seconds. The collected data includes: hot-dip galvanizing process data, including alloying temperature. (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%), and leveling process data including: strip entry thickness (mm), elongation (%), Yield strength (MPa), rolling kilometers of the work rolls (km). Subsequently, a roughness measuring instrument was used to measure three points on the working side, operating side and the middle of the strip surface in the width direction, and then the average value was taken as the roughness value Ra (µm) of the current strip. Some data are shown in Table 1.

[0025] Table 1. On-site data (partial) S2: Setting the target surface roughness index. Based on product specifications and customer requirements, set a specific target surface roughness index. This index can be a specific value (e.g., Ra=1.2µm) or an allowable range (e.g., 1.0µm≤Ra≤1.5µm).

[0026] According to step S2, the target surface roughness value is 1.25µm as required by the customer, with an allowable deviation range of [1.20, 1.30] (µm).

[0027] S3: Construction of LSTM-based Surface Roughness Prediction Model: To provide accurate evaluation criteria for the objective function of the PSO optimization algorithm, this invention constructs a deep learning prediction model for real-time prediction of surface roughness values ​​under different combinations of process parameters. The specific steps are as follows: S31: Data Preprocessing and Feature Sequence Construction Abnormal data cleaning: The collected raw process data is processed to remove non-steady-state process points such as belt breakage and deceleration stop sections in the production process, so as to ensure that the model training set reflects the normal production process.

[0028] Normalization: Due to the huge difference in dimensions of the collected process parameters (such as alloying temperature and strip thickness), the Min-Max normalization method is used to scale all the original data to the (0,1) interval to accelerate the convergence of the neural network.

[0029] Time series sliding window construction: Considering that the wear of the work rolls (rolling kilometers) in the leveling process is a continuous degradation process, this model does not only take single-point data, but also adopts the sliding window technique. The feature sequence is composed of parameters from the current time t and the previous n historical volumes. .

[0030] Input vector definition: The feature vector x at each time step includes: alloying temperature. (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%), strip entry thickness (mm), elongation (%), Yield strength (MPa), rolling kilometers of the work rolls (km). The feature matrix X1 consists of the following parameters: S32: Dataset Partitioning and Temporal Feature Isolation To ensure the model has good generalization ability and maintains prediction accuracy on unseen production data, the preprocessed feature sequences are subjected to strict set partitioning: Chronological Sequence Division: Considering the continuity of the hot-dip galvanizing production process and the evolving characteristics of equipment conditions (such as roll wear), this invention does not employ random sampling but strictly divides the sample according to the chronological order of strip steel production. The processed sample library is divided into a training set, a validation set, and a test set, with a ratio of [ratio to be filled in]. .

[0031] Sample isolation: Ensure that the timestamp of the training set is earlier than that of the validation set, and the timestamp of the validation set is earlier than that of the test set, so as to simulate the real-world application scenario of the model "learning from history and predicting the future" in actual production.

[0032] S33: Model Adaptive Hyperparameter Optimization To prevent underfitting or overfitting of the model due to manually set parameters, a Bayesian optimization algorithm is introduced to automatically optimize the core parameters of the LSTM network. The optimization objective is to minimize the mean squared error on the validation set.

[0033] The optimization space includes the number of hidden layer neurons (32-256), the initial learning rate (0.0001-0.01), the time step n (3-15), and the Dropout ratio (0.1-0.5). Through multiple iterations, the optimal combination of hyperparameters is determined to maximize model performance. .

[0034] S34: LSTM Network Architecture Design Input layer: receiving dimension is The tensor of n is the optimal time step determined through optimization.

[0035] LSTM Hidden Layers: The first LSTM layer captures the long-range dependence of the original process parameters in the time dimension (such as the evolution trend of the roll surface morphology on the rolling kilometer), and the second layer further extracts deep nonlinear features.

[0036] Fully connected layer: 1-2 fully connected layers are connected after the LSTM layer, and the extracted spatiotemporal features are mapped to the nonlinear roughness prediction space using the ReLU activation function.

[0037] Output layer: The output is a scalar value, i.e., the predicted surface roughness. .

[0038] S35: Model Training and Evaluation Metrics Loss function: The mean squared error (MSE) is used as the loss function. Performance evaluation metrics: Mean absolute percentage error (MAPE) and coefficient of determination (R²) are introduced to evaluate the model.

[0039] The model is considered to have the accuracy to guide PSO optimization only when R² reaches a preset threshold (e.g., above 0.92) and MAPE is below 5%. The training strategy uses the Adam optimizer with an early stopping mechanism, stopping training when the validation set loss no longer decreases within 10 consecutive epochs.

[0040] S36: Model Denormalization and Calling Mechanism Inverse normalization: the predicted values ​​output by the model. It needs to be restored to a roughness value with actual physical meaning through the inverse operation of normalization. (µm) to align with target metrics Compare them.

[0041] Online optimization call: The trained model is fixed as a function object. In the subsequent PSO search in stages (D) and (F), whenever a particle updates its position (generating a new combination of parameters), the algorithm automatically inputs the candidate parameter combination (such as elongation) into the model. The model quickly calculates the predicted roughness through forward propagation and feeds it back to the objective function Fitness.

[0042] Based on step S3, construct the LSTM model, as follows: Figure 3 As shown, the model, after Bayesian optimization, determined a time step of n=5 and 128 hidden layer neurons. On the test set, it achieved a MAPE of 2.35% and a coefficient of determination R² of 0.97, demonstrating the accuracy required to guide PSO optimization.

[0043] S4: First Stage Optimization: Preliminary PSO search for elongation in the leveling process. This stage focuses on the parameter that has the most significant impact on surface morphology in the leveling process—elongation (%). Other leveling process parameters (such as entry thickness, yield strength, and rolling mileage) can be considered as fixed input values ​​at this stage. The elongation value must be within the safe range allowed by mechanical properties to avoid strip breakage or other unacceptable defects.

[0044] Search range setting: Set an initial, relatively broad search range [Emin1, Emax1] for the elongation rate (%). This range should be set based on production experience, historical data, or preliminary model estimates, and should cover the elongation rate range corresponding to the vast majority of possible high-quality roughnesses.

[0045] Objective function definition: Establish the objective function Fitness1 to quantify the deviation between the surface roughness and the target value. The optimization objective is to minimize this deviation.

[0046] in, It is the predicted value obtained by inputting the elongation to be optimized along with other fixed process parameters into the LSTM prediction model.

[0047] PSO Search and Output: Initialize the PSO population, with each particle representing a candidate elongation value. Within the range [Emin1, Emax1], perform iterative updates of the PSO algorithm (velocity and position updates) until convergence. This stage outputs one or a set of preliminary optimal elongation values. .

[0048] According to step S4, as Figure 4 As shown, according to production safety regulations, the safe range for the elongation of this strip steel specification is set to [0.6%, 1.4%]. Then, a PSO search is initiated, with alloying temperature, alloying speed, and other parameters as constant inputs. After 30 iterations, the optimal elongation output by the PSO algorithm is 1.12%. Subsequently, 1.12% is substituted into the LSTM model, and the predicted value is 1.26µm.

[0049] S5: Judgment and Stage Switching. The optimal elongation value will be determined. The surface roughness value is obtained by applying it to the LSTM prediction model.

[0050] Conditional judgment: With target roughness index Compare them.

[0051] Condition met: If the target requirement has been met (within the allowable error range), the optimization is considered successful. The output is used as the final optimization parameters, and the process ends. The process ends when the final optimization parameters are used.

[0052] Condition not met: If If the target requirement is still not met, it indicates that relying solely on extension rate optimization is insufficient, and a more refined joint optimization stage is needed.

[0053] According to step S5, the predicted value is within the target interval [1.20, 1.30] (µm), which meets the condition. Therefore, the system determines that the first stage of optimization is successful and there is no need to enter the multi-parameter joint optimization.

[0054] S6: Second-stage optimization: Multi-parameter joint PSO search. Define a joint optimization variable set X2 containing five key parameters, including: elongation. (%), alloying temperature (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%).

[0055] Refined Search Range: To improve efficiency and accuracy, the search range is intelligently limited based on the results of the first stage and the current production status. Specifically, the search range for the elongation rate is limited to... in Typically 10%. And the parameters in the hot-dip galvanizing process... , , , The search range is limited to 10% above or below its current measurement value or the best value of the previous period.

[0056] Objective function definition: The objective function Fitness2 evaluates the roughness deviation caused by the combination of multiple parameters.

[0057] in, It is the joint prediction value obtained by inputting a five-dimensional vector, including elongation and hot-dip galvanizing process parameters, into the LSTM prediction model.

[0058] PSO Search and Output: Initialize the PSO population, where each particle is a five-dimensional vector containing five parameters. Within a defined joint search space, iteratively update the PSO algorithm until convergence. The output of this stage is the optimal combination of the five parameters after joint optimization, denoted as . .

[0059] S7: Output the final optimized parameters. This is the optimal combination of the five parameters obtained from the second-stage optimization. (i.e., final elongation, alloying temperature, alloying speed, and iron content on the upper and lower surfaces) are output as the final optimized control parameters.

[0060] Because the target interval requirement is met, the system skips S6 and S7 and proceeds directly to S8.

[0061] S8: Closed-loop control and feedback mechanism. The optimal parameter combination is sent to the underlying control system (Level 2 and PLC) via the industrial network, thereby driving the actuators to perform actions, including adjusting the rolling force or tension of the leveling mill to lock in the optimal elongation, adjusting the gas flow and fan frequency of the alloying furnace to match temperature and speed, and fine-tuning the power distribution of the induction heater to control the iron content; subsequently, the actual surface roughness is collected using an online measuring instrument at the leveling exit. And calculate its residual with respect to the target. If the residual exceeds the preset threshold (e.g., 0.05µm), a short-term compensation mechanism is triggered to fine-tune the parameters. At the same time, the system stores the measured data of each volume into the historical database. After accumulating to the preset increment (20 volumes), step S3 is automatically called to perform incremental training and hyperparameter re-optimization of the LSTM prediction model, thereby realizing closed-loop control and feedback of the model.

[0062] According to step S8, the optimal elongation value of 1.12% is sent to the leveling machine control system via the underlying PLC. After the strip is produced, the average Ra value measured by the online measuring instrument is 1.27µm, with an error of only 0.02µm. The system stores this data in the database, and the production of this roll is completed.

[0063] Example 2 Below, according to Figure 1 The overall flowchart illustrates in detail the surface roughness optimization and control method for hot-dip galvanized steel strip based on a hierarchical strategy in this invention.

[0064] According to step S1, in the production task of high-strength deep-drawing duplex steel DP590+ZF on the hot-dip galvanized steel strip production line, the data acquisition system on the production line collects data every 5 seconds. The collected data includes: hot-dip galvanizing process data, including alloying temperature. (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%), and leveling process data including: strip entry thickness (mm), elongation (%), Yield strength (MPa), rolling kilometers of the work rolls (km), some data are shown in Table 2. Subsequently, a roughness measuring instrument was used to measure three points on the working side, operating side and the middle of the strip surface in the width direction, and then the average value was taken as the roughness value Ra (µm) of the current strip.

[0065] Table 2. On-site data (partial) According to step S2, based on the customer's requirement, the target surface roughness value is 1.50µm, and the allowable deviation range is... µm.

[0066] Based on step S3, construct the LSTM model, as follows: Figure 3 As shown, the model, after Bayesian optimization, determined a time step of n=6 and the number of hidden layer neurons to be 128. On the test set, the MAPE reached 1.82% and the coefficient of determination R2 reached 0.98, which is sufficient to guide PSO optimization.

[0067] According to steps S4 and S5, as Figure 4 As shown, for high-strength steel, the upper limit of elongation is limited, and the search range is set to [0.4%, 1.0%]. Then, a PSO search is initiated. When the PSO reaches the physical upper limit of elongation (1.0%), the surface roughness predicted by LSTM is 1.32µm. Since... The target was not met. Step (F) will be automatically triggered to proceed to the second phase of optimization, as follows: Figure 5 As shown.

[0068] Based on steps S6 and S7, the system automatically constructs a five-dimensional optimization vector based on the current measured production values. And refine the search boundary based on the principle of "baseline value ±10%": Extension rate: Based on the optimal value of 1.0% in the first stage, the search range is set to [0.90%, 1.10%].

[0069] Alloying temperature: The current measured value is 520℃, and the search range is set to [468℃, 572℃].

[0070] Alloying speed: Current speed 115m / min, search range set to [103.5, 126.5]m / min.

[0071] Iron content on the upper and lower surfaces: currently 9.5% and 9.6% respectively, with the search range set to [8.55%, 10.56%].

[0072] Subsequently, the PSO algorithm was iteratively optimized. Population initialization: 50 particles were randomly generated, each a five-dimensional vector. Finally, in the 45th iteration, the global optimal solution was locked at: , ℃, m / min, , Under this parameter combination, the surface roughness predicted by LSTM is 1.49µm.

[0073] According to step S8, the optimal parameter combination is transmitted to the underlying control system (Level 2 and PLC) via the industrial network, thereby driving the actuators to perform actions, including locking the maximum safe tension of the leveling machine, increasing the power of the alloying furnace heating section, and adjusting the sensor power, achieving precise control of temperature and iron content. Combined with the hot-dip galvanizing process, the surface roughness was successfully controlled at 1.51µm, verifying its technical feasibility. Furthermore, the system stores sample data of these special working conditions in a historical database and automatically performs incremental training on the LSTM model after accumulating 20 rolls to fine-tune the model weights, thereby reducing the prediction error of the later rolls and realizing the self-learning and closed-loop iteration of the control system.

[0074] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy, characterized in that: Includes the following steps: S1: Data Acquisition; S2: Setting the target surface roughness index; S3: Construction of a surface roughness prediction model based on LSTM; S4: First stage optimization: Preliminary PSO search for the elongation rate of the leveling process; S5: Judgment and Phase Switching; S6: Second-stage optimization: Multi-parameter joint PSO search; S7: Output the final optimized parameters; S8: Closed-loop control and feedback mechanism.

2. The method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy according to claim 1, characterized in that: The data collection specifically includes: The hot-dip galvanizing process includes: alloying temperature (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%) The leveling process includes: strip entry thickness (mm), elongation (%), Yield strength (MPa), rolling kilometers of the work rolls (km); Surface roughness: The surface roughness of hot-dip galvanized iron strip is measured using a roughness measuring instrument. In order to avoid uneven roughness in the width direction of the strip, three points are measured in the width direction: the working side, the operating side, and the middle. The average value is then taken as the roughness value Ra (µm) of the current strip.

3. The method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy according to claim 2, characterized in that: The target surface roughness index can be a specific value (e.g., Ra = 1.2µm) or an allowable range (e.g., 1.0µm ≤ Ra ≤ 1.5µm).

4. The method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy according to claim 3, characterized in that: The construction of the LSTM-based surface roughness prediction model specifically includes: to provide an accurate evaluation basis for the objective function of the PSO optimization algorithm, a deep learning prediction model is constructed to predict surface roughness values ​​under different combinations of process parameters in real time; the specific steps are as follows: S31: Data Preprocessing and Feature Sequence Construction Abnormal data cleaning: The collected raw process data is processed to remove abnormal process points such as belt breakage and deceleration stop sections in the production process, so as to ensure that the model training set reflects the normal production process. Normalization: Due to the huge difference in dimensions of the collected process parameters (such as alloying temperature and strip thickness), the Min-Max normalization method is used to scale all the original data to the (0,1) interval to accelerate the convergence of the neural network. Time series sliding window construction: Considering that the wear of the work rolls (rolling kilometers) in the leveling process is a continuous degradation process, this model does not only take single-point data, but also adopts the sliding window technique; the parameters of the current time t and the previous n historical volumes are selected to form the feature sequence. ; Input vector definition: The feature vector x at each time step includes: alloying temperature. (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%), strip entry thickness (mm), elongation (%), Yield strength (MPa), rolling kilometers of the work rolls (km); where the feature matrix X1 consists of the following parameters: ； S32: Dataset Partitioning and Temporal Feature Isolation To ensure the model has good generalization ability and maintains prediction accuracy on unseen production data, the preprocessed feature sequences are subjected to strict set partitioning: Time-sequence partitioning: Considering the continuity of the hot-dip galvanizing production process and the evolutionary characteristics of equipment conditions (such as roll wear), random sampling is not used. Instead, the sample library is strictly partitioned according to the chronological order of strip steel production. The processed sample library is divided into training set, validation set, and test set, with a ratio of [ratio missing]. ; Sample isolation: Ensure that the timestamp of the training set is earlier than that of the validation set, and the timestamp of the validation set is earlier than that of the test set, thereby simulating the real-world application scenario of the model "learning from history and predicting the future" in actual production. S33: Model Adaptive Hyperparameter Optimization To prevent underfitting or overfitting of the model due to manually set parameters, a Bayesian optimization algorithm is introduced to automatically optimize the core parameters of the LSTM network; the optimization objective is to minimize the mean squared error on the validation set. The optimization space includes the number of hidden layer neurons (32-256), the initial learning rate (0.0001-0.01), the time step n (3-15), and the Dropout ratio (0.1-0.5). Through multiple iterations, the optimal combination of hyperparameters that maximizes model performance is determined. ; S34: LSTM Network Architecture Design Input layer: receiving dimension is The tensor, where n is the optimal time step determined through optimization; LSTM Hidden Layer: The first LSTM layer captures the long-range dependence of the original process parameters in the time dimension (such as the evolution trend of the roll surface morphology on the rolling kilometer), and the second layer further extracts deep nonlinear features. Fully connected layer: 1-2 fully connected network layers are connected after the LSTM layer, and the ReLU activation function is used to map the extracted spatiotemporal features to the nonlinear roughness prediction space; Output layer: The output is a scalar value, i.e., the predicted surface roughness. ; S35: Model Training and Evaluation Metrics Loss function: The mean squared error (MSE) is used as the loss function. Performance evaluation metrics: Mean absolute percentage error (MAPE) and coefficient of determination (R²) are introduced to evaluate the model; The model is considered to have the accuracy to guide PSO optimization only when R2 reaches a preset threshold (e.g., above 0.92) and MAPE is below 5%. The training strategy uses the Adam optimizer and sets an early stopping mechanism, stopping training when the validation set loss no longer decreases within 10 consecutive epochs. S36: Model Denormalization and Calling Mechanism Inverse normalization: the predicted values ​​output by the model. It needs to be restored to a roughness value with actual physical meaning through the inverse operation of normalization. (µm) to align with target metrics Perform a comparison; Online optimization call: The trained model is fixed as a function object. In the subsequent first-stage optimization: the initial PSO search stage for the elongation rate of the leveling process and the second-stage optimization: the multi-parameter joint PSO search stage, whenever the particle updates its position (generating a new parameter combination), the algorithm automatically inputs the candidate parameter combination (such as elongation rate) into the model. The model quickly calculates and predicts the roughness through forward propagation and feeds it back to the objective function Fitness.

5. The method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy according to claim 4, characterized in that: The first stage of optimization: the initial PSO search for the elongation rate of the leveling process specifically includes: The focus is on the elongation (%), which has the most significant impact on the surface morphology during the leveling process. Other leveling process parameters (such as entry thickness, yield strength, and rolling mileage) can be regarded as fixed input values ​​at this stage. The elongation value must be within the safe range allowed by mechanical properties to avoid strip breakage or other unacceptable defects. Search range setting: Set a preliminary, relatively broad search range [Emin1, Emax1] for the elongation rate (%); the setting of this range should be based on production experience, historical data or preliminary model prediction, and should be able to cover the elongation rate range corresponding to most possible high-quality roughness. Objective function definition: Establish the objective function Fitness1 to quantify the deviation between the surface roughness and the target value; the optimization objective is to minimize this deviation. in, It is the predicted value obtained by inputting the elongation to be optimized and other currently fixed process parameters into the LSTM prediction model; PSO Search and Output: Initialize the PSO population, with each particle representing a candidate value for elongation; within the range [Emin1, Emax1], perform iterative updates of the PSO algorithm (velocity and position updates) until convergence; this stage outputs one or a set of preliminary optimal elongation values. .

6. The method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy according to claim 5, characterized in that: The judgment and phase switching specifically include: The optimal elongation value The surface roughness value is obtained by applying it to the LSTM prediction model. Conditional judgment: With target roughness index Compare; Conditions met: If the target requirement has been met (within the allowable error range), the optimization is considered successful; the output is used as the final optimization parameters, and the process ends; Output The process ends when the final optimization parameters are used. Condition not met: If If the target requirement is still not met, it indicates that relying solely on elongation optimization is insufficient and a more refined joint optimization stage is needed. Whether to execute S6 depends on the result of S5.

7. The method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy according to claim 6, characterized in that: The second stage of optimization: multi-parameter joint PSO search specifically includes: Define a joint optimization variable set X2 containing five key parameters, including: elongation. (%), alloying temperature (°C), alloying speed (m / min), iron content of the upper surface coating (%), Iron content of the lower surface coating (%) Refined Search Range: To improve efficiency and accuracy, the search range is intelligently limited based on the results of the first stage and the current production status. Specifically, the search range for the elongation rate is limited to... in Typically 10%; while the parameters of the hot-dip galvanizing process... , , , The search range is limited to 10% above or below its current measurement value or the best value of the previous period. Objective function definition: The objective function Fitness2 evaluates the roughness deviation caused by the combination of multiple parameters; in, It is the joint prediction value obtained by inputting a five-dimensional vector, including elongation and hot-dip galvanizing process parameters, into the LSTM prediction model; PSO Search and Output: Initialize the PSO population, where each particle is a five-dimensional vector containing five parameters; within the defined joint search space, perform iterative updates of the PSO algorithm until convergence; the output of this stage is the optimal combination of the five parameters after joint optimization, denoted as . .

8. The method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy according to claim 7, characterized in that: The final optimized output parameters specifically include: The optimal combination of the five parameters obtained from the second stage optimization. (i.e., final elongation, alloying temperature, alloying speed, and iron content on the upper and lower surfaces) are output as the final optimized control parameters.

9. The method for optimizing and controlling the surface roughness of hot-dip galvanized steel strip based on a hierarchical strategy according to claim 8, characterized in that: The closed-loop control and feedback mechanism specifically includes: The optimal parameter combination is transmitted to the underlying control system (Level 2 and PLC) via the industrial network, thereby driving the actuators to perform actions, including adjusting the rolling force or tension of the leveling mill to lock in the optimal elongation, adjusting the gas flow rate and fan frequency of the alloying furnace to match the temperature and speed, and fine-tuning the power distribution of the induction heater to control the iron content; subsequently, the actual surface roughness is collected using an online measuring instrument at the leveling exit. And calculate its residual with respect to the target. If the residual exceeds the preset threshold (e.g., 0.05µm), a short-term compensation mechanism is triggered to fine-tune the parameters. At the same time, the system stores the measured data of each volume into the historical database. After accumulating to the preset increment (20 volumes), step S3 is automatically called to perform incremental training and hyperparameter re-optimization of the LSTM prediction model, thereby realizing closed-loop control and feedback of the model.

Images