An electromagnetic stirrer parameter intelligent optimization system and method

Abstract

The present application relates to the technical field of electromagnetic stirring, and discloses an intelligent optimization system and method for electromagnetic stirrer parameters, which comprises a data acquisition and preprocessing module, a defect quantification analysis module, a physical-guided neural network reverse mapping module, a multi-objective optimization module and a control execution and closed-loop feedback module.The defect quantification analysis module uses an improved YOLOv8 semantic segmentation network to segment the defect area of a low-power tissue image, and constructs five continuous quantitative indicators, namely, porosity, crack severity index, shrinkage cavity volume fraction, center segregation proportion and white band width.The physical-guided neural network reverse mapping module embeds the metallurgical solidification physical law as a hard constraint into the neural network, and realizes accurate reverse mapping of defect features to electromagnetic stirring parameters.The multi-objective optimization module searches for the global optimal parameter combination within the candidate parameter range based on an improved NSGA-II algorithm, and combines a segmented adaptive weight and a multi-objective optimization function with physical punishment to issue the global optimal parameter combination to the PLC for closed-loop control.

Description

Technical Field

[0001] This invention relates to the technical field of electromagnetic stirring, specifically to an intelligent optimization system and method for electromagnetic stirrer parameters, and more specifically to an intelligent optimization system for electromagnetic stirrer parameters based on a low-magnification microstructure defect rating map of continuously cast steel billets. Background Technology

[0002] Low-magnification structural defects in continuously cast billets (such as central porosity, central cracks, central shrinkage cavities, and central segregation) are core factors affecting the internal quality of steel, directly determining the performance and yield of subsequent rolled products. Electromagnetic stirring is an important means to improve the internal quality of continuously cast billets, and the settings of its parameters (stirring current, stirring frequency, stirring method, and installation position) directly affect the solidification structure and defect formation of the billet.

[0003] Currently, the industry generally uses the following two methods for defect control:

[0004] Firstly, manual adjustment based on experience. Operators manually adjust the electromagnetic stirring parameters based on their personal experience and the results of low-level pickling. This method suffers from problems such as adjustment lag, low precision, and heavy reliance on personal experience, making it difficult to achieve stable quality control.

[0005] Secondly, traditional AI optimization solutions. Some existing technologies employ a separate approach of "defect identification + parameter optimization," first identifying the defect level through deep learning, and then outputting parameters through database comparison or black-box optimization algorithms. However, this type of solution has the following shortcomings:

[0006] Several technical solutions already exist for automatic low-magnification structural evaluation. For example, Chinese patent application CN202310063154.4, "Automatic Identification and Grading System for Low-Magnification Defects in Continuous Casting Billets," achieves fully automated defect grading across the entire process of machining, corrosion, photography, and image recognition. Baosteel's related patents (CN121392315A, CN121640014A) also demonstrate a focus on deep learning-based low-magnification defect recognition. However, these technologies all focus solely on defect grading itself and do not address parameter optimization or closed-loop control based on the grading results.

[0007] In terms of electromagnetic stirring parameter optimization, existing technologies are mostly model-based offline optimizations, failing to achieve closed-loop reverse mapping based on defect images. For example, dual-loop control and PID iterative algorithms are used to correct the stirring position based on temperature field and billet growth data, but the input is model calculation data rather than a low-magnification defect rating map. Chinese patent CN110625079B (Beijing University of Science and Technology) discloses an intelligent online control system for continuous casting electromagnetic stirring, which adjusts parameters by offline modeling, online solution of thermal / flow / solute states, and comparison with a target database; the feedback logic is model calculation-driven rather than defect-driven. Chinese patent CN119098566A (Baosteel) determines parameters through solidification heat transfer models, electromagnetic field simulation, and offline sampling analysis; the feedback path is offline analysis. Chinese patent CN105935752A proposes a parameter range for vertical stirring, but only a fixed process parameter range, without addressing intelligent defect analysis.

[0008] In summary, none of the existing technologies have achieved an end-to-end closed loop of "defect quantification → physical-guided AI reverse mapping → parameter optimization", nor have they embedded metallurgical physical laws into AI models to solve the black box problem. Summary of the Invention

[0009] To address the aforementioned technical problems in related technologies, this invention provides an intelligent optimization system and method for electromagnetic stirrer parameters, which can solve the above problems.

[0010] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows:

[0011] An intelligent parameter optimization system for an electromagnetic stirrer includes:

[0012] The data acquisition and preprocessing module is used to acquire low-magnification pickling images and production process data of continuous casting billets, and to preprocess the low-magnification pickling images.

[0013] The defect quantification analysis module is used to segment defect regions in the preprocessed low-magnification pickling image and calculate continuous quantification indices covering all stirring-related defects. These continuous quantification indices include: central porosity P. porosity Central crack severity index C index Central shrinkage cavity volume fraction V shrink Central segregation percentage (Sseg), white bright band width (W) band ;

[0014] The physical-guided neural network inverse mapping module is used to take the continuous quantitative index and production process data as input, embed the physical laws of metallurgical solidification as hard constraints into the neural network, and output the candidate parameter range of the electromagnetic stirrer.

[0015] A multi-objective optimization module is used to search for the globally optimal parameter combination within the candidate parameter range based on the improved NSGA-II algorithm, combined with an objective function that includes piecewise adaptive weights and physical constraint penalty terms.

[0016] The control execution and closed-loop feedback module is used to send the optimal parameter combination to the PLC control system to adjust the execution parameters, and to collect the adjusted production data and feed it back to the physical guidance neural network for incremental updates.

[0017] Furthermore, in the defect quantification analysis module:

[0018] Central loose porosity , This represents the total area of ​​the central porous region extracted from the cross-sectional image of the cast billet. The total area of ​​the central inspection zone of the billet cross section, and the central porosity. The value range is 0~10%;

[0019] Central Crack Severity Index , This represents the total length of cracks per unit area within the cross-section at the center of the cast billet. This represents the maximum width of a single crack within the cross-section. The value range is 0 to 0.4;

[0020] Central shrinkage volume fraction V shrink The value range is 0~5%;

[0021] Central segregation ratio S seg Value range: 0~8%;

[0022] White band width W band The value range is 0~10mm.

[0023] Furthermore, the objective function of the multi-objective optimization module is:

[0024]

[0025] in , , , , For industry standard-based rating Figures 1-4 Dynamic weighting coefficients for segmented adjustment of threshold levels; This is a penalty term for physical constraints. This is the penalty coefficient.

[0026] Furthermore, the physical laws of metallurgical solidification are embedded as hard constraints in the neural network, including at least one of the following constraints:

[0027] Liquid core length ratio constraint: , The thickness of the liquid core inside the cast billet at the location of the agitator installation. This represents the total metallurgical thickness of the cast billet;

[0028] Solidification rate constraint: ;

[0029] Stirring current boundary constraints: The stirring current range is 200A~600A;

[0030] Stirring frequency boundary constraints: The stirring frequency range is 2Hz~12Hz;

[0031] Bright white strip width constraint: Bright white strip width shall not exceed 5mm.

[0032] Furthermore, the backbone network of the physical-guided neural network inverse mapping module adopts a ResNet-18 feature extraction network and a fully connected inverse mapping network, and the input layer includes: central loose porosity. Central Crack Severity Index Central shrinkage cavity volume fraction Central segregation ratio White bright band width The steel grade includes carbon equivalent, cross-sectional dimensions, drawing speed, secondary cooling strength, superheat, stirring current I, stirring frequency I, stirring method I, and installation position I; the output layer includes: stirring current II, stirring frequency II, stirring method II, and installation position II.

[0033] Furthermore, the training data for the physical-guided neural network reverse mapping module includes: more than 120,000 historical production measurement data and more than 80,000 virtual sample data generated through solidification heat transfer-electromagnetic field coupling simulation.

[0034] Furthermore, the physical-guided neural network inverse mapping module incorporates categorized parameter optimization quantization rules:

[0035] Central loose porosity : For every 1% increase, the stirring current increases by 80A to 120A, and 1% < When the concentration is ≤3%, slightly increase the stirring current. When the value is greater than 3%, the upper limit of the stirring current is significantly increased;

[0036] Central Crack Severity Index : 0.1 < When the value is ≤0.3, adjust the stirring frequency to 4Hz~5Hz; When the value is greater than 0.3, the stirring frequency is forcibly locked at 3Hz, and the stirring current is simultaneously reduced.

[0037] Central shrinkage volume fraction 0.5% < When the concentration is ≤2%, the stirring position should be locked at 25% of the liquid core. When the liquid content is greater than 2%, the stirring position should still be locked at 25% of the liquid core, and the stirring current should be increased appropriately.

[0038] Central segregation percentage 1.5% < When the concentration is ≤5%, adjust the stirring frequency to 6Hz~7Hz; When the concentration is >5%, the stirring frequency is locked at 8Hz;

[0039] White band width :2mm< When the thickness is ≤5mm, slightly reduce the stirring current; When the thickness is greater than 5mm, significantly reduce the stirring current to the equipment's safe lower limit of 200A.

[0040] Furthermore, the multi-objective optimization module uses the PINN model as a surrogate model to replace traditional simulation calculations, reducing the optimization iteration time to the second level, and uses the entropy weight method to automatically select the globally optimal parameter combination from the Pareto optimal solution.

[0041] A method for intelligent optimization of electromagnetic stirrer parameters includes the following steps:

[0042] Step 1: Acquire low-magnification pickling images of continuously cast billets and synchronous process data, and perform noise reduction, contrast enhancement and segmentation preprocessing on the low-magnification pickling images;

[0043] Step 2: Segment the defect region using an improved YOLOv8 semantic segmentation network, and calculate a continuous quantitative index covering all stirring-related defects: central porosity P. porosity Central crack severity index C index Central shrinkage cavity volume fraction V shrink Central segregation ratio S seg White bright band width W band ;

[0044] Step 3: Input the continuous quantitative indicators and process data into the physical guidance neural network, and use the physical laws of metallurgical solidification as hard constraints to deduce the candidate parameter range of the electromagnetic stirrer.

[0045] Step 4: Based on the improved NSGA-II algorithm, and combined with an objective function that includes piecewise adaptive weights and physical constraint penalty terms, search for the globally optimal parameter combination within the candidate parameter range;

[0046] Step 5: Send the optimal parameter combination to the PLC control system for adjustment;

[0047] Step 6: Collect the adjusted production data and feed it back to the physical guidance neural network for incremental model updates.

[0048] Furthermore, in the objective function of step four, the segmented adaptive weight coefficients are automatically adjusted according to the 1-4 level grading thresholds of the industry standard rating chart.

[0049] The beneficial effects of this invention are:

[0050] (1) By replacing the traditional discrete level with a continuous defect quantification index, the defect is accurately characterized, providing a clear and quantified target input for parameter optimization and solving the problem of insufficient accuracy of traditional methods.

[0051] (2) The first defect-parameter inverse mapping model based on physical guided neural network (PINN) was created, which embeds the physical laws of metallurgical solidification as hard constraints into the AI ​​model, and completely solves the problem that the output parameters of the pure data-driven black box model do not meet the process feasibility.

[0052] (3) A multi-objective optimization objective function with piecewise adaptive weights and physical penalties covering all stirring-related defects is proposed. The optimization priority can be automatically adjusted according to the defect rating level, while filtering out infeasible parameters that do not meet physical constraints, thus overcoming the limitations of traditional fixed-weight linear combination.

[0053] (4) The PINN proxy model is used to replace the traditional simulation calculation. Combined with the improved NSGA-Ⅱ algorithm with physical constraints, the optimization time is compressed from hours to seconds, which greatly improves the optimization efficiency.

[0054] (5) It realizes the end-to-end closed-loop fusion of defect identification and parameter optimization, replacing the traditional separate module architecture. It achieves continuous incremental updates of the model through production data feedback, which significantly improves the accuracy of reverse mapping and optimization effect. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0056] Figure 1 This is an overall architecture block diagram of the intelligent optimization system for electromagnetic stirrer parameters described in an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the improved YOLOv8 semantic segmentation network in the defect quantification analysis module described in this embodiment of the invention;

[0058] Figure 3 This is a schematic diagram of the network structure of the Physically Guided Neural Network (PINN) reverse mapping module described in an embodiment of the present invention;

[0059] Figure 4 This is a schematic diagram of the optimization process of the improved NSGA-II algorithm in the multi-objective optimization module of the present invention.

[0060] Figure 5 These are the key technical parameter charts described in the embodiments of the present invention. Detailed Implementation

[0061] 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 some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0062] like Figure 1-5 As shown, the intelligent optimization system for electromagnetic stirrer parameters based on the low-magnification microstructure defect rating map of continuously cast steel billets provided by the present invention includes a data acquisition and preprocessing module, a defect quantification analysis module, a physical-guided neural network reverse mapping module, a multi-objective optimization module, and a control execution and closed-loop feedback module.

[0063] Example 1:

[0064] The data acquisition and preprocessing module of this application acquires cross-sectional images of continuously cast billets after pickling using an industrial camera with a resolution of no less than 2048×2048 pixels. Simultaneously, it acquires process parameters such as steel composition, cross-sectional dimensions, casting speed, secondary cooling strength, and superheat from the production control system. The image preprocessing workflow is as follows: first, noise reduction is performed using a 3×3 Gaussian filter kernel; then, histogram equalization is used to enhance the contrast between the defect and background regions; finally, the Otsu adaptive threshold segmentation algorithm is used to achieve preliminary separation of the background and defect regions.

[0065] Example 2:

[0066] The defect quantification analysis module of this application receives the preprocessed image and uses an improved YOLOv8 semantic segmentation network to accurately segment the defect region. This network, based on the standard YOLOv8, is customized and optimized for the characteristics of low-magnification defects in continuous casting: a multi-scale attention mechanism is added to the feature fusion layer to improve the detection sensitivity for small-sized cracks and porosity defects. After segmentation, the following five quantification indicators are extracted:

[0067] Central loose porosity P porosity = Sdefect / S total S defect S represents the total area of ​​the central porous region extracted from the cross-sectional image of the cast billet. total The total area of ​​the detection zone at the center of the cross-section of the billet is 0-10%, which allows for continuous quantification of the severity of porosity.

[0068] Central Crack Severity Index C index =ρ crack ×W max , where ρ crack W represents the total length of cracks per unit area (crack length density) within the cross-section of the billet center. max This is the maximum width of a single crack within the cross-section; the value of this index ranges from 0 to 0.4.

[0069] Central shrinkage volume fraction V shrink The volume percentage of shrinkage cavities is estimated based on the area of ​​shrinkage cavities in the cross-section of the billet center and the statistical law along the length of the continuously cast billet, with a value ranging from 0 to 5%.

[0070] Central segregation ratio S seg The ratio of the area of ​​the solute enrichment center segregation region obtained by image segmentation within the cross-section of the billet to the total area of ​​the detection region at the center of the cross-section, with a value range of 0~8%, is used to characterize the severity of the center component segregation.

[0071] White band width W band The average width of the negative biased bright band region extracted from the image segmentation along the radial direction of the billet thickness is 0~10mm, which is used to characterize the severity of the side effects caused by electromagnetic stirring.

[0072] Based on the above quantitative indicators, a multi-objective optimization objective function is constructed:

[0073]

[0074] in For the central loose weight, For the center crack weight, For hole shrinkage weight, For center-based partial weighting, Weighting white highlights, in the physical constraint penalty term This is a penalty term for physical constraints, used to penalize candidate parameters that do not satisfy the physical laws of metallurgy. The penalty coefficient is used to balance the priority of defect optimization and constraint satisfaction. All weight coefficients are dynamically adjusted in segments based on the quantified value of the defect and the 1-4 level thresholds in the industry standard rating chart, covering all defects related to electromagnetic stirring. When a defect is in a level 1 minor state, the weight is low, and other defects are optimized first; when a defect escalates to a level 3 or 4 severe state, the weight increases exponentially, forcing the optimizer to prioritize solving the most serious quality problems, perfectly adapting to different defect scenarios and overcoming the limitations of traditional fixed weights.

[0075] Example 3:

[0076] This application constructs a Physically Guided Neural Network (PINN) reverse mapping module, which embeds the physical laws of metallurgical solidification as hard constraints into the neural network to achieve accurate reverse mapping from defect quantification features to electromagnetic stirring parameters, thus completely solving the defects of pure data-driven black box models.

[0077] (1) Model input

[0078] The input layer contains 14-dimensional features: porosity Crack Severity Index Volume fraction of shrinkage cavity Central segregation ratio White bright band width Steel grade, carbon equivalent, cross-sectional dimensions, drawing speed, secondary cooling strength, superheat, stirring current, stirring frequency, stirring method, and installation location.

[0079] (2) Embedding of physical hard constraints (mandatory constraints, not soft regularization)

[0080] This application incorporates the core physical laws of metallurgical solidification as mandatory constraints of the loss function into the training process of the neural network, ensuring that the parameters output by the model satisfy the physical laws:

[0081] Core length constraint: The proportion of the core length corresponding to the electromagnetic stirring position must meet the following requirements. This ensures that the stirring action occurs at the solidification front;

[0082] Solidification rate constraint: The solidification rate at the stirring position must meet the following requirements. This ensures that the stirring can effectively act on the paste-like area;

[0083] Parameter boundary constraints: stirring current range 200A~600A, frequency range 2Hz~12Hz, to meet the physical boundaries of equipment operation;

[0084] White band constraint: The stirring intensity should not be too high, and the width of the white band should not exceed 5mm to avoid side effects caused by stirring.

[0085] (3) Network structure

[0086] The backbone network uses ResNet-18 feature extraction + fully connected inverse mapping network:

[0087] The first 5 layers are feature extraction layers, which perform high-order feature extraction on the input defect and process features;

[0088] The middle three layers are reverse mapping layers, which learn the inverse mapping relationship from defect features to stirring parameters;

[0089] The output layer directly predicts the optimal parameters in four dimensions: stirring current, stirring frequency, stirring method, and installation location.

[0090] (4) Training data construction

[0091] The training dataset contains two classes of data to achieve high-precision training with a small sample size:

[0092] Historical production measurement data: 120,000 samples of continuous casting production data from more than 5 years;

[0093] Coupled simulation data: 80,000 virtual samples generated through solidification heat transfer-electromagnetic field coupled simulation, covering extreme process scenarios and making up for the lack of measured data.

[0094] (5) Classification parameter optimization quantization rules

[0095] For different defect types, the model incorporates precise quantification adjustment rules, solving the problem of existing technologies that only provide directional principles.

[0096] Loose center ( For every 1% increase in porosity, the stirring current increases by 80A to 120A, which can reduce porosity by 1.0% to 1.5%.

[0097] Central Crack Severity Index : 0.1 < When the value is ≤0.3, adjust the stirring frequency to 4Hz~5Hz; When the value is greater than 0.3, the stirring frequency is forcibly locked at 3Hz, and the stirring current is simultaneously reduced.

[0098] Central shrinkage volume fraction 0.5% < When the concentration is ≤2%, the stirring position should be locked at 25% of the liquid core. When the liquid content is greater than 2%, the stirring position should still be locked at 25% of the liquid core, and the stirring current should be increased appropriately.

[0099] Central segregation percentage 1.5% < When the concentration is ≤5%, adjust the stirring frequency to 6Hz~7Hz; When the concentration is >5%, the stirring frequency is locked at 8Hz;

[0100] Bright white band width: 2mm < When the thickness is ≤5mm, slightly reduce the stirring current; When the thickness is greater than 5mm, significantly reduce the stirring current to the equipment's safe lower limit of 200A.

[0101] Example 4:

[0102] This application's multi-objective optimization module is based on an improved NSGA-II algorithm, searching for the globally optimal parameter combination within a range of candidate parameters. This module uses the PINN model as a fast surrogate model, with the objective function... As shown in the aforementioned formula, the optimization algorithm flow is as follows: population initialization → physical constraint filtering → non-dominated sorting → selection / crossover / mutation → Pareto solution set → entropy weight method to select the optimal solution.

[0103] The control execution and closed-loop feedback module sends optimal parameters to the PLC control system via industrial Ethernet, and the PLC system automatically adjusts the parameters of the electromagnetic stirrer. Simultaneously, the module collects adjusted production data in real time (including new low-magnification images and process parameters) and periodically feeds it back to the PINN model for incremental learning, enabling continuous model optimization.

[0104] Example 5:

[0105] Taking the production of GCr15 bearing steel continuously cast square billets at a steel plant as an example, the system of the present invention was verified through practical application. The cross-sectional dimensions of the billet are 250mm×250mm. Under the original production process, the overall pass rate of low-magnification defects is only 75%, with prominent problems of central porosity, central segregation, and central cracks, as well as a slight white bright band.

[0106] Original process conditions: Electromagnetic stirring parameters were 350A current, 5Hz frequency, and 18% liquid core ratio at the stirring position. The corresponding defect conditions were: porosity 4.2% (corresponding to standard rating level 4, extremely severe), crack severity index 0.35 (corresponding to level 4, extremely severe), shrinkage cavity volume fraction 1.2% (corresponding to level 3, severe), center segregation ratio 4.5% (corresponding to level 3, severe), and white bright band width 3mm (corresponding to level 2, moderate).

[0107] The system of this invention is used for optimization: the low-magnification image of the original process and the process data are input into the system, the defect quantification analysis module outputs the above five quantitative indicators; the physical-guided neural network reverse mapping module combines process conditions and physical constraints to deduce the candidate parameter range (current 350~550A, frequency 5~10Hz, liquid core ratio 22%~28%); the multi-objective optimization module searches within the candidate range, and the optimal parameter combination output is: current 450A, frequency 8Hz, and liquid core ratio of 25% at the stirring position.

[0108] Optimized Results: Low-magnification inspection results of the production batch after parameter adjustment show that the porosity decreased to 0.7% (corresponding to Grade 1, slight), the crack severity index decreased to 0.04 (corresponding to Grade 1, slight), the shrinkage volume fraction decreased to 0.2% (corresponding to Grade 1, slight), the center segregation ratio decreased to 1.2% (corresponding to Grade 1, slight), and the width of the bright white band decreased to 1.8mm (corresponding to Grade 1, slight). The overall low-magnification defect pass rate increased from 75% to 93.2%.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent optimization system for electromagnetic stirrer parameters, characterized in that, include: The data acquisition and preprocessing module is used to acquire low-magnification pickling images and production process data of continuous casting billets, and to preprocess the low-magnification pickling images. The defect quantification analysis module is used to segment defect regions in the preprocessed low-magnification pickling image and calculate continuous quantification indices covering all stirring-related defects. These continuous quantification indices include: central porosity P. porosity Central crack severity index C index Central shrinkage cavity volume fraction V shrink Central segregation percentage (Sseg), white bright band width (W) band ; The physical-guided neural network inverse mapping module is used to take the continuous quantitative index and production process data as input, embed the physical laws of metallurgical solidification as hard constraints into the neural network, and output the candidate parameter range of the electromagnetic stirrer. A multi-objective optimization module is used to search for the globally optimal combination of parameters within the candidate parameter range based on the improved NSGA-II algorithm, combined with an objective function that includes piecewise adaptive weights and physical constraint penalty terms. The control execution and closed-loop feedback module is used to send the optimal parameter combination to the PLC control system to adjust the execution parameters, and to collect the adjusted production data and feed it back to the physical guidance neural network for incremental updates.

2. The intelligent parameter optimization system for an electromagnetic stirrer according to claim 1, characterized in that, In the defect quantification analysis module: Central loose porosity , This represents the total area of ​​the central porous region extracted from the cross-sectional image of the cast billet. The total area of ​​the central inspection zone of the billet cross section, and the central porosity. The value range is 0~10%; Central Crack Severity Index , This represents the total length of cracks per unit area within the cross-section at the center of the cast billet. This represents the maximum width of a single crack within the cross-section. The value range is 0 to 0.4; Central shrinkage volume fraction V shrink The value range is 0~5%; Central segregation ratio S seg Value range: 0~8%; White band width W band The value range is 0~10mm.

3. The intelligent parameter optimization system for an electromagnetic stirrer according to claim 1, characterized in that, The objective function of the multi-objective optimization module is: in , , , , The dynamic weighting coefficients are adjusted segmentally based on the industry standard rating chart's grading thresholds for levels 1-4. This is a penalty term for physical constraints. This is the penalty coefficient.

4. The intelligent parameter optimization system for an electromagnetic stirrer according to claim 1, characterized in that, The physical laws of metallurgical solidification are embedded as hard constraints in the neural network, including at least one of the following constraints: Liquid core length ratio constraint: , The thickness of the liquid core inside the cast billet at the location of the agitator installation. This represents the total metallurgical thickness of the cast billet; Solidification rate constraint: ; Stirring current boundary constraints: The stirring current range is 200A~600A; Stirring frequency boundary constraints: The stirring frequency range is 2Hz~12Hz; Bright white strip width constraint: Bright white strip width shall not exceed 5mm.

5. The intelligent parameter optimization system for an electromagnetic stirrer according to claim 1, characterized in that, The backbone network of the physical-guided neural network inverse mapping module adopts a ResNet-18 feature extraction network and a fully connected inverse mapping network. The input layer includes: central loose porosity. Central Crack Severity Index Central shrinkage cavity volume fraction Central segregation ratio White bright band width The steel grade includes carbon equivalent, cross-sectional dimensions, drawing speed, secondary cooling strength, superheat, stirring current I, stirring frequency I, stirring method I, and installation position I; the output layer includes: stirring current II, stirring frequency II, stirring method II, and installation position II.

6. The intelligent parameter optimization system for an electromagnetic stirrer according to claim 1, characterized in that, The training data for the physical-guided neural network reverse mapping module includes: more than 120,000 historical production measurement data and more than 80,000 virtual sample data generated through solidification heat transfer-electromagnetic field coupling simulation.

7. The intelligent parameter optimization system for an electromagnetic stirrer according to claim 1, characterized in that, The physical-guided neural network inverse mapping module incorporates categorical parameter optimization quantization rules: Central loose porosity : For every 1% increase, the stirring current increases by 80A to 120A, and 1% < When the concentration is ≤3%, slightly increase the stirring current. When the value is greater than 3%, the upper limit of the stirring current is significantly increased; Central Crack Severity Index : 0.1 < When the value is ≤0.3, adjust the stirring frequency to 4Hz~5Hz; When the value is greater than 0.3, the stirring frequency is forcibly locked at 3Hz, and the stirring current is simultaneously reduced. Central shrinkage volume fraction 0.5% < When the concentration is ≤2%, the stirring position should be locked at 25% of the liquid core. When the liquid content is greater than 2%, the stirring position should still be locked at 25% of the liquid core, and the stirring current should be increased appropriately. Central segregation percentage 1.5% < When the concentration is ≤5%, adjust the stirring frequency to 6Hz~7Hz; When the concentration is >5%, the stirring frequency is locked at 8Hz; White band width :2mm< When the thickness is ≤5mm, slightly reduce the stirring current; When the thickness is greater than 5mm, significantly reduce the stirring current to the equipment's safe lower limit of 200A.

8. The intelligent parameter optimization system for an electromagnetic stirrer according to claim 1, characterized in that, The multi-objective optimization module uses the PINN model as a surrogate model to replace traditional simulation calculations, reducing the optimization iteration time to the second level, and uses the entropy weight method to automatically select the globally optimal parameter combination from the Pareto optimal solution.

9. A method for intelligent optimization of electromagnetic stirrer parameters, characterized in that, Includes the following steps: Step 1: Acquire low-magnification pickling images of continuously cast billets and synchronous process data, and perform noise reduction, contrast enhancement and segmentation preprocessing on the low-magnification pickling images; Step 2: Segment the defect region using an improved YOLOv8 semantic segmentation network, and calculate a continuous quantitative index covering all stirring-related defects: central porosity P. porosity Central crack severity index C index Central shrinkage cavity volume fraction V shrink Central segregation ratio S seg White bright band width W band ; Step 3: Input the continuous quantitative indicators and process data into the physical guidance neural network, and use the physical laws of metallurgical solidification as hard constraints to deduce the candidate parameter range of the electromagnetic stirrer. Step 4: Based on the improved NSGA-II algorithm, and combined with an objective function that includes piecewise adaptive weights and physical constraint penalty terms, search for the globally optimal parameter combination within the candidate parameter range; Step 5: Send the optimal parameter combination to the PLC control system for adjustment; Step 6: Collect the adjusted production data and feed it back to the physical guidance neural network for incremental model updates.

10. The intelligent optimization method for electromagnetic stirrer parameters according to claim 9, characterized in that, In the objective function of step four, the segmented adaptive weight coefficients are automatically adjusted according to the grading thresholds of levels 1 to 4 in the industry standard rating chart.

Images