A method for precisely controlling thickness difference of float glass

Abstract

The present application relates to the technical field of float glass preparation, and specifically discloses a method for precisely controlling the thickness difference of float glass, comprising the following steps: S10, constructing a multi-layer composite heat preservation structure outside the water tank in the high-temperature zone of the tin tank of a float glass production line; S20, collecting the process parameters of the tin tank in real time, including the temperature field distribution, pressure parameters and the geometric parameters of the glass ribbon; S30, processing the process parameters using a field-adaptive deep learning algorithm to generate a control signal for the edge puller; S40, adjusting the angle, speed, strength and position parameters of the edge puller according to the control signal; S50, repeating steps S20 to S40 to form a closed-loop control cycle, and when the thickness difference of the glass is stabilized within the required range, entering a maintenance mode to reduce the calculation load. A stable temperature environment is provided by the heat preservation structure, the algorithm fully learns the special rules of glass forming, the glass process characteristics are deeply integrated into the control system design, and the precise control of the thickness difference is achieved.

Description

Technical Field

[0001] This invention relates to the field of float glass manufacturing technology, and in particular to a method for precise control of thickness difference in float glass. Background Technology

[0002] In the context of fierce global manufacturing competition and rapid technological iteration, float glass, as an indispensable basic material for multiple industries such as construction, automobiles, and electronics, directly affects the performance and cost of downstream products due to its quality stability and production efficiency. With the market's ever-increasing demands for float glass quality, especially the stringent standards for glass thickness uniformity in high-end applications, controlling thickness variation under current technological conditions has become a key factor restricting product competitiveness.

[0003] Currently, the control of thickness variation in float glass mainly relies on traditional automated control methods (such as PID control) and process optimization based on empirical models. These methods maintain production stability to a certain extent, but their limitations are becoming increasingly apparent as market demands for glass quality continue to rise, especially in the production of ultra-thin, ultra-thick, and specialty glass.

[0004] Traditional PID controllers rely on linearized models and fixed parameters, making it difficult to effectively handle the complex nonlinear, time-varying, and large hysteresis characteristics within the molten glass bath. For example, the viscosity and thermal conductivity of molten glass change drastically with temperature, causing the performance of controllers with fixed parameters to degrade under fluctuating operating conditions. This makes it difficult to stably control the thickness difference within the high-standard range of less than 0.05mm, failing to meet the demands of high-end applications.

[0005] In recent years, although some studies have attempted to introduce intelligent algorithms such as deep learning into process control, these algorithms are mostly general models that have failed to effectively integrate with the specific physical mechanisms of float glass forming (such as heat conduction, fluid dynamics, and glass transition). This "black box" application results in poor model interpretability, and the control strategy may violate physical laws under operating conditions outside the scope of training data, raising questions about generalization ability and reliability. Summary of the Invention

[0006] To address the problems of insufficient precision, poor adaptability, and inadequate integration of intelligent algorithms with the process in existing float glass thickness difference control technologies, this invention provides a method for precise control of float glass thickness difference.

[0007] The technical solution adopted in this invention is:

[0008] A method for precisely controlling the thickness difference of float glass includes the following steps:

[0009] S10. Construct a multi-layer composite insulation structure outside the water bath in the high-temperature zone of the tin bath in the float glass production line;

[0010] S20. Real-time acquisition of tin bath process parameters, including temperature field distribution, pressure parameters, and glass ribbon geometric parameters;

[0011] S30. Use a domain-adaptive deep learning algorithm to process the process parameters and generate control signals for the edge-pulling machine;

[0012] S40. Adjust the angle, speed, force, and position parameters of the edge-pulling machine according to the control signal;

[0013] S50. Repeat steps S20 to S40 to form a closed-loop control cycle. When the thickness difference of the glass stabilizes within the required range, enter the maintenance mode to reduce the computational load.

[0014] Preferably, in step S1, the multi-layer composite insulation structure includes a high-silica rope layer, a high-temperature resistant mud layer, and a stainless steel sleeve layer arranged sequentially from the inside to the outside.

[0015] Constructing a multi-layer composite insulation structure includes the following steps:

[0016] S11. A high-silica rope layer is formed by winding a high-silica rope around the surface of the water-filled package. The winding density is 5 to 8 turns per centimeter, and the winding thickness is 5 to 10 mm. The high-silica rope layer is doped with nano-zirconia particles at a doping ratio of 4% to 5% by weight.

[0017] S12. Apply high-temperature resistant clay to the outside of the high-silica rope layer to form a high-temperature resistant clay layer with a coating thickness of 3-5 mm. The high-temperature resistant clay includes aluminum silicate fiber and binder. Among them, silicon carbide fiber is added to the high-temperature resistant clay layer at a weight ratio of 2%-3%.

[0018] S13. A stainless steel sleeve is installed on the outside of the high-temperature resistant mud layer to form a stainless steel sleeve layer, and then fastened with bolts.

[0019] Preferably, in step S20, collecting process parameters includes the following steps:

[0020] S21. Set the initial multi-scale sampling strategy: Use the first sampling frequency for the molten tin temperature. A second sampling frequency is used to address the thickness difference. The third sampling frequency is used for the edge-pulling machine parameters. ,in, ;

[0021] S22. Real-time calculation of the characteristic time constant of the current molten glass. The calculation formula is: ,in, Current temperature The viscosity of the molten glass. Let be the elastic modulus of glass, where The preset threshold;

[0022] S23, The characteristic time constant With reference time constant If a comparison is made, This triggers a sampling frequency update based on the characteristic time constant. Recalculate and update the first sampling frequency Second sampling frequency and the third sampling frequency ;

[0023] S24. Use the updated first sampling frequency. Second sampling frequency and the third sampling frequency Perform multi-scale sampling.

[0024] Preferably, in step S30, the domain-adaptive deep learning algorithm is a hybrid neural network structure, including a physical guidance encoder, a spatiotemporal attention module, and a control policy generator:

[0025] Physically guided encoders are used to map temperature field features, flow features, and geometric features to a feature space;

[0026] The spatiotemporal attention module is used to calculate the attention weights for different regions;

[0027] The control strategy generator is used to output the four-dimensional control parameters of the edge-pulling machine.

[0028] Preferably, in step S30, the physical guidance encoder is trained using a constraint loss function, the formula of which is: ,in, and These are the weighting coefficients. For time, For the temperature field, The thermal diffusivity of glass, For the Laplace operator, This is the loss term for data fitting.

[0029] Preferably, in step S30, the operation of the spatiotemporal attention module includes the following steps:

[0030] S31. Real-time calculation of flow stability indices for each region ,in, For a region index, it represents the first... One region;

[0031] S32. Based on the flow stability index of each region Calculate the attention weights for each region The formula is ,in, For activation function, The transformation function is based on the relaxation time. For the total number of regions, For a region index, it represents the first... One region;

[0032] S33, When the maximum weight change At that time, the attention weights for all regions are recalculated and updated, where, This is a preset threshold.

[0033] Preferably, the control strategy generator employs a constrained optimization algorithm, as shown in the formula: , , ,in, For the actual output, To predict the output, Temperature of the molding zone The lowest molding temperature, The highest molding temperature, Viscosity, Minimum molding viscosity, This is the highest molding viscosity.

[0034] Preferably, step S30 further includes an intelligent annotation method for training data, comprising the following steps:

[0035] Different labeling strategies are used depending on the production stage;

[0036] Label confidence level ,in, This represents the average value of the thickness difference. The standard deviation of the thickness difference;

[0037] when Manual review will be initiated at that time, among which, This is the minimum confidence threshold.

[0038] Preferably, in step S40, the method for adjusting the parameters of the edge-pulling machine includes the following steps:

[0039] S41. Obtain the current relaxation time of the glass. ;

[0040] S42. Adopt a gradual adjustment strategy. ,in, For a moment The parameter adjustment amount, The maximum allowable adjustment amount, To adjust the duration of the process;

[0041] S43, When the relaxation time Changes exceeded At that time, the adjustment strategy parameters are recalculated and updated, where, This is the threshold for the rate of change.

[0042] Preferably, in step S50, the closed-loop control includes the following steps:

[0043] S51. Calculate the current thickness difference fluctuation range. ;

[0044] S52, Dynamically Adjust Control Cycle The formula is ,in, The characteristic time constant of the control system, For target thickness;

[0045] S53, when If the temperature continues to rise, activate emergency control mode.

[0046] The beneficial effects of this invention are:

[0047] The thermal insulation structure provides a stable temperature environment and reduces external interference; the algorithm fully learns the special laws of glass forming and achieves precise control of thickness difference by deeply integrating the glass process characteristics into the control system design; closed-loop control ensures long-term operational stability, and the systematic solution overcomes the limitations of traditional methods, significantly improving control accuracy and adaptability. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the method for precise control of thickness difference in float glass in an embodiment of the present invention.

[0049] Figure 2 This is a schematic diagram of the process for constructing a multi-layer composite insulation structure in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the process parameters acquisition in an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram illustrating the operation of the spatiotemporal attention module in an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram illustrating the process of adjusting the parameters of the edge-pulling machine in an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of the closed-loop control process in an embodiment of the present invention. Detailed Implementation

[0054] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0055] Example

[0056] Existing methods for controlling the thickness difference of float glass have algorithms that are too general and lack consideration for the specific characteristics of the glass forming process, resulting in insufficient control accuracy and poor adaptability. Therefore, this embodiment provides a precise control method for the thickness difference of float glass, comprising the following steps:

[0057] S10. Construct a multi-layer composite insulation structure outside the water bath in the high-temperature zone of the tin bath in the float glass production line;

[0058] S20. Real-time acquisition of tin bath process parameters, including temperature field distribution, pressure parameters, and glass ribbon geometric parameters;

[0059] S30. Use a domain-adaptive deep learning algorithm to process the process parameters and generate control signals for the edge-pulling machine;

[0060] S40. Adjust the angle, speed, force, and position parameters of the edge-pulling machine according to the control signal;

[0061] S50. Repeat steps S20 to S40 to form a closed-loop control cycle. When the thickness difference of the glass stabilizes within the required range, enter the maintenance mode to reduce the computational load.

[0062] This embodiment provides a stable process temperature environment for control by constructing a multi-layer composite insulation structure, reducing external interference; it acquires system status by collecting key process parameters in real time; it uses a deep learning algorithm specifically designed for float glass processing for intelligent decision-making, deeply integrating glass process characteristics into the control system design to achieve precise control of thickness differences; it achieves precise control by accurately adjusting the edge-pulling machine parameters; and finally, it forms a closed-loop system for continuous optimization, ensuring long-term operational stability. Each step is interconnected, forming a complete technical closed loop, overcoming the limitations of traditional methods and significantly improving control accuracy and adaptability.

[0063] To improve the consistency and reliability of the moisture retention effect of the multi-layer composite insulation structure, in one embodiment, in step S1, the multi-layer composite insulation structure includes a high-silica rope layer, a high-temperature resistant mud layer and a stainless steel sleeve layer arranged sequentially from the inside to the outside.

[0064] Constructing a multi-layer composite insulation structure includes the following steps:

[0065] S11. A high-silica rope layer is formed by winding a high-silica rope around the surface of the water-filled package. The winding density is 5 to 8 turns per centimeter, and the winding thickness is 5 to 10 mm. The high-silica rope layer is doped with nano-zirconia particles at a doping ratio of 4% to 5% by weight.

[0066] S12. Apply high-temperature resistant clay to the outside of the high-silica rope layer to form a high-temperature resistant clay layer with a coating thickness of 3-5 mm. The high-temperature resistant clay includes aluminum silicate fiber and binder. Among them, silicon carbide fiber is added to the high-temperature resistant clay layer at a weight ratio of 2%-3%.

[0067] S13. A stainless steel sleeve is installed on the outside of the high-temperature resistant mud layer to form a stainless steel sleeve layer, and then fastened with bolts.

[0068] High-density winding and uniform coating ensure the continuity of the insulation layer, while the stainless steel sleeve provides mechanical protection. By balancing insulation and durability through winding density and thickness, temperature uniformity is enhanced, indirectly reducing fluctuations in thickness.

[0069] Traditional fixed-frequency sampling strategies cannot adapt to the dynamic characteristics of glass rheological properties changing with temperature, resulting in sampling data that cannot accurately reflect the process status. Therefore, in one embodiment, step S20 of collecting process parameters includes the following steps:

[0070] S21. Set the initial multi-scale sampling strategy: Use the first sampling frequency for the molten tin temperature. A second sampling frequency is used to address the thickness difference. The third sampling frequency is used for the edge-pulling machine parameters. ,in, ;

[0071] S22. Real-time calculation of the characteristic time constant of the current molten glass. The calculation formula is: ,in, Current temperature Viscosity of molten glass (in Pa·s). This is the elastic modulus of glass (in Pa).

[0072] S23, The characteristic time constant With reference time constant If a comparison is made, This triggers a sampling frequency update based on the characteristic time constant. Recalculate and update the first sampling frequency Second sampling frequency and the third sampling frequency ,in, The preset threshold;

[0073] S24. Use the updated first sampling frequency. Second sampling frequency and the third sampling frequency Perform multi-scale sampling.

[0074] Based on the principles of glass rheology, the sampling frequency is dynamically adjusted by calculating the characteristic time constant in real time. When the glass properties change, the system automatically recalibrates the sampling strategy to ensure that the sampling interval matches the material's dynamic response characteristics. High-frequency sampling captures rapidly changing temperature parameters, mid-frequency sampling matches the inertial characteristics of thickness changes, and low-frequency sampling adapts to the response speed of the mechanical system. This adaptive sampling strategy can intelligently adjust the data acquisition frequency according to the actual rheological properties of the glass, avoiding resource waste caused by oversampling and information loss caused by undersampling. By correlating the sampling strategy with the intrinsic properties of the material, the efficiency and accuracy of data acquisition are ensured, providing a higher quality data foundation for subsequent control algorithms.

[0075] Deep learning algorithms may suffer from a "black box" problem in float glass control, failing to effectively utilize prior physical knowledge of glass forming, leading to poor model interpretability and unstable control performance. Therefore, in one embodiment, in step S30, the domain-adaptive deep learning algorithm is a hybrid neural network structure, including a physical guidance encoder, a spatiotemporal attention module, and a control policy generator.

[0076] Physically guided encoders are used to map temperature field features, flow features, and geometric features to a feature space;

[0077] The spatiotemporal attention module is used to calculate the attention weights for different regions;

[0078] The control strategy generator is used to output the four-dimensional control parameters of the edge-pulling machine.

[0079] By designing a specialized hybrid network structure, a physical model is combined with a data-driven approach. A physical-guided encoder ensures the learning process conforms to physical laws, a spatiotemporal attention module focuses on key regions, and a control strategy generator synthesizes various information to generate the optimal control strategy. These three modules work collaboratively to achieve intelligent decision-making. This hybrid architecture effectively combines the reliability of the physical model with the adaptability of the data-driven model, maintaining the powerful fitting ability of deep learning while possessing good physical interpretability. It is particularly suitable for industrial processes with well-defined physical mechanisms, such as float glass manufacturing, significantly improving the reliability and stability of control.

[0080] Purely data-driven deep learning models may learn feature representations that violate physical laws during training, potentially leading to poor performance outside the training data distribution and a lack of extrapolation ability. Therefore, in one embodiment, in step S30, the physical guidance encoder is trained using a constraint loss function, as shown in the formula... ,in, These are the weighting coefficients for the physical constraint terms. These are the weighting coefficients of the data fitting term. Time (in seconds). This represents the temperature field (unit: °C). is the thermal diffusivity of glass (unit: m² / s). For the Laplace operator (unit: m⁻²). This is the loss term for data fitting.

[0081] A physical constraint term is introduced into the loss function to force the network to learn feature representations that conform to the laws of heat conduction. By adjusting the weight coefficients, the requirements of physical constraints and data fitting are balanced, ensuring that the model both respects physical laws and can learn the specificities of actual working conditions from the data.

[0082] The introduction of physical constraints significantly improves the model's generalization ability and extrapolation performance, ensuring that model predictions remain physically plausible even in conditions not covered by training data. It also acts as a regularization mechanism, preventing overfitting and enhancing robustness under complex conditions.

[0083] Traditional attention mechanisms in float glass control applications rely solely on statistical data features, failing to effectively utilize process knowledge regarding glass flow stability. This may lead to a mismatch between attention allocation and actual process importance. Therefore, in one embodiment, the spatiotemporal attention module operates in step S30 by including the following steps:

[0084] S31. Real-time calculation of flow stability indices for each region ,in, For a region index, it represents the first... One region;

[0085] S32. Based on the flow stability index of each region Calculate the attention weights for each region The formula is ,in, For activation function, The transformation function is based on the relaxation time. For the total number of regions, For a region index, it represents the first... One region;

[0086] S33, When the maximum weight change At that time, the attention weights for all regions are recalculated and updated, where, This is a preset threshold.

[0087] By introducing a flow stability index, process knowledge is incorporated into attention calculations, allocating more attention resources to regions with poor stability. A weight change threshold is set, and the attention allocation strategy is automatically updated when operating conditions change significantly.

[0088] The dynamic attention allocation mechanism based on process knowledge significantly improves the practicality and effectiveness of the algorithm, ensuring that algorithm resources are concentrated in the areas that require the most attention, thereby improving the targeting and efficiency of control. In particular, when operating conditions fluctuate, it can quickly focus on key problem areas and enhance the system's emergency response capability.

[0089] To avoid the control algorithm generating control strategies that exceed the glass forming process window during optimization, potentially leading to product quality issues or even production failures, in one embodiment, the control strategy generator employs a constrained optimization algorithm, as shown in the formula: , , ,in, For the actual output, To predict the output, Temperature of the molding zone (in °C). This is the minimum molding temperature (in °C). The maximum molding temperature (in °C). Viscosity (in Pa·s) This is the lowest molding viscosity (in Pa·s). This is the highest molding viscosity (in Pa·s).

[0090] By introducing process constraints as hard constraints into the optimization objective, we ensure that all generated control strategies are within the process-feasible range. The optimization process minimizes the error between the control objective and the actual output while satisfying the basic physical constraints.

[0091] The introduction of process window constraints significantly improves the safety and reliability of the control system, avoids blind optimization of control strategies, and ensures the stability of the production process. In particular, when operating conditions fluctuate drastically, the constraints can prevent the control system from performing dangerous operations, providing an important guarantee for safe production.

[0092] Traditional annotation methods cannot adapt to the differences in data quality at different production stages in the float glass production environment, easily leading to problems such as high noise in training data and limited model performance. Therefore, in one embodiment, step S30 also includes an intelligent annotation method for training data, comprising the following steps:

[0093] Different labeling strategies are used depending on the production stage;

[0094] Label confidence level ,in, This represents the average thickness difference (in mm). This represents the standard deviation of the thickness difference (in mm).

[0095] when Manual review will be initiated at that time, among which, This is the minimum confidence threshold.

[0096] The labeling confidence level is dynamically calculated based on production stability. Different labeling strategies are used for data with different confidence levels. Low-confidence data are subject to manual review to ensure labeling quality.

[0097] Intelligent annotation methods significantly improve the quality and utilization of training data, providing a reliable data foundation for model training. The confidence mechanism provides a quantitative indicator of data quality, which can be used to adjust the contribution of different data points during model training, thereby improving training effectiveness.

[0098] To address the issue that excessively drastic parameter adjustments in traditional control methods can lead to quality defects such as vibration and streaks in the glass ribbon, affecting product consistency, in one embodiment, the method for adjusting the edge-pulling machine parameters in step S40 includes the following steps:

[0099] S41. Obtain the current relaxation time of the glass. ;

[0100] S42. Adopt a gradual adjustment strategy. ,in, For a moment The parameter adjustment amount, The maximum allowable adjustment amount, The duration of the adjustment process (in seconds);

[0101] S43, When the relaxation time Changes exceeded At that time, the adjustment strategy parameters are recalculated and updated, where, This is the threshold for the rate of change.

[0102] Based on the viscoelastic characteristics of glass materials, an exponential progressive adjustment strategy is adopted to ensure a smooth transition in the control quantity, avoid abrupt changes, prevent the impact of mechanical shock on product quality, and automatically update the adjustment parameters when the material properties change, ensuring that the adjustment strategy always matches the current material properties and improving the accuracy of control.

[0103] To address the problem that a fixed control cycle cannot adapt to changes in quality fluctuations, resulting in poor control performance, in one embodiment, the closed-loop control in step S50 includes the following steps:

[0104] S51. Calculate the current thickness difference fluctuation range. (Unit: mm);

[0105] S52, Dynamically Adjust Control Cycle (Unit: seconds), formula is: ,in, The characteristic time constant of the control system (in seconds). Target thickness (in mm);

[0106] S53, when If the temperature continues to rise, activate emergency control mode.

[0107] The control cycle is dynamically adjusted based on the fluctuation range of the thickness difference; the greater the fluctuation, the more frequent the control. An anomaly handling mechanism is set up so that when the fluctuation continues to increase, an emergency control mode is activated to handle the anomaly, such as manual verification or manual operation. The variable cycle control strategy improves control efficiency, ensuring control effectiveness while avoiding unnecessary control actions. The anomaly handling mechanism enhances the robustness of the system, ensuring that the system remains stable even under extreme operating conditions.

[0108] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are 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 precise control of thickness difference in float glass, characterized in that, Includes the following steps: S10. Construct a multi-layer composite insulation structure outside the water bath in the high-temperature zone of the tin bath in the float glass production line; S20. Real-time acquisition of tin bath process parameters, including temperature field distribution, pressure parameters, and glass ribbon geometric parameters; S30. Use a domain-adaptive deep learning algorithm to process the process parameters and generate control signals for the edge-pulling machine; S40. Adjust the angle, speed, force, and position parameters of the edge-pulling machine according to the control signal; S50. Repeat steps S20 to S40 to form a closed-loop control cycle. When the thickness difference of the glass stabilizes within the required range, enter the maintenance mode to reduce the computational load.

2. The method for precise control of thickness difference in float glass according to claim 1, characterized in that, In step S1, the multi-layer composite insulation structure includes a high-silica rope layer, a high-temperature resistant mud layer, and a stainless steel sleeve layer arranged sequentially from the inside to the outside. Constructing a multi-layer composite insulation structure includes the following steps: S11. A high-silica rope layer is formed by winding a high-silica rope around the surface of the water-filled package. The winding density is 5 to 8 turns per centimeter, and the winding thickness is 5 to 10 mm. The high-silica rope layer is doped with nano-zirconia particles at a doping ratio of 4% to 5% by weight. S12. Apply high-temperature resistant clay to the outside of the high-silica rope layer to form a high-temperature resistant clay layer with a coating thickness of 3-5 mm. The high-temperature resistant clay includes aluminum silicate fiber and binder. Among them, silicon carbide fiber is added to the high-temperature resistant clay layer at a weight ratio of 2%-3%. S13. A stainless steel sleeve is installed on the outside of the high-temperature resistant mud layer to form a stainless steel sleeve layer, and then fastened with bolts.

3. The method for precise control of thickness difference in float glass according to claim 2, characterized in that, In step S20, collecting process parameters includes the following steps: S21. Set the initial multi-scale sampling strategy: Use the first sampling frequency for the molten tin temperature. A second sampling frequency is used to address the thickness difference. The third sampling frequency is used for the edge-pulling machine parameters. ,in, ; S22. Real-time calculation of the characteristic time constant of the current molten glass. The calculation formula is: ,in, Current temperature The viscosity of the molten glass. The elastic modulus of glass; S23, The characteristic time constant With reference time constant If a comparison is made, This triggers a sampling frequency update based on the characteristic time constant. Recalculate and update the first sampling frequency Second sampling frequency and the third sampling frequency ,in, The preset threshold; S24. Use the updated first sampling frequency. Second sampling frequency and the third sampling frequency Perform multi-scale sampling.

4. The method for precise control of float glass thickness difference according to claim 3, characterized in that, In step S30, the domain-adaptive deep learning algorithm is a hybrid neural network structure, including a physical guidance encoder, a spatiotemporal attention module, and a control policy generator: Physically guided encoders are used to map temperature field features, flow features, and geometric features to a feature space; The spatiotemporal attention module is used to calculate the attention weights for different regions; The control strategy generator is used to output the four-dimensional control parameters of the edge-pulling machine.

5. The method for precise control of thickness difference in float glass according to claim 4, characterized in that, In step S30, the physical guided encoder is trained using a constraint loss function, the formula of which is: ,in, and These are the weighting coefficients. For time, For temperature field, The thermal diffusivity of glass, For the Laplace operator, This is the loss term for data fitting.

6. The method for precise control of float glass thickness difference according to claim 5, characterized in that, In step S30, the spatiotemporal attention module operates by including the following steps: S31. Real-time calculation of flow stability indices for each region ,in, For a region index, it represents the first... One region; S32. Based on the flow stability index of each region Calculate the attention weights for each region The formula is ,in, For activation function, The transformation function is based on the relaxation time. For the total number of regions, For a region index, it represents the first... One region; S33, When the maximum weight change At that time, the attention weights for all regions are recalculated and updated, where, This is a preset threshold.

7. The method for precise control of float glass thickness difference according to claim 6, characterized in that, The control strategy generator employs a constrained optimization algorithm, the formula of which is: , , ,in, For the actual output, To predict the output, Temperature of the molding zone The lowest molding temperature, The highest molding temperature, Viscosity, Minimum molding viscosity, This is the highest molding viscosity.

8. The method for precise control of thickness difference in float glass according to claim 7, characterized in that, Step S30 also includes an intelligent annotation method for training data, comprising the following steps: Different labeling strategies are used depending on the production stage; Label confidence level ,in, This represents the average value of the thickness difference. The standard deviation of the thickness difference; when Manual review will be initiated at that time, among which, This is the minimum confidence threshold.

9. The method for precise control of thickness difference in float glass according to claim 8, characterized in that, In step S40, the method for adjusting the parameters of the edge-pulling machine includes the following steps: S41. Obtain the current relaxation time of the glass. ; S42. Adopt a gradual adjustment strategy. ,in, For a moment The parameter adjustment amount, The maximum allowable adjustment amount, To adjust the duration of the process; S43, When the relaxation time Changes exceeded At that time, the adjustment strategy parameters are recalculated and updated, where, This is the threshold for the rate of change.

10. The method for precise control of float glass thickness difference according to claim 9, characterized in that, In step S50, the closed-loop control includes the following steps: S51. Calculate the current thickness difference fluctuation range. ; S52, Dynamically Adjust Control Cycle The formula is ,in, The characteristic time constant of the control system, For target thickness; S53, when If the temperature continues to rise, activate emergency control mode.

Images