Intelligent closed-loop control method for cooling water volume of continuous casting mold based on convolutional neural network

Abstract

The present application relates to the technical field of metallurgical continuous casting automation control, in particular to a continuous casting crystallizer cooling water intelligent closed-loop regulation and control method based on convolutional neural network; the method collects temperature and process parameters, fuses and constructs a three-dimensional input tensor; inputs a pre-trained CNN model to inverse the virtual heat flow density distribution of the solidification front; combines the runtime length and water temperature to calculate the real-time equivalent thermal resistance, generates a flow regulation instruction to drive an electric regulating valve to execute closed-loop control. The present application realizes real-time reconstruction of heat flow through a physical information neural network, introduces a fouling growth model and energy conservation constraints to solve the long-period operating condition drift problem, and combines variable parameter PID and anti-oscillation strategies to realize fine and smooth control, effectively solving the problems of traditional PID control hysteresis and poor adaptability of pure AI models, significantly improving the safety of the continuous casting process and the quality of the cast slab.

Description

Technical Field

[0001] This invention relates to the field of automated control technology for continuous casting in metallurgy, specifically to an intelligent closed-loop control method for cooling water volume in continuous casting crystallizers based on convolutional neural networks. Background Technology

[0002] The continuous casting crystallizer is the core component of the continuous casting machine, and its cooling effect directly determines the surface quality and internal structure of the billet. The control of the cooling water flow rate of the crystallizer is crucial: too small a flow rate will cause the surface temperature of the billet to be too high, leading to steel leakage accidents; too large a flow rate will lead to an increase in cooling cracks on the surface of the billet, and may even cause premature wear of the copper plate.

[0003] The existing cooling water control system for continuous casting molds has the following two significant drawbacks:

[0004] On the one hand, existing systems mainly rely on temperature feedback from thermocouples on the outside of copper plates and use traditional PID algorithms for flow regulation. However, due to the huge thermal resistance and significant hysteresis in the transfer of heat from the inside of the molten steel to the outer wall, this control method based on the temperature of the outer wall cannot perceive the real heat flow distribution at the solidification front in real time. It often only reacts after the risk of steel leakage or cracks occur, which is essentially a delayed "post-event remedy" and is difficult to meet the needs of high-precision process control.

[0005] On the other hand, many of the improvement solutions that incorporate artificial intelligence are purely data-driven models that lack integration with physical mechanisms. These models ignore the dynamic impact of key factors such as scale growth on the copper plate of the crystallizer and water temperature fluctuations on heat transfer efficiency. When the scale on the copper plate thickens or the cooling water temperature changes, the heat flow value predicted by the model will deviate significantly from the actual operating conditions, resulting in inaccurate flow regulation. This makes it impossible to maintain a high-precision dynamic balance during long-term operation, thus limiting the robustness and practicality of intelligent control under complex operating conditions.

[0006] Therefore, there is an urgent need for a method that can overcome the above-mentioned defects and realize intelligent closed-loop control of cooling water volume in continuous casting molds based on real-time heat flow reconstruction and physical mechanism compensation. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides an intelligent closed-loop control method for cooling water volume in continuous casting crystallizers based on convolutional neural networks. This method has advantages such as real-time heat flow reconstruction, dynamic compensation of physical mechanisms, and zoned anti-oscillation control. It solves the problems of lagging thermal field perception in traditional PID control and lack of physical adaptability in pure AI models, which lead to inaccurate long-cycle regulation.

[0009] (II) Technical Solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent closed-loop control of cooling water volume in continuous casting crystallizers based on convolutional neural networks, comprising the following steps:

[0011] Step S1: Real-time acquisition of temperature data from the thermocouple array arranged on the outside of the copper plate of the crystallizer and the current process parameters; conversion of temperature data into spatial image features and process parameters into time-series feature maps; fusion of spatial image features and time-series feature maps to construct a three-dimensional input tensor containing temperature field and operating condition information.

[0012] Step S2: Input the three-dimensional input tensor into the pre-trained convolutional neural network model, use the spatiotemporal feature extraction unit in the convolutional neural network model to perform convolution operation on the three-dimensional input tensor, and output the virtual heat flux density distribution matrix of the solidification front inside the crystallizer after aggregating the feature vector.

[0013] Step S3: Obtain the cumulative running time of the crystallizer copper plate and the current cooling water inlet temperature. Combine the virtual heat flux density distribution matrix to calculate the real-time equivalent thermal resistance of each zone. Generate a flow regulation command based on the deviation between the preset target heat flux value and the real-time equivalent thermal resistance. Drive the electric regulating valve to perform cooling water flow regulation and return to step S1 for the next iteration.

[0014] Preferably, in step S1, converting temperature data into spatial image features includes the following steps:

[0015] Step S1.1: Normalize the temperature data to generate the first grayscale image;

[0016] Step S1.2: Map the pulling speed value in the process parameters to the second single-channel feature map, and map the liquid level height value to the third single-channel feature map;

[0017] Step S1.3: Stack the first grayscale image, the second single-channel feature map, and the third single-channel feature map along the channel dimension to obtain a three-dimensional input tensor.

[0018] Preferably, in step S2, the process of performing convolution operations on the three-dimensional input tensor using the spatiotemporal feature extraction unit in the convolutional neural network model satisfies the following formula:

[0019]

[0020] in, This is the feature map after the convolution operation. For activation function, For 3D convolution kernel weights, This represents a 3D convolution operation. For the three-dimensional input tensor, For bias terms;

[0021] The convolutional neural network model also includes a residual connection structure, the output of which is... Satisfy the following formula:

[0022]

[0023] in, The output of the main path, The input features are directly copied from the bypass path.

[0024] Preferably, in step S2, the training process of the convolutional neural network model includes minimizing the loss function. The loss function It consists of data fitting terms and physical constraint terms, and the calculation formula is as follows:

[0025]

[0026] in, The virtual heat flux density distribution matrix obtained from the prediction, This is a true heat flux density label matrix. The divergence of the virtual heat flux density distribution matrix. Let α be the instantaneous rate of change of heat capacity calculated based on the law of conservation of energy, and β be the data fitting weight and physical constraint weight, respectively.

[0027] Preferably, in step S3, calculating the real-time equivalent thermal resistance of each partition includes: calculating the estimated function value of the copper plate scale thickness based on a preset scale growth model and the cumulative running time. The calculation formula is as follows:

[0028]

[0029] in, This represents the maximum theoretical scale thickness. Let be the scale growth rate constant. The cumulative runtime value, The erosion coefficient caused by Rabo wave. For a moment The Rapo value, This is the rated Rapo value; based on the first value in the virtual heat flux density distribution matrix. Heat flow values ​​for each zone Copper plate scale thickness prediction function value and the current cooling water inlet temperature value The real-time equivalent thermal resistance value of this partition is obtained using heat transfer calculation logic. The calculation formula is as follows:

[0030]

[0031] in, For the new copper coin state, the first The baseline thermal resistance value for each zone, This represents the thermal conductivity value of the scale layer. This is a correction function for the heat transfer coefficient based on water temperature.

[0032] Preferably, in step S3, generating the flow regulation command includes:

[0033] Read the preset first Target heat flux density threshold values ​​for each zone ;

[0034] Based on the target heat flux density threshold value With the Heat flow values ​​for each zone The difference between them, combined with the real-time equivalent thermal resistance value Calculate the first Traffic deviation in each partition The calculation formula is as follows:

[0035]

[0036] in, The value of the proportional gain coefficient is given; the cumulative deviation term from the previous moment is added to the flow deviation to obtain the value of the first time step. Final flow control instructions for each partition .

[0037] Preferably, in step S3, the generation of the final flow regulation command further includes an anti-oscillation constraint step:

[0038] The formula for calculating the change in flow regulation command between two adjacent time steps is as follows:

[0039]

[0040] like Then force Revised to sign ;

[0041] like Then keep Unchanged; among them, This represents the maximum permissible increment of valve actuation. It is a symbolic function.

[0042] Preferably, the proportional gain coefficient value It is a variable parameter, and its value is based on the current pulling speed. Dynamic adjustment, with the following adjustment rules:

[0043] when hour, ;

[0044] when hour, ;

[0045] when hour, ;

[0046] in, The minimum pulling speed threshold, The upper limit of the pulling speed threshold, .

[0047] Preferably, the combination of the number of longitudinal intervals and the number of transverse partitions on the inner surface of the crystallizer is consistent with the physical pipeline distribution boundary of the cooling water circuit; when the heat flux value at any position in the virtual heat flux density distribution matrix exceeds the preset steel leakage warning threshold... When this occurs, the emergency stop procedure is triggered, and all electric regulating valves are locked at their maximum opening.

[0048] Preferably, in step S2, the global pooling unit in the convolutional neural network model adopts average pooling or max pooling to reduce feature dimensionality and retain key thermal features.

[0049] The fully connected output unit maps the aggregated feature vector into a sequence of heat flux density values ​​corresponding to the number of crystallizer partitions.

[0050] (III) Beneficial Effects

[0051] Compared with existing technologies, this invention provides an intelligent closed-loop control method for cooling water volume in continuous casting crystallizers based on convolutional neural networks, which has the following beneficial effects:

[0052] 1. This invention maps multi-source temperature data into spatial image features through a convolutional neural network and integrates operating parameters such as casting speed and liquid level to directly invert the virtual heat flux density distribution at the solidification front inside the crystallizer. This mechanism overcomes the thermal hysteresis problem of traditional PID control relying on external wall temperature feedback and realizes real-time perception of the internal thermal field state. This enables the system to intervene before the risk of steel leakage or cracking occurs, changing the control mode from a lagging "post-event remedy" to an active "pre-event warning", effectively reducing the accident rate in continuous casting production.

[0053] 2. To address the issue that pure data models cannot adapt to copper plate scaling growth and water temperature fluctuations, this invention introduces a scaling growth model and heat transfer formulas to calculate the dynamic equivalent thermal resistance in real time and physically correct the AI ​​prediction results. This combination ensures that the flow regulation command is always generated based on the real heat transfer environment, effectively solving the model prediction deviation caused by scaling or water temperature changes, and guaranteeing the long-term stable operation and regulation accuracy of intelligent control under complex working conditions.

[0054] 3. This invention divides the cooling water circuit into independent control zones based on physical boundaries, and combines a variable parameter proportional gain coefficient that is dynamically adjusted with the casting speed to achieve fine and independent control of heat flow in different areas. This avoids local overcooling or overheating defects caused by global unified control, thereby improving the surface quality and internal structure uniformity of the billet. In addition, the introduction of valve action increment limiting logic (anti-oscillation constraint) effectively suppresses the risk of hydraulic oscillation caused by violent fluctuations in model output, ensures smooth actuator action, and reduces equipment failure rate. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to the present invention.

[0056] Figure 2 This is a diagram showing the overall control architecture of the intelligent closed-loop regulation method for cooling water volume in continuous casting crystallizers based on convolutional neural networks, as described in this invention.

[0057] Figure 3 This is a schematic diagram of input tensor construction provided in an embodiment of the present invention;

[0058] Figure 4 A schematic diagram of the virtual heat flux density prediction model based on physical information convolutional neural network provided in an embodiment of the present invention. Detailed Implementation

[0059] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Example 1

[0061] This embodiment mainly describes the hardware environment, data acquisition method, and input feature tensor construction process required to implement the method of the present invention.

[0062] like Figure 2As shown, the physical basis of this system includes a thermocouple array arranged on the outside of the copper plate of the continuous casting crystallizer, a PLC / DCS control system, and an electric regulating valve group.

[0063] This includes temperature acquisition, process parameter acquisition, and actuators;

[0064] Thermocouple arrays are evenly arranged along the longitudinal and transverse directions on the outer side of the crystallizer copper plate to collect temperature data at each monitoring point in real time. .

[0065] The PLC system synchronously acquires the current process parameters, including the billet pulling speed. Intermediate tank liquid level height Cooling water inlet temperature and the cumulative runtime of the crystallizer .

[0066] Each control zone corresponds to an independent electric regulating valve, which is used to receive flow regulation commands and change the cooling water flow rate.

[0067] In the edge computing unit (which uses an industrial-grade control computer with GPU acceleration capabilities and at least 8GB of video memory to meet the computing power requirements of real-time inference of 3D convolutional neural networks), the collected raw data is processed in real time to construct a 3D input tensor, specifically including the following steps:

[0068] Step S1.1 (Temperature Field Image): First, the temperature data collected by the thermocouples is normalized to eliminate the influence of range differences at different temperature measurement points, generating the first grayscale image. The grayscale values ​​of pixels in an image directly correspond to temperature levels, forming spatial distribution characteristics.

[0069] Step S1.2 (Operating Condition Characteristic Mapping): Map the pulling speed value in the process parameters. Mapped to a second single-channel feature map , the liquid level height value Mapped to a third single-channel feature map These two feature maps are expanded using a one-dimensional ordered signal or filled with two-dimensional space to ensure that their spatial resolution is consistent with that of the first grayscale image.

[0070] In essence, due to the pulling speed and liquid level As a global scalar parameter, this embodiment copies and fills its value to have the same length and width dimensions as the temperature image. In a two-dimensional matrix, a spatially uniform feature map is formed to ensure that the spatial dimensions of each channel of the input tensor are consistent.

[0071] Step S1.3 (Construction of 3D Tensor): Stack the first grayscale image, the second single-channel feature map, and the third single-channel feature map along the channel dimension to obtain a 3D input tensor. This tensor contains both the spatial information of the temperature field on the surface of the crystallizer and dynamic operating condition information such as pulling speed and liquid level, serving as the unified input for the subsequent convolutional neural network.

[0072] It should be noted that: scalar pulling speed Copy and fill the data into a two-dimensional matrix with the same resolution as the temperature map to form a spatially uniform pull speed feature map; process the liquid level height in the same way.

[0073] In addition, the system employs hardware triggering or software timestamping mechanisms to ensure that temperature acquisition, process parameter acquisition, and model inference are strictly synchronized on a microsecond-level timescale, avoiding control lag caused by data transmission delays.

[0074] Example 2

[0075] like Figure 3 As shown, this embodiment details the structural design, training strategy, and introduction of physical mechanisms of the core algorithm model.

[0076] The convolutional neural network (CNN) model constructed in this invention includes a spatiotemporal feature extraction unit, a residual connection structure, a global pooling unit, and a fully connected output layer.

[0077] Spatiotemporal feature extraction includes: using three-dimensional convolution kernels. For input tensor The convolution operation is performed using the following formula:

[0078]

[0079] in, For activation functions (such as ReLU). This represents a 3D convolution operation; this process can simultaneously extract spatial location features and time series features, and output a feature vector after aggregation.

[0080] The residual connection structure includes: to prevent gradient vanishing in deep networks and accelerate convergence, residual connections are introduced into the model; main path output Input directly copied from the bypass Add them together to get the final output. .

[0081] The output mapping includes: reducing the feature dimensionality using average pooling or max pooling while retaining key thermal features; and finally, mapping the aggregated feature vectors to a virtual heat flux density distribution matrix corresponding to the number of crystallizer partitions using a fully connected output unit. .

[0082] In this embodiment, the label data used to train the model... It is derived from high-precision finite element simulation (FEM) data of continuous casting process and offline calibration experimental data, and has been data augmented (such as adding Gaussian noise and simulating different scaling degrees) to cover a wide range of working conditions.

[0083] To address the vulnerability of purely data-driven models to operational condition drift, this embodiment introduces physical conservation laws as constraints during the training phase; loss function It consists of two parts:

[0084]

[0085] Data fitting term: First term Minimize predicted heat flow With real labels Errors (usually derived from historical high-precision simulations or offline calibration data) between different sources.

[0086] Physical constraint term: Second term Forced model output heat dissipation It satisfies the energy conservation equation, which is equal to the instantaneous rate of change of heat capacity. .

[0087] Among them, in the equation Specifically refers to the temperature field data collected by thermocouples on the outer wall of the crystallizer. Representing the divergence of the virtual heat flux density field, this constraint forces the heat flux distribution output by the model to conform to the law of conservation of energy, thereby improving the model's generalization ability under unknown conditions.

[0088] Weight balancing: α and β are the data fitting weight and physical constraint weight, respectively. By adjusting their ratio, the model is ensured to maintain high prediction accuracy while conforming to the physical laws of heat transfer. Models trained in this way maintain high prediction robustness even under unseen conditions such as thickened copper scale or water temperature fluctuations.

[0089] Example 3

[0090] like Figure 4 As shown, this embodiment describes the specific control logic of the system during online operation, including equivalent thermal resistance calculation, variable parameter PID flow regulation, and instruction generation.

[0091] In the control loop, the system reads the cumulative running time of the crystallizing copper plate in real time. and inlet water temperature And it is corrected by combining the virtual heat flux density distribution matrix output by the CNN:

[0092] Scale thickness prediction: Based on a preset scale growth model, the predicted scale thickness of the copper plate is calculated using a function value that calculates the cumulative operating time. :

[0093]

[0094] in, For the maximum theoretical scale thickness, Let be the scale growth rate constant. The scouring coefficient caused by speed fluctuations. For a moment The pulling speed, This is the rated pulling speed.

[0095] Partition thermal resistance update: combining the first part of the virtual heat flux density distribution matrix Heat flow values ​​for each zone The real-time equivalent thermal resistance value of this partition is obtained using heat transfer calculation logic. :

[0096]

[0097] in, For the new copper coin state, the first The baseline thermal resistance value for each zone, This represents the thermal conductivity value of the scale layer. This is a correction function for the heat transfer coefficient based on water temperature.

[0098] The system reads the preset first Target heat flux density threshold for each zone Calculate the flow deviation :

[0099]

[0100] Variable gain Proportional gain coefficient It is a variable parameter, and its value is based on the current pulling speed. Dynamic adjustment.

[0101] when (Lower limit pulling speed threshold) (Greater gain, improved sensitivity);

[0102] when (Within the normal pulling speed range) (Moderate gain);

[0103] when (At the maximum pulling speed threshold) (Lower gain to prevent overshoot).

[0104] in .

[0105] It should be noted that: introducing equivalent thermal resistance As a gain compensation factor, it exists because as fouling thickens and thermal resistance increases, the change in heat flux per unit flow rate decreases; multiplied by It can dynamically amplify the adjustment amount to offset the decrease in control sensitivity caused by changes in thermal resistance, ensuring stable control accuracy under different operating cycles.

[0106] Subsequently, the calculated flow deviation is added to the cumulative deviation term from the previous moment to obtain the result. Final flow control instructions for each partition .

[0107] Final flow regulation command The corresponding electric regulating valve is activated to change the cooling water flow rate. The system then returns to step S1 for the next iteration, forming a closed-loop control.

[0108] Example 4

[0109] This embodiment describes the system's boundary condition handling, anti-vibration strategy, and anomaly safety logic;

[0110] To prevent frequent operation of the electric regulating valve or hydraulic oscillation and water hammer effects caused by drastic fluctuations in model output, the system introduces incremental limiting logic:

[0111] Change calculation: Calculate the change in flow regulation command between two adjacent time steps. .

[0112] Correction logic: If (Maximum allowed increment), then force will Revised to .

[0113] like Then keep constant.

[0114] This mechanism ensures smooth actuator operation, extends equipment life, and reduces failure rate.

[0115] Furthermore, the system monitors the virtual heat flux density distribution matrix in real time. .

[0116] Zoning consistency: The combination of the number of longitudinal sections and the number of transverse sections on the inner surface of the crystallizer is strictly consistent with the physical pipeline distribution boundary of the cooling water circuit to ensure that control commands are accurately applied to specific physical areas.

[0117] Emergency Trigger: When the heat flux value at any location in the virtual heat flux density distribution matrix exceeds the preset steel leakage warning threshold. This threshold (indicating an excessively thin solidified shell or a very high risk of localized overheating) These are dynamic critical values ​​determined based on historical accident data and finite element simulation analysis. They are typically set at the boundary of the safety margin to avoid false alarms that could lead to unplanned downtime.

[0118] 1. Immediately trigger the emergency stop pouring procedure;

[0119] 2. Lock all electric regulating valves to their maximum opening to maximize cooling intensity and prevent the steel leakage accident from escalating (or open the corresponding valves according to the principle of prioritizing high-temperature areas to avoid misjudgment that could lead to excessive cooling in other areas and cause cracks).

[0120] 3. Simultaneously issue audible and visual alarms to the operator and record fault data for subsequent analysis.

[0121] In summary, through the synergistic cooperation of the four embodiments described above, this invention constructs a complete intelligent cooling control method for continuous casting crystallizers. This method not only utilizes convolutional neural networks to achieve a reverse mapping from the outer wall temperature to the internal solidification front heat flow, effectively eliminating the control delay caused by thermal hysteresis; it also introduces a scaling growth model and energy conservation physical constraints, solving the problem of decreased accuracy caused by operating condition drift in long-term operation of pure data-driven models; at the same time, through independent zone control, variable parameter gain strategy, and valve action anti-oscillation mechanism, it achieves refined and smooth control of the billet heat flow distribution, significantly improving the safety of the continuous casting process, the uniformity of billet quality, and the reliability of equipment operation.

[0122] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0123] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent closed-loop control of cooling water volume in continuous casting crystallizers based on convolutional neural networks, characterized in that, Includes the following steps: Step S1: Real-time acquisition of temperature data from the thermocouple array arranged on the outside of the copper plate of the crystallizer and the current process parameters; conversion of temperature data into spatial image features and process parameters into time-series feature maps; fusion of spatial image features and time-series feature maps to construct a three-dimensional input tensor containing temperature field and operating condition information. Step S2: Input the three-dimensional input tensor into the pre-trained convolutional neural network model, use the spatiotemporal feature extraction unit in the convolutional neural network model to perform convolution operation on the three-dimensional input tensor, and output the virtual heat flux density distribution matrix of the solidification front inside the crystallizer after aggregating the feature vector. Step S3: Obtain the cumulative running time of the crystallizer copper plate and the current cooling water inlet temperature. Combine the virtual heat flux density distribution matrix to calculate the real-time equivalent thermal resistance of each zone. Generate a flow regulation command based on the deviation between the preset target heat flux value and the real-time equivalent thermal resistance. Drive the electric regulating valve to perform cooling water flow regulation and return to step S1 for the next iteration.

2. The intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to claim 1, characterized in that: In step S1, converting temperature data into spatial image features includes the following steps: Step S1.1: Normalize the temperature data to generate the first grayscale image; Step S1.2: Map the pulling speed value in the process parameters to the second single-channel feature map, and map the liquid level height value to the third single-channel feature map; Step S1.3: Stack the first grayscale image, the second single-channel feature map, and the third single-channel feature map along the channel dimension to obtain a three-dimensional input tensor.

3. The intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to claim 1, characterized in that: In step S2, the process of performing convolution operations on the three-dimensional input tensor using the spatiotemporal feature extraction unit in the convolutional neural network model satisfies the following formula: ； in, This is the feature map after the convolution operation. For activation function, For the weights of the 3D convolution kernel, This represents a 3D convolution operation. For the three-dimensional input tensor, For bias terms; The convolutional neural network model also includes a residual connection structure, the output of which is... Satisfy the following formula: ； in, The output of the main path, The input features are directly copied from the bypass path.

4. The intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to claim 1, characterized in that: In step S2, the training process of the convolutional neural network model includes minimizing the loss function. The loss function It consists of data fitting terms and physical constraint terms, and the calculation formula is as follows: ； in, The virtual heat flux density distribution matrix obtained from the prediction, This is a label matrix representing the actual heat flux density. The divergence of the virtual heat flux density distribution matrix. Let α be the instantaneous rate of change of heat capacity calculated based on the law of conservation of energy, and β be the data fitting weight and physical constraint weight, respectively.

5. The intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to claim 1, characterized in that: In step S3, calculating the real-time equivalent thermal resistance of each partition includes: calculating the estimated function value of copper plate scale thickness based on a preset scale growth model and the cumulative running time. The calculation formula is as follows: ； in, This represents the maximum theoretical scale thickness. Let be the scale growth rate constant. The cumulative runtime value, The erosion coefficient caused by Rabo wave. For a moment The Rapo value, This is the rated Rapo value; based on the first value in the virtual heat flux density distribution matrix. Heat flow values ​​for each zone Copper plate scale thickness prediction function value and the current cooling water inlet temperature value The real-time equivalent thermal resistance value of this partition is obtained using heat transfer calculation logic. The calculation formula is as follows: ； in, For the new copper coin state, the first The baseline thermal resistance value for each zone, This represents the thermal conductivity value of the scale layer. This is a correction function for the heat transfer coefficient based on water temperature.

6. The intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to claim 1, characterized in that: In step S3, the generation of the flow regulation command includes: Read the preset first Target heat flux density threshold values ​​for each zone ; Based on the target heat flux density threshold value With the Heat flow values ​​for each zone The difference between them, combined with the real-time equivalent thermal resistance value Calculate the first Traffic deviation in each partition The calculation formula is as follows: ； in, The value of the proportional gain coefficient is given; the cumulative deviation term from the previous moment is added to the flow deviation to obtain the value of the first time step. Final flow control instructions for each partition .

7. The intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to claim 6, characterized in that: In step S3, the generation of the final flow regulation command also includes an anti-oscillation constraint step: The formula for calculating the change in flow regulation command between two adjacent time steps is as follows: ； like Then force Revised to sign ; like Then keep Unchanged; among them, This represents the maximum permissible increment of valve actuation. It is a symbolic function.

8. The intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to claim 1, characterized in that: The proportional gain coefficient value It is a variable parameter, and its value is based on the current pulling speed. Dynamic adjustment, with the following adjustment rules: when hour, ; when hour, ; when hour, ; in, The minimum pulling speed threshold, The upper limit of the pulling speed threshold, .

9. The intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to claim 1, characterized in that: The combination of the number of longitudinal and transverse partitions on the inner surface of the crystallizer is consistent with the physical piping distribution boundary of the cooling water circuit; when the heat flux value at any position in the virtual heat flux density distribution matrix exceeds the preset steel leakage warning threshold... When this occurs, the emergency stop procedure is triggered, and all electric regulating valves are locked at their maximum opening.

10. The intelligent closed-loop control method for cooling water volume of continuous casting crystallizer based on convolutional neural network according to claim 1, characterized in that: In step S2, the global pooling unit in the convolutional neural network model adopts average pooling or max pooling to reduce feature dimensionality and retain key thermal features. The fully connected output unit maps the aggregated feature vector into a sequence of heat flux density values ​​corresponding to the number of crystallizer partitions.

Images