Continuous casting process crystallizer liquid level initiative rapid control system based on mechanism fusion AI

Abstract

The invention discloses a continuous casting process crystallizer liquid level initiative rapid control system based on mechanism fusion AI. The system comprises a real-time data acquisition system, a PINN prediction model, a crystallizer liquid level BP-PID controller and a stopper rod automatic controller. Wherein the real-time data acquisition system is responsible for acquiring key parameters of a continuous casting process and sending the key parameters to the designed PINN prediction model and the crystallizer liquid level BP-PID controller. And further obtaining complete flow field information and nozzle flow velocity through the prediction model. And then PID parameters of the crystallizer liquid level BP-PID controller are changed based on BP neural network calculation according to the data and the collected parameters, a more reasonable stopper target set height is output, and a stopper quickly and accurately reaches a control position through a stopper position PID controller. The metallurgical reaction state of the real-time continuous casting process can be rapidly provided, the water inlet and outlet flow speed can be predicted according to the liquid level height and the stopper rod position, the parameters of the crystallizer liquid level controller are actively and automatically adjusted, and closed-loop optimization control over the metallurgical process is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of steel metallurgy artificial intelligence, and relates to a mechanism fusion AI-based continuous casting process crystallizer liquid level initiative rapid control system. BACKGROUND

[0002] Keeping the crystallizer liquid level stable is the key to improving the product quality of the continuous casting production process. The stability of the crystallizer liquid level is controlled by the balance of the pulling speed and the flow size from the tundish to the crystallizer, and the stopper is the most important flow control element in the process of ensuring the stability of the liquid level. Figure 1 As shown in the figure, the flow of molten steel at the water gap is directly related to the flow field near the stopper head region. The molten steel flow is fast near the stopper head, the flow field is uneven, and is affected by the design of the tundish, the remaining steel amount, and the flow speed of the ladle water gap. Therefore, it is necessary to obtain the flow field information near the stopper head during control. However, due to the high-temperature environment inside the tundish and the crystallizer and the high nonlinearity and dynamics of the metallurgical reaction, a “black box” that is difficult to directly explore is formed, so engineers cannot obtain the actual flow information, which greatly affects the precision and effect of flow control. This method not only has low efficiency, but also cannot capture real-time and comprehensive metallurgical information, and cannot realize immediate adjustment and control of the liquid level and the stopper.

[0003] At present, the main method used by the industry to solve the “black box” problem is still numerical simulation. Although this method has high precision, the calculation time is too long, sometimes even several days, and cannot provide real-time metallurgical reaction state, so it cannot meet the real-time demand of engineering control.

[0004] The rapid development of artificial intelligence technology brings new ideas for solving the “black box” problem of such metallurgical high-temperature containers as the tundish and the crystallizer. Deep learning, as an important branch of artificial intelligence, has strong non-linear mapping ability and data-driven learning mechanism, providing a new path for modeling and control of complex metallurgical systems. Through high-fidelity mechanism modeling, combined with the physical information neural network PINN (Physics-Informed Neural Networks), the internal working state of the metallurgical high-temperature container can be quickly and accurately predicted. PINN is a new type of neural network architecture that adds physical equations as constraints to the neural network, making the fitted results more consistent with physical laws, solving the real-time problem faced by traditional numerical methods in solving complex physical problems. At the same time, it also overcomes the problem that the traditional model may lead to results that violate physical laws when blindly using data-driven methods, affecting the generalization and interpretability of the model.

[0005] In the crystallizer liquid level control, there is a method of using BP neural network to predict the liquid level according to the metallurgical state data for control. However, the accuracy of this method is low, which cannot meet the engineering needs. Therefore, the application proposes a new idea of using BP neural network for adaptive control, and using physical information neural network PINN to predict whether the flow is balanced and whether the liquid level can be stable according to the stopper position, so as to realize active and refined control. SUMMARY

[0006] To solve the above problems, the application provides a continuous casting process crystallizer liquid surface rapid control system based on mechanism fusion AI. The system takes the intermediate ladle flow field prediction model based on PINN and the crystallizer liquid surface controller based on BP-PID as the core, which can not only quickly provide the real-time metallurgical reaction state of the continuous casting process, but also predict the water inlet flow rate according to the liquid level and the stopper position, actively and automatically adjust the crystallizer liquid level controller parameters, and realize closed-loop optimization control of the metallurgical process. This not only significantly improves the production efficiency and product quality stability, but also reduces the energy consumption and production cost, and promotes the intelligent transformation and sustainable development of the steel industry.

[0007] The application provides a continuous casting process crystallizer liquid surface active rapid control system based on mechanism fusion AI, which comprises a real-time data acquisition part, an intermediate ladle flow field PINN prediction part, a crystallizer liquid surface BP-PID control part and a stopper automatic control part.

[0008] The real-time data acquisition system is used for acquiring the key parameters of the continuous casting process in the continuous casting production system in real time, including the liquid level in the intermediate ladle, the crystallizer liquid surface height, the casting speed, the billet width, the tundish water inlet insertion depth, the intermediate ladle residual steel amount and the stopper height, and sends the intermediate ladle related parameters to the intermediate ladle flow field PINN prediction model, and sends the crystallizer related parameters, the equipment acquisition parameters and the continuous casting process setting parameters to the BP-PID controller.

[0009] The intermediate ladle flow field PINN prediction part adopts a PINN neural network, which is composed of multiple fully connected layers, has 3 neurons in the input layer, 32 neurons in each of 12 hidden layers, and 3 neurons in the output layer, and is constrained by the two-dimensional incompressible fluid Navier-Stokes equation. The training set used by the PINN neural network is a combination of water model experiment data and simulation data based on experiment verification. A composite loss function is used in the training process, including data loss and physical equation residual loss. The data loss is calculated by using mean square error MSE, and the equation residual loss is also calculated by using MSE, so that the model output satisfies the Navier-Stokes equation. At the same time, the pressure, flow rate data and corresponding partial derivative error are used to construct function, Function according to dynamic pressure, flow field , Directional velocity data and corresponding partial derivatives The mean squared error is constructed; and the Adam optimizer is used.

[0010] The tundish flow field PINN prediction module acquires data collected by the data acquisition system in real time, performs calculations, and outputs the prediction results as complete flow field information and the nozzle flow velocity calculated based on this flow field information. The nozzle flow velocity information is transmitted to the mold liquid level BP-PID control module through the data interface.

[0011] The crystallizer liquid level BP-PID control part includes two parts: BP neural network and liquid level PID controller. Among them, the output layer of BP neural network has three neurons, which correspond to the three parameters of PID respectively. The input of the BP neural network is the actual height of the current crystallizer liquid level, the water inlet flow rate predicted by the intermediate flow field PINN prediction model and the parameters set by the system.

[0012] BP neural network calculates PID parameters based on the current system state The PID parameters are updated in real time according to the changes in the liquid level, and the updated parameters are passed to the liquid level PID controller; the liquid level PID controller uses these parameters to adjust the stopper rod target setting height, output a more reasonable stopper rod target setting height, and transmit it to the stopper rod position automatic control part.

[0013] The stopper rod position automatic control part generates an accurate control signal for the stopper rod actuator after receiving the stopper rod target set height and the actual height of the stopper rod, and then the stopper rod actuator controls the stopper rod to reach the control position quickly and accurately.

[0014] The advantages of the present invention are:

[0015] 1. This invention, based on a mechanism-integrated AI-based rapid mold level control system for the continuous casting process, integrates the PINN model with a BP-PID controller to achieve real-time, accurate prediction of metallurgical reaction states and proactive closed-loop control optimization, significantly improving control accuracy and quality stability. The PINN model utilizes the Navier-Stokes equations as physical constraints, ensuring the scientific and accurate prediction results and providing a reliable basis for ensuring stable control of the molten steel level in the mold.

[0016] 2. This invention's rapid mold level control system for the continuous casting process, based on a mechanism-integrated AI system, automatically adjusts the control parameters of the mold level controller, effectively addressing the various dynamic impacts of frequent changes in production parameters during the production process and ensuring stable product quality. Compared to traditional control methods, this system can more flexibly adapt to different production conditions and requirements, reducing the delay of passive stopper control based on liquid level and the errors caused by traditional stopper control methods, significantly improving product qualification rate and consistency.

[0017] 3. This invention's rapid mold level control system for the continuous casting process, based on a mechanism-integrated AI system, also demonstrates outstanding energy conservation and consumption reduction. By accurately predicting and then precisely controlling the tundish outlet flow rate and mold level, it reduces unnecessary energy loss and optimizes energy efficiency. Furthermore, the improvements in production efficiency and product quality indirectly reduce production costs, bringing significant economic benefits to steel companies. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is the schematic diagram of the crystallizer liquid level control principle in the continuous casting process.

[0019] Figure 2 This is a schematic diagram of the active rapid control system for the liquid level of the crystallizer in the continuous casting process based on the mechanism fusion of AI in the present invention.

[0020] Figure 3 This is the neural network structure of the PINN fast prediction model for the tundish flow field. DETAILED DESCRIPTION

[0021] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0022] The present invention is based on the mechanism fusion AI continuous casting process mold liquid level active rapid control system, the main components include: real-time data acquisition part, tundish flow field PINN prediction part, mold liquid level BP-PID control part and stopper rod automatic control part, such as Figure 2 shown.

[0023] The real-time data acquisition system continuously collects key parameters of the continuous casting process from the continuous casting production system; the key parameters of the continuous casting process include:

[0024] ① Parameters collected by various sensors installed on the ladle, tundish, and crystallizer, including the liquid level in the tundish, the liquid level in the crystallizer, the insertion depth of the ladle nozzle, the amount of steel remaining in the tundish, and the height of the stopper rod;

[0025] ② Set parameters, including casting speed and billet width;

[0026] Further, the real-time data acquisition system sends the collected parameters to the tundish flow field PINN rapid prediction model and the controller, including transmitting the parameters collected by the tundish upper sensor (stopper height, tundish liquid level, etc.) and the data collected by the tundish upper sensor (tundish nozzle insertion depth) to the tundish flow field PINN prediction model, and transmitting the parameters collected by the crystallizer upper sensor (crystallizer liquid level, etc.) and the continuous casting process setting parameters to the BP-PID controller.

[0027] The tundish flow field PINN prediction part is the core part of the state prediction of the control system of the present application, and the physical information neural network adopts a PINN neural network.

[0028] As shown in Figure 3 , the PINN neural network is composed of multiple fully connected layers. In the present application, the specific structure is designed as [3, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 3], which means that the input layer has 3 neurons, each of the 12 hidden layers has 32 neurons, and the output layer has 3 neurons. At the same time, after each hidden layer, a hyperbolic tangent (tanh) activation function is used to increase the non-linear expression ability of the network, so as to strengthen the problem solving ability of the deep learning model, which is specifically represented as follows:

[0029]

[0030] The tundish metallurgical state rapid prediction model uses the two-dimensional incompressible fluid Navier-Stokes equation as a constraint, and the formula is:

[0031]

[0032]

[0033] In the formula, is the density of the molten steel; is the velocity vector, is the velocity in the direction, is the velocity in the direction, is the main flow direction of the molten steel along the length of the tundish; is the velocity in the direction, is the velocity in the direction perpendicular to the direction . is the dynamic pressure. is the dynamic viscosity of the molten steel fluid. is the gravity. The Reynolds number of the Navier-Stokes equation is calculated according to the actual equipment parameters of the crystallizer and the tundish.

[0034] ​The training set used by the tundish metallurgical state rapid prediction model is a combination of water model experimental data and simulation data based on experimental verification. The specific data structure is as follows:

[0035]

[0036] in, is the time data, is the stopper height, It is the tundish liquid level. is the complete flow field information, respectively Directional speed, Directional velocity and pressure field.

[0037] At the same time, in order to improve the training efficiency and stability of the model, the input data is normalized so that the data is distributed in the range [0 1]. Normalization is achieved by subtracting the mean of each input feature and dividing it by the standard deviation of the feature. The specific formula is:

[0038]

[0039] in, For the The result after normalization of the observations. is the mean of feature X, is the standard deviation of feature X.

[0040] The training of the tundish metallurgical state rapid prediction model adopts a composite loss function, which includes two parts: data loss and physical equation residual loss. Among them, the data loss uses the mean square error ( ) calculation, the residual loss of the equation is also used Calculate and ensure that the model output satisfies the Navier-Stokes equation, which is:

[0041]

[0042] Where, is the key parameter variable used to calculate MSE, which is the dynamic pressure in this model ( ) or flow field , Speed ​​data in direction ( , ). is the number of sampling points, is the predicted value, is the true value; For the sampling points.

[0043] Furthermore, the training of the rapid prediction model of the tundish metallurgical state is constructed based on the pressure, flow rate data and the corresponding partial derivative errors. function, Function according to dynamic pressure , flow field , Speed ​​data in direction ( , ) and the corresponding partial derivatives ( ), as shown in the following formula.

[0044]

[0045] in, 、 、 、 is a custom weight parameter.

[0046] Furthermore, the Adam optimizer is used for training the rapid prediction of the tundish metallurgical state. This optimizer adaptively adjusts the learning rate, which helps the model converge effectively in complex loss landscapes.

[0047] The pre-trained tundish metallurgical state rapid prediction model accesses data collected by the data acquisition system in real time via a data interface. After calculation, it outputs the predicted results, including the complete flow field information of the molten steel and the nozzle flow velocity calculated based on this flow field information. This complete flow field information is displayed in real-time as an image, facilitating manual monitoring of the metallurgical state within the tundish. It also visualizes parameters collected by the aforementioned sensors, such as the pressure field, tundish temperature, molten steel flow velocity, and inlet and outlet flow rates. The nozzle flow velocity information is transmitted to the mold level BP-PID control unit via the data interface.

[0048] The crystallizer liquid level control part includes two parts: BP neural network and liquid level PID controller. The output layer of the BP neural network has three neurons, which correspond to the three parameters of PID respectively. The input of the BP neural network is the actual height of the current crystallizer liquid level, the error between the nozzle flow rate predicted by the intermediate flow field PINN prediction model and the height set by the system, the casting speed, the casting width and other process parameters. The output is expressed as:

[0049]

[0050] in, for Time system error (the difference between the set liquid level height of the crystallizer and the actual liquid level height).

[0051] They are proportional, integral, and differential gains respectively.

[0052] The mold level control part uses the back propagation algorithm to adjust the weights and biases of the neural network to minimize the error signal of the mold level as the basis for updating parameters. The parameter update of the mold level BP-PID controller is as follows:

[0053]

[0054] Among them, the hidden layer activation function is the hyperbolic tangent function , . Output layer activation function is a non-negative hyperbolic tangent function , . is the matrix of hidden layer coefficients, is the inertia coefficient, is the learning rate. Input vector for the input layer, is the hidden layer output vector, k is the current iteration number, is the weight coefficient of input layer-hidden layer, The weight coefficient of the hidden layer-output layer, is the weight change after update, is the increment of system output, is the system error, y is the actual output value, is the input of the output layer, is the hidden layer input, and the superscript brackets are the dimensions.

[0055] During the control process, the BP neural network calculates the PID parameters according to the current system status (liquid level, nozzle flow rate, pulling speed, etc.) The PID controller uses these parameters to adjust the stopper target height, outputs a more reasonable stopper target height, and transmits it to the stopper position automatic control unit. After receiving the stopper target height and the actual stopper height, the stopper position PID controller generates an accurate control signal for the stopper actuator. The stopper actuator then controls the stopper to quickly and accurately reach the control position, changing the flow field near the stopper and the nozzle, thereby affecting the flow of molten steel entering the crystallizer through the nozzle, achieving the purpose of controlling the crystallizer liquid level. This effectively addresses the various dynamic impacts brought about by the frequent changes in production parameters during the production process, ensuring the stability of product quality.

[0056] The above examples describe specific embodiments of the present application, but it is understood that the specific description herein should not be construed to limit the spirit and scope of the present application, and various modifications to the above examples made by those skilled in the art after reading the description are within the scope of the present application.

Claims

1. A rapid active control system for the mold liquid level in a continuous casting process based on mechanism fusion AI, characterized by: It includes real-time data acquisition part, tundish flow field PINN prediction part, mold liquid level BP-PID control part and stopper position PID controller; The real-time data acquisition system is used to collect key parameters of the continuous casting process in the continuous casting production system in real time, including the liquid level in the tundish, the liquid level height of the crystallizer, the casting speed, the width of the cast billet, the insertion depth of the large ladle nozzle, the remaining steel amount in the tundish, and the height of the stopper rod, and send the parameters related to the large ladle and the tundish to the PINN prediction model of the tundish flow field, and send the parameters collected by the crystallizer-related parameter equipment and the continuous casting process setting parameters to the BP-PID controller; The PINN prediction part of the tundish flow field adopts a PINN neural network, which consists of multiple fully connected layers, with 3 neurons in the input layer, 32 neurons in each of the 12 hidden layers, and 3 neurons in the output layer. The two-dimensional Navier-Stokes equations of incompressible fluid are used as constraints. The training set used by the PINN neural network is a collection of water model experimental data and simulation data based on experimental verification. A composite loss function is used in the training process, including data loss and physical equation residual loss. At the same time, the dynamic pressure is used as the training set. , flow field , Speed ​​data in direction ( , ) and the corresponding partial derivative errors to construct the Loss function; The optimizer uses Adam; The PINN prediction part of the tundish flow field obtains data collected by the data acquisition system in real time, performs calculations, and outputs the prediction results as complete flow field information and the nozzle flow velocity calculated based on the flow field information. The nozzle flow velocity information is transmitted to the crystallizer liquid level BP-PID control part through the data interface. The crystallizer liquid level BP-PID control part includes two parts: BP neural network and liquid level PID controller; wherein, the output layer of BP neural network has three neurons, corresponding to the three parameters of PID respectively. The input of the BP neural network is the actual height of the current crystallizer liquid level, the water inlet flow rate predicted by the intermediate flow field PINN prediction model, and the parameters set by the system; BP neural network calculates PID parameters based on the current system state , updates the PID parameters in real time according to the changes in the stopper rod height and transmits the updated parameters to the liquid level PID controller; the liquid level PID controller uses these parameters to adjust the stopper rod target setting height, outputs a more reasonable stopper rod target setting height, and transmits it to the stopper rod position PID controller; the stopper rod position PID controller generates an accurate control signal for the stopper rod actuator after receiving the stopper rod target setting height and the actual stopper rod height, and then the stopper rod actuator controls the stopper rod to reach the control position quickly and accurately.

2. The active rapid control system for mold level in a continuous casting process based on mechanism fusion AI as claimed in claim 1, characterized in that: In the PINN neural network, after each hidden layer, the hyperbolic tangent activation function is used to increase the nonlinear expression ability of the network, specifically: 。 3. The active rapid control system for mold level in a continuous casting process based on mechanism fusion AI as claimed in claim 1, characterized in that: The Navier-Stokes equations for a two-dimensional incompressible fluid are: ； ； Where, is the density of molten steel; is the velocity vector, for Direction speed, The direction is the mainstream direction of molten steel along the length of the tundish; for Direction (perpendicular to direction) of the speed; is the dynamic pressure; is the dynamic viscosity of the molten steel fluid; is gravity; the Reynolds number of the Navier-Stokes equation is calculated based on the actual equipment parameters of the crystallizer and the tundish.

4. The active rapid control system for mold level in a continuous casting process based on mechanism fusion AI as claimed in claim 1, characterized in that: The data loss and physical equation residual loss in the composite loss function are both calculated using the mean square error (MSE) to ensure that the model output satisfies the Navier-Stokes equation. The formula is: ； Where, is the key parameter variable used to calculate MSE; is the number of sampling points, is the predicted value, is the true value; For the sampling points.

5. The active rapid control system for the mold liquid level in a continuous casting process based on mechanism fusion AI as claimed in claim 1, characterized in that: The function is specifically: ； in, 、 、 、 is a custom weight parameter.

6. The active rapid control system for the mold liquid level in a continuous casting process based on mechanism fusion AI as claimed in claim 1, characterized in that: The training data structure is: ； in, is the time data, is the stopper rod height, is the tundish liquid level; is the complete flow field information, respectively Directional speed, Directional velocity and pressure field.

7. The active rapid control system for mold level in a continuous casting process based on mechanism fusion AI as claimed in claim 1, characterized in that: The input data of the PINN prediction part of the intermediate package flow field is normalized so that the data is distributed in the [01] interval. The normalization is achieved by subtracting the mean of each input feature and dividing it by the standard deviation of the feature, specifically: ； in, is the normalized result of the i-th observation value; is the mean of feature X, is the standard deviation of feature X.

8. The active rapid control system for mold level in a continuous casting process based on mechanism fusion AI as claimed in claim 1, characterized in that: The BP-PID control part of the crystallizer liquid level adjusts the weight and bias of the BP neural network through the back propagation algorithm to minimize the error signal of the crystallizer liquid level as the basis for updating the parameters. The method is as follows: ； Among them, the hidden layer activation function is the hyperbolic tangent function , ; Output layer activation function is a non-negative hyperbolic tangent function , ; is the matrix of hidden layer coefficients, is the inertia coefficient, is the learning rate; Input vector for the input layer, is the hidden layer output vector, k is the current iteration number, is the weight coefficient of input layer-hidden layer, The weight coefficient of the hidden layer-output layer, is the weight change after update, is the increment of system output, is the system error, y is the actual output value, is the input of the output layer, is the hidden layer input, and the superscript brackets are the dimensions.

Images