Method and system for intelligently controlling argon blowing flow in LF furnace heating process

Abstract

The invention provides a method and system for intelligently controlling the argon blowing flow in the LF furnace heating process, and belongs to the field of steelmaking. The method comprises the steps that a training data set is formed based on the flow and liquid level height of argon blown into a steel ladle in the bottom blowing stage in the LF furnace heating process; based on the data set, performing iterative training through a gradient descent optimization algorithm in a PyTorch built-in optimizer to obtain a prediction curve model; determining the argon flow corresponding to the target liquid level height based on the prediction curve model; in the heating process, the liquid level height is monitored in real time, and the argon flow is adjusted through the liquid level height. The invention avoids the loss of molten steel and the reduction of heating efficiency, ensures the safety and stability of production, reduces the production cost and relieves the labor intensity of workers.

Description

Technical Field

[0001] The present invention belongs to the field of steelmaking, and particularly relates to a method and system for intelligently controlling the argon blowing flow rate during the heating process of an LF furnace. Background Art

[0002] With the rapid development of the steel industry, significant progress has been made in converter smelting technology, secondary refining technology, and continuous casting technology. As a key transitional link between converter steelmaking and continuous casting processes, the LF furnace has been widely used. LF refining has important functions such as adjusting the temperature and composition of molten steel and regulating the production rhythm, and has become the main secondary refining means.

[0003] During the heating process of LF refining, argon blowing and stirring are indispensable. A reasonable argon flow rate can ensure sufficient stirring work, enabling the molten steel in the upper and lower parts of the ladle to be fully mixed and achieving effective heating. However, there are many problems in the actual production during the heating of the LF furnace. When intense heating occurs with submerged arc discharge, if the argon blowing and stirring control is improper, the molten steel level may fluctuate violently, leading to problems such as molten steel foaming, overflow of molten steel, overflow of slag, and splashing, which not only result in low heating efficiency but also cause a large amount of metal charge loss.

[0004] Currently, the flow control level during the heating process of the LF furnace is limited. The "Automatic Control System and Control Method for Soft Argon Blowing at the Bottom of the Ladle in an LF Furnace" proposed in Chinese Patent CN201911194214.6 in 2019 is only applicable to the soft blowing process of LF furnace refining and cannot meet the control requirements for the argon blowing flow rate during the heating process, with limited application scope. Currently, the flow control during the heating process of the LF furnace is mostly in the fuzzy control stage, making it difficult to accurately regulate the argon blowing flow rate, and the situation of excessive slag overflow due to too rapid increase in the liquid level occurs frequently. In domestic steel mills, in order to solve the problem of slag overflow, measures such as adjusting the composition of the refining slag, the viscosity of the slag, the thickness of the slag layer, or changing the number of bottom blowing gas pipelines are often taken, but these methods increase the process complexity and workload and also raise the production cost. Therefore, there is an urgent need for a method that can reasonably control the argon blowing flow rate during the LF heating process to improve production efficiency, reduce costs, and reduce metal charge loss. Summary of the Invention

[0005] The purpose of the present invention is to realize a method for controlling the argon blowing flow rate during the heating process of an LF furnace, which can automatically control a reasonable argon blowing flow rate during the heating process and play a role in providing the best metallurgical reaction conditions for the LF heating process.

[0006] The technical solution of the present invention is as follows:

[0007] A method for intelligently controlling the argon blowing flow rate during the heating process of an LF furnace, the method comprising:

[0008] Constructing a training data set based on the magnitude of the argon flow rate blown into the ladle during the bottom blowing stage and the liquid level height during the heating process of the LF furnace;

[0009] An iterative training is performed on the basis of the training dataset by using the gradient descent optimization algorithm in the built-in optimizer of PyTorch to obtain a prediction curve model;

[0010] Based on the prediction curve model, the argon flow rate corresponding to the target liquid level height is determined;

[0011] During the heating process, the liquid level height is monitored in real time, and the argon flow rate is adjusted according to the liquid level height.

[0012] Further, the iterative training based on the dataset by using the gradient descent optimization algorithm in the built-in optimizer of PyTorch to obtain a prediction curve model is specifically as follows:

[0013] Define a linear model using PyTorch, with the argon flow rate as the input feature and the liquid level height as the target output;

[0014] The mean squared error is used as the loss function, and the stochastic gradient descent optimizer is used for parameter update;

[0015] Set the number of iterative training times. In each iteration, a certain number of samples are randomly selected from the training dataset as a mini-batch for training, and finally a prediction curve model is obtained.

[0016] Further, the gradient descent optimization algorithm is specifically as follows:

[0017] x t = x t-1 - α * ▽f Bt (x t-1 )

[0018] where x t represents the gradient at time step t, x t-1 represents the gradient at time step t - 1, α represents the learning rate, and ▽f Bt (x t-1 ) represents the mini-batch stochastic gradient obtained by repeated sampling.

[0019] Further, it also includes that when it is monitored that the liquid level height exceeds the critical liquid level of the LF furnace, the argon blowing flow rate is reduced.

[0020] Further, the liquid level height is the sum of the molten steel liquid level and the upper slag level height of the molten steel.

[0021] Further, it also includes a manual control mode: manually set the argon flow rate corresponding to the target liquid level height according to the prediction curve model to realize the manual adjustment of the argon blowing flow rate.

[0022] A system for intelligently controlling the argon blowing flow rate during the heating process of an LF furnace, the system includes:

[0023] Data acquisition module: used to obtain the argon gas flow rate and liquid level height data blown into the ladle during the bottom blowing stage in the LF furnace heating process, and construct a training data set;

[0024] Model training module: used to iteratively train through the gradient descent optimization algorithm in the built-in optimizer of PyTorch based on the data set to obtain a prediction curve model;

[0025] Flow rate determination module: determine the argon gas flow rate corresponding to the target liquid level height based on the prediction curve model;

[0026] Liquid level monitoring and flow rate control module: during the heating process, monitor the liquid level height in real time, and adjust the argon gas flow rate according to the liquid level height.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] The present invention uses a deep learning optimization algorithm to achieve precise control of the argon gas blowing flow rate, effectively avoiding problems such as slag overflow. By collecting the argon gas flow rate and liquid level height data during the bottom blowing stage to construct a training data set, and iteratively training through the gradient descent algorithm of the built-in optimizer of PyTorch to obtain a prediction curve model, the argon gas flow rate corresponding to the target liquid level is determined accordingly. At the same time, the liquid level height is monitored in real time. Once the critical liquid level is exceeded, the system automatically reduces the argon gas blowing flow rate. This method precisely regulates the argon gas blowing flow rate, maintains the liquid level stability, reduces the phenomena of slag overflow and slag splashing, avoids steel liquid loss and reduction of heating efficiency, ensures the safe and stable production, and also reduces the production cost and lightens the labor intensity of workers. Brief description of the drawings

[0029] The drawings generally illustrate various embodiments by way of example rather than limitation, and are used together with the description of the specification and the claims to illustrate the embodiments of the invention. Where appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be an exhaustive or exclusive embodiment of the device or method.

[0030] Figure 1 Shows a schematic diagram of the LF furnace of the present invention;

[0031] Figure 2 Shows a schematic diagram of the method flow of the present invention.

[0032] Reference numerals:

[0033] 1. High-temperature metal liquid level sensor; 2. Ladle furnace, 3. On-site control box, 4. On-site computer, 5. On-site display console, 6. Gas control cabinet, 7. Flow control valve, 8. Ladle cover, 9. Graphite electrode is 9. Detailed implementation manners

[0034] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0035] As Figure 1 - Figure 2 shown, an embodiment of the present invention provides a method for intelligently controlling the argon blowing flow rate during the heating process of an LF furnace, including the following steps:

[0036] (1) Collect the argon gas flow rate and the liquid level height value blown into the ladle during the bottom blowing stage in the LF heating and refining process during a large number of previous production processes.

[0037] (2) Take the argon gas flow rate and the liquid level height as two core training data parameters, and after 19,000 iterations of gradient descent optimization processing by the internal optimizer of PyTorch, a prediction curve representing the heating bottom blowing stage is generated.

[0038] The principle formula of the gradient descent algorithm is calculated according to the relationship shown in Equation (1):

[0039]

[0040] In the formula, x t represents the gradient at time step t, a represents the learning rate, represents the gradient of the loss function J(x) with respect to the parameter x t , represents the change trend and speed of J(x t ) at the current point, and t is set to 0 at the time step before the iteration starts.

[0041] Based on Equation 1, the mini-batch stochastic gradient descent algorithm is selected

[0042]

[0043] In the formula, t represents the time step, Bt represents the mini-batch composed of the training data sample indices, |B| represents the size of the batch, that is, the number of samples in the mini-batch, represents the gradient at x t-1 , represents the mini-batch stochastic gradient obtained by repeated sampling.

[0044] Substitute Equation (2) into Equation (1) to obtain the applied general gradient descent formula (3).

[0045]

[0046] In the formula, x t represents the gradient at time step t, x t-1 represents the gradient at time step t - 1, α represents the learning rate, It represents the mini-batch stochastic gradient obtained from repeated sampling.

[0047] In the present invention, the built-in optm module and troch.nn module in PyTorch are mainly adopted.

[0048] In the torch.nn module, nn.Linear is used to define a linear layer of a neural network to linearly transform the input features. The loss function Loss is the mean squared error between the model output and the target value.

[0049] The Stochastic Gradient Descent (SGD) optimizer provided by optim is used for parameter update, and the number of training epochs is set to 19000.

[0050] After the training is completed, the fitted curve model image is displayed in the PyCharm development environment. From this image, the relationship between the blowing rate and the liquid level can be intuitively seen.

[0051] (3) Through the function in PyTorch for obtaining training data, the training data is retrieved and stored in an array.

[0052] (4) The control program analyzes according to the predicted data generated in the heating bottom-blowing stage, and sends the preferred predicted gas flow information to the on-site control box 3 and the gas control cabinet 6.

[0053] (5) The control program sends the generated predicted image information to the on-site display console 5. The on-site operator can manually set the gas flow corresponding to the expected target liquid level height according to the predicted image on the display, and send the set value to the gas control cabinet 6 through the on-site control box 3.

[0054] (6) If there is no on-site worker for manual regulation, the control program based on the PyTorch deep learning framework will automatically send the gas flow information that best conforms to the actual normal bottom-blowing stage to the gas control cabinet.

[0055] In the present invention, the implementation schemes are divided into three groups, and the ladle furnace 2 with different slag levels is heated and blown respectively.

[0056] In the implementation scheme, first, the argon blowing flow required in the heating stage is set, and then after a certain heating time, it is tested whether there are harmful phenomena such as liquid level overflow and splashing inside the ladle; if there is a slag overflow phenomenon, whether the ladle intelligent control system can effectively handle the slag overflow situation in a short time.

[0057] In this implementation scheme, the argon blowing flow rate during the entire heating process is interacted by a computer based on deep learning software through control signals with the gas control cabinet 6, the on-site control box 3, and the flow control valve 7. Except for special abnormal situations, the entire heating process does not require manual intervention. During the heating time, the argon blowing flow rate is continuously adjusted as the deep learning software continuously self-learns.

[0058] In the present invention, the high-temperature metal liquid level sensor 1 sends an alarm signal to the on-site control box 3. The control box 3 converts the digital signal into a control signal and sends it to the on-site computer 4. After receiving the signal, the on-site computer 4 converts the signal into a computer instruction. After the control program based on deep learning receives the instruction, it generates the current slag overflow alarm log, and the log content includes the alarm time, the blowing flow rate, and the liquid level height. The control program performs self-learning optimization based on the training data, selects the optimal gas flow rate, and sends the flow rate information to the on-site control box 3 for emergency flow rate adjustment to avoid the overflow of molten steel.

[0059] In the present invention, the control program generates an initial prediction curve based on the optimal gas flow rate and the molten steel liquid level height value in a large amount of past production data. The control program transmits the prediction curve image information to the on-site display console 5. The on-site workers estimate the current molten steel height and the stirring degree of the molten steel based on the prediction image of the on-site display console 5, and further select the on-site control box 3 for real-time flow rate adjustment. Set the starting argon blowing flow rate through the on-site control box 3. The deep learning software starts training based on the starting flow rate. After 19,000 iterations, it generates the predicted flow rate and the predicted liquid level height. The program sets the predicted argon blowing flow rate required for this production again according to the predicted values.

[0060] If the worker does not perform further operations, the on-site control box will directly execute the predicted flow rate provided by the control program. Set the newly trained predicted flow rate as the initial argon flow rate for a new round.

[0061] Table 1 of the implementation scheme of this embodiment

[0062]

[0063] It can be seen from the implementation scheme table that there are 2 times of slag overflow in furnace 1, while there is no slag overflow in furnace 2 and furnace 3. The total duration of slag overflow treatment in furnace 1 is 4 minutes, indicating that the system can make real-time flow rate adjustment in case of slag overflow, and the liquid level adjustment is below the critical liquid level within 2 minutes each time.

[0064] As shown in Table 1 of the refining parameters of each furnace in the on-site implementation scheme

[0065] In Example 1, the bottom blowing time is 20 minutes, the initial height of the molten steel is 3013 mm, and the first gas flow rate is set to 160 L / min.

[0066] During the bottom blowing process of Example 1, the generated prediction data is displayed on the monitor. Workers set the bottom blowing parameters respectively by operating the on-site controller. The deep learning software automatically regulates the argon blowing flow rate throughout the process, records the data after the bottom blowing ends, and compares the real value with the predicted height on the monitor.

[0067] During the process of Example 1, the slag overflow treatment time is 4 minutes, and the ladle capacity is 100t.

[0068] Table 1 of the examples shows the parameters and slag overflow results of three groups of examples.

[0069] It can be seen from Table 1 of the implementation plan that in Example 1, after 20 minutes, there are 2 times of slag overflow and 1 time of slag overflow treatment. However, there are no abnormal situations in ladle 2 and ladle 3, indicating that the self-learning control program has the function of adjusting the flow rate in real time.

[0070] During the implementation process, the implementation plans of Examples 1 to 3 were successfully completed. According to the data in Table 1, only in Example 1, the situation of exceeding the critical height occurred during heating, and it was necessary to carry out slag pressing treatment twice. At this time, the computer in the main control room received an alarm signal, and the gas control cabinet carried out slag pressing treatment on it.

[0071] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered within the protection scope of the present invention.

Claims

1. A method for intelligently controlling the argon blowing flow rate during the heating process of an LF furnace, characterized in that, The method includes: Constructing a training data set based on the magnitude of the argon gas flow rate and the liquid level height blown into the ladle during the bottom blowing stage in the LF furnace heating process; Iteratively training through the gradient descent optimization algorithm in the built-in optimizer of PyTorch based on the training data set to obtain a prediction curve model; Determining the argon gas flow rate corresponding to the target liquid level height based on the prediction curve model; During the heating process, the liquid level height is monitored in real time, and the argon gas flow rate is adjusted according to the liquid level height.

2. The method for intelligently controlling the argon blowing flow rate during the heating process of the LF furnace according to claim 1, wherein, The specific process of iteratively training through the gradient descent optimization algorithm in the built-in optimizer of PyTorch based on the data set to obtain a prediction curve model is as follows: Defining a linear model using PyTorch, with the argon gas flow rate as the input feature and the liquid level height as the target output; Adopting the mean square error as the loss function and using the stochastic gradient descent optimizer to update the parameters; Setting the number of iterative training times. In each iteration, a certain number of samples are randomly selected from the training data set as a mini-batch for training, and finally a prediction curve model is obtained.

3. The method for intelligently controlling the argon blowing flow rate during the heating process of the LF furnace according to claim 1, wherein, The specific gradient descent optimization algorithm is as follows: x t = x t-1 -α*▽f Bt (x t-1 ) where x t represents the gradient at time step t, x t-1 represents the gradient at time step t - 1, α represents the learning rate, and ▽f Bt (x t-1 ) represents the mini - batch stochastic gradient obtained by repeated sampling.

4. The method for intelligently controlling the argon blowing flow rate in the heating process of an LF furnace according to claim 1, wherein, It also includes reducing the argon gas blowing flow rate when it is monitored that the liquid level height exceeds the critical liquid level of the LF furnace.

5. The method for intelligently controlling the argon blowing flow rate during the heating process of an LF furnace according to claim 1, characterized in that, The liquid level height is the sum of the molten steel liquid level and the upper slag level height of the molten steel.

6. The method for intelligently controlling the argon blowing flow rate in the LF furnace heating process according to claim 1, characterized in that, It also includes a manual control mode: manually setting the argon gas flow rate corresponding to the target liquid level height according to the prediction curve model to achieve manual adjustment of the argon gas blowing flow rate.

7. An intelligent system for controlling the argon blowing flow rate during the heating process of an LF furnace, characterized in that, The system includes: A data acquisition module: used to obtain the data of the argon gas flow rate magnitude and the liquid level height blown into the ladle during the bottom blowing stage in the LF furnace heating process and construct a training data set; A model training module: used to iteratively train through the gradient descent optimization algorithm in the built-in optimizer of PyTorch based on the data set to obtain a prediction curve model; A flow rate determination module: determining the argon gas flow rate corresponding to the target liquid level height based on the prediction curve model; A liquid level monitoring and flow rate control module: during the heating process, the liquid level height is monitored in real time, and the argon gas flow rate is adjusted according to the liquid level height.

Images