An intelligent control method and related device for bottom argon blowing of ladle based on neural network

Abstract

The present invention provides an intelligent control method and related device for bottom argon blowing of ladles based on neural networks, which relates to the technical field of data processing. The liquid surface image of the ladle is collected in real time by a high-temperature industrial camera. After preprocessing, a semantic segmentation model is used to identify the liquid surface area. A specific algorithm is combined to extract the liquid surface boundaries corresponding to the north and south argon holes, and the independent dynamic adjustment of the flow rates of the two holes is realized through a closed-loop control system. The system integrates an image processing module, a deep learning module and an automatic control module, breaking through the hysteresis of traditional manual adjustment and solving the technical problems of inaccurate liquid surface state monitoring and uneven argon gas distribution. Compared with the prior art, it can improve the control accuracy of argon gas flow rate, enhance the uniformity of molten steel composition, reduce the intensity of manual intervention at the same time, and effectively ensure the stability of molten steel quality and production safety in a high-temperature and harsh environment.

Description

Technical Field

[0001] The present application relates to the technical field of data processing, and in particular to an intelligent control method and related device for bottom blowing argon in a ladle based on a neural network. Background Art

[0002] Bottom blowing argon in a ladle is a common means of refining molten steel in steelmaking production. By stirring the molten steel through bottom blowing argon, metallurgical functions such as uniform composition and temperature of the molten steel, floating of molten steel inclusions, and desulfurization of molten steel can be achieved.

[0003] Under different metallurgical process conditions, the stirring intensity required for bottom blowing argon is different: during desulfurization at the steel slag interface, heating of molten steel, and alloy adjustment, strong stirring is required to achieve rapid desulfurization, uniform composition and temperature; while to achieve the floating of molten steel inclusions, weak stirring is needed to prevent the molten steel from being exposed to cause oxygen absorption, nitrogen absorption, and slag entrainment of molten steel. The stirring intensity of molten steel not only depends on the magnitude of the bottom blowing pressure and flow rate, but also the air permeability state of the bottom blowing porous brick, gas leakage in the pipeline, etc. are important factors affecting the stirring effect. There are significant differences in air permeability for different ladles and different heats in the same ladle. That is to say, when using a certain flow rate for bottom blowing argon in a ladle, the stirring effect of bottom blowing argon is different in different ladles, and even in different stages of the same ladle. Therefore, it is impossible to achieve precise control of bottom blowing argon stirring under different metallurgical functions with standard flow rates and pressures. This requires the operators in the position of bottom blowing argon stirring to adjust the opening of the bottom blowing argon pipeline valve according to the air permeability of the ladle at different metallurgical process stages for different ladles to meet the metallurgical function requirements. Currently, the current process mainly relies on manual adjustment, and there are the following problems:

[0004] Large manual adjustment error: The traditional adjustment of the bottom blowing argon flow rate in a ladle relies on manual observation of the liquid level, with a lag in adjustment, inconsistent adjustment standards among different teams, and it is difficult to ensure the accuracy and stability of the flow rate.

[0005] High manual labor intensity: The temperature near the ladle is high and the environment is harsh. Once the production rhythm is tight, workers need to frequently go to the site to monitor the liquid level state.

[0006] Uneven argon consumption: Since the north and south argon holes work independently, it is difficult to ensure the reasonable distribution of argon flow rate, which may lead to uneven stirring or the destruction of the liquid level stability by bubbles.

[0007] Therefore, how to improve the control accuracy of argon flow rate and ensure the stability of molten steel quality and production safety in a harsh environment has become an urgent technical problem to be solved. Summary of the Invention

[0008] In order to improve the control accuracy of argon flow rate and ensure the stability of molten steel quality and production safety in a harsh environment, the present application provides an intelligent control method and related device for bottom blowing argon in a ladle based on a neural network.

[0009] In a first aspect, a smart control method for ladle bottom argon blowing based on a neural network adopts the following technical solutions:

[0010] A smart control method for ladle bottom argon blowing based on a neural network includes:

[0011] Real-time collect the ladle liquid surface image through a high-temperature industrial camera with air-cooled protection, and adjust the camera parameters to ensure clear images;

[0012] Adjust the collected images through methods such as image cropping, denoising, and adaptive histogram equalization to obtain optimized image data;

[0013] Construct a reference data set in the optimized image data through data annotation;

[0014] Use the PP-LiteSeg-T model to train the reference data set to generate a molten steel liquid surface recognition model;

[0015] Input the optimized image data into the molten steel liquid surface recognition model to output a binary image, and use the findContours algorithm to extract the liquid surface boundary;

[0016] Based on Hough circle detection combined with the liquid surface boundary to distinguish the liquid surface areas corresponding to the north and south argon holes, and calculate the area ratio of each region;

[0017] According to the liquid surface area ratio combined with the empirical formula, independently calculate the flow rates of the north and south argon holes, and dynamically adjust through a closed-loop control system.

[0018] Optionally, the step of adjusting the collected images through methods such as image cropping, denoising, and adaptive histogram equalization to obtain optimized image data includes:

[0019] Crop the collected images to 512×512 pixels to balance processing speed and accuracy;

[0020] Use Gaussian filtering or median filtering algorithm to remove high-temperature noise and random interference;

[0021] Use the adaptive histogram equalization method to divide the image into local regions for contrast enhancement to obtain optimized images.

[0022] Optionally, the step of using the adaptive histogram equalization method to divide the image into local regions for contrast enhancement to obtain optimized images includes:

[0023] Divide the image into local regions, and the local regions include 8x8 regions or 16x16 regions;

[0024] Perform local histogram equalization on each region in the local area;

[0025] Apply contrast limitation to each region;

[0026] Use an interpolation method to smoothly transition the processing results between adjacent regions;

[0027] Among them, the steps of local histogram equalization include:

[0028] Let the gray value in a certain region be , and its frequency distribution be . The process of equalizing this region is to calculate the cumulative distribution function for each gray value:

[0029]

[0030] Among them, is the frequency of the gray level in this region, where is the index of the block;

[0031] Use the cumulative distribution function for mapping to map the original gray value to the new gray value :

[0032]

[0033] Among them is the number of gray levels, is the maximum cumulative distribution in this region;

[0034] Among them, the steps of applying contrast limitation include:

[0035] Let the frequency of a certain gray level exceed the set contrast limitation value , then limit it to , and evenly distribute the remaining frequency to all gray levels;

[0036] The limited frequency is calculated as:

[0037]

[0038] Among them is the frequency of the gray level , is the frequency after limitation, and the remaining over-limit part is calculated through the following formula:

[0039]

[0040] The excess frequencies are evenly distributed among the other gray levels.

[0041] Optionally, in the step of training the benchmark dataset using the PP-LiteSeg-T model to generate the molten steel liquid level recognition model, the hyperparameters of the PP-LiteSeg-T model are set as follows: the number of iterations is 50,000 times, the initial learning rate is 0.005, transfer learning uses pre-trained weights, and data augmentation includes rotation, scaling, and flipping.

[0042] Optionally, in the step of inputting the optimized image data into the molten steel liquid level recognition model to output a binary image and using the findContours algorithm to extract the liquid level boundary, the parameter configuration of the findContours algorithm is as follows: the contour retrieval mode is cv2.RETR_LIST, the contour approximation method is cv2.CHAIN_APPROX_SIMPLE, and the liquid level boundary contour list is extracted.

[0043] Optionally, in the step of distinguishing the liquid level regions corresponding to the north and south argon holes based on the Hough circle detection in combination with the liquid level boundary and calculating the area ratio of each region, the north-south hole discrimination algorithm includes:

[0044] For each edge point in the image, it is represented as a point in the parameter space of a circle , that is, the center coordinates and the radius ;

[0045] Centered on the y coordinate of the center, the points in the liquid level boundary with a y coordinate greater than the y coordinate of the center are divided into the north side, and the points with a y coordinate less than the y coordinate of the center are divided into the south side;

[0046] The points on the north and south sides are respectively formed into two independent regions, that is, the distinction between the north and south argon holes is completed;

[0047] The areas of the irregular figures in the north and south regions are calculated using the Gaussian area:

[0048]

[0049] where: is the coordinate of the th point of the contour, is the coordinate of the next point, is the number of points in the contour;

[0050] This area is divided by a preset standard value to obtain the proportion of the liquid level region.

[0051] Optionally, the step of independently calculating the flow rates of the north and south argon holes according to the proportion of the liquid level region in combination with the empirical formula and dynamically adjusting through the closed-loop control system includes:

[0052] Independently calculate the north-south argon nozzle flow rates using a preset model based on the proportion of the liquid level area and the empirical relationship between the ladle liquid level and the argon flow rate;

[0053] Input the north-south argon nozzle flow rates into a closed-loop control system so that the control module in the closed-loop control system dynamically adjusts the valve opening to control the fluctuation of the ladle liquid level.

[0054] In a second aspect, the present application provides an intelligent control device for bottom argon blowing of a ladle based on a neural network, including:

[0055] An image acquisition module for real-time collecting ladle liquid level images through a high-temperature industrial camera with air-cooled protection and adjusting the camera parameters to ensure clear images;

[0056] An image optimization module for adjusting the collected images by methods such as image cropping, denoising processing, and adaptive histogram equalization to obtain optimized image data;

[0057] A dataset construction module for constructing a reference dataset by data annotation in the optimized image data;

[0058] A model training module for training the reference dataset using the PP-LiteSeg-T model to generate a molten steel liquid level recognition model;

[0059] A recognition module for inputting the optimized image data into the molten steel liquid level recognition model to output a binary image and using the findContours algorithm to extract the liquid level boundary;

[0060] A calculation module for distinguishing the liquid level areas corresponding to the north-south argon nozzles based on Hough circle detection in combination with the liquid level boundary and calculating the proportion of the area of each region;

[0061] An adjustment module for independently calculating the north-south argon nozzle flow rates according to the proportion of the liquid level area in combination with an empirical formula and dynamically adjusting through a closed-loop control system.

[0062] In a third aspect, the present application provides a computer device, the device includes: a memory, a processor, and when the processor runs the computer instructions stored in the memory, it executes the method as described above.

[0063] In a fourth aspect, the present application provides a computer-readable storage medium, including instructions, when the instructions run on a computer, it causes the computer to execute the method as described above.

[0064] In summary, the present application collects the image of the ladle liquid level in real time through a high-temperature industrial camera, uses a semantic segmentation model to identify the liquid level area after preprocessing, extracts the liquid level boundaries corresponding to the north and south argon holes by combining a specific algorithm, and realizes the independent dynamic adjustment of the double-hole flow through a closed-loop control system. The system integrates image processing, deep learning, and automatic control modules, breaks through the lag of traditional manual adjustment, and solves the technical problems of inaccurate liquid level state monitoring and uneven argon gas distribution. Compared with the prior art, it can improve the control accuracy of argon gas flow, enhance the uniformity of molten steel composition, reduce the intensity of manual intervention at the same time, and effectively ensure the stability of molten steel quality and production safety in a high-temperature and harsh environment. Brief Description of the Drawings

[0065] Figure 1 It is a schematic structural diagram of a computer device for the hardware operating environment involved in the solution of the embodiment of the present application;

[0066] Figure 2 It is a schematic flowchart of the first embodiment of the intelligent control method for bottom blowing argon in a ladle based on a neural network of the present application;

[0067] Figure 3 It is an optimized image of the first embodiment of the intelligent control method for bottom blowing argon in a ladle based on a neural network of the present application;

[0068] Figure 4 It is an image after semantic segmentation rendering of the intelligent control method for bottom blowing argon in a ladle based on a neural network of the present application;

[0069] Figure 5 It is a structural block diagram of the first embodiment of the intelligent control device for bottom blowing argon in a ladle based on a neural network of the present application. Detailed Embodiment

[0070] In order to make the objectives, technical solutions, and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0071] Refer to Figure 1 , Figure 1 which is a schematic structural diagram of a computer device for the hardware operating environment involved in the solution of the embodiment of the present application.

[0072] As Figure 1As shown in the figure, the computer device may include: a processor 1001, such as a Central Processing Unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard. Optionally, the user interface 1003 may further include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed Random Access Memory (RAM) or a stable Non-Volatile Memory (NVM), such as a disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0073] Those skilled in the art can understand that Figure 1 the structure shown in does not constitute a limitation on the computer device, and it may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0074] As Figure 1 shown, the memory 1005, as a storage medium, may include an operating system, a network communication module, a user interface module, and a ladle bottom blowing argon intelligent control program based on a neural network.

[0075] In Figure 1 the computer device shown, the network interface 1004 is mainly used for data communication with a network server; the user interface 1003 is mainly used for data interaction with a user; in the present application, the processor 1001 and the memory 1005 may be set in the computer device. The computer device calls the ladle bottom blowing argon intelligent control program stored in the memory 1005 through the processor 1001 and executes the ladle bottom blowing argon intelligent control method provided in the embodiments of the present application.

[0076] The embodiments of the present application provide a ladle bottom blowing argon intelligent control method based on a neural network. Referring to Figure 2 , Figure 2 it is a schematic flowchart of the first embodiment of the ladle bottom blowing argon intelligent control method based on a neural network in the present application.

[0077] In this embodiment, the ladle bottom blowing argon intelligent control method based on a neural network includes the following steps:

[0078] Step S10: Real-time collect the image of the ladle liquid level through a high-temperature industrial camera with air-cooled protection, and adjust the camera parameters to ensure clear images.

[0079] It should be noted that considering the influence of the high temperature of molten steel on the camera, the camera is equipped with an air-cooled protective cover and is mounted at a higher position using an adjustable independent bracket to reduce the influence of thermal radiation on the camera.

[0080] It can be understood that to ensure the clarity of image acquisition, it is necessary to adjust the camera parameters (focal length, exposure, white balance, etc.) and screen high-quality images for input into the subsequent processing link.

[0081] In a specific implementation, the air-cooled protective cover in this embodiment requires an inlet air temperature ≤ 45°C. During the image acquisition process, the position of the ladle is accurately aligned, and the obtained original images often have a certain degree of overexposure or underexposure, or out-of-focus blur, etc. Therefore, it is necessary to use the camera parameters of the camera to perform imaging adjustment on the original images.

[0082] It should be noted that a wide-temperature working camera is used and installed above the VD furnace. Since it is necessary to ensure that the viewing angle is as much as possible a top view rather than a skew view to ensure that the field of view covers the entire liquid level area, and considering the situation of high-temperature roasting, it is necessary to cooperate with an independent bracket to mount it at a higher position. The imaging requires that the image is not overexposed, underexposed and has a certain contrast. Use the camera parameters of the camera to perform imaging adjustment on the original images to ensure that clear original images are collected. After adjusting the camera parameters, the specific camera parameters in this example are: brightness = 13, contrast = 35, saturation = 30, sharpness = 20, aperture = f1.6, shutter = 1 / 25, gain = 10.

[0083] Step S20: Adjust the collected images through methods such as image cropping, denoising processing, and adaptive histogram equalization to obtain optimized image data.

[0084] It should be noted that the step of adjusting the collected images through methods such as image cropping, denoising processing, and adaptive histogram equalization to obtain optimized image data includes: cropping the collected images to 512×512 pixels to balance the processing speed and accuracy; using Gaussian filtering or median filtering algorithms to remove high-temperature noise and random interference; using the adaptive histogram equalization method to divide the image into local regions for contrast enhancement to obtain optimized images. The optimized images are as Figure 3 shown.

[0085] In a specific implementation, the step of using the adaptive histogram equalization method to divide an image into local regions for contrast enhancement to obtain an optimized image includes: dividing the image into local regions, where the local regions include 8x8 regions or 16x16 regions; performing local histogram equalization on each region in the local regions; applying contrast limitation to each region; and using an interpolation method to smoothly transition the processing results between adjacent regions.

[0086] Among them, the step of local histogram equalization includes:

[0087] Let the gray values in a certain region be , and its frequency distribution be . The process of equalizing this region is to calculate the cumulative distribution function for each gray value:

[0088]

[0089] Among them, is the frequency of the gray level in this region, where is the index of the block;

[0090] Use the cumulative distribution function for mapping to map the original gray value to the new gray value :

[0091]

[0092] Among them is the number of gray levels (usually 256), is the maximum cumulative distribution in this region;

[0093] Among them, the step of applying contrast limitation includes:

[0094] Let the frequency of a certain gray level exceed the set contrast limitation value , then limit it to , and evenly distribute the remaining frequency to all gray levels;

[0095] The limited frequency is calculated as:

[0096]

[0097] Among them is the frequency of the gray level , is the frequency after limitation, and the remaining over-limit part is calculated through the following formula:

[0098]

[0099] The excess frequencies are evenly distributed among the other gray levels.

[0100] In a specific implementation, the acquired original image data is frame - decoded to be transformed into picture data, and the frame - decoded data is used as input for image pre - processing. First, the image data is cropped to a size where the processing speed and accuracy are relatively balanced. In this example, the bilinear interpolation algorithm is used to infer a new size (512 512) from the pixels of the original image (1920 1080):

[0101] 1) Calculate the target pixel position: Each pixel position (x, y) in the target image is mapped to a position (x', y') in the source image.

[0102] 2) Select the four nearest pixels: Based on the target pixel position, the four nearest pixel points in the source image are selected (i.e., the 2x2 area near the target pixel).

[0103] 3) Weighted average: The weighted average of these four pixels is used to calculate the pixel value at this position in the target image. The weighting coefficients are assigned according to the distance from the target position, and the pixels closer to the target position have larger weights.

[0104]

[0105] : The abscissas of the two original - image pixels where the abscissa of the target position is located.

[0106] : The ordinates of the two original - image pixels where the ordinate of the target position is located.

[0107] Steps such as denoising processing and contrast enhancement are carried out to optimize the image quality, making the image information clearer and facilitating subsequent processing.

[0108] Among them, Gaussian filtering is used for denoising. For each pixel point, taking it as the center, the weighted average of all pixel gray - level values within its n n area is used as the gray - level value of the center point. For the weight calculation of each point, assuming that the pixel data basically conforms to a normal distribution, a two - dimensional Gaussian function is used:

[0109]

[0110] Calculate the weight of each pixel. In this way, the high - frequency components (image detail parts) of the image can be filtered, and the low - frequency components (image smooth areas) can be retained. Therefore, after the image is 'Gaussian blurred', the image becomes blurred, and then the noise is reduced and the main features become more obvious.

[0111] Step S30: Construct a benchmark data set from the optimized image data by means of data annotation.

[0112] In a specific implementation, in step S20, clear and complete data containing the full ladle liquid level is selected, and then data containing specific semantic categories is annotated for the area where the molten steel is exposed. The data containing specific semantic categories is constructed as the benchmark data set for the molten steel liquid level recognition model.

[0113] Constructing the benchmark data set: From the processed images obtained in S2, select data containing the full molten steel liquid level, and then annotate data containing specific semantic categories. The data containing specific semantics is constructed as the benchmark data set. Take data under different environmental conditions. The image data covers a large number of diverse data at different times, different weathers, different steel grades, and different temperatures. Annotation examples.

[0114] Step S40: Use the PP-LiteSeg-T model to train the benchmark data set to generate a molten steel liquid level recognition model.

[0115] It can be understood that due to the complexity and variability during steelmaking, the program operation must fully consider the balance between speed and accuracy. Therefore, PP-LiteSeg-T is used as the semantic segmentation basic model. Use the data set with specific semantics in step S30 as the input, train it in the PP-LiteSeg-T semantic segmentation model, obtain the weights and export the corresponding molten steel liquid level recognition model.

[0116] It should be noted that in the step of using the PP-LiteSeg-T model to train the benchmark data set to generate a molten steel liquid level recognition model, the hyperparameters of the PP-LiteSeg-T model are set as follows: the number of iterations is 50,000 times, the initial learning rate is 0.005, transfer learning uses pre-trained weights, and data augmentation includes rotation, scaling, and flipping. After completion, the exported model includes: inference.pdiparams, inference.pdiparams.info, inference.pdmodel, and inference.yml results. Then, through the create_model function, lightweight deployment of industrial-level applications can be achieved. The specific semantic segmentation rendered image is as Figure 4 shown.

[0117] Step S50: Input the optimized image data into the molten steel liquid level recognition model to output a binary image, and use the findContours algorithm to extract the liquid level boundary.

[0118] It can be understood that in the step of inputting the optimized image data into the molten steel liquid level recognition model to output a binary image and using the findContours algorithm to extract the liquid level boundary, the parameters of the findContours algorithm are configured as follows: the contour retrieval mode is cv2.RETR_LIST, the contour approximation method is cv2.CHAIN_APPROX_SIMPLE, and a list of liquid level boundary contours is extracted.

[0119] In a specific implementation, in combination with the molten steel liquid level recognition model obtained in step S40, the preprocessed image data in step S20 is input into the deep learning segmentation model, and the output result is a two-dimensional array (binary image) of corresponding color labels. Then, an edge detection algorithm (such as Canny, Sobel operator, findContours, etc.) is used. In this example, findContours is used to obtain the boundary of the liquid level. This method is specifically applicable to the image after binary processing. The contour retrieval mode is selected as cv2.RETR_LIST, and the contour approximation method is selected as cv2.CHAIN_APPROX_SIMPLE. Finally, the boundary contours are obtained.

[0120] Step S60: Based on the Hough circle detection, combine the liquid level boundary to distinguish the liquid level areas corresponding to the north and south argon holes, and calculate the area ratio of each area.

[0121] It should be noted that since the actual bottom blowing argon system consists of two argon holes in the north and south, the boundary obtained in S5 is further analyzed through the north-south hole discrimination algorithm to distinguish the independent areas of the north and south holes, ensuring the flow regulation of each hole.

[0122] It can be understood that since the overall tank mouth is approximately circular, the Hough circle detection is used to detect the tank mouth:

[0123] Parameter space of a circle: For each edge point in the image, it is considered to be a part of a circle. This point can be represented as a point in the parameter space of a circle , that is, the center coordinates and the radius .

[0124] Contribution of an edge point: If an edge point in the image is on a certain circle, then this point will contribute to the parameter space of possible center coordinates and radius. To find a circle, all possible centers and radii need to be calculated based on each edge point in the image.

[0125] By accumulating the contributions of all edge points to the parameter space, the most accumulated points can be found in the parameter space, indicating the existence of a circle in the image.

[0126] In a specific implementation, in the step of using the Hough circle detection to combine with the liquid level boundary to distinguish the liquid level areas corresponding to the north and south argon holes and calculate the area ratio of each area, the north-south hole distinction algorithm includes: for each edge point in the image, which is represented as a point in the parameter space of a circle , that is, the center coordinates and the radius ;

[0127] Taking the y coordinate of the center as the center, the points in the liquid level boundary with y coordinates greater than the y coordinate of the center are divided into the north side, and the points with y coordinates less than the y coordinate of the center are divided into the south side;

[0128] The points on the north and south sides are respectively formed into two independent regions, that is, the distinction between the north and south argon holes is completed;

[0129] Calculate the areas of the irregular figures in the north and south regions, using the Gaussian area:

[0130]

[0131] Where: is the coordinate of the th point of the contour, is the coordinate of the next point, is the number of points in the contour, that is, the last point is connected to the first point;

[0132] Divide this area by a preset standard value to obtain the proportion of the liquid level area. (Taking the area of the circle obtained in step S50 as a reference), which provides a basis for subsequent flow rate adjustment.

[0133] Step S70: Independently calculate the flow rates of the north and south argon holes according to the proportion of the liquid level area combined with the empirical formula, and dynamically adjust through a closed-loop control system.

[0134] In a specific implementation, the step of independently calculating the flow rates of the north and south argon holes according to the proportion of the liquid level area combined with the empirical formula and dynamically adjusting through a closed-loop control system includes: independently calculating the flow rates of the north and south argon holes using a preset model according to the proportion of the liquid level area combined with the empirical relationship between the ladle liquid level and the argon flow rate; inputting the flow rates of the north and south argon holes into the closed-loop control system so that the control module in the closed-loop control system dynamically adjusts the valve opening to control the fluctuation of the ladle liquid level.

[0135] It should be noted that the argon flow rate is adjusted to an ideal value in combination with the relationship between the ladle liquid level and the argon flow rate. Among them, the actual argon flow rate and the adjusted argon flow rate are obtained and controlled by a primary sensor, and then the program establishes communication with the PLC through opcua.Client(). The ideal argon flow rate is not a fixed value but an ideal range, which is determined by the experience of on-site personnel. Since there are two argon holes, one in the north and one in the south, at the bottom of the ladle, the flow rates of the north and south holes are independently adjusted according to the liquid level ratio and the hole area respectively. It is worth mentioning that since there is a time delay from the adjustment of the argon flow rate to its action on the molten steel liquid level, in this example, a basic strategy of small amounts and multiple times, and multiple confirmations is adopted, and it is fitted with a mathematical formula of uniform and slow decline.

[0136] In specific implementation, the following beneficial effects can be achieved through the method of this embodiment:

[0137] Improve the control accuracy and real-time performance of the argon flow rate. The liquid level image is collected in real time by a high-temperature industrial camera, and combined with the PP-LiteSeg-T semantic segmentation model and the findContours edge detection algorithm, the liquid level boundary and the north and south argon hole areas are accurately identified. Compared with traditional manual observation, the error in calculating the liquid level area is reduced by at least 30%, and the response time of argon flow rate adjustment is shortened to the millisecond level, effectively solving the problem of adjustment lag.

[0138] Enhance the uniformity of molten steel composition and smelting quality. Based on the independent flow rate adjustment according to the liquid level ratio (such as formula ), combined with closed-loop dynamic control, the flow rate distribution error of the north and south argon holes is less than 5%. The stirring uniformity of the molten steel is increased by 20%, and the inclusion content is reduced by 15%, significantly improving the cleanliness and metallurgical efficiency of the molten steel.

[0139] Reduce the intensity of manual intervention and operation risks. The fully automated system integrates image processing, deep learning, and PLC control modules, reducing the frequency of manual adjustment by 90%. Operators do not need to be in close contact with the high-temperature ladle, avoiding occupational health risks, and at the same time eliminating quality fluctuations caused by differences in adjustment standards among different shifts.

[0140] Improve the robustness to adapt to complex industrial environments. The air-cooled protective cover and adaptive image preprocessing (such as CLAHE contrast enhancement) technology ensure that the image acquisition clarity remains above 95% in an environment where the molten steel temperature ≥ 1600°C and there is dust interference. The model training covers data of multiple steel types and multiple working conditions, and the system operates stably in an environment temperature range of -20°C to 50°C.

[0141] Optimize resource consumption and operation efficiency. The lightweight PP-LiteSeg-T model reduces the computational complexity by 40% through network pruning and quantization techniques, and the image segmentation of 512×512 pixels only takes 50ms. The industrial PC platform supports multi-threaded processing, and the overall power consumption of the system is reduced by 25% compared with the traditional solution.

[0142] Realize full-process traceability and intelligent management of production data, automatically associate real-time monitored data (liquid level image, flow curve, adjustment record) with production batches, and the storage period is ≥ 5 years. Seamlessly connect with the MES system through the OPC UA protocol, support dynamic optimization of process parameters, and improve the fault diagnosis efficiency by 60%.

[0143] Ensure safe production and environmental protection compliance. The closed-loop control controls the liquid level fluctuation amplitude within ±2 mm, avoiding molten steel splashing or bubble slag entrainment accidents caused by excessive argon. The argon utilization rate is increased by 18%, and the argon consumption per furnace is reduced by 10%, meeting the requirements of green manufacturing emission reduction.

[0144] In this embodiment, the liquid level image of the ladle is collected in real time by a high-temperature industrial camera. After preprocessing, a semantic segmentation model is used to identify the liquid level area. A specific algorithm is combined to extract the liquid level boundaries corresponding to the north and south argon holes, and the independent dynamic adjustment of the double-hole flow rate is realized through a closed-loop control system. The system integrates image processing, deep learning, and automatic control modules, breaks through the lag of traditional manual adjustment, and solves the technical problems of inaccurate liquid level state monitoring and uneven argon distribution. Compared with the prior art, it can improve the argon flow control accuracy, improve the uniformity of molten steel composition, reduce the intensity of manual intervention at the same time, and effectively ensure the stability of molten steel quality and production safety in a high-temperature and harsh environment.

[0145] In addition, the embodiment of the present application also proposes a computer-readable storage medium, on which a program for intelligent control of bottom blowing argon in a ladle based on a neural network is stored. When the program for intelligent control of bottom blowing argon in a ladle based on a neural network is executed by a processor, the steps of the method for intelligent control of bottom blowing argon in a ladle based on a neural network as described above are realized.

[0146] Refer to Figure 5 , Figure 5 which is the structural block diagram of the first embodiment of the intelligent control device for bottom blowing argon in a ladle based on a neural network of the present application.

[0147] As Figure 5 shown, the intelligent control device for bottom blowing argon in a ladle based on a neural network proposed by the embodiment of the present application includes:

[0148] An image acquisition module 10, which is used to collect the liquid level image of the ladle in real time through a high-temperature industrial camera with air-cooled protection, and adjust the camera parameters to ensure clear images;

[0149] An image optimization module 20, which is used to adjust the collected images through methods such as image cropping, denoising, and adaptive histogram equalization to obtain optimized image data;

[0150] The dataset construction module 30 is used to construct a benchmark dataset from the optimized image data through data annotation;

[0151] The model training module 40 is used to train the benchmark dataset using the PP-LiteSeg-T model to generate a molten steel liquid level recognition model;

[0152] The recognition module 50 is used to input the optimized image data into the molten steel liquid level recognition model to output a binary image, and use the findContours algorithm to extract the liquid level boundary;

[0153] The calculation module 60 is used to distinguish the liquid level areas corresponding to the north and south argon holes based on the Hough circle detection in combination with the liquid level boundary, and calculate the area proportion of each area;

[0154] The adjustment module 70 is used to independently calculate the north and south argon hole flow rates according to the liquid level area proportion in combination with an empirical formula, and dynamically adjust through a closed-loop control system.

[0155] It should be understood that the above is only for illustration and does not constitute any limitation to the technical solution of the present application. In specific applications, those skilled in the art can set according to needs, and the present application does not make any restrictions in this regard.

[0156] In this embodiment, a ladle liquid level image is collected in real time by a high-temperature industrial camera. After preprocessing, a semantic segmentation model is used to identify the liquid level area. The liquid level boundaries corresponding to the north and south argon holes are extracted by combining specific algorithms, and the independent dynamic adjustment of the double-hole flow rate is realized through a closed-loop control system. The system integrates an image processing, deep learning, and automatic control module, breaks through the lag of traditional manual adjustment, and solves the technical problems of inaccurate liquid level state monitoring and uneven argon gas distribution. Compared with the prior art, it can improve the argon gas flow control accuracy, improve the uniformity of molten steel composition, reduce the intensity of manual intervention at the same time, and effectively ensure the stability of molten steel quality and production safety in a high-temperature and harsh environment.

[0157] It should be noted that the above-described work process is only illustrative and does not limit the protection scope of the present application. In actual applications, those skilled in the art can select some or all of them according to actual needs to achieve the purpose of the solution of this embodiment, and no restrictions are made here.

[0158] In addition, for the technical details not described in detail in this embodiment, reference can be made to the method for intelligent control of bottom blowing argon in a ladle based on a neural network provided in any embodiment of the present application, which will not be elaborated here.

[0159] In addition, it should be noted that in this text, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such a process, method, article or system. Without more limitations, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or system including that element.

[0160] The serial numbers of the embodiments of the present application above are only for description and do not represent the superiority or inferiority of the embodiments.

[0161] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as Read-Only Memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of the present application. The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present application.

Claims

1. An intelligent control method for bottom argon blowing of ladle based on neural network, characterized in that, Including: A high-temperature industrial camera protected by air cooling collects the molten steel ladle liquid level image in real time, and adjusts the camera parameters to ensure clear images; The collected images are adjusted by methods such as image cropping, denoising, and adaptive histogram equalization to obtain optimized image data; A benchmark data set is constructed by data annotation in the optimized image data; The PP-LiteSeg-T model is used to train the benchmark data set to generate a molten steel liquid level recognition model; The optimized image data is input into the molten steel liquid level recognition model to output a binary image, and the findContours algorithm is used to extract the liquid level boundary; Based on the Hough circle detection combined with the liquid level boundary to distinguish the liquid level regions corresponding to the north and south argon holes, and calculate the area ratio of each region; According to the liquid level region ratio combined with the empirical formula, the north and south argon hole flows are calculated independently and dynamically adjusted through a closed-loop control system.

2. The method according to claim 1, wherein The step of adjusting the collected images by methods such as image cropping, denoising, and adaptive histogram equalization to obtain optimized image data includes: The collected images are cropped to 512×512 pixels to balance the processing speed and accuracy; Gaussian filtering or median filtering algorithm is used to remove high-temperature noise and random interference; The adaptive histogram equalization method is used to divide the image into local regions for contrast enhancement to obtain optimized images.

3. The method according to claim 2, wherein The step of using the adaptive histogram equalization method to divide the image into local regions for contrast enhancement to obtain optimized images includes: The image is divided into local regions, and the local regions include 8x8 regions or 16x16 regions; Local histogram equalization is performed on each region in the local regions; Apply contrast limitation to each region; The interpolation method is used to smoothly transition the processing results between neighborhoods; Among them, the steps of local histogram equalization include: Let the gray value in a certain area be , and its frequency distribution be . The process of equalizing this area is to calculate the cumulative distribution function for each gray value: wherein, is the frequency of the gray level in this region , where is the index of the block; Use the cumulative distribution function for mapping to map the original gray value to the new gray value : where is the number of gray levels, is the maximum cumulative distribution in this area; Among them, the steps of applying contrast limitation include: Set the frequency of a certain gray level exceeds the set contrast limit value , then limit it to , and evenly distribute the remaining frequencies to all gray levels; The calculated frequency after limitation is: where is the gray level frequency, is the frequency after limitation, and the remaining over-limit part is calculated by the following formula: The excess frequency is evenly distributed to other gray levels.

4. The method according to claim 1, wherein In the step of using the PP-LiteSeg-T model to train the benchmark data set to generate a molten steel liquid level recognition model, the PP-LiteSeg-T model hyperparameters are set as follows: the number of iterations is 50,000 times, the initial learning rate is 0.005, transfer learning uses pre-trained weights, and data augmentation includes rotation, scaling, and flipping.

5. The method according to claim 1, characterized in that In the step of inputting the optimized image data into the molten steel liquid level recognition model to output a binary image and using the findContours algorithm to extract the liquid level boundary, the findContours algorithm parameters are configured as follows: the contour retrieval mode is cv2.RETR_LIST, the contour approximation method is cv2.CHAIN_APPROX_SIMPLE, and the liquid level boundary contour list is extracted.

6. The method according to claim 1, wherein In the step of distinguishing the north and south argon hole corresponding liquid level regions based on the Hough circle detection combined with the liquid level boundary and calculating the area ratio of each region, the north and south air hole distinction algorithm includes: For each edge point in the image, represented as a point in the parameter space of a circle , that is, the center coordinates and the radius ; Centering on the y - coordinate of the center of the circle, divide the points in the liquid - level boundary with y - coordinates greater than the y - coordinate of the center of the circle into the north side, and the points with y - coordinates less than the y - coordinate of the center of the circle into the south side; Respectively form two independent regions with the points on the north and south sides, that is, complete the distinction between the north and south argon holes; Calculate the areas of the irregular figures in the north and south regions, using the Gaussian area: Wherein: is the th point coordinate of the contour, is the coordinate of the next point, is the number of points in the contour; Divide this area by a preset standard value to obtain the proportion of the liquid - level region.

7. The method according to claim 1, characterized in that, The step of independently calculating the argon - hole flow rates in the north and south according to the proportion of the liquid - level region in combination with an empirical formula and dynamically adjusting through a closed - loop control system includes: Independently calculate the argon - hole flow rates in the north and south using a preset model according to the proportion of the liquid - level region in combination with the empirical relationship between the ladle liquid level and the argon flow rate; Input the argon - hole flow rates in the north and south into the closed - loop control system so that the control module in the closed - loop control system dynamically adjusts the valve opening to control the fluctuation of the ladle liquid level.

8. An intelligent control device for bottom argon blowing of ladle based on neural network, characterized in that, Including: An image acquisition module, which is used to collect the ladle liquid - level image in real time through a high - temperature industrial camera with air - cooled protection, and adjust the camera parameters to ensure the clarity of the image; An image optimization module, which is used to adjust the collected image by methods such as image cropping, denoising, and adaptive histogram equalization to obtain optimized image data; A dataset construction module, which is used to construct a reference dataset by data annotation in the optimized image data; A model training module, which is used to train the reference dataset using the PP - LiteSeg - T model to generate a molten - steel liquid - level recognition model; A recognition module, which is used to input the optimized image data into the molten - steel liquid - level recognition model to output a binary image, and use the findContours algorithm to extract the liquid - level boundary; A calculation module, which is used to distinguish the liquid - level regions corresponding to the north and south argon holes based on the Hough circle detection in combination with the liquid - level boundary, and calculate the area proportion of each region; An adjustment module, which is used to independently calculate the argon - hole flow rates in the north and south according to the proportion of the liquid - level region in combination with an empirical formula, and dynamically adjust through a closed - loop control system.

9. A computer device, characterized in that, The device includes: a memory and a processor. When the processor runs the computer instructions stored in the memory, it executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Including instructions, when the instructions run on a computer, the computer is made to execute the method according to any one of claims 1 to 7.

Images