Adaptive image segmentation converter steelmaking furnace mouth pressure intelligent balancing method

Abstract

The invention relates to the technical field of intelligent control in the converter steelmaking process, and particularly discloses a converter steelmaking furnace mouth pressure intelligent balancing method with self-adaptive image segmentation, which comprises the following steps: image acquisition: acquiring a furnace mouth flame image in real time through an industrial camera; image preprocessing: adopting a dynamic gamma correction algorithm to adaptively adjust a correction coefficient according to an R channel mean value of the image; in the flame feature extraction stage, the corrected image is subjected to bilateral filtering to remove noise, and then an improved OTSU segmentation algorithm is input to extract a flame main reaction area; in the pressure intelligent control stage, a flame characteristic-furnace pressure nonlinear model is constructed, and finally the rotating speed of an ID fan is dynamically adjusted through a fuzzy self-adaptive controller; the furnace mouth flame image is intelligently monitored, then the furnace mouth pressure is analyzed, the system pressure fluctuation standard deviation is controlled to be reduced, the gas recovery rate is improved, the false detection rate is reduced, the energy efficiency of converter steelmaking is improved, and the production safety is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control in the converter steelmaking process, and particularly relates to an intelligent balance method for the converter steelmaking furnace mouth pressure with adaptive image segmentation. Background Technique

[0002] In the converter steelmaking process, the precise control of the differential pressure at the furnace mouth is the core link to ensure gas recovery efficiency and safe production. Traditional methods mainly rely on pressure sensors installed at the end of the flue for detection. However, this indirect measurement method has significant defects: the pressure signal needs to be transmitted through a pipeline up to 12 - 18 meters long, resulting in a detection delay of up to 5 - 8 seconds, and it is unable to capture the transient pressure fluctuations during the smelting process. Taking the actual measurement data of a steel group in 2022 as an example, when the silicon content in hot metal suddenly increases by 0.3%, the pressure regulation response of the traditional sensor system lags behind by 6.2 seconds, causing an instantaneous decrease of 11% in the gas recovery rate. In recent years, although machine vision technology has been tried to be introduced into this field, the reliability of existing solutions still faces severe challenges under complex working conditions. For example, the fixed - threshold flame segmentation method has a misdetection rate of over 35% when the flue gas concentration at the furnace mouth fluctuates, and it is unable to distinguish the main reaction area of the flame from the splashing sparks; when using the YOLOv5 deep - learning model for flame recognition, although the segmentation accuracy is improved to 85%, the model training requires more than 5,000 sets of labeled samples, the hardware deployment cost is as high as 230,000 yuan, and the GPU power consumption exceeds 250W, making it difficult to meet the electrical safety standards in the metallurgical explosion - proof area. In addition, existing control strategies generally use PID controllers with fixed parameters. When the raw material composition fluctuates or the oxygen - blowing intensity changes, the system is prone to overshoot and oscillation phenomena. A case of a steel plant shows that when the carbon content in hot metal fluctuates by ±0.25%, the pressure overshoot of the traditional PID controller increases by 120%, and the steady - state regulation time is extended to more than 12 seconds. These technical bottlenecks together constitute an impossible triangle of "detection accuracy - response speed - implementation cost", severely restricting the energy efficiency improvement and intelligent transformation of the steelmaking process. Therefore, it is necessary to design an intelligent balance method for the converter steelmaking furnace mouth pressure with adaptive image segmentation to solve the problem of low energy efficiency in the existing converter steelmaking process. Summary of the Invention

[0003] Aiming at the problems existing in the prior art, the purpose of the present invention is to provide an intelligent balance method for the converter steelmaking furnace mouth pressure with adaptive image segmentation.

[0004] The technical solution adopted by the present invention to solve its technical problems is: an intelligent balance method for the converter steelmaking furnace mouth pressure with adaptive image segmentation, including the following steps:

[0005] 1) Image acquisition: Real - time acquisition of the furnace mouth flame image through an industrial camera;

[0006] 2) Image preprocessing: The dynamic gamma correction algorithm is used to adaptively adjust the correction coefficient according to the average value of the R channel of the image:

[0007] γ = 0.5 + 0.1·tanh(64128 - μR);

[0008] 3) Flame feature extraction stage: After the corrected image is denoised by bilateral filtering, it is input into the improved OTSU segmentation algorithm to extract the main reaction area of the flame. The objective function of the improved OTSU segmentation algorithm is defined as:

[0009] σw 2 (t) = ω0(t)σ02(t) + ω1(t)σ12(t) + 0.85·N_totalN_connected(t);

[0010] 4) Intelligent pressure control stage: A non-linear model of flame characteristics - furnace pressure is constructed, and finally the rotational speed of the ID fan is dynamically adjusted through a fuzzy adaptive controller.

[0011] Specifically, the industrial camera in step 1) is installed 6 meters to the side of the furnace mouth and is equipped with a high-temperature resistant protective cover. The image data collected by the industrial camera is transmitted to the edge computing unit through optical fiber to execute the image processing algorithm.

[0012] Specifically, in the correction coefficient formula in step 2), γ is the gamma correction coefficient, which is used to adjust the image brightness contrast; μR is the average value of the R channel pixels, reflecting the flame brightness level; 64128 is the constant term in the formula, a normalization parameter, representing the linear mapping of the gray value range 64 - 128; tanh is the hyperbolic tangent function, which is used to smoothly adjust the γ value.

[0013] Specifically, the correction coefficient formula ensures that when the illuminance is low, i.e., μR < 90, the enhancement coefficient γ drops to 0.45 to display the dark details, and when the brightness is high, i.e., μR > 160, γ rises to 0.65 to suppress overexposure; the corrected image is denoised by bilateral filtering, with the spatial domain standard deviation σd = 3 and the value domain standard deviation σr = 15, reducing the noise power by more than 12 dB while retaining the flame edge features.

[0014] Specifically, σw in step 3) 2 (t) is the weighted between-class variance, which is used to measure the quality of the segmentation threshold t; ω0(t), ω1(t) are the pixel proportion weights of the foreground (flame) and background; σ02(t), σ12(t) are the within-class variances of the foreground and background, measuring the dispersion degree of their respective regions; N_total is the total number of image pixels; 0.85 is the weight coefficient of the connectivity constraint term, obtained through experimental optimization;

[0015] N_connected is the number of connected pixels in the segmented region. The constraint on the number of connected pixels eliminates the interference of discrete noise points. After traversing the gray levels to calculate the optimal threshold t*, a morphological opening operation is performed using a 3×3 circular structuring element. Finally, the segmentation IoU reaches 92.7%. Based on the segmentation result, the flame area ratio η = N_flame / N_total and the morphological compactness F = 4πA / P² are extracted as key feature parameters; F is the morphological compactness of the flame, measuring the degree of closeness to a perfect circle, and F = 1 represents a perfect circle; A is the area of the flame region, in pixels; P is the perimeter of the flame region, the length of the pixel boundary.

[0016] Specifically, the flame feature-furnace pressure nonlinear model in step 4) is as follows:

[0017] ΔP = 8.71n(η / η0) + 0.35||F - F_ref||² + 1.2σT;

[0018] In the formula, ΔP is the change in furnace mouth pressure, in Pa, which is the target parameter to be dynamically adjusted; η is the flame area ratio, η = N_flame / N_total, where N_flame is the number of flame pixels; η0 = 0.42 is the reference area ratio; F is the current morphological compactness of the flame; F_ref is the reference compactness; σT is the temperature field uniformity index, reflecting the standard deviation of the temperature distribution in the furnace; 8.71, 0.35, and 1.2 are all model coefficients determined by multiple regression fitting; the fuzzy adaptive controller takes the pressure deviation e and the deviation change rate ec as inputs, and generates the rotational speed adjustment amount Δf using 49 expert rules. An example of the rule is: if e is negative large and ec is positive small, then Δf is positive medium. The fuzzy adaptive controller dynamically adjusts the rule weights through an online learning mechanism. When the silicon content of the hot metal fluctuates by more than ±0.2%, the priority of the compensation-type rules is automatically enhanced.

[0019] Specifically, the fuzzy adaptive controller is implemented in the TIA Portal platform using structured text language, and three-level safety responses are set through the PLC program: when η > 0.65, a yellow warning is triggered and the rotational speed is increased by 5%; when η > 0.7, a red alarm is triggered and the emergency smoke exhaust system is started.

[0020] The present invention has the following beneficial effects:

[0021] The intelligent balance method for the converter steelmaking furnace mouth pressure designed by the present invention intelligently monitors the furnace mouth flame image, and then analyzes the furnace mouth pressure. The standard deviation of the control system pressure fluctuation decreases, the gas recovery rate increases, the misdetection rate decreases, the energy efficiency of converter steelmaking is improved, and production safety is enhanced. Specific embodiments

[0022] The technical solutions in the embodiments of the present invention will be further described clearly and completely below. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0023] An intelligent balance method for the converter steelmaking furnace mouth pressure with adaptive image segmentation constructs a full-process closed-loop system from flame dynamic feature extraction to precise pressure regulation through the deep integration of multi-dimensional visual perception and intelligent control technologies. The core of the method lies in innovatively combining machine vision technology with fuzzy adaptive control, breaking through the spatio-temporal limitations of traditional pressure detection and the static limitations of control strategies. The system first collects the furnace mouth flame images in real time through an industrial camera, and uses a dynamic gamma correction algorithm to enhance the details in the low-illumination area and eliminate backlight interference. After the corrected image is denoised by bilateral filtering, it is input into an improved OTSU segmentation algorithm to extract the main reaction area of the flame. This algorithm introduces a spatial connectivity constraint term on the basis of the traditional between-class variance, significantly improving the segmentation accuracy. Based on the segmentation results, characteristic parameters such as the flame area ratio and morphological compactness are calculated, and a non-linear mapping model between them and the furnace mouth pressure is established. Finally, the rotational speed of the ID fan is dynamically adjusted through a fuzzy adaptive controller to achieve precise balance of the differential pressure.

[0024] Specifically, it includes the following steps.

[0025] 1. Image acquisition: The furnace mouth flame images are collected in real time through an industrial camera; the hardware system deployment adopts a hierarchical architecture design: The Basler acA2000-50gm industrial camera collects the furnace mouth flame images at a resolution of 1920×1080 and a frame rate of 50fps, and is installed 6 meters to the side of the furnace mouth and equipped with a high-temperature resistant protective cover. The image data is transmitted to the edge computing unit through optical fiber. This unit is equipped with 64GB of memory and a 256-core GPU, and is responsible for executing image processing algorithms. The control instructions are received by the Siemens S7-1500PLC through the Profinet protocol and converted into a 4-20mA signal to drive the ABB ACS880 frequency converter, and the control period is strictly limited within 50ms.

[0026] 2. Image preprocessing: The dynamic gamma correction algorithm is used to adaptively adjust the correction coefficient according to the mean value of the R channel of the image:

[0027] γ = 0.5 + 0.1·tanh(64128 - μR).

[0028] In the correction coefficient formula, γ is the gamma correction coefficient, which is used to adjust the brightness contrast of the image; μR is the average value of the pixels in the R channel, reflecting the flame brightness level; 64128 is the constant term in the formula, a normalization parameter, representing the linear mapping of the gray value range 64 - 128; tanh is the hyperbolic tangent function, which is used to smoothly adjust the γ value. In the correction coefficient formula, μR is the average value of the pixels in the R channel. The formula ensures that when the illuminance is low, i.e., μR < 90, the enhancement coefficient γ drops to 0.45 to enhance the details in the dark areas, and when the brightness is high, i.e., μR > 160, γ rises to 0.65 to suppress overexposure; the corrected image is subjected to noise suppression through bilateral filtering. The standard deviation in the spatial domain σd = 3, and the standard deviation in the value domain σr = 15, reducing the noise power by more than 12 dB while retaining the flame edge features.

[0029] During the implementation of dynamic gamma correction, the average value μR of the R channel of the image is calculated in real time, and the non - linear correction mode is activated when μR < 100. Taking γ = 0.55 as an example, the amount of detail information in the corrected image at the gray level [0, 80] is increased to 3.2 times that of the original image (verified by information entropy calculation). The bilateral filtering is implemented using the bilateralFilter function in the OpenCV library, and the processing time is controlled within 8 ms / frame, meeting the real - time requirements.

[0030] 3. Flame feature extraction stage: After the corrected image is denoised by bilateral filtering, it is input into the improved OTSU segmentation algorithm to extract the main reaction area of the flame. The objective function of the improved OTSU segmentation algorithm is defined as:

[0031] σw 2 (t) = ω0(t)σ02(t) + ω1(t)σ12(t) + 0.85·N_totalN_connected(t).

[0032] σw 2(t) is the weighted between-class variance, used to measure the quality of the segmentation threshold t; ω0(t) and ω1(t) are the pixel proportion weights of the foreground (flame) and background; σ02(t) and σ12(t) are the within-class variances of the foreground and background, measuring the dispersion of their respective regions; N_total is the total number of pixels in the image; 0.85 is the weight coefficient of the connectivity constraint term, obtained through experimental optimization; N_connected is the number of connected pixels in the segmented region. The connectivity pixel number constraint eliminates the interference of discrete noise points. After traversing the gray levels to calculate the optimal threshold t*, a morphological opening operation is performed using a 3×3 circular structuring element. Finally, the segmentation IoU reaches 92.7%. IoU is the intersection over union, a segmentation accuracy metric, IoU = overlapping region / (flame region + segmented region - overlapping region). Based on the segmentation result, the flame area ratio η = N_flame / N_total and the morphological compactness F = 4πA / P2 are extracted as key feature parameters; F is the flame morphological compactness, measuring the degree of closeness to a perfect circle, and F = 1 represents a perfect circle; A is the area of the flame region, in pixels; P is the perimeter of the flame region, the length of the pixel boundary.

[0033] The improved OTSU segmentation algorithm optimizes the execution efficiency by pre-computing the connectivity mapping table: the number of connected regions N_connected corresponding to each gray level is established in advance, reducing the single-segmentation time consumption from 120 ms of the traditional algorithm to 68 ms. In the morphological post-processing stage, first, an erosion operation is performed using a 3×3 rectangular kernel to eliminate isolated noise points, and then the same kernel is used for dilation to restore the main shape of the flame. The processed binary image extracts the precise boundary through contour analysis.

[0034] 4. Intelligent pressure control stage: Construct a non-linear model of flame characteristics - furnace pressure, and finally dynamically adjust the ID fan speed through a fuzzy adaptive controller.

[0035] The non-linear model of flame characteristics - furnace pressure is:

[0036] ΔP = 8.71n(η / η0) + 0.35 / / F - F_ref / / 2 + 1.2σT.

[0037] In the formula, ΔP is the change in the furnace mouth pressure, with the unit of Pa, which is the target parameter to be dynamically adjusted; η is the proportion of the flame area, η = N_flame / N_total, where N_flame is the number of flame pixels; η0 = 0.42 is the reference area ratio; F is the current compactness of the flame shape; F_ref is the reference compactness; σT is the temperature field uniformity index, reflecting the standard deviation of the temperature distribution in the furnace; 8.71, 0.35, and 1.2 are all model coefficients determined by multiple regression fitting; the fuzzy adaptive controller takes the pressure deviation e and the deviation change rate ec as inputs, and uses 49 expert rules to generate the rotational speed adjustment amount Δf. An example of the rule is: if e is negative large and ec is positive small, then Δf is positive medium, indicating that when the pressure is lower than the set value and the deviation decreases rapidly, the rotational speed needs to be increased moderately. The fuzzy adaptive controller dynamically adjusts the rule weights through an online learning mechanism. When the fluctuation of the molten iron silicon content exceeds ±0.2%, the priority of the compensation rules is automatically enhanced.

[0038] The fuzzy adaptive controller is implemented in the TIA Portal platform (Siemens industrial automation programming platform) using structured text (ST) language. The rule base is stored in a two-dimensional matrix. The domain of the input variable e is [-30Pa, +30Pa], the domain of ec is [-15Pa / s, +15Pa / s], and the domain of the output variable Δf is [-15Hz, +15Hz]. The defuzzification uses the center of area method, and the control quantity is updated every 10ms. To cope with sudden working conditions, a three-level safety response is set through the PLC program: when η > 0.65, a yellow warning is triggered and the rotational speed is increased by 5%, and when η > 0.7, a red alarm is triggered and the emergency smoke exhaust system is started.

[0039] Industrial verification data shows that in the continuous 90-day test of a 210t converter in a certain iron and steel enterprise, the standard deviation of the system pressure fluctuation decreased from ±9.8Pa of the traditional method to ±3.1Pa, the gas recovery rate increased by 2.7 percentage points to 89.2%, and the misdetection rate was less than 0.3 times per furnace. The system achieved 1,500 hours of fault-free operation in an environment of 65°C high temperature and 4kV / 5kHz electromagnetic interference, meeting the requirements of the SIL-2 safety level certification. SIL-2 is the safety integrity level 2, which complies with industrial safety standards.

[0040] The present invention is not limited to the above embodiments. Anyone should know that structural changes made under the inspiration of the present invention, as long as they have the same or similar technical solutions as the present invention, fall within the protection scope of the present invention.

[0041] The technologies, shapes, and structures not detailed in the present invention are all well-known technologies.

Claims

1. An intelligent balancing method for the pressure at the converter steelmaking furnace mouth with adaptive image segmentation, characterized in that It includes the following steps: 1) Image acquisition: The industrial camera is used to collect the flame image at the furnace mouth in real time; 2) Image preprocessing: The dynamic gamma correction algorithm is adopted to adaptively adjust the correction coefficient according to the mean value of the R channel of the image: γ = 0.5 + 0.1·tanh(64128 - μR); 3) Flame feature extraction stage: After the corrected image is denoised by bilateral filtering, it is input into the improved OTSU segmentation algorithm to extract the main reaction area of the flame. The objective function of the improved OTSU segmentation algorithm is defined as: σw 2 (t) = ω0(t)σ02(t) + ω1(t)σ12(t) + 0.85·N_total N_connected(t); 4) Intelligent pressure control stage: A non-linear model of flame feature - furnace pressure is constructed, and finally the rotational speed of the ID fan is dynamically adjusted by a fuzzy adaptive controller.

2. The intelligent balance method for the converter steelmaking furnace mouth pressure of adaptive image segmentation according to claim 1, characterized in that, The industrial camera in step 1) is installed 6 meters to the side of the furnace mouth and is equipped with a high-temperature resistant protective cover. The image data collected by the industrial camera is transmitted to the edge computing unit through optical fiber to execute the image processing algorithm.

3. The intelligent balance method for the converter steelmaking furnace mouth pressure of adaptive image segmentation according to claim 1, characterized in that, In the correction coefficient formula in step 2), γ is the gamma correction coefficient, which is used to adjust the image brightness contrast; μR is the mean value of the R channel pixels, which reflects the flame brightness level; 64128 is the constant term in the formula, which is a normalization parameter representing the linear mapping of the gray value range of 64 - 128; tanh is the hyperbolic tangent function, which is used to smoothly adjust the γ value.

4. The intelligent balance method for the converter steelmaking furnace mouth pressure of adaptive image segmentation according to claim 3, characterized in that, The correction coefficient formula ensures that when the illuminance is low, i.e., μR < 90, the enhancement coefficient γ drops to 0.45 to display the dark details. When the brightness is high, i.e., μR > 160, γ rises to 0.65 to suppress overexposure. The corrected image is denoised by bilateral filtering, with the spatial domain standard deviation σd = 3 and the value domain standard deviation σr = 15, reducing the noise power by more than 12 dB while retaining the flame edge features.

5. The intelligent balance method for the converter steelmaking furnace mouth pressure of adaptive image segmentation according to claim 1, characterized in that, The σw in step 3) above 2 (t) is the weighted between-class variance, which is used to measure the quality of the segmentation threshold t; ω0(t) and ω1(t) are the pixel proportion weights of the foreground (flame) and background; σ02(t) and σ12(t) are the within-class variances of the foreground and background, which measure the dispersion of their respective regions; N_total is the total number of pixels in the image; 0.85 is the weight coefficient of the connectivity constraint term, which is obtained through experimental optimization; N_connected is the number of connected pixels in the segmented area. The constraint of the number of connected pixels eliminates the interference of discrete noise points. After traversing the gray levels to calculate the optimal threshold t*, a morphological opening operation is performed using a 3×3 circular structuring element. Finally, the segmentation IoU reaches 92.7%. Based on the segmentation result, the flame area ratio η = N_flame / N_total and the morphological compactness F = 4πA / P2 are extracted as key feature parameters; F is the flame morphological compactness, which measures the degree of closeness to a perfect circle. When F = 1, it is a perfect circle; A is the area of the flame region, in pixels; P is the perimeter of the flame region, in pixel boundary length.

6. The intelligent balance method for the converter steelmaking furnace mouth pressure of adaptive image segmentation according to claim 1, characterized in that The non-linear model of flame feature - furnace pressure in step 4) is: ΔP = 8.71n(η / η0) + 0.35 / / F - F_ref / / 2 + 1.2σT; In the formula, ΔP is the change in furnace mouth pressure, in Pa, which is the target parameter to be dynamically adjusted; η is the flame area ratio, η = N_flame / N_total, and N_flame is the number of flame pixels; η0 = 0.42 is the reference area ratio; F is the current flame morphological compactness; F_ref is the reference compactness; σT is the temperature field uniformity index, which reflects the standard deviation of the temperature distribution in the furnace; 8.71, 0.35, and 1.2 are all model coefficients determined by multiple regression fitting. The fuzzy adaptive controller takes the pressure deviation e and the deviation change rate ec as inputs, and uses 49 expert rules to generate the rotational speed adjustment amount Δf. An example of the rule representation is: if e is negative large and ec is positive small, then Δf is positive medium. The fuzzy adaptive controller dynamically adjusts the rule weights through an online learning mechanism. When the molten iron silicon content fluctuates by more than ±0.2%, the priority of the compensation type rules is automatically enhanced.

7. The intelligent balance method for the converter steelmaking furnace mouth pressure of adaptive image segmentation according to claim 6, characterized in that, The said fuzzy adaptive controller is implemented in the TIA Portal platform using structured text language, and three-level safety responses are set through the PLC program: when η > 0.65, a yellow warning is triggered and the rotational speed is increased by 5%, and when η > 0.7, a red alarm is triggered and the emergency smoke exhaust system is started.