Blast furnace tapping slag iron flow detection method, system and equipment based on computer vision and medium

Abstract

The invention provides a blast furnace tapping slag iron flow detection method, system and equipment based on computer vision and a medium, and belongs to the technical field of blast furnace smelting, infrared and visible light images of a tapping hole are synchronously collected through a double-spectrum camera, and stream contour extraction and instantaneous sectional area calculation are achieved through HSV-SVM separation, YOLOv5 boundary calibration and three-dimensional point cloud reconstruction. A drop point track is tracked by combining an improved LK optical flow method, a turbulence coefficient is dynamically corrected through an LSTM network, and the accuracy of a flow velocity component is optimized. And finally, calculating the real-time flow based on the mass flow, judging the tapping operation state by using a ResNet34 network, and outputting flow data and abnormal early warning. Multispectral fusion improves the molten iron area separation accuracy, three-dimensional point cloud sectional area calculation avoids perspective errors, accurate flow metering and real-time state early warning are achieved, molten iron residual loss is reduced, and tapping operation automation and reliability are improved.

Description

Technical Field

[0001] This invention belongs to the field of blast furnace smelting technology, specifically relating to a method, system, equipment, and medium for detecting the flow rate of blast furnace tapping slag based on computer vision. Background Technology

[0002] In the iron and steel smelting industry, blast furnace ironmaking is the core step in producing pig iron. The blast furnace tapping process refers to the periodic discharge of molten iron from the taphole into a ladle or main trough for further processing. The stability and controllability of this process directly affect the smooth operation of the blast furnace, equipment safety, and the scheduling of subsequent steelmaking processes. With the development of industrial automation and intelligent manufacturing, intelligent sensing technologies based on image recognition, sensor fusion, and artificial intelligence are gradually being applied to the metallurgical field.

[0003] Dual-spectrum imaging technology, capable of simultaneously capturing the surface morphology and temperature distribution characteristics of molten iron, has been widely explored for monitoring high-temperature melts. While visible light cameras capture the brightness characteristics of molten iron, the brightness of iron oxide scale and molten iron tends to converge during blast furnace tapping, making them easily misidentified as molten iron areas, leading to significant deviations in stream contour extraction. Dust obstruction can also cause uneven brightness, resulting in a high rate of lost tracking of the landing point trajectory.

[0004] Related technologies typically set the turbulence coefficient to 1.0 or 2.0. However, during blast furnace tapping, the flow stream is affected by tapping pressure and ambient wind speed, resulting in dynamic changes in turbulence intensity (e.g., weak turbulence in the initial stage, strong turbulence in the middle stage, and weak turbulence in the later stage). Setting the turbulence coefficient to 1.0 or 2.0 leads to high errors in flow velocity calculation. Furthermore, estimating the cross-sectional area based on a fixed value for the flow stream shape will result in morphological distortion due to pressure changes, leading to large errors in cross-sectional area and failing to meet the requirements for detecting slag and iron flow rates during blast furnace tapping. Summary of the Invention

[0005] This invention provides a computer vision-based method for detecting the flow rate of slag and iron during blast furnace tapping. By accurately measuring the flow rate, determining the status in real time, and issuing early warnings of anomalies, the method improves the efficiency of tapping operations.

[0006] The methods include: S101: Acquire bispectral image data of the blast furnace taphole; S102: Separating the effective flow area of ​​molten iron based on dual-spectral image data; S103: Divide the effective flow area of ​​molten iron into streams to extract the stream outline; S104: Calculate the instantaneous cross-sectional area of ​​the stream based on the stream profile; S105: Track the trajectory of molten iron drop points based on image sequences and calculate the velocity component; S106: Dynamically correct the turbulence coefficient in velocity calculation to optimize the accuracy of velocity components; S107: Calculate the mass flow rate based on the instantaneous cross-sectional area and the corrected velocity component to achieve flow output; S108: Determines the iron tapping operation status based on bispectral image features, outputs flow data and anomaly warnings, and completes full-process analysis.

[0007] Preferably, S102 specifically includes the following steps: The infrared image data is processed by calculating an adaptive threshold and performing binary segmentation to obtain preliminary results for the molten iron region outside the infrared region. The visible light image data is processed, and the molten iron region is calibrated based on a preset temperature isotherm using a pre-trained YOLOv5 segmentation network to obtain the segmentation result of the molten iron region on the visible light side. The visible light side molten iron region segmentation result is converted to the HSV color space, and the iron oxide scale pixels are removed by color feature matching to obtain the corrected visible light side molten iron region. The preliminary result of the outer red molten iron region is combined with the corrected visible light molten iron region by performing a pixel-level intersection operation, retaining common pixels to form the final effective molten iron flow area.

[0008] Preferably, S103 specifically includes the following steps: Temperature thresholding is performed on the infrared image, and the segmentation threshold is dynamically determined based on the highest temperature in the iron tapping area and the background temperature to extract the boundary of the molten iron area. Based on a deep learning model, the boundary of the molten iron stream is identified by the 1600℃ isotherm feature in visible light images, and the outline of the molten iron stream is obtained. By using color space analysis, interference areas caused by iron oxide scale are identified and removed, thereby optimizing the integrity of the molten iron flow profile. The segmentation results of infrared and visible light images are spatially matched and fused to generate a unified flow contour. Based on camera calibration parameters, the flow contours in the image are converted into physical dimensions to ensure the physical accuracy of the contours.

[0009] Preferably, S104 specifically includes the following steps: Set the sampling frequency of the instantaneous cross-sectional area, and correspond each sampling time with the cross-sectional position of the flow stream determined in step S103; It calls upon the pre-established mapping relationship between pixels and physical dimensions to calculate the actual physical area represented by a single pixel; For the cross-sectional position at each sampling time, the contour pixel coordinates are extracted from the flow contour obtained in step S103 and organized into a closed sequence. The instantaneous physical cross-sectional area is calculated by determining the number of pixels located inside the contour sequence, counting the total number of valid pixels, and combining this with the pixel physical area. Outlier detection and correction are performed on the instantaneous physical cross-sectional area sequence recorded in chronological order, and the corrected cross-sectional area sequence is output for subsequent calculations.

[0010] Preferably, S105 specifically includes the following steps: In the initial frame of the image sequence, a specific area is delineated for tracking during the falling section of the molten iron flow; Within a specific area, points with image features are selected as tracking points; Between consecutive image frames, the positional movement of the tracking point is traced, and its pixel displacement is calculated; Based on the pixel displacement and image acquisition time interval, combined with the spatial coordinate mapping relationship, the actual movement speed of the tracking point is calculated; Statistical analysis was performed on the velocities of all tracking points to determine the mainstream velocity component of the molten iron stream in the vertical direction.

[0011] Preferably, S106 specifically includes the following steps: Collect the vertical velocity components, instantaneous cross-sectional area, and manually labeled true values ​​of turbulence coefficients from historical frames during blast furnace tapping to construct a dataset; The vertical velocity and instantaneous cross-sectional area in the dataset are normalized using Min-Max to map their numerical range to the interval between 0 and 1. Build an LSTM long short-term memory network, with the normalized vertical velocity and instantaneous cross-sectional area of ​​the current frame as inputs, and the output as the correction value of the turbulence coefficient; An LSTM network was trained using labeled data, with the mean squared error between the predicted corrected value and the true value as the loss function. The network weights were iteratively optimized using the Adam optimizer. The real-time calculated vertical velocity and instantaneous cross-sectional area are input into the trained LSTM network to obtain the dynamically corrected turbulence coefficient, which is used to adjust the vertical velocity component.

[0012] Preferably, S107 specifically includes the following steps: The instantaneous cross-sectional area data from step S104, the vertical velocity component data from step S105, and the turbulence coefficient data from step S106 are obtained and aligned by timestamp to form a real-time calculation parameter set. The temperature of the molten iron region is sampled from the infrared image in step S101, the average temperature is calculated, and the density parameter of the molten iron is determined based on industrial standards. Using a real-time calculation parameter set and molten iron density parameters, the instantaneous mass flow rate is calculated using the mass flow rate formula, and the average mass flow rate is obtained by averaging the continuous calculation results. The average mass flow rate is compared with similar data in the blast furnace historical database. If it exceeds the fluctuation range, the real-time parameters are corrected and recalculated until the data is valid. The effective average mass flow rate and related parameters are displayed in real time and stored in partitions.

[0013] This application also provides a computer vision-based system for detecting the flow rate of blast furnace tapping slag and iron, the system comprising: The data acquisition module is used to acquire bispectral image data from the blast furnace taphole. The region separation module separates the effective flow area of ​​molten iron based on dual-spectral image data; The flow segmentation module is used to segment the effective flow area of ​​molten iron into flow segments in order to extract the flow outline; The cross-sectional area calculation module calculates the instantaneous cross-sectional area of ​​the stream based on the stream profile; The flow velocity tracking module tracks the trajectory of molten iron drop points based on image sequences and calculates the flow velocity components; The velocity correction module is used to dynamically correct the turbulence coefficient in velocity calculation to optimize the accuracy of velocity components. The flow calculation module calculates the mass flow rate based on the instantaneous cross-sectional area and the corrected flow velocity component, and then outputs the flow rate. The judgment and early warning module determines the iron tapping operation status based on bispectral image features and outputs flow data and abnormal warnings, completing the full process analysis.

[0014] According to another embodiment of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the computer vision-based blast furnace tapping slag and iron flow detection method.

[0015] According to another embodiment of this application, a storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the computer vision-based blast furnace tapping slag and iron flow detection method.

[0016] As can be seen from the above technical solutions, the present invention has the following advantages: The blast furnace tapping slag and iron flow detection method provided by this invention improves the separation accuracy of the effective flow area of ​​molten iron through dual-spectral image fusion; and reduces flow detection error by combining LSTM network to dynamically correct the turbulence coefficient and dynamically calculate the molten iron density, thus meeting the requirements of industrial-grade accuracy.

[0017] This invention uses dual-spectral image features to quantitatively determine the timing of tapping initiation, slag appearance, and tapping blockage based on molten iron brightness, temperature distribution, and flow morphology, improving accuracy and preventing residual iron loss due to untimely tapping. Through timestamp synchronization, parameters such as instantaneous flow rate, average flow rate, cross-sectional area, flow velocity, and turbulence coefficient are stored in daily partitions. Output process data such as slag-to-iron ratio and cumulative flow rate can be directly used for furnace process optimization, providing quantitative support for process improvement. This enhances the intelligence and stability of blast furnace tapping. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The flowchart shows a computer vision-based method for detecting the flow rate of blast furnace tapping slag and iron. Figure 2 This is a schematic diagram of a computer vision-based blast furnace tapping slag and iron flow detection system. Figure 3 This is a schematic diagram of an electronic device. Detailed Implementation

[0020] The following describes in detail the computer vision-based method for detecting blast furnace tapping slag and iron flow rate. Specific details, such as particular system structures and technologies, are presented for illustrative purposes and not for limitation, to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details.

[0021] It should be understood that, when used in this specification, terms include indicating the presence of a described feature, integral, step, operation, element, and / or component, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms include, encompass, have, and variations thereof mean including but not limited to, unless otherwise specifically emphasized.

[0022] The statements such as "one embodiment" or "some embodiments" described in this application mean that one or more embodiments of this application include the specific features, structures, or characteristics described in that embodiment. Therefore, the statements such as "in one embodiment," "in some embodiments," "in other embodiments," and "in still other embodiments" in this application do not necessarily refer to the same embodiment, but rather mean one or more, but not all, embodiments, unless otherwise specifically emphasized.

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 The diagram shows a flowchart of a blast furnace tapping slag and iron flow detection method based on computer vision in a specific embodiment. The method includes: S101: Acquire bispectral image data of the blast furnace taphole.

[0025] In some embodiments, a visible light camera with a 450nm blue laser light source and an infrared thermal imager are installed directly in front of the blast furnace taphole to form a dual-spectrum acquisition system. The acquisition software is run by an industrial control computer to synchronize the clocks of the two cameras according to the NTP protocol and acquire dual-spectrum image data in real time. The visible light image is saved in JPEG format and the infrared image is saved in TIFF format. At the same time, a timestamp and furnace number are added to each frame of the image.

[0026] S102: Separating the effective flow area of ​​molten iron based on dual-spectral image data.

[0027] S102 specifically includes the following execution methods: S1021: The dual-spectral image data acquired in step S101 is split into visible light image data and infrared image data, wherein the visible light image data is used to capture the brightness characteristics of molten iron, and the infrared image data is used to capture the temperature characteristics of slag.

[0028] In some embodiments, the temperature of a substance corresponds to the brightness and color temperature of the emitted visible light. A temperature difference of approximately 200-400°C exists between molten iron (approximately 1600°C) and molten slag (approximately 1200-1400°C), resulting in significantly different brightness and color characteristics in the visible light spectrum. Molten iron is brighter and its color leans towards a bright yellowish-white, while molten slag is relatively darker and its color may lean towards red or dark red.

[0029] In this embodiment, the pixel area corresponding to the 1600℃ isotherm is marked on the visible light image as the calibration basis for the boundary of the main channel of molten iron. The pixel area corresponding to the 1600℃ isotherm is not a direct temperature measurement, but rather a correspondence between specific brightness / color characteristics and the 1600℃ molten iron temperature established through prior calibration.

[0030] Specifically, a high-temperature thermometer is used to measure and record the brightness and color hue of the molten iron at a temperature of 1600℃ in a visible light image. This process is equivalent to creating a visual feature template for the molten iron at 1600℃. In subsequent real-time detection, the YOLOv5 segmentation network learns and identifies pixel regions with brightness values ​​and color hues that match this template, and classifies these regions as molten iron, thus achieving segmentation.

[0031] It can be seen that visible light images do not directly measure temperature, but rather use the differences in brightness and color between molten iron and slag to identify and segment molten iron regions by utilizing pre-calibrated relationships between brightness and temperature, and between color and temperature.

[0032] S1022: Process the split infrared image data, determine the highest temperature Tmax in the iron tap area and the background temperature Tbg outside the iron tap area by pixel traversal, and then calculate the segmentation threshold Tadapt according to Tadapt=0.6Tmax+0.4Tbg. Use the segmentation threshold Tadapt to perform binary segmentation on the infrared image data to obtain the preliminary result of the molten iron area outside the infrared region.

[0033] In some embodiments, the iron tap region and the background region are delineated on the infrared image. All pixels in the iron tap region are traversed, and the maximum temperature value is recorded as Tmax. All pixels in the background region are traversed, and the average temperature is calculated as Tbg. Tmax and Tbg are substituted into the mathematical model to calculate Tadapt. Then, all pixels in the infrared image are traversed again, and pixels with a temperature value ≥ Tadapt are marked as candidate molten iron pixels, and pixels with a temperature value < Tadapt are marked as background pixels, forming a preliminary result of the molten iron region on the outer infrared side.

[0034] S1023: Process the split visible light image data, mark the pixel region corresponding to the 1600℃ isotherm on the visible light image as the calibration basis for the boundary of the main channel of molten iron, and then input the marked visible light image data into the pre-trained YOLOv5 segmentation network to obtain the segmentation result of the molten iron region on the visible light side.

[0035] In some embodiments, the infrared thermal imager can directly measure the temperature value of each pixel. The 1600°C isotherm is extracted from the infrared image data through threshold segmentation. In this embodiment, all pixels with a temperature value equal to or close to 1600°C are found in the infrared image, and the set of these points constitutes the 1600°C isotherm.

[0036] To achieve high-precision segmentation, the temperature information from the infrared image is combined with the high spatial resolution and texture of the visible light image. This is based on registration using dual-spectrum cameras. During system installation and calibration, a coordinate transformation relationship was established between the infrared and visible light cameras, ensuring that a point in the infrared image corresponds to a point in the visible light image. Through this registration relationship, the 1600℃ isotherm found in the infrared image is synchronously mapped onto the visible light image acquired at the same time.

[0037] This embodiment involves annotating 1600℃ isotherms on visible light images. It is based on infrared images, which provide accurate temperature data. The 1600℃ isotherms are extracted and mapped to the visible light image through registration parameters to provide high-resolution edges, which serve as training labels or prior knowledge for the YOLOv5 network.

[0038] It should be further explained that, through on-site measurement using a high-temperature thermometer, the brightness and color hue of the molten iron at a temperature of 1600℃ were recorded in the visible light image. Based on this, the boundary pixels of the main channel of molten iron corresponding to the 1600℃ isotherm were marked on multiple sets of visible light images under different molten iron tapping conditions, forming a training dataset. The dataset was divided into a training set and a validation set in an 8:2 ratio and input into the YOLOv5 segmentation network. The batch size was set to 16 and the number of iterations was set to P for training. Every 100 iterations, the segmentation accuracy was verified using the validation set until the accuracy stabilized above the preset accuracy and training stopped. The visible light image segmented in step S1021 was input into the trained YOLOv5 segmentation network, and the network output the boundary pixel coordinates of the main channel of molten iron, forming the segmentation result of the molten iron region on the visible light side.

[0039] S1024: Convert the visible light side molten iron region segmentation result to the HSV color space, count the distribution range of HSV color components in the region, compare with the HSV color characteristics of iron oxide scale, remove pixels that match the color characteristics of iron oxide scale, and obtain the corrected visible light side molten iron region.

[0040] In some embodiments, by utilizing the characteristic differences between iron oxide scale and molten iron in the HSV color space, optionally, the H value of molten iron is mostly 0-10, the S value is 0.6-0.9, and the V value is 0.7-1.0, which are significantly different from iron oxide scale. Through space transformation and feature matching, the interference of iron oxide scale on the molten iron area can be accurately identified and removed.

[0041] S1025: Perform pixel-level intersection operation between the preliminary result of the outer infrared molten iron region and the corrected visible light side molten iron region, retaining pixels that exist in both regions simultaneously, to form the final effective molten iron flow area.

[0042] In some embodiments, by utilizing the complementarity of the high temperature of molten iron and the specific isotherm of molten iron in infrared images, and the absence of iron oxide scale interference, only pixels that simultaneously satisfy both features are retained through intersection fusion, thereby eliminating high temperature backgrounds that are misjudged on the infrared outer side under a single feature and iron oxide scale that has not been removed on the visible light side.

[0043] This embodiment combines the advantages of both dual-spectral images for complementary verification, further reducing segmentation errors and ensuring that the final effective flow area of ​​molten iron conforms to both temperature and color / boundary characteristics, thereby improving the accuracy and reliability of region separation.

[0044] It can be seen that the high temperature of molten iron and the low temperature of the background in infrared images are used to achieve preliminary segmentation. The segmentation results are optimized by using the brightness of visible light images and specific isotherms of molten iron, as well as the absence of iron oxide scale interference. By fusion of the intersection of the two spectral regions, interference from slag, iron oxide scale, and background is eliminated, and the molten iron region is accurately separated.

[0045] S103: Divide the effective flow area of ​​molten iron into streams to extract the stream outline.

[0046] In some embodiments, the image of the effective flow area of ​​molten iron is smoothed and denoised using a 3×3 Gaussian filter template, and the edge pixels are filled with mirror images.

[0047] In this embodiment, Canny edge detection can be used to extract edges and mark the edge segments of the molten iron outlet and the edge areas of the molten iron landing point. Parallel lines with 5-pixel intervals are drawn perpendicular to the direction of molten iron flow. The intersection points of the parallel lines with the left and right edges of the molten iron area are recorded, and the midpoints of the intersection points are connected to form the center line of the flow stream. Cross-sectional sampling positions are divided every 10 pixels along the center line, and straight lines perpendicular to the center line are drawn. The intersection points of the straight lines with the left and right edges are taken as the contour endpoints, and the endpoints of adjacent sampling positions are connected to form the initial contour. If the distance between adjacent endpoints is >20 pixels, linear interpolation is used to supplement; if the endpoints deviate from the overall trend by >15 pixels, they are discarded to obtain the corrected flow stream contour.

[0048] It can be seen that the centerline reference and contour correction ensure the symmetry and continuity of the contour, and the deviation of the stream contour extraction is ≤3 pixels, providing an accurate contour basis for instantaneous cross-sectional area calculation.

[0049] S104: Calculate the instantaneous cross-sectional area of ​​the stream based on the stream profile.

[0050] Step S104 specifically includes the following methods: S1041: Set the sampling parameters for the instantaneous cross-sectional area, determine the position of the flow cross-section corresponding to each sampling moment, and the position of the flow cross-section corresponds one-to-one with the sampling position of the flow cross-section divided in step S103.

[0051] In some embodiments, the sampling frequency is set according to the industrial requirement of real-time monitoring of blast furnace tapping flow. Each sampling moment is bound to the sampling position of the stream cross-section determined in step S103 by a timestamp, ensuring that the area of ​​the same cross-sectional position is calculated for each sampling. If a brief delay occurs in the acquisition of the dual-spectral image, the sampling moment is synchronously adjusted to match the image frame, ensuring the temporal consistency between sampling and image data.

[0052] S1042: Retrieve the established pixel-physical size mapping relationship, and calculate the actual physical area corresponding to a single pixel through this mapping relationship, that is, the real physical area represented by one pixel in the image on the plane where the flow cross-section is located.

[0053] In some embodiments, a manually calibrated mapping relationship is retrieved. During the initial tapping stage, the actual width of the main channel of molten iron is measured using a high-temperature-resistant measuring ruler. The pixel width corresponding to the main channel in the image is statistically analyzed, and the pixel equivalent is calculated; thus, the physical area of ​​a single pixel is obtained. The mapping relationship is verified every 2 hours: the actual width of the main channel is measured again and compared with the image pixel width. If the pixel width changes by more than 5 pixels, k and S0 are recalculated to ensure the accuracy of the mapping relationship.

[0054] S1043: For the cross-sectional position of the stream corresponding to each sampling time, extract all the contour pixel coordinates of the cross-section from the corrected stream contour obtained in step S103, and arrange the contour pixel coordinates in clockwise order to form a closed cross-sectional contour pixel sequence.

[0055] In some embodiments, for each sampling time corresponding to the cross-sectional position, the coordinates (x, y) of all contour pixels at the cross-sectional position are selected from the corrected flow contour in step S103. An x-coordinate priority sorting method is used: first, the pixels are sorted in ascending order of x-coordinate; if the x-coordinates are the same, they are sorted in descending order of y-coordinate. The sorted coordinates are then adjusted to a clockwise order to ensure that the coordinates of the first and last pixels are the same. If there are gaps in the extracted contour pixels, linear interpolation is used to calculate and supplement the coordinates of the missing pixels, ensuring the closure and continuity of the contour sequence.

[0056] S1044: Perform pixel traversal on the region enclosed by the closed cross-sectional contour pixel sequence, determine whether each pixel is located inside the contour, count the total number of valid pixels inside, multiply the total number of valid pixels by the pixel equivalent calculated in step S1042, and obtain the instantaneous physical cross-sectional area at the sampling moment.

[0057] In some embodiments, a ray-mapping method is used to determine whether a pixel is located inside the contour. A horizontal ray is emitted from the pixel to be judged (x0, y0) to the right, and the number of intersections between the ray and the contour boundary is counted. If the number of intersections is odd, it is determined to be an internal valid pixel; if the number is even, it is an external background pixel. All pixels within the cross-sectional contour area are traversed, and the total number of valid pixels N is counted. N is multiplied by S0 obtained in step S1042. The instantaneous physical cross-sectional area A is calculated.

[0058] In this way, the pixel scale of the contour region is obtained by counting pixels, and then converted into physical area by combining pixel equivalents. The cross-sectional area is quantified from the image level to the physical level using geometric statistics.

[0059] S1045: Record all instantaneous physical cross-sectional areas in the order of sampling time. Compare the cross-sectional area values ​​of three consecutive sampling times. If the difference between the cross-sectional area at a certain time and the values ​​at the two adjacent times both exceed the preset difference threshold, it is determined to be an outlier. Replace the outlier with the average value of the values ​​at the two adjacent times and output the corrected instantaneous cross-sectional area sequence.

[0060] In some embodiments, the instantaneous physical cross-sectional area sequence (A1, A2, A3, ..., An) is recorded in the order of sampling time; for each intermediate time point Ai (2≤i≤n-1), the difference between Ai and Ai-1, |Ai-Ai-1|, and the difference between Ai and Ai+1, |Ai-Ai+1|, are calculated.

[0061] If both differences exceed ΔA, Ai is determined to be an outlier. Replace Ai with (Ai-1+Ai+1) / 2 to obtain the corrected Ai'; if A1 is outlier at the first time step, replace A1 with A2.

[0062] If An is abnormal at the last moment, An is replaced with An-1; the final output is the corrected sequence (A1,A2',A3,...,An), which is directly used for the mass flow rate calculation in step S107.

[0063] Thus, based on the physical characteristics of blast furnace tapping, the cross-sectional area of ​​the stream should change continuously and slowly due to the influence of gravity and tapping pressure. Outliers are identified by the difference between adjacent data and corrected by the average value of adjacent data, ensuring that the data sequence conforms to the actual physical laws.

[0064] S105: Track the trajectory of molten iron drop points based on image sequences and calculate the velocity component.

[0065] In some embodiments, for dual-spectral image sequences, ORB feature point matching is used to align consecutive frames to eliminate camera shake. A rectangular tracking area with a side length of 200 pixels is defined. Within the tracking area for each frame, the brightness peak point of the visible light image and the temperature peak point of the infrared image are taken, and the average coordinates of the two points are calculated as the trajectory point. 50 frames of trajectory points are collected, and outliers that deviate from adjacent points by more than 10 pixels are removed. A smooth trajectory is obtained by fitting a quadratic polynomial. Two points P1(x1,y1,t1) and P2(x2,y2,t2) with a spacing of ≥50 pixels on the trajectory are selected. The actual vertical distance is calculated by Δy=|y2-y1|×k, and the time difference is calculated by Δt=|t2-t1|, to obtain the vertical velocity component v. y =Δy / Δt.

[0066] This embodiment uses ORB feature point alignment to ensure inter-frame offset ≤ 1 pixel, dual-spectral peak points to improve the anti-interference capability of trajectory points, and parabolic fitting to improve trajectory smoothness.

[0067] S106: Dynamically corrects the turbulence coefficient in velocity calculations to optimize the accuracy of velocity components.

[0068] In some embodiments, real-time data from S101-S105 are integrated with historical data from multiple stable furnace cycles, and aligned by timestamps to form a basic database. The turbulence coefficients of historical stable periods are statistically analyzed, and a weighted average is used to obtain the initial coefficient μ0.

[0069] For real-time images, calculate the edge fluctuation, visible light grayscale variance, and infrared temperature variance for 20 frames. If the edge fluctuation is ≤3 pixels / frame, the grayscale variance is ≤20, and the temperature variance is ≤50℃, keep μ0; otherwise, adjust according to the excess ratio.

[0070] This embodiment uses historical data to determine the initial coefficients to ensure the rationality of the foundation, and adjusts the coefficients using real-time turbulence characteristics to adapt to changes in turbulence. It corrects for deviations between the coefficients and actual operating conditions to ensure that the coefficients optimize the accuracy of flow velocity calculations.

[0071] S107: Calculate the mass flow rate based on the instantaneous cross-sectional area and the corrected velocity component to achieve flow output.

[0072] S107 specifically includes the following execution methods: S1071: Integrate the real-time parameters required for mass flow rate calculation, obtain the instantaneous cross-sectional area data from step S104, the corrected vertical velocity component data from step S105, and the dynamically corrected turbulence coefficient data from step S106, align the three sets of data according to the timestamp, and form a real-time calculation parameter set.

[0073] In some embodiments, the cross-sectional area, velocity, and turbulence coefficient required for the calculation of mass flow rate at the same time are eliminated by aligning the timestamps to eliminate spatiotemporal misalignment of parameters, and the uniformity of the format avoids type errors in numerical calculations, thus ensuring the consistency of the basic calculation parameters.

[0074] S1072: Determine the current molten iron density parameter. From the real-time frame of the infrared image in step S101, select five 3×3 pixel sampling points (top, middle, bottom, left, and right) within the effective flow area of ​​the molten iron. Calculate the average temperature of each sampling point and take the overall average value T. Retrieve the correspondence between molten iron density and temperature in the industrial standard and calculate the current molten iron density ρ based on the average temperature T. If the temperature of the sampling point exceeds the normal temperature range of the molten iron and there are less than 3 effective sampling points, enable the default density and mark the density status as pending verification.

[0075] In some embodiments, sampling points avoid the edge of the flow stream, and the average temperature of each sampling point is taken as 3×3 pixels to reduce the impact of noise; if the temperature of a sampling point is >1600℃ or <1400℃, the point is removed and a replacement is selected nearby; the average temperature T is calculated when there are ≥3 valid sampling points, otherwise the default density is enabled; the density is recalculated to ensure that it matches the real-time temperature of the molten iron.

[0076] S1073: Perform mass flow calculation. For each set of data in the real-time calculation parameter set, substitute it into the mass flow formula to calculate the instantaneous mass flow Q_instantaneous. Calculate the instantaneous flow 10 times consecutively. After removing outliers that exceed the adjacent instantaneous values, take the arithmetic mean of the remaining values ​​as the average mass flow Q_average for that period. Record the instantaneous and average Q_average values ​​for each calculation, as well as the corresponding A, v, μ, and ρ parameters.

[0077] In some embodiments, the instantaneous flow rate formula is Q_instantaneous = ρ × A × v × μ, and the average flow rate formula is Q_average = (Q_instantaneous1 + Q_instantaneous2 + ... + Q_instantaneousn) / n, where n is the number of effective instantaneous flow rates.

[0078] By capturing real-time changes through instantaneous calculation and smoothing out interference through average calculation, the real-time performance of flow data is ensured, while avoiding jumps in single instantaneous values ​​caused by dust and light interference, which meets the needs of industrial monitoring.

[0079] S1074: Verify and correct the average mass flow rate data. Select the average mass flow rate data of the same 10 furnace batches and the same tapping stage from the blast furnace historical database, and calculate the historical average value Q_history and its fluctuation range. If the current Q level is within the fluctuation range, it is considered valid; if it is outside the range, check the real-time parameters. If the instantaneous cross-sectional area A changes by more than 15%, replace it with the average value of the previous 3 A values; if the flow velocity v changes by more than 20%, recalculate the flow velocity. If μ is abnormal, replace it with μ from the previous moment, correct it, and recalculate Q until it meets the fluctuation range. If the correction is ineffective after 3 times, mark the data as abnormal.

[0080] In some embodiments, based on the industrial pattern that the iron output flow rate is similar under the same process conditions, a reasonable fluctuation range is established through historical data, anomalies are judged by comparing with the current data, and then the flow rate data is corrected by parameter replacement to ensure that the flow rate data conforms to the actual process logic.

[0081] S1075: Display and store valid Q-level values. Storage is partitioned by day, with fields including timestamp, Q-level, Q-instant 1-Q-instant 10, A, v, μ, ρ, temperature, and status.

[0082] S108: Determines the iron tapping operation status based on dual-spectral image features, outputs flow data and anomaly warnings, and completes full-process analysis. In some embodiments, the image of the molten iron area in S102 and the labeled data on the start of tapping / slag appearance / blockage status are input into the ResNet34-CNN model. Color, shape, and edge features are extracted through convolution, and the operation status is classified and predicted.

[0083] This embodiment also pre-sets abnormal thresholds. Optionally, an abnormal warning can be triggered if the flow rate fluctuation is >15% / min, the slag appearance time is advanced / delayed by >5 minutes, or the residual iron content is >5% when the nozzle is blocked.

[0084] The system marks the start time of tapping, the time of slag appearance, and the time of tapping blockage; calculates the slag-to-iron ratio, cumulative flow rate, and Q-square integral for a single furnace; and outputs these values ​​to the process optimization system.

[0085] In one embodiment of the present invention, based on step S103, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S103 specifically includes the following steps: S1031: Perform temperature thresholding on the infrared image, dynamically determine the segmentation threshold based on the highest temperature in the iron tapping area and the background temperature, and extract the boundary of the molten iron area.

[0086] In some embodiments, by measuring the highest temperature and background temperature in the taphole area, a temperature threshold for the taphole area is calculated, and the infrared image is divided into molten iron and non-molten iron areas. This method can adapt to the temperature fluctuations during blast furnace tapping and improve the accuracy of molten iron extraction.

[0087] S1032: Based on a deep learning model, the boundary of the molten iron stream is identified by the 1600℃ isotherm feature in the visible light image, and the outline of the molten iron stream is obtained.

[0088] In some embodiments, a deep learning model is used to learn the features of molten iron streams in visible light images, and the 1600℃ isotherm is used as a key feature point to extract the contour boundary of the molten iron streams.

[0089] S1033: By using color space analysis, identify and remove interference areas caused by iron oxide scale, thereby optimizing the integrity of the molten iron flow profile.

[0090] In some embodiments, the visible light image is converted to the HSV color space, and the interference areas of iron oxide scale are identified and removed by analyzing the three channels of hue, saturation and lightness.

[0091] This embodiment utilizes the difference in color characteristics between iron oxide scale and molten iron to separate iron oxide scale from the molten iron region using a color space clustering method.

[0092] S1034: Spatial matching of the segmentation results of infrared and visible light images, and fusion to generate a unified flow contour.

[0093] In some embodiments, the molten iron region in the infrared image and the molten iron flow boundary in the visible light image are spatially aligned and fused to generate a unified flow profile. Using the calibration parameters of a dual-spectrum camera, the temperature information from the infrared image and the brightness information from the visible light image are matched in a spatial coordinate system, eliminating the limitations of a single-spectrum image and generating a more complete flow profile.

[0094] S1035: Based on camera calibration parameters, the flow contour in the image is converted into physical dimensions to ensure the physical accuracy of the contour.

[0095] In some embodiments, the flow stream contours in the image are converted into physical dimensions based on the calibration parameters of the dual-spectrum camera, ensuring the physical accuracy of the contours. Using the camera's calibration parameters, a mapping relationship between pixel coordinates and physical dimensions is established, converting the flow stream contours in the image into actual physical dimensions.

[0096] In one embodiment of the present invention, based on step S105, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S105 specifically includes the following steps: Step S1051: In the initial frame of the image sequence, a specific area for tracking is delineated in the falling section of the molten iron flow.

[0097] In some embodiments, by analyzing the morphology of the molten iron stream, a rectangular or custom-shaped region is delineated using edge detection along the parabolic segment after it leaves the taphole and before impacting the surface of the molten iron trough. This region should avoid the area where the stream adheres to the taphole and the area of ​​intense splashing at the point of impact. This avoids interference from areas with complex motion patterns and ensures the validity of the tracking data.

[0098] Step S1052: Within a specific area, select points with image features as tracking points.

[0099] In some embodiments, local abrupt changes in grayscale or texture in an image are utilized. Even if such points undergo slight movements in subsequent frames, the patterns of the surrounding pixel blocks remain relatively stable, making them easy to re-identify and match.

[0100] Step S1053: Track the position movement of the tracking point between consecutive image frames and calculate its pixel displacement.

[0101] In some embodiments, for each feature point selected in the previous frame, a small pixel window is defined around it. In the next frame, the algorithm searches for a region within a preset search area that is most similar to the content of that window image. The coordinate difference between the center points of the two regions is the pixel displacement vector of the feature point between the two frames.

[0102] Step S1054: Based on the pixel displacement and image acquisition time interval, and combined with the spatial coordinate mapping relationship, calculate the actual movement speed of the tracking point.

[0103] In some embodiments, the pixel displacement obtained in step S1053 is converted into a real-world physical displacement using the camera's calibration parameters. This physical displacement is then removed by the acquisition time interval between two camera frames to obtain the instantaneous velocity at that point.

[0104] In this embodiment, the pixel displacement in the image domain is mapped to the physical world through the camera's geometric model, and combined with time information, the motion speed is finally obtained.

[0105] Step S1055: Perform statistical analysis on the velocities of all tracking points to determine the mainstream velocity component of the molten iron stream in the vertical direction.

[0106] In some embodiments, due to turbulence and fluctuations on the surface of the stream, the velocity of a single feature point may not be representative of the whole. This embodiment collects the velocities of all successfully tracked feature points in the vertical direction, then calculates the median of these velocities or identifies their main distribution range through cluster analysis. This statistical value is used as the representative vertical velocity component of the stream at that cross-section to find a velocity value that can represent the motion trend of the main body of the stream.

[0107] In one embodiment of the present invention, based on step S106, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S106 specifically includes the following steps: S1061: Collect the vertical velocity components, instantaneous cross-sectional area, and manually labeled true values ​​of turbulence coefficients from historical frames during blast furnace tapping to construct a dataset; In some embodiments, every 10 frames are extracted from historical iron tapping videos, and the vertical flow velocity and instantaneous cross-sectional area calculated by S104 in that frame are manually read. The corresponding true values ​​of the turbulence coefficients are then labeled in conjunction with sensor data. The dataset contains at least 1,000 sets of flow velocity-cross-sectional area-correction coefficient samples.

[0108] S1062: Perform Min-Max normalization on the vertical velocity and instantaneous cross-sectional area in the dataset, mapping their numerical range to the interval between 0 and 1; In some embodiments, the LSTM network is sensitive to the range of input values. Normalization can eliminate dimensional differences, allowing the network to learn the relationships between features more efficiently; it accelerates the training convergence speed of the LSTM network and avoids gradient explosion or training instability caused by excessively large numerical ranges.

[0109] S1063: Construct an LSTM long short-term memory network. The input is the normalized vertical velocity and instantaneous cross-sectional area of ​​the current frame, and the output is the correction value of the turbulence coefficient. In some embodiments, the input layer involves two nodes, corresponding to the normalized flow velocity and cross-sectional area, the hidden layer involves 64 LSTM units, and the output layer involves one node, corresponding to the turbulence coefficient correction value. The LSTM units in the hidden layer are used to capture the changing trends of flow velocity and cross-sectional area over time; the output layer uses a linear activation function to directly output the correction value.

[0110] S1064: Train an LSTM network using labeled data, use the mean squared error between the predicted corrected value and the true value as the loss function, and iteratively optimize the network weights using the Adam optimizer. In some embodiments, the number of training epochs is set to 60, the batch size is 32, the network weights are adjusted using the Adam optimizer, and the learning rate is set to 0.001. The network performance is verified every 15 epochs, and training is stopped early if the loss no longer decreases.

[0111] It should be noted that after feature vector extraction, the results are predicted using a classification approach. The difference between the predicted and actual results is then calculated, and this difference is the formula for the loss function, specifically the cross-entropy function. The formula is as follows:

[0112] Where i represents the sample; M represents the number of categories; The sign function takes a value of 1 if the true class of sample i is equal to c, and a value of 0 otherwise. This represents the probability that observed sample i belongs to category c.

[0113] This embodiment minimizes the mean square error between the predicted and actual values, enabling the network to learn the mapping relationship between flow velocity, cross-sectional area, and turbulence coefficient correction values.

[0114] S1065: Input the real-time calculated vertical velocity and instantaneous cross-sectional area into the trained LSTM network to obtain the dynamically corrected turbulence coefficient, which is used to adjust the vertical velocity component.

[0115] In some embodiments, the real-time calculated vertical velocity and instantaneous cross-sectional area are input into a trained LSTM network to obtain the predicted correction value μ. corr ; using μ corr Adjusting the initial turbulence coefficient μ0, we obtain the dynamically corrected turbulence coefficient μ = μ0 × μ corr Based on supervised learning and LSTM dynamic correction methods, historical data is used to drive model learning, enabling adaptive adjustment of the turbulence coefficient and improving the robustness and accuracy of flow detection.

[0116] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0117] The following are embodiments of the blast furnace tapping slag and iron flow detection system based on computer vision provided in this disclosure. This system and the blast furnace tapping slag and iron flow detection method based on computer vision in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the blast furnace tapping slag and iron flow detection system based on computer vision, please refer to the embodiments of the blast furnace tapping slag and iron flow detection method based on computer vision.

[0118] like Figure 2 As shown, the system includes: Data acquisition module 201 is used to acquire bispectral image data of the blast furnace taphole; The region separation module 202 separates the effective flow area of ​​molten iron based on dual-spectral image data; The flow segmentation module 203 is used to segment the effective flow area of ​​molten iron into flow segments in order to extract the flow outline; The cross-sectional area calculation module 204 calculates the instantaneous cross-sectional area of ​​the stream based on the stream profile; The flow velocity tracking module 205 tracks the trajectory of molten iron drop points based on image sequences and calculates the flow velocity components; The velocity correction module 206 is used to dynamically correct the turbulence coefficient in the velocity calculation to optimize the accuracy of the velocity components. The flow calculation module 207 calculates the mass flow rate based on the instantaneous cross-sectional area and the corrected flow velocity component, and realizes the flow output. The judgment and early warning module 208 determines the iron tapping operation status based on the dual-spectral image features, and outputs flow data and abnormal warnings to complete the full process analysis.

[0119] like Figure 3 As shown, this application also provides an electronic device, including a display module 103, a memory 102, a processor 101, a communication module 104, and a computer program stored in the memory and executable on the processor 101. When the processor 101 executes the program, it implements the steps of a computer vision-based method for detecting the flow rate of blast furnace tapping slag.

[0120] In embodiments of the present invention, electronic devices include, but are not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments described and / or claimed herein.

[0121] In this embodiment, processor 101 may be implemented using at least one of an application-specific integrated circuit, a programmable logic device, a field-programmable gate array, a processor, a controller, a microcontroller, a microprocessor, or an electronic unit designed to perform the functions described herein. In some cases, such an implementation may be implemented within a controller. For software implementation, implementations such as processes or functions may be implemented with separate software modules that allow the performance of at least one function or operation. Software code may be implemented by a software application (or program) written in any suitable programming language, and the software code may be stored in memory and executed by the controller.

[0122] The display module 103 is used to display information input by the user or information provided to the user. The display module 103 may include a display panel, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like.

[0123] The memory 102 can be used to store software programs and various data. The memory 102 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0124] The communication module 104 transmits radio signals to and / or receives radio signals from at least one of a base station, an external terminal, and a server. Such radio signals may include voice call signals, video call signals, or various types of data sent and / or received according to text and / or multimedia messages.

[0125] The present invention also provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the computer vision-based blast furnace tapping slag and iron flow detection method.

[0126] The storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0127] The storage medium stores a program product capable of implementing the methods described above in this specification. In some possible implementations, various aspects of this disclosure can also be implemented as a program product comprising program code that, when run on a terminal device, causes the terminal device to perform the steps described in the exemplary methods section of this specification according to various exemplary embodiments of this disclosure.

[0128] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A computer vision-based method for detecting slag-iron flow rate in a blast furnace tapping, characterized by, The method comprises: S101: collecting dual-spectrum image data of the blast furnace tap hole; S102: separating the molten iron effective flow-through area based on the dual-spectrum image data; S103: performing flow segmentation on the molten iron effective flow-through area to extract the flow profile; S104: calculating the instantaneous cross-sectional area of the flow based on the flow profile; S105: tracking the molten iron landing point trajectory based on the image sequence and calculating the flow velocity component; S106: dynamically correcting the turbulence coefficient in the flow velocity calculation to optimize the accuracy of the flow velocity component; S107: calculating the mass flow based on the instantaneous cross-sectional area and the corrected flow velocity component to realize flow output; S108: determining the tapping operation state based on the dual-spectrum image features, and outputting the flow data and abnormality warning to complete the whole-process analysis.

2. The computer vision-based detection method of slag iron flow rate of a blast furnace tap hole according to claim 1, characterized in that, S102 specifically comprises the following steps: Split the dual-spectrum image data collected in step S101 into visible light image data and infrared image data; Process the infrared image data, calculate the adaptive threshold, and perform binary segmentation to obtain the preliminary result of the infrared molten iron area; Process the visible light image data, calibrate the molten iron area based on the preset temperature isotheral through the pre-trained YOLOv5 segmentation network to obtain the visible light molten iron area segmentation result; Convert the visible light molten iron area segmentation result to HSV color space and remove the iron oxide scale pixels through color feature matching to obtain the corrected visible light molten iron area; Perform pixel-level intersection operation on the preliminary result of the infrared molten iron area and the corrected visible light molten iron area, retain the common pixels, and form the final molten iron effective flow-through area.

3. The computer vision-based detection method of slag iron flow rate of a blast furnace tap hole according to claim 1, characterized in that, S103 specifically comprises the following steps: Perform temperature threshold processing on the infrared image, dynamically determine the segmentation threshold according to the highest temperature of the tap hole area and the background temperature, and extract the molten iron area boundary; In the visible light image, identify the molten iron flow boundary through the 1600℃ isotheral feature based on the deep learning model to obtain the molten iron flow profile; Optimize the integrity of the molten iron flow profile by identifying and removing the interference area of the iron oxide scale through color space analysis method; Spatially match the segmentation results of the infrared image and the visible light image to generate a unified flow profile; Convert the flow profile in the image to physical size based on the camera calibration parameters to ensure the physical accuracy of the profile.

4. The computer vision-based detection method of slag iron flow rate of a blast furnace tap hole according to claim 1, characterized in that, S104 specifically comprises the following steps: Set the sampling frequency of the instantaneous cross-sectional area, and correspond each sampling time to the flow cross-sectional position determined in step S103; Call the pre-established mapping relationship between pixels and physical size to calculate the actual physical area represented by a single pixel; For the cross-sectional position of each sampling time, extract the profile pixel coordinates from the flow profile obtained in step S103 and organize them into a closed sequence; Statistically count the total number of effective pixels by judging the number of pixels located inside the profile sequence, and calculate the instantaneous physical cross-sectional area combined with the pixel physical area; Perform outlier detection and correction on the time-sequentially recorded instantaneous physical cross-sectional area sequence, and output the corrected cross-sectional area sequence for subsequent calculation.

5. The computer vision-based detection method of slag iron flow rate of a blast furnace tap hole according to claim 1, characterized in that, S105 specifically comprises the following steps: In the initial frame of the image sequence, a specific area for tracking is demarcated in the falling section of the molten iron flow; Within the specific area, points with image features are selected as tracking points; Between consecutive image frames, the position movement of the tracking points is tracked, and their pixel displacements are calculated; Based on the pixel displacements and the image acquisition time interval, and in combination with the spatial coordinate mapping relationship, the actual motion speed of the tracking points is calculated; The speeds of all tracking points are statistically analyzed to determine the main flow speed component of the molten iron stream in the vertical direction.

6. The computer vision-based detection method of slag iron flow rate of a blast furnace tap hole according to claim 1, characterized in that, S106 specifically includes the following steps: Collect the vertical flow speed component, instantaneous cross-sectional area, and artificially labeled true value of the turbulent flow coefficient of the historical frame during the blast furnace tapping process to construct a data set; Perform Min-Max normalization processing on the vertical flow speed and instantaneous cross-sectional area in the data set to map their numerical ranges to the interval of 0 to 1; Build an LSTM long short-term memory network, with the input being the normalized current frame vertical flow speed and instantaneous cross-sectional area, and the output being the correction value of the turbulent flow coefficient; Train the LSTM network using the labeled data, with the mean square error between the predicted correction value and the true value serving as the loss function, and the network weights are iteratively optimized by the Adam optimizer; Input the real-time calculated vertical flow speed and instantaneous cross-sectional area into the trained LSTM network to obtain the dynamically corrected turbulent flow coefficient, which is used to adjust the vertical flow speed component.

7. The computer vision-based detection method of slag iron flow rate of a blast furnace tap hole according to claim 1, characterized in that, S107 specifically includes the following steps: Obtain the instantaneous cross-sectional area data of step S104, the vertical flow speed component data of step S105, and the turbulent flow coefficient data of step S106, and align them by timestamp to form a real-time calculation parameter set; Sample the temperature of the molten iron region from the infrared image of step S101, calculate the average temperature, and determine the molten iron density parameter based on the industry standard; Use the real-time calculation parameter set and the molten iron density parameter to calculate the instantaneous mass flow rate through the mass flow rate formula, and take the average of the consecutive calculation results to obtain the average mass flow rate; Compare the average mass flow rate with the same type of data in the blast furnace historical database, and if it exceeds the fluctuation range, correct the real-time parameters and recalculate until the data is valid; Display and store the valid average mass flow rate and related parameters in real time and in partitions.

8. A blast furnace tap hole slag and iron flow detection system based on computer vision, characterized in that, The system is used to implement the blast furnace tapping slag-iron flow detection method based on computer vision as claimed in any one of claims 1 to 7; The system comprises: a data acquisition module for acquiring dual-spectrum image data of the blast furnace tapping hole; a region separation module for separating the effective molten iron flow area based on the dual-spectrum image data; a stream segmentation module for segmenting the molten iron effective flow area to extract the stream profile; a cross-sectional area calculation module for calculating the instantaneous cross-sectional area of the stream based on the stream profile; a flow speed tracking module for tracking the molten iron falling point trajectory based on the image sequence and calculating the flow speed component; a flow speed correction module for dynamically correcting the turbulent flow coefficient in the flow speed calculation to optimize the accuracy of the flow speed component; a flow rate calculation module for calculating the mass flow rate based on the instantaneous cross-sectional area and the corrected flow speed component to realize flow rate output; a judgment and early warning module for determining the tapping operation state based on the dual-spectrum image features and outputting the flow rate data and abnormal early warning to complete the whole process analysis.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the blast furnace slag-iron flow rate detection method based on computer vision according to any one of claims 1 to 7 when executing the program.

10. A storage medium having stored thereon a computer program, characterized in that The computer program implements the steps of the blast furnace slag-iron flow rate detection method based on computer vision according to any one of claims 1 to 7 when executed by the processor.

Images