A visual-based method and system for identifying slag in molten iron of a blast furnace

Abstract

This invention discloses a vision-based method and system for identifying slag and iron in blast furnace molten iron flow. By extracting the molten iron flow contour boundary region from the molten iron flow image, dust region identification is performed on the molten iron flow contour boundary region to obtain a dust distribution map of the molten iron flow contour boundary region. Based on the molten iron flow contour boundary region and the dust distribution map, the ROI region of the molten iron flow is obtained, and slag and iron identification is performed on the ROI region to obtain the slag-to-iron ratio of the blast furnace molten iron flow. This solves the technical problem of low accuracy in existing blast furnace molten iron flow slag-to-iron identification methods. Through preprocessing of the molten iron flow image and adaptive neighborhood clustering image segmentation, adaptive identification of slag and iron pixels in the molten iron flow is achieved, thereby obtaining real-time statistical values ​​of the slag-to-iron ratio. This method has advantages such as strong real-time performance, high accuracy, strong anti-interference, and no dependence on other detection equipment, and is suitable for long-term stable monitoring of slag and iron discharge status at the tapping site.

Description

Technical Field

[0001] This invention mainly relates to the field of steel industry technology, specifically a vision-based method and system for identifying blast furnace molten iron flow slag and iron. Background Technology

[0002] Blast furnace ironmaking is a core process in the iron and steel metallurgical industry. Iron ore, coke, and other materials undergo complex physical and chemical reactions under high-temperature conditions inside the furnace, ultimately forming molten iron and slag, which are discharged from the taphole. Molten iron and slag are the main products of blast furnace ironmaking, and the slag / iron discharge during the tapping process directly reflects the internal state of the hearth; the slag / iron ratio is a crucial technical indicator. However, the blast furnace's high-temperature, sealed nature makes it impossible to directly observe the slag and iron discharge within the hearth. Online detection of the slag / iron ratio during the tapping process can reflect important information such as the hearth slag / iron level, hearth fluidity, and slag / iron discharge rate. This helps to promptly detect and address abnormal conditions during tapping, preventing serious accidents such as slag accumulation, increased blast pressure, and abnormal furnace conditions. Therefore, achieving blast furnace molten iron flow slag-iron identification is of vital practical significance for the stable and smooth operation of the blast furnace tapping process.

[0003] This patent targets the high-temperature molten iron flow and slag during the blast furnace tapping process. The inspected objects exhibit characteristics such as high temperature, strong light, and strong radiation, accompanied by strong dust, vibration, and noise interference, severely hindering online slag / iron identification of the molten iron flow. Existing technologies mainly fall into two categories: indirect sensor measurement and direct visual detection. Indirect measurement based on mass sensors relies on data from the track scale and the conveyor belt weighing system during tapping, but both data exhibit significant lag, and the moisture content of the slag particles cannot be accurately estimated, resulting in poor slag-to-iron ratio detection accuracy. Visual detection methods primarily utilize short-wave infrared images of the molten iron flow to identify the slag / iron ratio, thereby enabling real-time calculation. However, the high temperature and strong radiation characteristics of the molten iron flow itself, along with the harsh on-site environment, pose significant challenges to slag-to-iron identification from molten iron flow images. Therefore, this patent proposes an adaptive neighborhood clustering slag / iron identification method and system based on short-wave infrared images of the molten iron flow.

[0004] Patent publication number CN101240357A discloses a method for monitoring slag and iron emissions from an extra-large blast furnace and a method for monitoring the amount of slag and iron in the hearth. This method collects slag and iron mass data during the tapping process using a multi-channel tapping track scale and a multi-channel washed slag weighing system, and calculates the final slag and iron emissions using a multi-level computer management system. However, the distance from the tapping spout to the track scale is relatively long, and the tracks contain a significant amount of iron and slag, resulting in a large monitoring lag. Furthermore, the moisture content of the washed slag is difficult to determine, leading to significant errors in its mass measurement. While the system's calculations can be used as a reference, they are insufficient to reflect the real-time tapping status.

[0005] Patent publication number CN110184401A discloses an online detection method for the slag and iron status of a blast furnace. Imaging and lighting devices are installed above the blast furnace tapping area and the slag conveyor belt. Video images of the tapping area's iron trough and the slag conveyor belt are processed, and the slag / iron quantity is calculated using the maximum connected area. A threshold is set to determine the tapping / blocking / slag arrival status. However, this patent's observation targets are subject to significant objective uncertainties and interference. For example, the detection target is highly uncertain when the iron trough is covered by insulation material or an oxide layer of molten iron. Furthermore, the slag undergoes granulation and dehydration processes from the tapping point to the slag conveyor belt, resulting in significant lag. This patent can only identify three different tapping states and cannot provide more detailed slag / iron discharge data during the tapping process online. Its practical guidance for actual operation needs further verification. Summary of the Invention

[0006] The present invention provides a vision-based method and system for identifying blast furnace molten iron flow, slag, and iron, which solves the technical problem of low accuracy in existing blast furnace molten iron flow, slag, and iron identification.

[0007] To address the aforementioned technical problems, the present invention proposes a vision-based blast furnace molten iron flow slag and iron identification method, which includes:

[0008] Capture keyframes from video;

[0009] Extract the molten iron flow contour boundary region from the keyframes of the video image;

[0010] Dust region identification is performed on the boundary region of the molten iron flow contour to obtain a dust distribution map of the boundary region of the molten iron flow contour.

[0011] Based on the outline boundary region of the molten iron flow and the dust distribution map, the ROI region of the molten iron flow is obtained;

[0012] Slag and iron ratios of blast furnace molten iron flow are obtained by identifying the ROI region of the molten iron flow.

[0013] Furthermore, the extraction of the molten iron flow contour boundary region from the molten iron flow image in the video keyframes includes:

[0014] Convert the molten iron flow image to grayscale;

[0015] Bilateral filtering is applied to denoise the grayscale image of molten iron flow.

[0016] The Canny operator is used to extract the rough molten iron flow boundary for the bilaterally filtered molten iron flow image;

[0017] Based on the rough molten iron flow boundary, the Hough transform is used to detect the upper and lower boundary parabolas corresponding to the upper and lower boundaries of the molten iron flow stream.

[0018] The contour boundary region of the molten iron flow is obtained by using the closed region formed by the upper and lower boundary parabolas and the left and right boundaries of the molten iron flow image.

[0019] Furthermore, dust region identification is performed on the boundary region of the molten iron flow contour to obtain a dust distribution map of the boundary region of the molten iron flow contour, including:

[0020] A symmetric zero-area algorithm is used to find peaks in the boundary region of the molten iron flow profile;

[0021] Based on the peak finding results, determine whether the molten iron flow contour boundary region exhibits a single-peak characteristic. If so, the molten iron flow contour boundary region is directly regarded as a dust-free region. If not, determine whether the molten iron flow contour boundary region exhibits a bimodal characteristic. If so, use the OTSU algorithm to perform threshold segmentation on the bimodal peaks, retaining the region corresponding to the peak with the larger gray value as a dust-free region. The dust distribution map is as follows:

[0022]

[0023] Where F dust (i,j) represents the dust state of pixel I(i,j), where 1 indicates dust coverage and -1 indicates no dust coverage. D and R represent the sets of dusty and dust-free regions, respectively.

[0024] Furthermore, based on the molten iron flow contour boundary region and dust distribution map, the ROI region of the molten iron flow is obtained as follows:

[0025] Based on the parabolic equations corresponding to the upper and lower boundaries, the upper and lower boundary points corresponding to the pixels of the molten iron flow in the contour boundary region are obtained by iterative method. If the upper boundary point belongs to the upper boundary of the molten iron flow region and the lower boundary point belongs to the lower boundary of the molten iron flow region, then the upper boundary point is determined as the starting point of the molten iron flow ROI region. With the starting point of the molten iron flow ROI region as the upper left corner, a rectangular region with a preset length threshold and a preset width threshold is constructed.

[0026] Traverse the dust status within the rectangular area. When all the molten iron flow pixels within the rectangular area belong to the dust-free area set and the preset width threshold is less than the maximum width threshold, increment the preset width threshold by 1. Repeat this process until the preset width threshold equals the maximum width threshold.

[0027] Using the starting point of the molten iron flow ROI region as the top left corner, with a preset length threshold as the length, and a final preset width threshold as the width, construct a rectangular region as the molten iron flow ROI region.

[0028] Furthermore, the specific coordinates of the pixel point representing the molten iron flow as the starting point of the molten iron flow ROI region are as follows:

[0029] P s(i, 0.5*(a1(i-b1)) 2 +a2(i-b2) 2 +(c1+c2-k))),

[0030] Where i is the x-coordinate of the pixel of the molten iron flow, (a1,b1,c1) are the parabola equation parameters corresponding to the upper boundary parabola, (a2,b2,c2) are the parabola equation parameters corresponding to the lower boundary parabola, and k is the preset width threshold of the molten iron flow ROI region.

[0031] Furthermore, slag and iron identification in the ROI area of ​​molten iron flow includes:

[0032] Extract grayscale and texture features of the ROI region of molten iron flow;

[0033] Based on grayscale and texture features, an adaptive density peak clustering algorithm is used to perform clustering analysis on the pixels to be clustered in the neighborhood of different sliding windows to obtain the membership list vector of the pixels to be clustered.

[0034] The neighborhood cluster centers of the pixels to be clustered are obtained based on the membership list vector of the pixels to be clustered. The specific formula is as follows:

[0035]

[0036] Where C ij Let ξ() be the neighborhood cluster center of pixel (i,j), and let [M] be the binarization function. 1 i,j M 2 i,j ...M n i,j [] is the membership list vector of pixel (i,j), M n i,j Let ξ(M) be a linked list of the membership degrees of pixel (i,j) to the nth clustering target. n i,j ) is M n i,j The binarized value;

[0037] Find the cluster center points of all sliding windows in the ROI region of molten iron flow;

[0038] Select a sliding window containing three cluster centers, and sort the three cluster centers in descending order according to the mean membership degree of different categories;

[0039] Based on the three cluster center points arranged in descending order, the cluster center reference points for the slag region, iron region, and background region are obtained sequentially.

[0040] The category to which the pixel to be clustered belongs is obtained based on the Euclidean distance between the neighborhood cluster center and the reference point of the cluster center.

[0041] Furthermore, an adaptive density peak clustering algorithm is employed to perform clustering analysis on the pixels to be clustered within different sliding window neighborhoods, obtaining the membership list vector of the pixels to be clustered, including:

[0042] Calculate the discrete Bundle Ferroni index and obtain the optimal cutoff distance based on the discrete Bundle Ferroni index;

[0043] Calculate the cluster center decision coefficient, and obtain the cluster centers based on the global maximum value of the cluster center decision coefficient. The specific formula for calculating the cluster center decision coefficient is as follows:

[0044]

[0045] Where, ω m Let ω be the decision coefficient of the m-th data sample in descending order. max(m) Let ω represent the maximum decision coefficient of the m-th sample. m When obtaining the global maximum value, the first m data samples are determined as cluster centers, ρ i and ρ m+1 Let δ represent the local density formed by the i-th data sample and the (m+1)-th data sample around the data sample with a preset radius, respectively. i and δ m+1 d represents the distance from the i-th data sample and the (m+1)-th data sample to the higher local density point, respectively. j,(m+1) Let N be the Euclidean distance between the j-th data sample and the (m+1)-th data sample in descending order, and N be the total number of samples in descending order.

[0046] Based on the distance from the pixel to be clustered to the cluster center, obtain the membership list vector of the pixel to be clustered.

[0047] Furthermore, obtaining the slag-to-iron ratio of blast furnace molten iron flow includes:

[0048] Based on the slag and iron identification results, the number of slag pixels and the number of molten iron pixels in the ROI area of ​​the molten iron flow are counted.

[0049] The real-time slag-to-iron ratio is calculated based on the number of slag pixels and the number of molten iron pixels.

[0050] The vision-based blast furnace molten iron flow slag and iron identification system provided by this invention includes:

[0051] The present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the vision-based blast furnace molten iron flow slag and iron identification method provided by the present invention.

[0052] Compared with the prior art, the advantages of the present invention are as follows:

[0053] The vision-based blast furnace molten iron flow slag and iron identification method provided in this invention extracts the molten iron flow contour boundary region from the molten iron flow image, identifies the dust region within the molten iron flow contour boundary region, obtains the dust distribution map of the molten iron flow contour boundary region, obtains the molten iron flow ROI region based on the molten iron flow contour boundary region and performs slag and iron identification on the molten iron flow ROI region, and obtains the blast furnace molten iron flow slag and iron ratio. This solves the technical problem of low accuracy in existing blast furnace molten iron flow slag and iron identification methods. By preprocessing the molten iron flow image and adaptive neighborhood clustering image segmentation, adaptive identification of molten iron flow slag and iron pixels is achieved, thereby obtaining real-time slag and iron ratio statistics. This method has advantages such as strong real-time performance, high accuracy, strong anti-interference, and no dependence on other detection equipment, and is suitable for long-term stable monitoring of slag and iron discharge status at the iron tapping site.

[0054] The purpose of this invention:

[0055] The purpose of this invention is to provide a method and system for real-time monitoring of slag and iron based on visual detection for real-time molten iron flow images.

[0056] The purpose of this invention is to provide a method for rapid localization of molten iron streams in high-speed images of blast furnace molten iron flow and accurate extraction of the ROI region of the molten iron flow.

[0057] The purpose of this invention is to provide an adaptive clustering algorithm to achieve rapid and refined identification of low-contrast, irregular and complex textured slag and iron in high-speed molten iron flow images.

[0058] The purpose of this invention is to provide a method for detecting the slag discharge time of molten iron and for real-time calculation of the slag-to-iron ratio during the flow. Attached Figure Description

[0059] Figure 1 This is a structural block diagram of the vision-based blast furnace molten iron flow slag and iron identification device according to Embodiment 2 of the present invention;

[0060] Figure 2 This is a schematic diagram of pixel neighborhood adaptive clustering in Embodiment 2 of the present invention;

[0061] Figure 3 This is a field schematic diagram of the vision-based blast furnace molten iron flow slag and iron identification device according to Embodiment 3 of the present invention;

[0062] Figure 4 This is a structural block diagram of a vision-based blast furnace molten iron flow slag and iron identification system according to an embodiment of the present invention.

[0063] Figure label:

[0064] U1, High-speed camera hardware unit; U2, Video acquisition and transmission unit; U3, Video preprocessing unit; U4, Image preprocessing unit; U5, Molten iron flow slag and iron identification unit; U11, Near-infrared filter; U12, Zoom lens group; U13, CMOS high-speed photosensitive element; U14, Compressed gas purging module; U15, Cooling protective cover; U21, A / D conversion module; U22, Video compression module; U23, Photoelectric signal conversion module; U31, Video encoding and decoding module; U32, Video keyframe extraction module; U41, Molten iron flow contour boundary extraction module; U42, Dust area identification module; U43, Molten iron flow ROI area extraction module; U51, Image feature extraction module; U52, Image classification and segmentation module; U53, Slag-iron ratio weighted statistics module; 10, Memory; 20, Processor. Detailed Implementation

[0065] To facilitate understanding of the present invention, the present invention will be described more fully and in detail below with reference to the accompanying drawings and preferred embodiments, but the scope of protection of the present invention is not limited to the following specific embodiments.

[0066] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0067] Example 1

[0068] The vision-based blast furnace molten iron flow slag and iron identification method provided in Embodiment 1 of the present invention includes:

[0069] Step S101: Capture video keyframes;

[0070] Step S102: Extract the molten iron flow contour boundary region from the molten iron flow image in the video keyframe;

[0071] Step S103: Identify the dust region in the boundary area of ​​the molten iron flow outline to obtain a dust distribution map of the boundary area of ​​the molten iron flow outline.

[0072] Step S104: Obtain the ROI region of molten iron flow based on the molten iron flow contour boundary region and dust distribution map;

[0073] Step S105: Slag and iron identification is performed on the ROI region of the molten iron flow to obtain the slag and iron ratio of the blast furnace molten iron flow.

[0074] The vision-based blast furnace molten iron flow slag and iron identification method provided in this invention extracts the molten iron flow contour boundary region from the molten iron flow image, identifies the dust region within the molten iron flow contour boundary region, obtains the dust distribution map of the molten iron flow contour boundary region, obtains the molten iron flow ROI region based on the molten iron flow contour boundary region and performs slag and iron identification on the molten iron flow ROI region, and obtains the blast furnace molten iron flow slag and iron ratio. This solves the technical problem of low accuracy in existing blast furnace molten iron flow slag and iron identification methods. By preprocessing the molten iron flow image and adaptive neighborhood clustering image segmentation, adaptive identification of molten iron flow slag and iron pixels is achieved, thereby obtaining real-time slag and iron ratio statistics. This method has advantages such as strong real-time performance, high accuracy, strong anti-interference, and no dependence on other detection equipment, and is suitable for long-term stable monitoring of slag and iron discharge status at the iron tapping site.

[0075] Example 2

[0076] This invention enables the accurate calculation of the slag discharge time in the molten iron flow and the calculation of slag / iron discharge statistical indicators by identifying the image features of slag / iron in the blast furnace molten iron flow.

[0077] This invention embodiment achieves molten iron flow slag and iron identification through a vision-based blast furnace molten iron flow slag and iron identification device. Specifically, the vision-based blast furnace molten iron flow slag and iron identification device consists of a blast furnace molten iron flow object, a high-speed camera hardware unit U1, a video acquisition and transmission unit U2, a video preprocessing unit U3, an image preprocessing unit U4, and a molten iron flow slag and iron identification unit U5, as follows: Figure 1 As shown. Its detailed workflow is as follows: The molten iron flow radiation information enters the CMOS high-speed photosensitive element U13 through the near-infrared filter U11 and zoom lens group U12, realizing real-time acquisition of high-speed molten iron video stream information; the acquired real-time video stream sequence is compressed and encoded by the video acquisition and transmission unit U2 and transmitted remotely; the video preprocessing unit U3 decodes the video stream, converts it into a time-series image, and extracts key frames; the image preprocessing unit U4 extracts the boundaries of the molten iron flow key frames and identifies and fits the edges of the molten iron flow. Dust areas are identified within the molten iron flow boundary area, and the final low-dust region ROI of the molten iron flow is extracted; the molten iron slag and iron identification unit U5 uses a sliding window method to traverse and identify the image, obtaining the slag / iron membership degree of each pixel. Finally, based on the slag / iron membership degree of each pixel, the slag / iron information of the image is statistically calculated to complete the slag and iron identification process. The main constituent units are described below:

[0078] High-speed camera hardware unit U1

[0079] This unit includes a near-infrared filter U11, a zoom lens U12, a CMOS high-speed image sensor U13, a compressed gas purging system U14, and a cooling protective cover U15. There is a significant difference in infrared emissivity between slag and molten iron. The near-infrared filter effectively filters out visible light interference, thereby improving the image quality of the molten iron flow and the contrast between the slag and iron targets. The high-speed CMOS sensor, in conjunction with the zoom lens U12, effectively acquires clear images of the high-speed molten iron flow. Due to the harsh environment of industrial sites, including high temperatures, strong radiation, and heavy dust, the compressed gas purging module U14 cools the internal sensor and cleans the lens front, preventing dust from covering the lens and affecting image quality. Simultaneously, the cooling protective cover U15 is externally wrapped with heat-insulating material, effectively protecting the sensor from dust and providing heat insulation to ensure long-term stable operation of the equipment in harsh industrial environments.

[0080] Video capture and transmission unit U2

[0081] This unit uses the A / D conversion module U21 of the video capture card to convert the analog signal of the video stream acquired by U1 into a digital signal, and performs real-time buffering of the video stream in the onboard memory. Further, the video stream information is encoded and compressed by the video compression module U22, thereby achieving high-speed real-time processing of the video stream. The encoded and compressed video stream is then converted from electrical image information signals to optical signals by the photoelectric signal conversion module U23, and the signals are transmitted over long distances at high speed via optical fiber.

[0082] Video preprocessing unit U3

[0083] This unit primarily performs preprocessing of high-speed video streams, providing reliable data for subsequent algorithms. The photoelectric conversion module converts the optical signal from unit U2 into an electrical signal. Simultaneously, it decompresses the compressed video stream, and after passing through the video encoding / decoding module U31, converts the video stream information into a time-series image sequence. For high-speed video streams, there is random dust, noise, and other interference; noisy data entering subsequent algorithms will reduce their accuracy. Since high-speed video streams contain a lot of redundancy within short time intervals, redundant data will increase the burden on subsequent algorithms and severely affect their real-time performance. The video keyframe extraction module U32 extracts keyframes from the molten iron flow based on motion analysis. To improve the real-time performance and system portability of video stream information processing, the video encoding / decoding module U31 and the video keyframe extraction module U32 are integrated into an FPGA programmable hardware computing platform to achieve high-speed parallel preprocessing of video stream information.

[0084] Image preprocessing unit U4

[0085] This unit primarily preprocesses the keyframe image sequence extracted by unit U3, including the molten iron flow contour boundary extraction module U41, the dust region recognition module U42, and the molten iron flow ROI region extraction module U43. It achieves primary feature detection and extraction of high-speed molten iron flow image sequences, removes areas with severe dust and noise interference, and extracts reliable molten iron flow ROI regions.

[0086] The descriptions of each sub-functional module are as follows:

[0087] Iron Flow Profile Boundary Extraction Module U41

[0088] This module primarily handles the preprocessing of molten iron flow images, including image grayscale conversion, image denoising, and image boundary extraction and fitting. To facilitate subsequent processing, a weighted average algorithm is first used to convert the RGB image to grayscale.

[0089] I gray =0.3×I R +0.59×I G +0.11×I B (1)

[0090] For boundary extraction of molten iron flow images, the image features of the molten iron flow boundary are valid information, while the texture information inside the molten iron flow is interference information. To achieve better molten iron flow boundary extraction results, bilateral filtering is used to obtain image I. filt The calculation formula for achieving noise removal while preserving the boundary features of molten iron flow images is as follows:

[0091]

[0092]

[0093] In the formula, q and p represent the input pixel and the center pixel, respectively, i and j are the coordinates of the corresponding pixel, and σs and σr are the control parameters for the spatial domain and value domain filtering of the pixel, respectively.

[0094] Image I of molten iron flow after bilateral filtering filt The Canny operator is used to extract the boundary of the molten iron flow, and image morphological opening operations are used to remove burr features and contour noise within the molten iron flow stream from the contour image. The binarized contour image is I. binary (i,j). However, the outline of the molten iron flow is relatively rough at this point and may exhibit discontinuity. Since the free motion trajectory of a rigid body under gravity is a parabola, the trajectory of the fluid jet is approximated by a parabola in the molten iron flow. The Hough transform is used to detect the parabolic trajectory of the molten iron flow.

[0095] For a binary image of a molten iron stream, the parabolic equation of the trajectory curve of the molten iron stream boundary is defined as follows:

[0096] y = -a(xb) 2 +c (4)

[0097] The boundary pixels in the coordinate system (x,y) space are transformed to the parameter space (a,c) using the Hough transform:

[0098] c = a(xb) 2 -y (5)

[0099] A single pixel representing the boundary of a molten iron flow in the coordinate (x,y) space is transformed into a complete parabola in the (a,b,c) parameter space. The intersection points of different parabolas in the (a,b,c) parameter space represent points in the corresponding (x,y) space belonging to the same parabola. By traversing the number of intersection points in the (a,b,c) parameter space, the optimal parabolic equation in the (x,y) space can be solved. Since the upper and lower edges of the molten iron flow boundary image have the highest pixel density, a threshold is used to select the two coordinate points with the most votes in the (a,b,c) parameter space. The upper and lower edges of the molten iron flow corresponding to the two parabolas are identified by judging them on the y-axis. If the points corresponding to the two parabolas in the (a,b,c) parameter space are (a1,b1,c1) and (a2,b2,c2), respectively, the formula for determining their upper and lower edges is as follows:

[0100]

[0101] In the formula, C1 represents the category of the first parabola, 1 indicates that the edge is the upper edge, and -1 indicates that the boundary is the lower edge. The second parabola has the opposite value to the first, C2 = -C1. Therefore, the two parabolas corresponding to the upper and lower boundaries of the molten iron flow are C1, C2, and C3, C4, C5, C6, C7, C8, C9 ... u C d The closed region formed by the left and right boundaries of the image is the blast furnace molten iron flow region. This eliminates complex background noise interference and improves the efficiency and robustness of subsequent algorithm execution.

[0102] Dust Area Recognition Module U42

[0103] As the tapping process proceeds, trace elements such as sulfur (S) and phosphorus (P) within the furnace hearth are ejected into the air along with the molten iron. Under high-temperature conditions, these elements undergo violent combustion reactions, generating smoke and dust. Simultaneously, the high-speed molten iron flow washes over the iron trough, creating intermittent, complex dust. In the molten iron stream region between the two parabolic curves, there are still numerous uncertain dust interference factors. Since the emissivity of molten iron is 0.4, and the emissivity of slag varies from 0.4 to 0.96 with slag layer thickness, the gray-scale histogram distribution of molten iron and slag under dust-free conditions approximates a normal distribution. However, the dust-covered area has a significant attenuation effect on the infrared radiation of the slag / iron, causing its gray-scale values ​​to cluster in a lower gray-scale range. The image gray-scale is also significantly lower than that of the molten iron region, resulting in a bimodal distribution in the image. Therefore, a symmetric zero-area algorithm is used to find the peak in the ROI region. If the image exhibits a "single-peak" characteristic, it indicates that the molten iron stream region is less affected by dust, and no processing is required for the molten iron stream image. If the image exhibits a "bimodal" pattern, the OTSU algorithm is used to perform threshold segmentation on the bimodal peaks, retaining the region corresponding to the peak with the larger grayscale value. This region represents the area where the molten iron flow is less affected by dust. The dusty region and the dust-free region are represented by sets D and R, respectively. The final dust distribution map is shown below:

[0104]

[0105] In the formula, F dust (i,j) represents the dust state of pixel I(i,j), where 1 indicates dust coverage and -1 indicates no dust coverage.

[0106] Iron Flow ROI Region Extraction Module U43

[0107] For the low-dust areas extracted above, a sliding window traversal method was used to extract the ROI region. Since the dust coverage area is random, and the morphology of the taphole changes with the molten iron flow during the tapping cycle, the ROI region will vary depending on the tapping site conditions. Furthermore, different ROI regions of the same size may be selected at the same time. However, the molten iron stream exhibits a well-defined, regular shape at the taphole. With the intervention of air and the influence of gravity, the end of the molten iron stream gradually diverges, which is not conducive to the identification and statistical analysis of molten iron slag and iron. Therefore, the ROI region of the molten iron stream is set near the taphole, with the upper bound of the ROI length and width set as (l, w), and the dust coverage state of the molten iron stream F... dust (i,j) are used as constraints. The steps for solving the ROI region of molten iron flow are as follows:

[0108] Step 1: Set the x and y coordinates of the pixel to (i,j), and set the maximum length of the ROI region to l. max The upper and lower boundary points (i, a1(i-b1)) are calculated using the parabolic curves of the molten iron flow. 2+c1) and (i,a2(i-b2)) 2 +c2), and determine whether the current boundary point is the edge of the iron flow stream. If I binary (i, a1(i-b1)) 2 If +c1)=1, then record the current starting point P as the upper edge of the iron outlet. u Similarly, if I binary (i, a2(i-b2)) 2 If +c2)=1, then record the current starting point P as the lower edge of the iron outlet. d When P is not detected simultaneously u P d When i = i + 1, Step 1 is repeated.

[0109] Step 2: Determine the starting point P of the ROI rectangular region based on the detected starting points of the upper and lower boundaries of the iron tap. s (i, 0.5*(a1(i-b1)) 2 +a2(i-b2) 2 +(c1+c2-k))) with P s Construct a rectangular region R with length l and width k, centered at the top left corner. iron Traverse the dust state F within the rectangular area. dust (i,j). When all pixels in the region satisfy condition F dust (i,j)=-1&l<l max At that time, l = l + 1 and Step 2 is repeated.

[0110] Step 3: Based on the starting point coordinates P s (i, 0.5*(a1(i-b1)) 2 +a2(i-b2) 2 +(c1+c2-k))), with length l and width k, in the filtered image I filt The ROI region of the molten iron flow is obtained by cutting the middle section.

[0111] Iron slag and iron identification unit U5

[0112] The slag-iron identification unit U5 in this embodiment includes an image feature extraction module U51, an image classification and segmentation module U52, and a slag-iron ratio weighted statistics module U53. Specifically, slag-iron identification is performed on the ROI region image of the molten iron flow. Since the infrared emissivity of molten iron is 0.4, while the emissivity of slag varies between 0.4 and 0.96 with the slag layer thickness, and the temperatures of slag and iron also show significant differences in different regions of the molten iron flow, the grayscale ranges of slag and iron overlap, resulting in a unimodal distribution in the grayscale histogram. Segmenting the image solely based on global grayscale is insufficient to achieve ideal results. Therefore, this unit proposes an adaptive neighborhood weighted clustering slag / iron identification algorithm, the detailed process of which is as follows:

[0113] Adaptive Clustering Slag and Iron Recognition Algorithm

[0114] The grayscale information of pixels in molten iron flow is simultaneously determined by changes in molten iron temperature, dust interference, and slag layer thickness, exhibiting a large range of random variations. This makes it difficult to determine the slag / iron segmentation threshold for high-precision slag / iron identification. Therefore, this invention proposes an Adaptive Density Peak Clustering (ADPC) algorithm for dynamic slag / iron identification in molten iron flow. Traditional DPC algorithms primarily determine the clustering results using parameters ρ and σ, defined as follows:

[0115]

[0116] In the formula ρ i Represents the area around each data sample with a radius of d. c The local density is calculated using a Gaussian kernel to increase the algorithm's robustness. The algorithm defines the distance δ to points with higher local density. i Defined as:

[0117]

[0118] δ of data samples with global maximum density i Set as:

[0119] δ k =max(d ij (10)

[0120] However, the cutoff distance d of this algorithm c And the number of cluster categories is difficult to determine, d c The value of ρ directly determines i With δ i The distribution of the target area affects the clustering results. Furthermore, for slag / iron identification, the cluster categories may include any combination of slag / iron / background regions, and the number of clusters will dynamically change with the target area and cannot be manually determined. Therefore, the number of clusters also needs to be adaptively determined.

[0121] First, to determine the optimal cutoff distance dc Introducing the discrete Bundle Ferroni index:

[0122]

[0123] In the formula B n Let P be the discrete Bundle Feroni index, n be the total number of intervals divided by the sample, i be the corresponding sample interval sequence number, and P be the discrete Bundle Feroni index. i With Q i Defined as:

[0124]

[0125] In the formula p j ρ represents the probability that each data sample is selected as a cluster center. i δ i The larger the value of ρ, the greater the probability of it being selected as a cluster center. i δ i Normalized value:

[0126]

[0127] Because of B n When the global maximum value is obtained, all sample parameters ρ i δ i The highest degree of order is most conducive to image clustering and segmentation, thus achieving the optimal truncation distance d. c The choice.

[0128] In the clustering process, the determination of the number of clusters directly affects the clustering results. Since the number of target clusters in a target region is uncertain, setting the wrong number of clusters will lead to serious category identification errors. To achieve automatic determination of the number of clusters, a cluster center decision coefficient is proposed:

[0129]

[0130] Where, ω m Let ω be the decision coefficient of the m-th data sample in descending order. max(m) Let ω represent the maximum decision coefficient of the m-th sample. m When obtaining the global maximum value, the first m data samples are determined as cluster centers, ρ i and ρ m+1 Let δ represent the local density formed by the i-th data sample and the (m+1)-th data sample around the data sample with a preset radius, respectively. i and δ m+1 d represents the distance from the i-th data sample and the (m+1)-th data sample to the higher local density point, respectively. j,(m+1) Let N be the Euclidean distance between the j-th data sample and the (m+1)-th data sample in descending order, and N be the total number of samples in descending order.

[0131] When ω m When the global maximum value is obtained, the difference between two data points is maximized. The first m data samples are then selected as cluster centers, thus achieving adaptive determination of the number of categories. The membership relationship between pixels and clusters is defined by the distance from each pixel to different cluster centers.

[0132]

[0133] In the formula μ ij d represents the membership degree of pixel i to cluster center j, m is the total number of clusters, and d ij Let d be the Euclidean distance between pixel i and cluster center j. ik Let be the Euclidean distance between pixel i and pixel k. It should be noted that, in this embodiment, the clustered points represent data samples corresponding to each pixel, specifically a one-dimensional vector composed of the pixel's grayscale value and the extracted features.

[0134] Slag / Iron Neighborhood Adaptive Recognition

[0135] Due to the uneven distribution of dust, temperature, and slag layer thickness throughout the molten iron flow, the grayscale distribution of slag / iron pixels overlaps, making high-precision slag / iron identification difficult using a global image. However, for local images, the temperature variation is smaller, the dust distribution is approximately uniform, and the difference in emissivity between slag and iron results in higher image discrimination. Therefore, a slag / iron neighborhood adaptive identification method is proposed, such as... Figure 2 As shown in the figure, each point represents an image pixel, and the central black pixel represents the pixel to be classified. To ensure the consistency of original temperature and dust interference within the cluster neighborhood, p×p slag / iron cluster neighborhoods are constructed. Each neighborhood contains the central pixel to be clustered, and the sliding step size of each neighborhood is q. Therefore, the pixel to be clustered belongs to (p / q). 2 A sliding window is used, and then cluster analysis is performed on the pixels in different sliding windows.

[0136] Because slag / iron exhibits variations in brightness and darkness in images, and also possesses complex texture information, to fully utilize the image's grayscale and texture information, the image feature extraction module U51 extracts the image's grayscale value domain information and spatial domain texture information as input parameters for slag / iron identification. The grayscale co-occurrence matrix is ​​used to extract the image's four-dimensional features as follows:

[0137]

[0138]

[0139]

[0140]

[0141] Where Asm, Ent, Con, and Idm are the second-order angular moment, entropy, contrast, and inverse difference matrix, respectively.

[0142] The ADPC algorithm is used to perform clustering analysis on the above features plus image grayscale features in the neighborhood of different sliding windows, and the membership list vector [M] of each pixel is obtained. 1 i,j M 2 i,j ...M n i,j ], where M 1 i,j Let represent the membership list of pixel (i,j) to the first clustering target. Based on this, the neighborhood cluster center corresponding to pixel (i,j) is defined as:

[0143] C ij =Max(∑ξ(M) 1 i,j ),∑ξ(M 2 i,j )...∑ξ(M n i,j ))

[0144]

[0145] In the formula C ij The parameters of the cluster center with the highest membership degree for pixel (i,j) are obtained.

[0146] Since the different sliding windows have different combinations of slag / iron / background target categories, it is impossible to determine the specific category to which it belongs (slag / iron / background). Therefore, pixel category mapping is achieved by solving for the cluster center points of all sliding windows in the global ROI image. First, target blocks containing three target categories are selected from all sliding windows, and their corresponding cluster center points [c 1 ,c 2 ,c 3 Then, the cluster centers are sorted in descending order according to the mean membership degree of each category to obtain [c 1 ',c 2 ',c 3 From the emissivity characteristics of slag / iron, it can be seen that slag has the highest emissivity, followed by molten iron, and the background region has the lowest. Therefore, the cluster centers of different categories of pixels also follow this pattern. Thus, the reference points for the cluster centers of the slag / iron / background regions are as follows:

[0147]

[0148] In the formula c slag ciron c backg These represent the cluster center reference points for the slag / iron / background regions, respectively. For different pixels, the corresponding neighborhood cluster center C... ij With c slag c iron c backg The distance is represented by the weighted Euclidean distance:

[0149]

[0150] In the formula, (a, b) correspond to the neighborhood cluster centers C, respectively. ij With c slag c iron c backg w i The corresponding weights assigned to features of different dimensions are used to calculate the corresponding distances. Ultimately, the pixel is classified into the category with the smallest distance.

[0151] By counting the number of slag pixels (n) across the entire ROI region. slag And the number of molten iron pixels n iron Calculate the real-time slag / iron ratio n slag / n iron Meanwhile, when slag pixels are detected in the molten iron stream, the corresponding time is recorded, enabling real-time monitoring of slag removal time.

[0152] The vision-based blast furnace molten iron flow slag and iron identification method provided in this invention extracts the molten iron flow contour boundary region from the molten iron flow image, identifies the dust region within the molten iron flow contour boundary region, obtains the dust distribution map of the molten iron flow contour boundary region, obtains the molten iron flow ROI region based on the molten iron flow contour boundary region and performs slag and iron identification on the molten iron flow ROI region, and obtains the blast furnace molten iron flow slag and iron ratio. This solves the technical problem of low accuracy in existing blast furnace molten iron flow slag and iron identification methods. By preprocessing the molten iron flow image and adaptive neighborhood clustering image segmentation, adaptive identification of molten iron flow slag and iron pixels is achieved, thereby obtaining real-time slag and iron ratio statistics. This method has advantages such as strong real-time performance, high accuracy, strong anti-interference, and no dependence on other detection equipment, and is suitable for long-term stable monitoring of slag and iron discharge status at the iron tapping site.

[0153] Example 3

[0154] This embodiment, in conjunction with the accompanying drawings, provides a detailed description of the specific implementation scheme of the present invention. The present invention is applied to 1500m... 3 / 2650m 3 / 4800m 3 Multiple types of blast furnaces, at the blast furnace taphole according to Figure 3Install a corresponding high-speed image detection device. Due to the high flow rate of molten iron on site, to prevent motion artifacts, the frame rate is required to be above 300fps, and the exposure time is set to be below 3ms. Simultaneously, the slag / iron emissivity has high distinguishability in short-wave infrared, so a long-pass filter of 930nm or higher is used to filter out visible light interference. To improve system accuracy, the system's unit pixel resolution should be higher than 1mm. The specific implementation steps for completing the real-time slag-iron ratio detection process of molten iron flow are as follows:

[0155] 1. Adjust the camera's position, lens aperture, and focal length by setting parameters such as installation distance and angle on-site. Set parameters such as sensor exposure time and gain. Also, calibrate the high-speed camera acquisition unit U1 to correct image distortion.

[0156] 2. During the tapping process, U1 acquires a video of the molten iron flow for a specified duration, which is then compressed and encoded by U2 and transmitted to subsequent units for algorithm execution. Simultaneously, the current buffer is cleared, and the image acquisition operation is repeated.

[0157] 3. The video preprocessing unit U3 transcodes the compressed video stream transmitted in real time. The U32 module extracts the image key frame sequence corresponding to the video stream based on the motion state of the molten iron flow to reduce information redundancy and improve the computational efficiency of subsequent algorithms.

[0158] 4. The image preprocessing unit U4 reads the keyframe sequence of the image. The molten iron flow contour boundary extraction module U41 performs grayscale conversion, bilateral filtering, Canny operator boundary extraction, and Hough transform flow stream boundary detection on the image. The dust region recognition module U42 performs dust recognition on the image and calculates the dust mask image. The molten iron flow ROI region extraction module extracts the ROI mask of the dust-free region and extracts it from the filtered image to obtain the complete ROI region of the molten iron flow.

[0159] 5. For the extracted ROI image of molten iron flow, a sliding window is used to segment the image into its neighborhood. The image feature extraction module U51 extracts the spatial and texture features of the segmented image to construct a neighborhood image feature set. The image clustering and segmentation module U52 performs adaptive clustering recognition on the neighborhood feature set, automatically calculating the truncation distance and the number of cluster centers. The membership degree of each pixel to different cluster centers forms a matrix linked list. Statistical calculations are performed on the corresponding cluster center blocks of each sliding window to obtain the category center reference for the global image. Then, the category of different pixels is calculated through probability statistics.

[0160] 6. The slag-iron ratio weighted statistics module U53 performs statistical calculations on the slag / iron pixels of the molten iron flow image based on the category calculation results, monitors the real-time slag discharge time, and obtains the real-time slag / iron ratio.

[0161] The vision-based blast furnace molten iron flow slag and iron identification method provided in this invention extracts the molten iron flow contour boundary region from the molten iron flow image, identifies the dust region within the molten iron flow contour boundary region, obtains the dust distribution map of the molten iron flow contour boundary region, obtains the molten iron flow ROI region based on the molten iron flow contour boundary region and performs slag and iron identification on the molten iron flow ROI region, and obtains the blast furnace molten iron flow slag and iron ratio. This solves the technical problem of low accuracy in existing blast furnace molten iron flow slag and iron identification methods. By preprocessing the molten iron flow image and adaptive neighborhood clustering image segmentation, adaptive identification of molten iron flow slag and iron pixels is achieved, thereby obtaining real-time slag and iron ratio statistics. This method has advantages such as strong real-time performance, high accuracy, strong anti-interference, and no dependence on other detection equipment, and is suitable for long-term stable monitoring of slag and iron discharge status at the iron tapping site.

[0162] Reference Figure 4 The vision-based blast furnace molten iron flow slag and iron identification system proposed in this embodiment of the invention includes:

[0163] The system includes a memory 10, a processor 20, and a computer program stored on the memory 10 and executable on the processor 20. When the processor 20 executes the computer program, it implements the steps of the vision-based blast furnace molten iron flow slag and iron identification method proposed in this embodiment.

[0164] The specific working process and working principle of the vision-based blast furnace molten iron flow slag and iron identification system in this embodiment can be referred to the working process and working principle of the vision-based blast furnace molten iron flow slag and iron identification method in this embodiment.

[0165] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A vision-based method for identifying blast furnace molten iron flow slag and iron, characterized in that, The method includes: Capture keyframes from video; Extract the molten iron flow contour boundary region from the keyframes of the video image; Dust region identification is performed on the boundary region of the molten iron flow contour to obtain a dust distribution map of the boundary region of the molten iron flow contour. Based on the outline boundary region of the molten iron flow and the dust distribution map, the ROI region of the molten iron flow is obtained; Slag and iron identification is performed on the ROI region of the molten iron flow to obtain the slag-to-iron ratio of the blast furnace molten iron flow. This slag and iron identification on the ROI region of the molten iron flow includes: Extract grayscale and texture features of the ROI region of molten iron flow; Based on grayscale and texture features, an adaptive density peak clustering algorithm is used to perform clustering analysis on the pixels to be clustered within different sliding window neighborhoods, obtaining the membership list vector of the pixels to be clustered. The membership list vector obtained by using the adaptive density peak clustering algorithm to perform clustering analysis on the pixels to be clustered within different sliding window neighborhoods includes: Calculate the discrete Bundle Ferroni index and obtain the optimal cutoff distance based on the discrete Bundle Ferroni index; Calculate the cluster center decision coefficient, and obtain the cluster centers based on the global maximum value of the cluster center decision coefficient. The specific formula for calculating the cluster center decision coefficient is as follows: Where, ω m Let ω be the decision coefficient of the m-th data sample in descending order. max(m) Let ω represent the maximum decision coefficient of the m-th sample. m When obtaining the global maximum value, the first m data samples are determined as cluster centers, ρ i and ρ m+1 Let δ represent the local density formed by the i-th data sample and the (m+1)-th data sample around the data sample with a preset radius, respectively. i and δ m+1 d represents the distance from the i-th data sample and the (m+1)-th data sample to the higher local density point, respectively. j,(m+1) Let N be the Euclidean distance between the j-th data sample and the (m+1)-th data sample in descending order, and N be the total number of samples in descending order. Based on the distance from the pixel to be clustered to the cluster center, obtain the membership list vector of the pixel to be clustered; The neighborhood cluster centers of the pixels to be clustered are obtained based on the membership list vector of the pixels to be clustered. The specific formula is as follows: Where C ij Let ξ(·) be the neighborhood cluster center of pixel (i,j), and let [M] be the binarization function. 1 i,j M 2 i,j ...M n i,j [] is the membership list vector of pixel (i,j), M n i,j Let ξ(M) be a linked list of the membership degrees of pixel (i,j) to the nth clustering target. n i,j ) is M n i,j The binarized value; Find the cluster center points of all sliding windows in the ROI region of molten iron flow; Select a sliding window containing three types of cluster centers, and sort the three types of cluster centers in descending order according to the mean membership degree of different categories; Based on the three cluster center points arranged in descending order, the cluster center reference points for the slag region, iron region, and background region are obtained sequentially. The category to which the pixel to be clustered belongs is obtained based on the Euclidean distance between the neighborhood cluster center and the reference point of the cluster center.

2. The vision-based blast furnace molten iron flow slag and iron identification method according to claim 1, characterized in that, The molten iron flow contour boundary region extracted from the molten iron flow image in the video keyframes includes: Convert the molten iron flow image to grayscale; Bilateral filtering is applied to denoise the grayscale image of molten iron flow. The Canny operator is used to extract the rough molten iron flow boundary for the bilaterally filtered molten iron flow image; Based on the rough molten iron flow boundary, the Hough transform is used to detect the upper and lower boundary parabolas corresponding to the upper and lower boundaries of the molten iron flow stream. The contour boundary region of the molten iron flow is obtained by using the closed region formed by the upper and lower boundary parabolas and the left and right boundaries of the molten iron flow image.

3. The vision-based blast furnace molten iron flow slag and iron identification method according to claim 2, characterized in that, Dust region identification was performed on the boundary region of the molten iron flow profile to obtain a dust distribution map of the boundary region, including: A symmetric zero-area algorithm is used to find peaks in the boundary region of the molten iron flow profile; Based on the peak finding results, determine whether the molten iron flow contour boundary region exhibits a single-peak characteristic. If so, the molten iron flow contour boundary region is directly regarded as a dust-free region. If not, determine whether the molten iron flow contour boundary region exhibits a bimodal characteristic. If so, use the OTSU algorithm to perform threshold segmentation on the bimodal peaks, retaining the region corresponding to the peak with the larger gray value as a dust-free region. The dust distribution map is as follows: Where F dust (i,j) represents the dust state of pixel I(i,j), where 1 indicates dust coverage and -1 indicates no dust coverage. D and R represent the sets of dusty and dust-free regions, respectively.

4. The vision-based blast furnace molten iron flow slag and iron identification method according to claim 3, characterized in that, Based on the molten iron flow contour boundary region and dust distribution map, the ROI region of the molten iron flow is obtained as follows: Based on the parabolic equations corresponding to the upper and lower boundaries, the upper and lower boundary points corresponding to the pixels of the molten iron flow in the contour boundary region are obtained by iterative method. If the upper boundary point belongs to the upper boundary of the molten iron flow region and the lower boundary point belongs to the lower boundary of the molten iron flow region, then the upper boundary point is determined as the starting point of the molten iron flow ROI region. With the starting point of the molten iron flow ROI region as the upper left corner, a rectangular region with a preset length threshold and a preset width threshold is constructed. Traverse the dust status within the rectangular area. When all molten iron flow pixels within the rectangular area belong to the dust-free area set and the preset width threshold is less than the maximum width threshold, increment the preset width threshold by 1. Repeat this process until the preset width threshold equals the maximum width threshold. Using the starting point of the molten iron flow ROI region as the top left corner, with a preset length threshold as the length, and a final preset width threshold as the width, construct a rectangular region as the molten iron flow ROI region.

5. The vision-based blast furnace molten iron flow slag and iron identification method according to claim 4, characterized in that, The specific coordinates of the pixel point representing the molten iron flow as the starting point of the molten iron flow ROI region are as follows: P s (i,0.5*(a1(i-b1) 2 +a2(i-b2) 2 +(c1+c2-k))), Where i is the x-coordinate of the pixel of the molten iron flow, (a1,b1,c1) are the parabola equation parameters corresponding to the upper boundary parabola, (a2,b2,c2) are the parabola equation parameters corresponding to the lower boundary parabola, and k is the preset width threshold of the molten iron flow ROI region.

6. The vision-based blast furnace molten iron flow slag and iron identification method according to claim 5, characterized in that, Obtaining the slag-to-iron ratio of blast furnace molten iron flow includes: Based on the slag and iron identification results, the number of slag pixels and the number of molten iron pixels in the ROI area of ​​the molten iron flow are counted. The real-time slag-to-iron ratio is calculated based on the number of slag pixels and the number of molten iron pixels.

7. A vision-based blast furnace molten iron flow slag and iron identification system, the system comprising: The memory (10), the processor (20), and the computer program stored in the memory (10) and executable on the processor (20) are characterized in that the processor (20) implements the steps of the method according to any one of claims 1 to 6 when executing the computer program.

Images