A method for detecting a ladle liquid level based on AI image recognition of a ladle profile

Abstract

The present application relates to the technical field of computer vision, specifically to a ladle liquid level detection method based on AI image recognition of ladle profile, comprising the following steps: generating an overlapping area gradient histogram by vector synthesis of pixel gray scale, constructing a solid-liquid interface profile decoupling set by analyzing extreme values, generating a wear bias array by calculating edge coordinate distance, constructing a polygon cross-section boundary to obtain a measured volume by accumulating microelements, and generating a detection result by correlating a dynamic reference with a physical allowance, in the present application, the interface profile is stripped by analyzing extreme values for gray scale differences and overcoming optical interference, the radial distance is calculated by extracting coordinates and calculating the spatial centroid to generate a bias array, the inner wall wear error caused by the metallurgical process is compensated, the cross-section boundary is constructed based on the bias array and spatial microelement accumulation is performed, the planar view depth is mapped to a real three-dimensional physical volume, and the state is judged by obtaining a dynamic penalty multiplier and linking a static reference.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method for detecting liquid level in slag bags based on AI image recognition of slag bag contours. Background Technology

[0002] Computer vision technology is a discipline that studies how to enable machines to have visual perception capabilities. It mainly includes core aspects such as image acquisition, feature extraction, object detection, and image segmentation. This field captures optical information from the real world using camera equipment, transforming continuous images into a multi-dimensional matrix pixel array. Then, it uses a deep neural network architecture to perform convolution operations and feature mapping on the pixel set, completing the analysis of the objective physical scene at the image level. Among these methods, the traditional AI-based image recognition method for slag ladle contour detection refers to the technique of monitoring the filling position of molten copper slag inside the slag ladle during metallurgical operations. This traditional method typically involves fixing the slag ladle contour above the slag furnace. High-definition cameras are installed to acquire real-time two-dimensional image sequences containing slag bags. Each frame of the two-dimensional image is then input into a lightweight convolutional neural network for forward propagation to select the region of interest (ROI) of the pixels containing the slag bag. Next, a grayscale thresholding formula is used within the ROI to filter out high-brightness pixel blocks representing high-temperature molten copper slag. An edge detection operator is used to extract the top edge pixels of these high-brightness pixel blocks and connect them to form a two-dimensional liquid level contour line. Finally, the vertical coordinate pixel value of this two-dimensional liquid level contour line in the image coordinate system is directly extracted and compared with a pre-set fixed vertical coordinate pixel constant. When the vertical coordinate pixel value exceeds the pixel constant, a closing command is sent to the control circuit board of the alarm device to trigger an alarm.

[0003] Traditional methods rely on pixel extraction mechanisms based on fixed perspectives in two-dimensional images and perform target segmentation through grayscale threshold comparison. In complex metallurgical environments with strong light flares and dust obscuring the surface, misjudgment of contour edges is prone to occur. The judgment logic of extracting a single ordinate for static threshold comparison is divorced from the actual three-dimensional volume change law in physical space and does not take into account the irregular wear and deformation of the inner wall of the slag bag caused by high-temperature erosion. This makes the static benchmark unable to adapt to the fluctuation of physical capacity margin during dynamic filling process, leading to false alarms and missed alarms and endangering on-site operation safety. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a method for detecting the liquid level in a slag bag based on AI image recognition of the slag bag contour, comprising the following steps: S1: Filter the pixels in the overlapping areas of the image, perform vector synthesis calculation on the horizontal and vertical gray level differences to obtain the gradient magnitude and gradient direction angle, and generate the gradient histogram of the overlapping areas based on the cumulative distribution characteristics. S2: Calculate the local maximum diffusion trend of the gradient histogram of the overlapping region to obtain the attenuation ratio of the main peak and the attenuation ratio of the secondary peak, compare the attenuation deviation of the two peaks, determine the pixel belonging and perform isolation division, and construct the solid-liquid interface contour decoupling set. S3: Calculate the top edge coordinate set and spatial centroid position coordinate of the actual slag bag mouth edge contour in the solid-liquid interface contour decoupling set, calculate the difference between the radial straight distance between edge points and the standard slag bag mouth reference radius, and generate a circumferential wear offset array. S4: Calculate the vertical span of the actual liquid surface edge contour in the solid-liquid interface contour decoupling set to obtain the horizontal thin layer micro-element, multiply the corresponding height ratio multiplier with the circumferential wear offset array to generate a local offset array, superimpose the standard inner radius to construct the polygonal cross-sectional boundary, calculate the product of area and thickness to obtain the volume of the layer micro-element and perform spatial accumulation to generate the real-time solid measurement volume. S5: Calculate the volume change gradient by differentiating the real-time entity's measured volume, calculate the product of the dynamic penalty multiplier and the static safety benchmark as the real-time dynamic early warning benchmark, compare the two to obtain the critical state index, determine the trigger state based on the physical capacity margin, and generate the dynamic liquid level detection result of the slag bag.

[0005] As a further embodiment of the present invention, the gradient histogram of the overlapping region specifically includes the horizontal axis interval, the vertical axis frequency, and the peak coordinate; the decoupling set of the solid-liquid interface contour includes the slag pot opening pixel group, the liquid surface pixel group, and the overlapping boundary line; the circumferential wear offset array specifically refers to the offset angle, the wear depth scalar, and the radial deviation sequence; the real-time entity measured volume specifically includes the bottom reference capacity, the single-layer additional volume, and the total volume scalar; and the dynamic liquid level detection result of the slag pot includes the safety warning sign, the remaining capacity space, and the molten slag liquid level height.

[0006] As a further aspect of the present invention, the step of obtaining the gradient histogram of the overlapping region specifically includes: S101: Collect grayscale image arrays of slag bags at the high-temperature molten slag filling site, perform image morphology analysis and contour boundary stripping on the grayscale image array of slag bags, extract pixels in overlapping areas, perform grayscale subtraction operation on adjacent pixels in overlapping areas along the horizontal and vertical pixel array arrangement directions, extract horizontal and vertical change features, and obtain information sets of horizontal and vertical grayscale differences. S102: Call the information set of horizontal and vertical grayscale differences, construct a two-dimensional spatial orthogonal vector mapping system for the horizontal and vertical grayscale differences, perform square root operation on the two-dimensional spatial orthogonal vectors, extract the gradient magnitude, calculate the ratio of vertical grayscale difference to horizontal grayscale difference, and perform arctangent trigonometric function operation to analyze the direction attribute and obtain the gradient direction angle. S103: Divide the gradient direction angle into multiple uniformly distributed statistical intervals, map the gradient magnitude corresponding to the gradient direction angle to the assigned statistical interval, perform numerical summation clustering operation on all gradient magnitudes contained in each statistical interval, extract the cumulative distribution features, and construct a two-dimensional distribution mapping system with the direction interval as the horizontal axis and the cumulative magnitude as the vertical axis, and generate a gradient histogram of the overlapping region.

[0007] As a further aspect of the present invention, the process of dividing multiple uniformly distributed statistical intervals specifically comprises: The values ​​of all angle elements contained in the gradient direction angle are traversed and compared to obtain the minimum and maximum global angle span. The feature extraction resolution parameter in the data record is called, and a standard angle step scalar is generated based on the feature extraction resolution parameter. The difference between the maximum value of the global angle span and the minimum value of the global angle span is calculated to extract the absolute angle span parameter. A division operation is performed between the absolute angle span parameter and the standard angle step scalar, and the floating-point result generated by the division operation is rounded up to obtain the total number of statistical intervals. The minimum global angle span is used as the starting boundary node. The starting boundary node is continuously offset by the standard angle step size scalar and the total number of statistical intervals to generate a sequence of angle boundary nodes that are connected end to end. The multiple uniformly distributed statistical intervals are then segmented and output based on the angle boundary node sequence.

[0008] As a further aspect of the present invention, the step of obtaining the solid-liquid interface contour decoupling set specifically includes: S201: Perform local maximum diffusion trend extraction operation on the gradient histogram of the overlapping region, identify the first extreme value and the second extreme value in the histogram distribution interval, calculate the attenuation rate for the adjacent statistical area crossing feature corresponding to the first extreme value, obtain the main peak attenuation ratio, calculate the attenuation rate for the adjacent statistical area crossing feature corresponding to the second extreme value, obtain the secondary peak attenuation ratio, summarize and output the bimodal attenuation dataset. S202: Based on the bimodal attenuation dataset, perform a difference comparison operation on the attenuation ratio of the main peak and the attenuation ratio of the secondary peak, extract the bimodal deviation amplitude data, generate the bimodal attenuation deviation and construct an adaptive segmentation threshold for pixel gradient, call the pixels in the overlapping area, compare and analyze the gradient amplitude of each independent pixel with the adaptive segmentation threshold, determine the belonging status of the pixels in the overlapping area based on the comparison results, and obtain the category determination features. S203: Based on the boundary attribution labels of the category determination feature record, perform category isolation processing on the pixels in the overlapping area, divide the pixel array belonging to the molten slag surface attribute and the pixel array belonging to the solid furnace mouth attribute, and perform spatial array combination on the pixel coordinate data after stripping and dividing to establish a solid-liquid boundary contour decoupling set.

[0009] As a further aspect of the present invention, the step of obtaining the circumferential wear bias array specifically includes: S301: Read the pixel matrix data of the solid furnace mouth attribute recorded in the solid-liquid interface contour decoupling set, extract the actual slag bag mouth edge contour, obtain the preset perspective transformation matrix, perform three-dimensional space coordinate system transformation and projection mapping operation on the actual slag bag mouth edge contour and perspective transformation matrix, extract three-dimensional mapping features, and generate the top edge coordinate set. S302: Call the top-level edge coordinate set, perform an average calculation on the horizontal and vertical coordinates of multiple coordinate nodes contained in the top-level edge coordinate set, extract the geometric mean feature, obtain the spatial centroid position coordinates, and calculate the length of the line connecting each independent edge point and the spatial centroid position coordinate in the top-level edge coordinate set to obtain the radial straight-line distance. S303: Call the radial straight distance to obtain the pre-stored standard slag bag mouth reference radius in the data record, perform numerical subtraction between the radial straight distance and the standard slag bag mouth reference radius, extract the wear deviation data corresponding to each independent edge point, and perform circumferential serialization splicing and array combination arrangement of multiple wear deviation data according to the edge distribution angle to generate a circumferential wear offset array.

[0010] As a further aspect of the present invention, the process of obtaining the preset perspective transformation matrix is ​​specifically as follows: The system calls upon the three-dimensional spatial reference coordinate data of multiple fixed physical calibration target points obtained through actual surveying and mapping, and performs spatial array combination processing on the multiple three-dimensional spatial reference coordinate data to establish a standard three-dimensional coordinate set. Collect a field panoramic calibration grayscale image containing the multiple fixed physical calibration target points, perform corner detection and edge feature extraction operations on the field panoramic calibration grayscale image, parse the two-dimensional pixel coordinate data mapped by the multiple fixed physical calibration target points inside the image, and establish a reference two-dimensional coordinate set; The standard three-dimensional coordinate set and the reference two-dimensional coordinate set are aligned with feature points to construct a sequence of matching two-dimensional and three-dimensional point pairs. Perspective projection mapping is performed on the sequence of matching two-dimensional and three-dimensional point pairs to extract spatial dimension transformation parameters and construct an associated numerical array structure to generate the perspective transformation matrix.

[0011] As a further aspect of the present invention, the step of obtaining the real-time entity volume measurement is specifically as follows: S401: Read the pixel matrix data belonging to the slag liquid surface attribute in the solid-liquid interface contour decoupling set, extract the actual liquid surface edge contour, obtain the preset reference bottom coordinates, perform vertical span calculation for the depth correlation between the reference bottom coordinates and the actual liquid surface edge contour, obtain the horizontal thin layer micro element, perform proportional conversion for the elevation position of the horizontal thin layer micro element, and obtain the height proportional multiplier. S402: Perform a numerical multiplication operation between the height ratio multiplier and the circumferential wear offset array to generate a local offset array, obtain the preset standard inner radius corresponding to the depth, perform radial numerical addition and combination of the preset standard inner radius and the local offset array in the polar coordinate system, and perform a rectangular coordinate system transformation operation to extract the cross-sectional node coordinate parameters, connect them into a closed figure based on the cross-sectional node coordinate parameters, and construct the polygonal cross-sectional boundary. S403: Call the polygonal cross-section boundary, perform two-dimensional planar area calculation on the closed area enveloped by the polygonal cross-section boundary, extract area features, obtain pre-set inter-layer slice thickness data, extract thickness features, perform numerical multiplication operation on the area features and thickness features, obtain the layer micro-element volume, perform continuous three-dimensional spatial dimension accumulation operation on multiple layer micro-element volumes arranged vertically, and generate real-time entity measurement volume.

[0012] As a further aspect of the present invention, the process of obtaining the preset baseline bottom coordinates specifically includes: The three-dimensional geometric model data of the slag bag is called from the data record, and the spatial morphology is analyzed based on the three-dimensional geometric model data of the slag bag to peel off the bottom curved surface model of the internal cavity structure. The lowest spatial position extraction operation is performed on the bottom curved surface model to obtain the coordinates of the global lowest spatial nodes; Set the coordinates of the global lowest spatial node as the height reference zero point and perform three-dimensional spatial anchor point association calibration to generate the preset reference bottom coordinates; The process of obtaining the preset standard inner radius corresponding to the depth is as follows: The three-dimensional geometric model data of the slag bag is called from the data record. A vertical scale axis is established based on the center vertical line of the three-dimensional geometric model data of the slag bag. The internal cavity structure of the three-dimensional geometric model data of the slag bag is horizontally segmented along the vertical scale axis to extract the horizontal cross-sectional contour boundary corresponding to multiple discrete elevations. Geometric fitting calculations are performed on the horizontal cross-sectional contour boundary to obtain multiple depth elevation parameters and horizontal inner radius parameters. Associative topology construction is performed on the multiple depth elevation parameters and the horizontal inner radius parameters to establish a numerical lookup table for depth and inner radius dimensions. The depth span parameter corresponding to the actual liquid surface edge contour is called, and interpolation matching operation is performed on the depth span parameter in the numerical lookup table to extract the target size features and generate the preset standard inner radius of the corresponding depth.

[0013] As a further aspect of the present invention, the step of obtaining the dynamic liquid level detection result of the slag bag specifically includes: S501: Perform time-series evolution feature analysis on the data fluctuation curve of the real-time entity's measured volume on a continuous time axis, perform continuous difference derivative operation on the time-series evolution features, extract the volume increment change rate, obtain the volume change gradient, and convert the volume change gradient by mapping a preset attenuation weight coefficient to obtain the dynamic penalty multiplier. S502: Call the dynamic penalty multiplier and the real-time entity measured volume to obtain the preset static safety benchmark in the data record, perform a numerical multiplication operation on the dynamic penalty multiplier and the static safety benchmark, extract the warning boundary parameters of dynamic fluctuations, generate a real-time dynamic warning benchmark, perform a difference comparison analysis on the real-time entity measured volume and the real-time dynamic warning benchmark, extract the spatial volume difference between the two, and obtain the critical state index. S503: Call the critical state index, extract and calculate the remaining loadable volume associated with the critical state index in the internal space of the slag bag, obtain the physical capacity margin, determine the safety trigger state based on the physical capacity margin associated with the critical state index, summarize the liquid level data records, and generate the dynamic liquid level detection result of the slag bag.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, gradient vector synthesis and bimodal attenuation extremum analysis are performed on pixel grayscale differences to effectively strip the solid-liquid interface contour and overcome extremely severe optical interference. The top edge coordinates are extracted and combined with the spatial centroid position to calculate the radial distance to generate an offset array, which compensates for the irreversible wear error of the inner wall caused by the metallurgical process. Based on the local offset array, a polygonal cross-sectional boundary is constructed and a hierarchical micro-element spatial accumulation operation is performed to map the planar feature depth from the two-dimensional perspective to the real three-dimensional physical volume. The temporal evolution is analyzed to extract the volume change rate to obtain the dynamic penalty multiplier. Dynamic early warning parameters are generated by linking the static benchmark and the physical capacity margin is associated to perform comprehensive state judgment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] Please see Figure 1 This invention provides a method for detecting the liquid level in a slag bag based on AI image recognition of the slag bag contour, comprising the following steps: S1: Filter the pixels in the overlapping areas of the image, perform vector synthesis calculation on the horizontal and vertical gray level differences to obtain the gradient magnitude and gradient direction angle, and generate the gradient histogram of the overlapping areas based on the cumulative distribution characteristics. S2: Calculate the local maximum diffusion trend of the gradient histogram of the overlapping region to obtain the attenuation ratio of the main peak and the attenuation ratio of the secondary peak, compare the attenuation deviation of the two peaks, determine the pixel belonging and perform isolation division, and construct the solid-liquid interface contour decoupling set. S3: Calculate the top edge coordinate set and spatial centroid position coordinate of the actual slag port edge profile in the solid-liquid interface profile decoupling set, calculate the difference between the radial straight distance between edge points and the standard slag port reference radius, and generate the circumferential wear offset array. S4: Calculate the vertical span of the actual liquid surface edge contour in the solid-liquid interface decoupling set to obtain the horizontal thin layer micro-element. Multiply the corresponding height ratio multiplier with the circumferential wear offset array to generate a local offset array. Superimpose the standard inner radius to construct the polygonal cross-sectional boundary. Calculate the product of area and thickness to obtain the volume of the hierarchical micro-element and perform spatial accumulation to generate the real-time solid measurement volume. S5: Obtain the volume change gradient by differentiating the real-time entity volume, calculate the product of the dynamic penalty multiplier and the static safety benchmark as the real-time dynamic early warning benchmark, compare the two to obtain the critical state index, determine the trigger state based on the physical capacity margin, and generate the dynamic liquid level detection result of the slag bag.

[0020] The gradient histogram of the overlapping region specifically includes the horizontal axis interval, vertical axis frequency, and peak coordinates. The decoupling set of the solid-liquid interface contour includes the slag pot opening pixel group, the liquid surface pixel group, and the overlapping boundary line. The circumferential wear offset array specifically refers to the offset angle, wear depth scalar, and radial deviation sequence. The real-time entity measurement volume specifically includes the bottom reference capacity, single-layer additional volume, and total volume scalar. The dynamic liquid level detection results of the slag pot include safety warning signs, remaining capacity space, and molten slag liquid level height.

[0021] Please see Figure 2 The specific steps for obtaining the gradient histogram of the overlapping region are as follows: S101: Collect grayscale image arrays of slag bags at the high-temperature molten slag filling site, perform image morphology analysis and contour boundary stripping on the grayscale image array of slag bags, extract pixels in overlapping areas, perform grayscale subtraction operation on adjacent pixels in overlapping areas along the horizontal and vertical pixel array arrangement directions, extract horizontal and vertical change features, and obtain information sets of horizontal and vertical grayscale differences. A grayscale image array of the slag bag at the high-temperature molten slag filling site was acquired with a resolution of 1920×1080, at a frame rate of 30 frames per second to ensure dynamic capture accuracy. For the acquired grayscale image array of the slag bag, an edge detection operator was applied to identify high-brightness closed regions in the image. The outer envelope rectangle of the high-brightness closed region was used as the reference plane for image morphology analysis. Background pixels outside the reference plane were removed to complete the contour boundary stripping. Within the stripped reference plane, pixels with grayscale values ​​greater than a set brightness threshold of 200 and located within the preset geometric intersection range between the slag bag opening and the liquid surface were marked as overlapping region pixels. After extracting the overlapping region pixels, a pixel coordinate system was constructed, with the horizontal pixel array arrangement direction set as the X-axis and the vertical pixel array arrangement direction as the Y-axis. Each target pixel in the overlapping region pixels was traversed, and the grayscale value of that target pixel was read. The grayscale values ​​of its adjacent pixels within one pixel span in the positive X-axis direction and one pixel span in the positive Y-axis direction were also read. The grayscale values ​​of adjacent pixels along the positive X-axis are subtracted from the grayscale values ​​of the target pixel to extract the lateral variation features at that coordinate. Similarly, the grayscale values ​​of adjacent pixels along the positive Y-axis are subtracted from the grayscale values ​​of the target pixel to extract the lateral variation features at that coordinate. This process of subtracting grayscale values ​​from adjacent pixels is repeated for all overlapping pixels. All lateral and lateral variation features are then summarized to form a set of lateral and lateral grayscale difference information containing complete microscopic pixel gradient data. This set provides a quantified underlying pixel difference basis for subsequently constructing an orthogonal vector mapping system.

[0022] S102: Call the information set of horizontal and vertical grayscale differences, construct a two-dimensional spatial orthogonal vector mapping system for the horizontal and vertical grayscale differences, perform square root operation on the two-dimensional spatial orthogonal vectors, extract the gradient magnitude, calculate the ratio of vertical grayscale difference to horizontal grayscale difference, and perform arctangent trigonometric function operation to parse the direction attribute and obtain the gradient direction angle. The system retrieves information sets on horizontal and vertical grayscale differences, reading the horizontal and vertical change feature scalars corresponding to each coordinate point. A two-dimensional orthogonal vector mapping system is constructed with the horizontal axis as the horizontal axis and the vertical axis as the vertical axis. The read horizontal change feature scalars are used as the horizontal axis coordinates, and the vertical change feature scalars as the vertical axis coordinates, thus forming the two-dimensional orthogonal vector corresponding to that pixel. For each established two-dimensional orthogonal vector, its horizontal and vertical axis coordinate values ​​are squared. The two squared values ​​are then summed, and the square root of the sum is performed to extract the gradient magnitude representing the severity of local brightness changes. This value has a valid range between 0 and 255. Based on the gradient magnitude, the directional attribute of the feature vector is further analyzed. The vertical axis coordinate value is used as the dividend, and the horizontal axis coordinate value is used as the divisor. Division is performed with the divisor not equal to 0 to calculate the ratio of vertical grayscale difference to horizontal grayscale difference. After extracting this ratio, it is substituted into the arctangent trigonometric function to calculate the radian value. The calculated radian value is multiplied by 180 and divided by pi (3.14) to convert it into an angle, obtaining the gradient direction angle ranging from 0 to 360 degrees. When the divisor is 0, i.e., there is no grayscale change in the horizontal direction, the gradient direction angle is directly assigned a vertical attribute value of 90 degrees or 270 degrees by default to ensure the completeness of the overall algorithm logic. After converting all pixels, the scalar grayscale difference of a single pixel is converted into vector feature parameters with a clear physical direction.

[0023] S103: Divide the gradient direction angle into multiple uniformly distributed statistical intervals, map the gradient magnitude corresponding to the gradient direction angle to the assigned statistical interval, perform numerical summation clustering operation on all gradient magnitudes contained in each statistical interval, extract the cumulative distribution features, and construct a two-dimensional distribution mapping system with the direction interval as the horizontal axis and the cumulative magnitude as the vertical axis, and generate a gradient histogram of the overlapping region. For the gradient orientation angle spanning the entire angle range from 0 degrees to 360 degrees, multiple uniformly distributed statistical intervals are divided using a preset angle step size as the segmentation granularity. Setting the angle step size to 20 degrees, the 360-degree span is divided into 18 consecutive and non-overlapping angle intervals. The gradient orientation angle value corresponding to each overlapping pixel is extracted, and its value is matched and mapped to the aforementioned 18 assigned statistical intervals, simultaneously carrying the gradient magnitude value corresponding to that pixel. For all gradient magnitude values ​​distributed within each statistical interval, a full-value summation clustering operation is performed, that is, all gradient magnitudes falling within the same angle interval are summed, and the cumulative distribution feature value of each interval is extracted. Subsequently, a two-dimensional distribution mapping system is constructed with the orientation interval as the horizontal axis and the cumulative gradient magnitude value within that interval as the vertical axis.

[0024] Table 1 Statistical Interval Mapping Parameter Table

[0025] As shown in Table 1, the statistical interval mapping parameter table intuitively displays the clustering numerical differences of each interval. Based on the constructed two-dimensional distribution mapping system and the cumulative distribution characteristic parameters of each interval, a gradient histogram of the overlapping region with both visualization and quantification is generated. This structured data form provides a reliable discrete data support architecture for the subsequent extraction of bimodal extreme value decay.

[0026] Please see Figure 3 The specific steps for obtaining the solid-liquid interface contour decoupling set are as follows: S201: Perform local maximum diffusion trend extraction operation on the gradient histogram of the overlapping region, identify the first extreme value and the second extreme value in the histogram distribution interval, calculate the attenuation rate for the adjacent statistical area crossing feature corresponding to the first extreme value, obtain the attenuation ratio of the main peak, calculate the attenuation rate for the adjacent statistical area crossing feature corresponding to the second extreme value, obtain the attenuation ratio of the secondary peak, summarize and output the bimodal attenuation dataset. To extract the local maximum diffusion trend of the cumulative distribution characteristics of the gradient histogram in the overlapping region, a one-dimensional sliding window is first established to traverse the vertical axis numerical array of the entire histogram, with the sliding window step size set to one statistical interval. By comparing the cumulative vertical axis amplitude of the window center node with its adjacent nodes, all local maxima that are simultaneously greater than the values ​​of their two adjacent nodes are selected. These maxima are then sorted in descending order of absolute value, and the highest first extreme value and the second highest second extreme value are extracted. These two extreme values ​​represent the dominant directions of the two most prominent solid-liquid texture boundaries in the image. The attenuation rate is calculated for the adjacent statistical regions corresponding to the first extreme value. Specifically, the peak amplitude data of the first extreme value is read, and the amplitude data of one interval immediately on each side of it is read. The average of the amplitude data on both sides is divided by the peak amplitude data to obtain the main peak attenuation ratio parameter between 0 and 1. Similarly, for the attenuation rate calculation of the adjacent statistical region corresponding to the second extreme value, the peak amplitude data of the second extreme value and the amplitude data of its two adjacent sides are read, and the same division and mean operation is performed to obtain the secondary peak attenuation ratio parameter. The advantage of this attenuation ratio calculation method is that the numerical value directly reflects the concentration in the edge direction; the lower the value, the sharper and more concentrated the texture direction. Finally, the parameters including the first extreme value, the second extreme value, and their corresponding primary peak attenuation ratio and secondary peak attenuation ratio are structurally combined, summarized, and output as a bimodal attenuation dataset.

[0027] S202: Based on the bimodal attenuation dataset, the difference between the attenuation ratio of the main peak and the attenuation ratio of the secondary peak is compared and calculated to extract the bimodal deviation amplitude data, generate the bimodal attenuation deviation and construct an adaptive segmentation threshold for pixel gradient, call the pixels in the overlapping area, compare and analyze the gradient amplitude of each independent pixel with the adaptive segmentation threshold, determine the belonging status of the pixels in the overlapping area based on the comparison results, and obtain the category determination features. Based on the specific numerical records contained in the bimodal attenuation dataset, a difference comparison calculation is performed between the attenuation ratio of the primary peak and the attenuation ratio of the secondary peak. Specifically, the logic involves reading the decimal feature value of the primary peak attenuation ratio, subtracting it from the decimal feature value of the secondary peak attenuation ratio, and taking the absolute value to extract the bimodal deviation amplitude data, which characterizes the difference in sharpness between the two, thus generating the bimodal attenuation deviation. This deviation value reflects the degree of inconsistency in gradient distribution characteristics between the solid furnace mouth boundary and the liquid surface fluctuation boundary. After obtaining the bimodal attenuation deviation, it is linearly weighted and fused with the pre-configured base grayscale reference parameter 120 to construct a dynamic adaptive segmentation threshold for the pixel gradient of the current image. Subsequently, all previously acquired overlapping region pixel data are called, and the gradient amplitude value of each independent pixel is read sequentially. Each read gradient amplitude value is then rigorously compared and analyzed with the constructed adaptive segmentation threshold. If the gradient magnitude of a target pixel is greater than or equal to the adaptive segmentation threshold, the pixel is determined to have high-intensity physical boundary features and its classification is assigned to the category representing a stable structure. If the gradient magnitude is strictly less than the adaptive segmentation threshold, the pixel is determined to exhibit diffuse reflection or low contrast features and its classification is assigned to another category representing liquid fluctuations. By comparing and determining each pixel in the overlapping region, a set of category determination features containing the binary classification labels of all pixels is obtained.

[0028] S203: Based on the boundary attribution labels of the category determination feature records, perform category isolation processing on the pixels in the overlapping area, divide the pixel matrix belonging to the molten slag surface attribute and the pixel matrix belonging to the solid furnace mouth attribute, and perform spatial array combination on the pixel coordinate data after stripping and dividing to establish a solid-liquid boundary contour decoupling set. Based on the boundary attribution labels assigned to each independent pixel in the category determination feature record, a dual-category isolation process (both logical and physical) is performed on the massive overlapping pixel areas. Two independent empty data array buffers are constructed, and the labels attached to each pixel are traversed and read. For pixels labeled as representing liquid fluctuation characteristics, their coordinates and grayscale parameters are precisely routed and written into the first data array buffer, thus defining the pixel array belonging to the molten slag surface attribute. This array macroscopically corresponds to the dynamic reflective and rippled areas of the high-temperature molten slag surface. For pixels labeled as representing stable structural characteristics, their coordinates and grayscale parameters are routed and written into the second data array buffer, defining the pixel array belonging to the solid furnace mouth attribute. This array corresponds to the bright and relatively static metal envelope edge of the slag pot inner wall. After completely separating these two types of pixels with different physical attributes, their X-axis and Y-axis pixel coordinate data in the original image's two-dimensional plane are retained. For these stripped and segmented pixel coordinate data with attribute scalars, a gridded spatial array combination mapping is performed based on their spatial relative positions to ensure that the originally entangled and adhered primitives are decomposed into two independent topological layers that are continuous and do not overlap. After data recombination operations, a high-purity solid-liquid interface contour decoupling set is finally established and output at the memory level, eliminating the interference of strong liquid surface light on the recognition of entity edges.

[0029] Please see Figure 4 The specific steps for obtaining the circumferential wear bias array are as follows: S301: Read the pixel matrix data of the solid furnace mouth attribute recorded in the solid-liquid interface contour decoupling set, extract the actual slag bag mouth edge contour, obtain the preset perspective transformation matrix, perform three-dimensional space coordinate system transformation and projection mapping operation on the actual slag bag mouth edge contour and perspective transformation matrix, extract three-dimensional mapping features, and generate the top edge coordinate set. The pixel data belonging to the solid furnace mouth attribute recorded in the solid-liquid interface contour decoupling set are read. An edge-tracking fitting algorithm is applied to these discretely distributed pixel data points, connecting them into a closed continuous geometric curve, thereby extracting the pixel coordinate set of the actual slag pot mouth edge contour within the two-dimensional image plane. Due to the fixed pitch angle tilt of the camera, this two-dimensional plane contour exhibits elliptical distortion. A preset perspective transformation matrix containing nine internal parameter elements, generated during the initial calibration stage, is retrieved. This matrix is ​​pre-solved using a direct linear transformation method based on four physical reference points installed on-site. For the horizontal and vertical coordinate parameters of each pixel in the actual slag pot mouth edge contour, they are expanded to homogeneous coordinate format and then multiplied with the preset perspective transformation matrix. After the multiplication, the first two components of the resulting new vector are divided by the third homogeneous scaling component, completing the three-dimensional spatial coordinate system transformation and projection mapping operation. This calculation process eliminates the perspective distortion of the camera lens and extracts the spatial geometric three-dimensional mapping features after correcting distortion errors. All discrete three-dimensional node coordinates corrected by reverse perspective projection are stored and serialized, and the original elliptical point set with perspective error is discarded. Finally, a top-level edge coordinate set corresponding to the actual horizontal physical shape of the slag bag opening is generated in the virtual three-dimensional coordinate system.

[0030] S302: Call the top-level edge coordinate set, perform average calculation on the horizontal and vertical coordinates of multiple coordinate nodes contained in the top-level edge coordinate set, extract the geometric mean feature, obtain the spatial centroid position coordinates, and calculate the length of the line connecting each independent edge point and the spatial centroid position coordinate in the top-level edge coordinate set to obtain the radial straight-line distance. The top-level edge coordinate set is invoked, and all valid coordinate nodes within the set are traversed and read. The total number of current coordinate nodes is counted, and the X-axis abscissa values ​​of all nodes are summed sequentially, while the Y-axis ordinate values ​​of all nodes are summed simultaneously. The sum of the obtained abscissa values ​​is divided by the total number of coordinate nodes to extract the geometric mean feature of the abscissa; the sum of the ordinate values ​​is also divided by the total number of nodes to extract the geometric mean feature of the ordinate. The calculated geometric mean of the abscissa and ordinate is recombined into a new two-dimensional coordinate parameter point, which is the spatial centroid position coordinate representing the physical center symmetry position of the entire actual slag bag opening. After establishing the central reference anchor point, each independent edge point in the top-level edge coordinate set is traversed again. For the coordinates of each independent edge point and the obtained spatial centroid position coordinates, the square of the difference between the abscissa and the square of the difference between the ordinate are calculated respectively. These two squared values ​​are added together, and the sum is then subjected to a standard square root operation to rigorously calculate the length of the line connecting the two points. By using Euclidean distance calculation, the radial straight-line distance sequence of each profile section radiating from the central anchor point to the surrounding boundaries is accurately obtained. This numerical array objectively maps the actual physical size span distribution characteristics of the slag port at various angles in the plane.

[0031] S303: Call the radial straight distance, obtain the standard slag bag mouth reference radius stored in the data record, perform numerical subtraction between the radial straight distance and the standard slag bag mouth reference radius, extract the wear deviation data corresponding to each independent edge point, and perform circumferential serialization splicing and array combination arrangement of multiple wear deviation data according to the edge distribution angle to generate a circumferential wear offset array. The system calls upon the radial linear distance array parameters covering all contour nodes obtained from previous calculations, and simultaneously accesses the internally preset database of fundamental physical constants to retrieve the pre-entered reference radius value of 1200 mm, representing the standard slag pot opening in its factory state unaffected by any erosion. For the radial linear distance corresponding to each independent edge point, its current value is compared and subtracted from the reference radius value of the standard slag pot opening one by one. By subtracting the standard reference radius from the actual measured radial linear distance, local wear deviation data reflecting the degree of physical erosion of the inner wall of the metal container under high-temperature molten conditions is extracted.

[0032] Table 2 Wear Deviation Array Data Table

[0033] As shown in Table 2, the wear degradation of each cross-section at various angles can be intuitively obtained through basic subtraction operations. Multiple wear deviation data corresponding to all nodes were collected and summarized. Using the spatial centroid coordinates as the origin and poles, and the edge distribution angles from 0 degrees to 360 degrees as the sorting index, these wear deviation data were sequentially stitched together in a clockwise direction. After stitching, an array combination arrangement at the spatial topology level was performed, ultimately generating a circumferential wear offset array containing complete inner wall concave-convex deformation features. This numerical array demonstrates the asymmetric wear state of the container wall, providing a core spatial compensation basis for the accurate correction of subsequent three-dimensional physical volume calculations.

[0034] Please see Figure 5 The specific steps for obtaining the real-time entity volume measurement are as follows: S401: Read the pixel matrix data belonging to the slag liquid surface attribute in the solid-liquid interface contour decoupling set, extract the actual liquid surface edge contour, obtain the preset reference bottom coordinates, perform depth correlation vertical span calculation on the reference bottom coordinates and the actual liquid surface edge contour, obtain the horizontal thin layer micro element, perform proportional conversion on the elevation position of the horizontal thin layer micro element, and obtain the height proportional multiplier. The system reads pixel data belonging to the slag surface attributes recorded in the solid-liquid interface contour decoupling set. Morphological closing operations are used to eliminate isolated noise and voids within this pixel data. Then, the actual liquid surface edge contour in a two-dimensional plane is extracted along the outer pixel lines. Using a preset camera extrinsic matrix and the slag bag's three-dimensional geometric model data, the system performs reverse spatial ray tracing on the two-dimensional pixel coordinates of the actual liquid surface edge contour. The spatial intersection point of the ray tracing with the inner wall surface of the slag bag's three-dimensional geometric model is calculated, and the Z-axis elevation of this intersection point is extracted as the Z-axis spatial height parameter corresponding to the actual liquid surface edge contour. The system accesses a three-dimensional spatial calibration file to obtain the preset reference bottom coordinates representing the lowest physical extreme point of the slag bag's internal cavity; the Z-axis elevation of this reference point is set to an absolute value of 0. The Z-axis spatial height parameter corresponding to the actual liquid surface edge contour is subtracted from the Z-axis coordinate value of the reference bottom to obtain the current cumulative total liquid level depth. Along the total depth direction, the entire vertical space is divided into equally spaced slices according to a preset fixed height step of 10 mm, obtaining a large number of discretely stacked horizontal thin-layer micro-elements. Each thin-layer micro-element represents a volume integral node of an independent dimension. For each extracted horizontal thin-layer micro-element, its relative height elevation value is extracted. This elevation value is used as the numerator, and the maximum design height constant of the overall physical cavity of the slag bag is used as the denominator. A division calculation is performed to obtain a height ratio multiplier that fluctuates between 0 and 1. This multiplier value maps the nonlinear diffusion physical law of the high-temperature wear effect continuously intensifying from bottom to top with the depth of the slag bag.

[0035] S402: Perform numerical multiplication of the height proportional multiplier and the circumferential wear offset array to generate a local offset array, obtain the preset standard inner radius at the corresponding depth, perform radial numerical addition and combination of the preset standard inner radius and the local offset array in the polar coordinate system, and perform rectangular coordinate system transformation to extract the cross-sectional node coordinate parameters. Based on the cross-sectional node coordinate parameters, connect them into a closed figure to construct the polygonal cross-sectional boundary. For the height proportional multiplier scalar obtained from the analysis of the current horizontal thin-layer micro-element, the complete circumferential wear offset array generated in the previous steps is retrieved. The height proportional multiplier is then multiplied by the wear deviation data in each angular direction of the offset array. By scaling, the wear characteristics at the bottom are weakened while the true wear magnitude at the top is preserved, thus generating a local offset array that strictly fits the current thin-layer depth. Simultaneously, the 3D geometric standard model database is queried to obtain the preset standard inner radius of the factory geometric parameters corresponding to the depth of the current thin-layer micro-element. In a virtual polar coordinate system with the centroid of the container as the pole, the extracted constant value of the preset standard inner radius is radially added to the floating deviation values ​​at each angle in the aforementioned local offset array to obtain the actual polar diameter length sequence after correcting for physical erosion. Subsequently, a transformation operation from the polar coordinate system to the Cartesian coordinate system is performed, that is, the polar diameter length value is multiplied by the cosine of the angle to obtain the X-axis coordinate, and multiplied by the sine of the angle to obtain the Y-axis coordinate, accurately extracting the coordinate parameters of the discrete section nodes distributed on the same horizontal elevation surface. Based on the coordinate parameters of these spatially correlated discrete cross-section nodes, they are connected sequentially by straight line segments in ascending order of circumferential angle to form a smoothly transitioning closed geometric mesh, thereby constructing a polygonal cross-sectional boundary that truly reflects the irregular cross-sectional shape of the inner wall of the slag bag at that depth.

[0036] S403: Call the polygonal section boundary, perform two-dimensional planar area calculation on the closed area enclosed by the polygonal section boundary, extract area features, obtain pre-set inter-layer slice thickness data, extract thickness features, perform numerical multiplication operation on area features and thickness features, obtain hierarchical micro-element volume, perform continuous three-dimensional spatial dimension accumulation operation on multiple hierarchical micro-element volumes arranged vertically, and generate real-time entity measurement volume. The system utilizes a finely constructed irregular polygonal cross-sectional boundary to perform high-precision two-dimensional planar area calculations on the closed region formed by the physical envelope of this polygonal cross-sectional boundary. An integral estimation method, which decomposes the polygonal mesh into multiple connected triangular elements, is employed to sum the coordinate points within the boundary, accurately extracting the area features of the depth section, with the unit set to square millimeters. The system retrieves the interlayer slice thickness data between two adjacent thin-layer micro-elements pre-set in the initialization parameter table, setting this fixed thickness constant to 10 millimeters, and extracts this value as a thickness feature parameter. A basic numerical multiplication operation is performed on the calculated dynamic area feature value and the fixed thickness feature value to obtain the physical capacity of the micro-entity represented by the specific slice, i.e., the volume of a single independent three-dimensional hierarchical micro-element. As the liquid level rises vertically, a large number of bottom-up discrete slice datasets are generated during the calculation. For all the multiple hierarchical micro-element volume data regularly arranged along the vertical elevation direction, a continuous numerical summation operation based on the Riemann summation principle in three-dimensional space is performed. By accumulating from the bottom reference 0 height point all the way to the highest level micro-element where the current actual liquid level is located, all two-dimensional slices are expanded and converged to finally generate real-time solid measurement volume parameters that include wear dynamic compensation characteristics.

[0037] Please see Figure 6 The specific steps for obtaining the dynamic liquid level detection results of the slag bag are as follows: S501: Perform time-series evolution feature analysis on the data fluctuation curve of the real-time entity's measured volume on a continuous time axis, perform continuous difference derivative operation on the time-series evolution feature, extract the volume increment change rate, obtain the volume change gradient, and convert the volume change gradient by mapping the preset decay weight coefficient to obtain the dynamic penalty multiplier. For the real-time entity volume measurement records continuously output by the aforementioned cumulative calculation, a fixed sampling period of 50 milliseconds is used to write them into a continuous time axis buffer array with a length of 3 seconds, forming a data fluctuation curve representing the dynamic trend of capacity increase, and performing time-series evolution feature analysis. To eliminate high-frequency data spikes caused by molten slag surface rolling, a 5-point mean filtering method is used to smooth the data fluctuation curve. For the time-series evolution features after smoothing, the volume data difference between two adjacent sampling periods is extracted. This difference is divided by the fixed time interval of 50 milliseconds, and a strict continuous difference derivative operation is performed to extract the rate of change of volume increment representing the intensity of liquid injection, i.e., the volume change gradient at that sampling moment. The larger the gradient value, the larger the amount of high-temperature molten slag injected per unit time. After obtaining the volume change gradient scalar, an inverse proportional mapping correlation configuration is set for the magnitude of the volume change gradient value. When the volume change gradient exceeds a preset safety experience threshold of 50 liters per second, a preset attenuation weight coefficient ranging from 0.5 to 0.9, decreasing with the gradient, is retrieved from the parameter table for conversion, resulting in a real-time dynamically changing penalty multiplier. The advantage of this parameter lies in its innovative introduction of transient flow acceleration parameters into a comprehensive rating rule system for future physical overflow risks, achieving dynamic self-balancing of the safety redundancy benchmark.

[0038] S502: Call the dynamic penalty multiplier and the real-time entity measurement volume, obtain the preset static safety benchmark in the data record, perform numerical multiplication on the dynamic penalty multiplier and the static safety benchmark, extract the warning boundary parameters of dynamic fluctuation, generate the real-time dynamic warning benchmark, perform difference comparison and analysis on the real-time entity measurement volume and the real-time dynamic warning benchmark, extract the spatial volume difference between the two, and obtain the critical state index. The system retrieves the dynamic penalty multiplier scalar obtained within the current sampling period and the real-time entity measurement volume parameters generated by 3D accumulation. It then accesses the refueling operation safety rule management library to obtain the static safety baseline capacity value preset in the data records, such as specifying a safe physical loading limit of 15,000 liters for the current type of slag bag. For the obtained dynamic penalty multiplier, it performs a conventional multiplication operation directly with the static safety baseline value.

[0039] Table 3 Comparison Table of Dynamic Early Warning Benchmark and Penalty Multiplier

[0040] As shown in Table 3, by employing a multiplicative decay mechanism and driven by data, the warning boundary parameters for downward-floating and tightening dynamic fluctuations can be extracted based on the actual flow velocity environment on site, generating a real-time dynamic warning benchmark that can adapt to the physical risks of the current refueling condition. Subsequently, the current value of the calculated real-time entity volume is extracted and compared with the extracted real-time dynamic warning benchmark value. The difference between the calculated dynamic warning benchmark value and the entity volume value is subtracted to extract the scalar of the residual spatial volume difference between the two at the current operation time. The absolute value of this volume difference accurately establishes the critical state index characterizing the urgency of the overflow risk, providing a numerical intervention red line distance for the operation of automatic shut-off control valves.

[0041] S503: Call the critical state index, extract and calculate the remaining loadable volume associated with the critical state index in the internal space of the slag bag, obtain the physical capacity margin, determine the safety trigger state based on the physical capacity margin associated with the critical state index, summarize the liquid level data records, and generate the dynamic liquid level detection result of the slag bag. The system invokes a critical state indicator containing specific volume difference records. It then extracts and calculates the remaining loadable volume actually associated with this critical state indicator in the three-dimensional physical space inside the slag bag. Specifically, it reads the theoretical total volume parameter of the container and subtracts the currently corrected physical volume value to obtain a physical load margin scalar data with physical quantifiable significance, such as 350 liters. Based on the specific physical load margin value recorded in this critical state indicator, a step-distribution threshold classification logic is executed to determine the final safety trigger state. The specific judgment thresholds are set as follows: when the physical load margin value is greater than or equal to 1000 liters, the safety trigger state is determined and assigned to the normal loading category; when the physical load margin value is less than 1000 liters but greater than or equal to 300 liters, the safety trigger state is determined to be a critical risk level category, and a control code command requiring a reduction in the injection flow rate is output to the communication bus; when the physical load margin value is strictly less than the extreme value of 300 liters, the safety trigger state is determined to be a highly dangerous red zone category, and a trigger command to directly shut off the physical injection valve is immediately generated. After completing the stage status assessment step, the acquired multi-dimensional data, such as the trigger status identifier number, precise timestamp, and physical capacity margin value, are packaged into parameters. All underlying time-series liquid level data records are summarized and synchronously written to the log database of the backend server. Finally, the dynamic liquid level detection results of the slag bag that can be read by the visualization terminal are output.

[0042] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting the liquid level in a slag package based on AI image recognition of the slag package outline, characterized in that, Includes the following steps: S1: Filter the pixels in the overlapping areas of the image, perform vector synthesis calculation on the horizontal and vertical gray level differences to obtain the gradient magnitude and gradient direction angle, and generate the gradient histogram of the overlapping areas based on the cumulative distribution characteristics. S2: Calculate the local maximum diffusion trend of the gradient histogram of the overlapping region to obtain the attenuation ratio of the main peak and the attenuation ratio of the secondary peak, compare the attenuation deviation of the two peaks, determine the pixel belonging and perform isolation division, and construct the solid-liquid interface contour decoupling set. S3: Calculate the top edge coordinate set and spatial centroid position coordinate of the actual slag bag mouth edge contour in the solid-liquid interface contour decoupling set, calculate the difference between the radial straight distance between edge points and the standard slag bag mouth reference radius, and generate a circumferential wear offset array. S4: Calculate the vertical span of the actual liquid surface edge contour in the solid-liquid interface contour decoupling set to obtain the horizontal thin layer micro-element, multiply the corresponding height ratio multiplier with the circumferential wear offset array to generate a local offset array, superimpose the standard inner radius to construct the polygonal cross-sectional boundary, calculate the product of area and thickness to obtain the volume of the layer micro-element and perform spatial accumulation to generate the real-time solid measurement volume. S5: Calculate the volume change gradient by differentiating the real-time entity's measured volume, calculate the product of the dynamic penalty multiplier and the static safety benchmark as the real-time dynamic early warning benchmark, compare the two to obtain the critical state index, determine the trigger state based on the physical capacity margin, and generate the dynamic liquid level detection result of the slag bag.

2. The slag bag level detection method based on AI image recognition of slag bag contour as described in claim 1, characterized in that, The gradient histogram of the overlapping region specifically includes the horizontal axis interval, the vertical axis frequency, and the peak coordinate. The decoupling set of the solid-liquid interface contour includes the slag pot opening pixel group, the liquid surface pixel group, and the overlapping boundary line. The circumferential wear offset array specifically refers to the offset angle, the wear depth scalar, and the radial deviation sequence. The real-time entity measured volume specifically includes the bottom reference capacity, the single-layer additional volume, and the total volume scalar. The dynamic liquid level detection result of the slag pot includes the safety warning sign, the remaining capacity space, and the molten slag liquid level height.

3. The slag bag level detection method based on AI image recognition of slag bag contour as described in claim 1, characterized in that, The specific steps for obtaining the gradient histogram of the overlapping region are as follows: S101: Collect grayscale image arrays of slag bags at the high-temperature molten slag filling site, perform image morphology analysis and contour boundary stripping on the grayscale image array of slag bags, extract pixels in overlapping areas, perform grayscale subtraction operation on adjacent pixels in overlapping areas along the horizontal and vertical pixel array arrangement directions, extract horizontal and vertical change features, and obtain information sets of horizontal and vertical grayscale differences. S102: Call the information set of horizontal and vertical grayscale differences, construct a two-dimensional spatial orthogonal vector mapping system for the horizontal and vertical grayscale differences, perform square root operation on the two-dimensional spatial orthogonal vectors, extract the gradient magnitude, calculate the ratio of vertical grayscale difference to horizontal grayscale difference, and perform arctangent trigonometric function operation to analyze the direction attribute and obtain the gradient direction angle. S103: Divide the gradient direction angle into multiple uniformly distributed statistical intervals, map the gradient magnitude corresponding to the gradient direction angle to the assigned statistical interval, perform numerical summation clustering operation on all gradient magnitudes contained in each statistical interval, extract the cumulative distribution features, and construct a two-dimensional distribution mapping system with the direction interval as the horizontal axis and the cumulative magnitude as the vertical axis, and generate a gradient histogram of the overlapping region.

4. The slag bag level detection method based on AI image recognition of slag bag contour as described in claim 3, characterized in that, The process of dividing the statistical intervals into multiple uniformly distributed intervals is as follows: The values ​​of all angle elements contained in the gradient direction angle are traversed and compared to obtain the minimum and maximum global angle span. The feature extraction resolution parameter in the data record is called, and a standard angle step scalar is generated based on the feature extraction resolution parameter. The difference between the maximum value of the global angle span and the minimum value of the global angle span is calculated to extract the absolute angle span parameter. A division operation is performed between the absolute angle span parameter and the standard angle step scalar, and the floating-point result generated by the division operation is rounded up to obtain the total number of statistical intervals. The minimum global angle span is used as the starting boundary node. The starting boundary node is continuously offset by the standard angle step size scalar and the total number of statistical intervals to generate a sequence of angle boundary nodes that are connected end to end. The multiple uniformly distributed statistical intervals are then segmented and output based on the angle boundary node sequence.

5. The slag bag level detection method based on AI image recognition of slag bag contour as described in claim 3, characterized in that, The specific steps for obtaining the solid-liquid interface contour decoupling set are as follows: S201: Perform local maximum diffusion trend extraction operation on the gradient histogram of the overlapping region, identify the first extreme value and the second extreme value in the histogram distribution interval, calculate the attenuation rate for the adjacent statistical area crossing feature corresponding to the first extreme value, obtain the main peak attenuation ratio, calculate the attenuation rate for the adjacent statistical area crossing feature corresponding to the second extreme value, obtain the secondary peak attenuation ratio, summarize and output the bimodal attenuation dataset. S202: Based on the bimodal attenuation dataset, perform a difference comparison operation on the attenuation ratio of the main peak and the attenuation ratio of the secondary peak, extract the bimodal deviation amplitude data, generate the bimodal attenuation deviation and construct an adaptive segmentation threshold for pixel gradient, call the pixels in the overlapping area, compare and analyze the gradient amplitude of each independent pixel with the adaptive segmentation threshold, determine the belonging status of the pixels in the overlapping area based on the comparison results, and obtain the category determination features. S203: Based on the boundary attribution labels of the category determination feature record, perform category isolation processing on the pixels in the overlapping area, divide the pixel array belonging to the molten slag surface attribute and the pixel array belonging to the solid furnace mouth attribute, and perform spatial array combination on the pixel coordinate data after stripping and dividing to establish a solid-liquid boundary contour decoupling set.

6. The slag bag level detection method based on AI image recognition of slag bag contour as described in claim 5, characterized in that, The specific steps for obtaining the circumferential wear bias array are as follows: S301: Read the pixel matrix data of the solid furnace mouth attribute recorded in the solid-liquid interface contour decoupling set, extract the actual slag bag mouth edge contour, obtain the preset perspective transformation matrix, perform three-dimensional space coordinate system transformation and projection mapping operation on the actual slag bag mouth edge contour and perspective transformation matrix, extract three-dimensional mapping features, and generate the top edge coordinate set. S302: Call the top-level edge coordinate set, perform an average calculation on the horizontal and vertical coordinates of multiple coordinate nodes contained in the top-level edge coordinate set, extract the geometric mean feature, obtain the spatial centroid position coordinates, and calculate the length of the line connecting each independent edge point and the spatial centroid position coordinate in the top-level edge coordinate set to obtain the radial straight-line distance. S303: Call the radial straight distance to obtain the pre-stored standard slag bag mouth reference radius in the data record, perform numerical subtraction between the radial straight distance and the standard slag bag mouth reference radius, extract the wear deviation data corresponding to each independent edge point, and perform circumferential serialization splicing and array combination arrangement of multiple wear deviation data according to the edge distribution angle to generate a circumferential wear offset array.

7. The slag bag level detection method based on AI image recognition of slag bag contour as described in claim 6, characterized in that, The process of obtaining the preset perspective transformation matrix is ​​as follows: The system calls upon the three-dimensional spatial reference coordinate data of multiple fixed physical calibration target points obtained through actual surveying and mapping, and performs spatial array combination processing on the multiple three-dimensional spatial reference coordinate data to establish a standard three-dimensional coordinate set. Collect a field panoramic calibration grayscale image containing the multiple fixed physical calibration target points, perform corner detection and edge feature extraction operations on the field panoramic calibration grayscale image, parse the two-dimensional pixel coordinate data mapped by the multiple fixed physical calibration target points inside the image, and establish a reference two-dimensional coordinate set; The standard three-dimensional coordinate set and the reference two-dimensional coordinate set are aligned with feature points to construct a sequence of matching two-dimensional and three-dimensional point pairs. Perspective projection mapping is performed on the sequence of matching two-dimensional and three-dimensional point pairs to extract spatial dimension transformation parameters and construct an associated numerical array structure to generate the perspective transformation matrix.

8. The method for detecting liquid level in a slag bag based on AI image recognition of the slag bag contour according to claim 6, characterized in that, The specific steps for obtaining the real-time entity volume measurement are as follows: S401: Read the pixel matrix data belonging to the slag liquid surface attribute in the solid-liquid interface contour decoupling set, extract the actual liquid surface edge contour, obtain the preset reference bottom coordinates, perform vertical span calculation for the depth correlation between the reference bottom coordinates and the actual liquid surface edge contour, obtain the horizontal thin layer micro element, perform proportional conversion for the elevation position of the horizontal thin layer micro element, and obtain the height proportional multiplier. S402: Perform a numerical multiplication operation between the height ratio multiplier and the circumferential wear offset array to generate a local offset array, obtain the preset standard inner radius corresponding to the depth, perform radial numerical addition and combination of the preset standard inner radius and the local offset array in the polar coordinate system, and perform a rectangular coordinate system transformation operation to extract the cross-sectional node coordinate parameters, connect them into a closed figure based on the cross-sectional node coordinate parameters, and construct the polygonal cross-sectional boundary. S403: Call the polygonal cross-section boundary, perform two-dimensional planar area calculation on the closed area enveloped by the polygonal cross-section boundary, extract area features, obtain pre-set inter-layer slice thickness data, extract thickness features, perform numerical multiplication operation on the area features and thickness features, obtain the layer micro-element volume, perform continuous three-dimensional spatial dimension accumulation operation on multiple layer micro-element volumes arranged vertically, and generate real-time entity measurement volume.

9. The method for detecting the liquid level in a slag bag based on AI image recognition of the slag bag contour as described in claim 8, characterized in that, The process of obtaining the preset baseline bottom coordinates is as follows: The three-dimensional geometric model data of the slag bag is called from the data record, and the spatial morphology is analyzed based on the three-dimensional geometric model data of the slag bag to peel off the bottom curved surface model of the internal cavity structure. The lowest spatial position extraction operation is performed on the bottom curved surface model to obtain the coordinates of the global lowest spatial nodes; Set the coordinates of the global lowest spatial node as the height reference zero point and perform three-dimensional spatial anchor point association calibration to generate the preset reference bottom coordinates; The process of obtaining the preset standard inner radius corresponding to the depth is as follows: The three-dimensional geometric model data of the slag bag is called from the data record. A vertical scale axis is established based on the center vertical line of the three-dimensional geometric model data of the slag bag. The internal cavity structure of the three-dimensional geometric model data of the slag bag is horizontally segmented along the vertical scale axis to extract the horizontal cross-sectional contour boundary corresponding to multiple discrete elevations. Geometric fitting calculations are performed on the horizontal cross-sectional contour boundary to obtain multiple depth elevation parameters and horizontal inner radius parameters. Associative topology construction is performed on the multiple depth elevation parameters and the horizontal inner radius parameters to establish a numerical lookup table for depth and inner radius dimensions. The depth span parameter corresponding to the actual liquid surface edge contour is called, and interpolation matching operation is performed on the depth span parameter in the numerical lookup table to extract the target size features and generate the preset standard inner radius of the corresponding depth.

10. The method for detecting the liquid level in a slag bag based on AI image recognition of the slag bag contour as described in claim 8, characterized in that, The specific steps for obtaining the dynamic liquid level detection results of the slag bag are as follows: S501: Perform time-series evolution feature analysis on the data fluctuation curve of the real-time entity's measured volume on a continuous time axis, perform continuous difference derivative operation on the time-series evolution features, extract the volume increment change rate, obtain the volume change gradient, and convert the volume change gradient by mapping a preset attenuation weight coefficient to obtain the dynamic penalty multiplier. S502: Call the dynamic penalty multiplier and the real-time entity measured volume to obtain the preset static safety benchmark in the data record, perform a numerical multiplication operation on the dynamic penalty multiplier and the static safety benchmark, extract the warning boundary parameters of dynamic fluctuations, generate a real-time dynamic warning benchmark, perform a difference comparison analysis on the real-time entity measured volume and the real-time dynamic warning benchmark, extract the spatial volume difference between the two, and obtain the critical state index. S503: Call the critical state index, extract and calculate the remaining loadable volume associated with the critical state index in the internal space of the slag bag, obtain the physical capacity margin, determine the safety trigger state based on the physical capacity margin associated with the critical state index, summarize the liquid level data records, and generate the dynamic liquid level detection result of the slag bag.

Images