Online intelligent detection method and system based on low magnification detection technology

Abstract

The application relates to the technical field of industrial vision intelligence, and discloses an online intelligent detection method and system based on low-magnification detection technology, which comprises the following steps: acquiring a continuous casting billet macro image and extracting a boundary contour; determining a dendrite growth direction vector and constructing a global vector matrix; extracting an asymmetric shrinkage feature of a corner region; quantifying a diamond segregation degree to obtain a segregation severity index; converting a concentration distribution scalar field according to the segregation severity index, depicting a solidification state, and obtaining an overall solidification front distribution; matching and judging a front advancing speed and a heat flow distribution, adjusting a vector distribution field, and obtaining an optimized intersection point; extracting interference data according to the optimized intersection point, quantitatively evaluating diamond segregation according to the interference data, and obtaining a final diamond segregation quantitative evaluation result. The method can solve the problems of the continuous casting billet quality caused by uneven dendrite growth boundaries, diamond segregation degrees and solidification front distributions.

Description

Technical Field

[0001] This invention relates to the field of industrial vision intelligence technology, and in particular to an online intelligent detection method and system based on low-magnification detection technology. Background Technology

[0002] Currently, continuous casting billet production is a crucial step in achieving efficient and high-quality billet production in the steel industry. The internal quality of the billet directly determines the performance and yield of subsequent rolled products. For the production of high-quality special steel, quantitative assessment of rhomboid segregation defects during the solidification process of continuously cast billets through visual inspection is a key step in ensuring stable and controllable internal quality of the billets.

[0003] In existing technologies, most detection methods rely on offline sampling analysis or low-resolution imaging.

[0004] However, in actual production, these low-resolution imaging methods are difficult to accurately extract the spatial distribution and convergence intensity of dendrite vectors in industrial noise images due to the variable dendrite growth direction, dynamic evolution of the solidification front, and the influence of propulsion speed and heat flow distribution. This results in the inability to accurately depict macroscopic texture changes, significant deviations in judging the position and severity of rhomboid distortion, and the inability to guarantee stable and controllable internal quality of the billet.

[0005] In summary, existing technologies are unable to accurately characterize the directionality of dendrite growth and its convergence features, resulting in the inability to precisely locate solidification junctions and quantify rhombic segregation defects. Summary of the Invention

[0006] This invention provides an online intelligent detection method and system based on low-magnification detection technology to solve the billet quality problems caused by dendrite growth boundaries, rhombic segregation degree and uneven distribution of solidification front.

[0007] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an online intelligent detection method based on low-magnification detection technology, comprising:

[0008] A macroscopic image of a continuously cast square billet is acquired, and dendrite growth boundaries are extracted based on the macroscopic image of the continuously cast square billet to obtain the boundary contour;

[0009] Extract the discrete coordinate points of the boundary contour and calculate the dendrite growth direction vector. Generate a set of directed vectors based on the dendrite growth direction vector and the discrete coordinate points, and perform spatial mapping to obtain the vector distribution field.

[0010] Effective dendrite junction clusters are selected based on the vector distribution field, potential junction points are calculated based on the effective dendrite junction clusters, a set of junction point coordinates is obtained, and a morphological feature matrix is ​​constructed based on the set of junction point coordinates. The convergence intensity is calculated based on the morphological feature matrix.

[0011] If the convergence intensity exceeds a preset intensity threshold, a rhombus segregation candidate region is extracted to obtain a rhombus segregation candidate region image. A gray-level co-occurrence matrix is ​​constructed based on the rhombus segregation candidate region image. The feature values ​​of the gray-level co-occurrence matrix are extracted and mapped to obtain a segregation severity index.

[0012] According to the preset physical quantity mapping rules, the concentration distribution scalar field is transformed for the segregation severity index to obtain the concentration distribution scalar field, and the solidification state is characterized according to the concentration distribution scalar field to obtain the overall solidification front distribution.

[0013] Based on the overall solidification front distribution, a matching judgment is made between the front advancement velocity and the heat flux distribution. If the front advancement velocity and the heat flux distribution do not match, the vector distribution field is adjusted to obtain an optimized intersection point.

[0014] Interference data is extracted based on the optimized intersection point, and a quantitative assessment of rhombus transformation is performed based on the interference data to obtain the final quantitative assessment result of rhombus transformation.

[0015] Secondly, the present invention provides an online intelligent detection system based on low-magnification detection technology, comprising:

[0016] The boundary data acquisition module is used to acquire a macroscopic image of the continuously cast billet and extract the dendrite growth boundary based on the macroscopic image of the continuously cast billet to obtain the boundary contour.

[0017] The vector distribution construction module is used to extract the discrete coordinate points of the boundary contour and calculate the dendrite growth direction vector. Based on the dendrite growth direction vector and the discrete coordinate points, a set of directed vectors is generated and spatially mapped to obtain the vector distribution field.

[0018] The convergence intensity extraction module is used to screen out effective dendritic convergence clusters based on the vector distribution field, calculate potential convergence points based on the effective dendritic convergence clusters, obtain a set of convergence point coordinates, construct a morphological feature matrix for the corresponding position based on the set of convergence point coordinates, and calculate the convergence intensity based on the morphological feature matrix.

[0019] The segregation degree quantification module is used to extract a diamond-shaped segregation candidate region if the convergence intensity exceeds a preset intensity threshold, obtain a diamond-shaped segregation candidate region image, construct a gray-level co-occurrence matrix based on the diamond-shaped segregation candidate region image, extract the feature values ​​of the gray-level co-occurrence matrix, and perform mapping calculation to obtain a segregation severity index.

[0020] The front distribution reconstruction module is used to perform a concentration distribution scalar field transformation on the segregation severity index according to a preset physical quantity mapping rule to obtain a concentration distribution scalar field, and to characterize the solidification state based on the concentration distribution scalar field to obtain the overall solidification front distribution.

[0021] The leading edge matching optimization module is used to determine the matching between the leading edge advance velocity and the heat flux distribution based on the overall solidification leading edge distribution. If the leading edge advance velocity and the heat flux distribution do not match, the vector distribution field is adjusted to obtain an optimized intersection point.

[0022] The rhombus transformation quantitative assessment module is used to extract interference data based on the optimized intersection point, and to perform rhombus transformation quantitative assessment based on the interference data to obtain the final rhombus transformation quantitative assessment result.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] (1) This invention obtains macroscopic images of continuously cast billets, and sequentially performs dendrite growth boundary extraction, vector distribution field construction, intersection point positioning and convergence intensity calculation, and finally completes the quantitative evaluation of rhombic transformation; this series of operations constitutes a complete analysis chain from image feature extraction to physical defect quantification, which can systematically diagnose the internal quality of the billet; thus solving the comprehensive problem that the billet quality is difficult to control accurately online due to unclear dendrite growth boundaries and uneven distribution of solidification front.

[0025] (2) This invention constructs a vector distribution field that characterizes the macroscopic growth trend by extracting discrete coordinate points of the boundary contour and calculating dendrite growth direction vector, and selects effective dendrite intersection clusters based on its local divergence to locate potential intersection points; this operation realizes an intuitive and quantitative characterization of dendrite growth direction and its spatial convergence characteristics; thus directly solving the problem that the existing technology is difficult to accurately extract its vector distribution and convergence characteristics due to the variable dendrite growth direction.

[0026] (3) The present invention extracts candidate regions by setting an intensity threshold and extracts features using the gray-level co-occurrence matrix to quantify the severity index of segregation, and then maps it into a concentration distribution scalar field according to physical rules. This operation transforms visual texture features into quantifiable and traceable physical field parameters, thereby realizing an objective assessment of segregation defects. This solves the problem that the existing technology has difficulty in quantifying and assessing rhomboid segregation defects, and realizes an accurate determination of their severity. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of an online intelligent detection method based on low-magnification detection technology provided in the first embodiment of the present invention;

[0028] Figure 2This is a schematic diagram of an online intelligent detection system based on low-magnification detection technology provided in the second embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Reference Figure 1 The first embodiment of the present invention provides an online intelligent detection method based on low-magnification detection technology, comprising the following steps:

[0031] S11, Obtain a macroscopic image of the continuously cast square billet, and extract the dendrite growth boundary based on the macroscopic image of the continuously cast square billet to obtain the boundary contour;

[0032] S12, extract the discrete coordinate points of the boundary contour and calculate the dendrite growth direction vector. Generate a set of directed vectors based on the dendrite growth direction vector and the discrete coordinate points and perform spatial mapping to obtain the vector distribution field.

[0033] S13, based on the vector distribution field, select effective dendrite junction clusters, calculate potential junction points based on the effective dendrite junction clusters, obtain a set of junction point coordinates, and construct a morphological feature matrix for the corresponding position based on the set of junction point coordinates, and calculate the convergence intensity based on the morphological feature matrix.

[0034] S14, if the convergence intensity exceeds a preset intensity threshold, a rhombus segregation candidate region is cropped to obtain a rhombus segregation candidate region image, and a gray-level co-occurrence matrix is ​​constructed based on the rhombus segregation candidate region image. The feature values ​​of the gray-level co-occurrence matrix are extracted and mapped to obtain a segregation severity index.

[0035] S15. According to the preset physical quantity mapping rules, the concentration distribution scalar field is transformed for the segregation severity index to obtain the concentration distribution scalar field, and the solidification state is characterized according to the concentration distribution scalar field to obtain the overall solidification front distribution.

[0036] S16. Based on the overall solidification front distribution, a matching judgment is made between the front advancement velocity and the heat flow distribution. If the front advancement velocity and the heat flow distribution do not match, the vector distribution field is adjusted to obtain an optimized intersection point.

[0037] S17. Extract interference data based on the optimized intersection point, and perform a quantitative assessment of the rhombus transformation based on the interference data to obtain the final quantitative assessment result of the rhombus transformation.

[0038] In step S11, it is necessary to obtain a macroscopic image of the continuously cast billet and extract the dendrite growth boundary based on the macroscopic image of the continuously cast billet to obtain the boundary contour.

[0039] It should be noted that this step utilizes high-resolution imaging and multi-stage image processing algorithms to accurately extract dendrite growth boundaries from the macroscopic image of the continuously cast billet, constructing a continuous and complete boundary contour. The acquisition of the macroscopic image of the continuously cast billet is achieved using high-resolution imaging equipment, such as an industrial camera with a resolution of 20 megapixels. In this embodiment, multi-angle light source illumination technology is also incorporated, projecting LED light sources of different intensities from three directions—top, 45-degree side angle, and low angle—to enhance the contrast of the acquired macroscopic image of the continuously cast billet.

[0040] In one implementation, the step of extracting dendrite growth boundaries based on the macroscopic image of the continuously cast billet to obtain boundary contours includes:

[0041] The macroscopic image of the continuously cast billet is converted to grayscale to obtain grayscale image data;

[0042] Edge detection is performed on the grayscale image data to obtain boundary edge data;

[0043] Boundary continuity analysis is performed on the boundary edge data. If boundary breaks exist, boundary break repair is performed to obtain the boundary contour of the dendrite growth boundary.

[0044] It should be noted that the macroscopic image of the continuously cast billet is a color image captured by an industrial camera, while grayscale processing is the process of combining the RGB three-channel information of the color image into a single-channel grayscale value according to specific weights. The grayscale value ranges from 0 to 255, with a higher value indicating higher pixel brightness. Grayscale image data is the result of grayscale processing. For example, grayscale processing can be performed using a weighted average method, specifically by linearly combining the RGB three components of the pixels in the macroscopic image of the continuously cast billet with weight coefficients of 0.30, 0.59, and 0.11 to obtain grayscale image data. This weight coefficient configuration is optimized based on human visual sensitivity, maximizing the preservation of the detailed contrast of the dendritic structure.

[0045] In this embodiment, edge detection refers to the image processing process of identifying pixels as edge points by calculating the grayscale gradient of the image using a differential operator. This is used to obtain the dendrite growth boundary formed by the solid-liquid interface during solidification. When performing edge detection on the grayscale image data, the Sobel operator algorithm based on gradient calculation is used. Specifically, Sobel edge detection is performed on a 1024×1024 pixel grayscale image. By calculating the gradient values ​​in the horizontal and vertical directions within a 3×3 neighborhood, regions with gradient values ​​greater than a threshold are selected as boundary locations. Furthermore, Gaussian filtering is used to remove noise, generating clear and reliable boundary edge data. The gradient threshold is adaptively set based on the statistical characteristics of the gradient amplitude of the grayscale image and dynamically adjusted in conjunction with the signal-to-noise ratio and dendrite contrast of the macroscopic image of the continuously cast billet, ensuring effective differentiation between true boundary gradient peaks and noise pseudo-edges.

[0046] Specifically, boundary continuity refers to the characteristic that edge pixels maintain spatial adjacency; boundary breakpoints refer to interruptions in the edge chain with a length greater than 1 pixel. When performing boundary continuity analysis, an eight-neighbor contour tracing algorithm is used to determine connectivity. A chain search is performed starting from the initial edge point, recording the sequence of continuous edge points. Isolated point sets that do not form closed loops are identified as boundary breakpoints, and these breakpoints are repaired.

[0047] Furthermore, boundary breakpoint repair refers to the process of generating virtual edge pixels between breakpoints using mathematical interpolation methods to restore the geometric continuity of the boundary; the boundary contour refers to the complete dendritic boundary geometric pixel data containing a continuous coordinate sequence. In one embodiment, a linear interpolation method is used for boundary breakpoint repair. Specifically, by calculating the Euclidean distance and orientation angle between the two ends of the breakpoint, intermediate interpolation points are generated at 1-pixel intervals. The generated new pixel coordinate sequence is then merged into the original boundary edge data to form a closed and continuous boundary contour.

[0048] In step S12, discrete coordinate points of the boundary contour are extracted and dendrite growth direction vectors are calculated. A directed vector set is generated based on the dendrite growth direction vectors and the discrete coordinate points, and spatial mapping is performed to obtain a vector distribution field, including:

[0049] Traverse the coordinate sequence of the boundary contour, extract discrete coordinates, and obtain discrete coordinate points;

[0050] The normal direction is calculated based on the discrete coordinate points to obtain the dendrite growth direction vector extending from the boundary to the liquid core side;

[0051] By combining the discrete coordinate points and the dendrite growth direction vector, a set of directed vectors is generated, and the set of directed vectors is spatially mapped to obtain the vector distribution field characterizing the macroscopic dendrite growth trend.

[0052] It should be noted that this step is based on the boundary profile and constructs a global vector distribution field that characterizes the growth trend of dendrites from the solidification front to the liquid core side through discrete point sampling, normal direction calculation and spatial mapping interpolation.

[0053] Here, the coordinate sequence refers to the array of point coordinates arranged in pixel connection order of the boundary contour; discrete coordinate points refer to the set of isolated coordinate points extracted from the continuous boundary contour curve at fixed intervals. When extracting discrete coordinates, a fixed-interval sampling strategy is adopted. Specifically, by traversing the boundary contour, a sampling point is extracted every 3 to 5 coordinate points to form a discrete coordinate point set. For example, for a boundary contour containing 800 pixels, setting the sampling interval to 4 will extract 200 discrete coordinate points. This sampling density reflects the dendrite arm bending morphology while reducing the computational load to 25% of the original data.

[0054] In this embodiment, the normal direction refers to the unit vector direction perpendicular to the tangent direction and pointing towards the liquid core; the dendrite growth direction vector refers to the unit vector direction pointing from the solidification front towards the unsolidified liquid core region, representing the dendrite growth orientation, and is determined by the normal direction and the inner and outer sides of the billet cross section. Specifically, when calculating the normal direction, the three-point estimation method is used to calculate the local tangent slope. Specifically, the current discrete point and its adjacent points are taken, and the tangent slope k is obtained through the difference method. The normal slope is -1 / k. When the tangent slope k is zero, the normal direction is directly defined as the perpendicular direction. For example, at the coordinate point (620, 480), its predecessor point is (616, 478) and its successor point is (624, 483). The calculated tangent slope is k = (483-478) / (624-616) = 5 / 8 = 0.625, and the normal slope is -1 / 0.625 = -1.6. Considering the geometric position of this point on the right side of the billet, the dendrite growth direction vector is determined to be a unit vector pointing to the center of the cross section (-0.85, 0.53).

[0055] Furthermore, a directed vector set refers to a paired dataset consisting of each discrete coordinate point and its corresponding dendrite growth direction vector. That is, each discrete coordinate point is assigned a vector with a direction. For example, in the region slightly to the right of the center of the billet section, the coordinates of a certain point are (620, 480), and the angle of its growth direction vector is -45 degrees. Then, the directed vector can be represented as an arrow extending 0.8 units to the lower left from that point. This vector set completely records the local growth orientation of dendrites at various points on the boundary.

[0056] Specifically, the vector distribution field refers to a two-dimensional vector matrix in which each grid node is assigned a direction and a magnitude. The process of spatially mapping the set of directed vectors involves dividing the physical dimensions of the billet cross-section into a grid of the continuously cast billet cross-section, mapping the discrete dendrite growth direction vectors to the grid of the continuously cast billet cross-section according to their spatial positions, and generating a continuous vector distribution field covering the entire cross-section by performing inverse distance weighted interpolation on the grid nodes, thereby realizing the spatial visualization and matrix representation of the macroscopic dendrite growth trend.

[0057] For example, the physical dimensions of the billet cross-section are divided into 200×200=40000 grid cells. For the center point of each grid, the inverse distance weighted interpolation method is used to calculate the weighted average of the 5 to 8 nearest directed vectors. At the center point coordinates (580, 510), the 6 nearest directed vectors are found with distances of 3.2, 4.5, 5.1, 6.8, 7.3, and 8.0 pixels, respectively, with corresponding weights of 0.28, 0.14, 0.11, 0.06, 0.05, and 0.04. After weighted averaging, the dendrite growth direction vector of this grid is (-0.81, -0.59), and the modulus is normalized to 1. Finally, the vector distribution field shows a clear centripetal convergence pattern in the central region of the billet.

[0058] In step S13, effective dendrite junction clusters are selected based on the vector distribution field, potential junction points are calculated based on the effective dendrite junction clusters, a set of junction point coordinates is obtained, and a morphological feature matrix is ​​constructed based on the set of junction point coordinates. The convergence intensity is calculated based on the morphological feature matrix.

[0059] It should be noted that this step identifies high-density aggregation regions through local divergence calculation, and accurately locates the coordinates of dendrite growth intersection points through spatial connectivity analysis and density clustering; then, it extracts the morphological features of the corner regions around the intersection points, and quantifies the degree of asymmetric contraction through multi-directional differential operations and sector dispersion calculation to obtain the aggregation intensity.

[0060] In one implementation, effective dendrite junction clusters are selected based on the vector distribution field, and potential junction points are calculated based on the effective dendrite junction clusters to obtain a set of junction point coordinates, including:

[0061] Calculate the local divergence of the vector distribution field. If the local divergence is greater than a preset divergence threshold, mark it as a high-density aggregation candidate point to obtain a high-density aggregation candidate point set.

[0062] Based on the set of high-density aggregated candidate points, spatial connectivity is constructed to obtain connected sub-regions, and density feature analysis is performed on the connected sub-regions to obtain highly cohesive effective dendritic junction clusters;

[0063] Calculate the spatial geometric center of the effective dendrite cluster to obtain the set of intersection point coordinates that characterize the macroscopic dendrite growth intersection position.

[0064] It is worth noting that local divergence refers to the sum of the spatial rates of change of vector components in all directions within a small region of the vector field. A divergence greater than 0 indicates that the vectors are converging. High-density aggregation candidate points refer to grid points where the centripetal aggregation of dendrite growth direction vectors exceeds a preset divergence threshold. The preset divergence threshold is set based on the statistical distribution characteristics of the vector distribution field and the physical laws of dendrite growth and convergence to ensure that the threshold can effectively distinguish between normal dendrite growth regions and abnormal convergence regions; here, it is set to 0.8. When calculating the local divergence of the vector distribution field, the discrete divergence calculation formula is used. Specifically, for each grid point, its four neighboring vector components (up, down, left, and right) are taken, and the sum of the partial derivatives of the horizontal component with respect to x and the vertical component with respect to y is calculated. For example, at the grid point (580, 510) in the central region of the vector distribution field, the horizontal components of the four neighboring vectors are -0.81 and -0.85, and the vertical components are -0.59 and -0.62, respectively, and the calculated divergence value is -0.07, which does not exceed the threshold. However, at the point (582, 512), the divergence value is calculated to be 0.92, which is greater than 0.8, and this point is included in the high-density aggregation candidate point set.

[0065] In this embodiment, a connected sub-region refers to a continuous set of points formed by spatial adjacency between candidate points. When constructing spatial connectivity for a high-density aggregated candidate point set, a neighborhood search method is used to merge two points with a distance less than the connectivity threshold into the same connected sub-region. The connectivity threshold directly affects the accuracy of spatial connectivity analysis and the identification precision of effective dendrite junction clusters. It is adaptively set based on the spatial scale characteristics of dendrite growth during the solidification process of continuously cast billets and the image acquisition resolution. This ensures that the threshold can accurately capture the continuous convergence pattern of dendrite arms while avoiding misjudging different junction areas as the same region due to an excessively large threshold, or causing over-segmentation of a single junction area due to an excessively small threshold. Here, it is set to 5 grid units.

[0066] In one example, an effective dendrite cluster refers to a high-density vector aggregation region with highly convergent dendrite growth directions and physical authenticity. Density feature analysis of connected sub-regions requires identifying effective dendrite clusters with high cohesion based on the density distribution characteristics of points. This is achieved by setting a density threshold of at least 10 points per unit area and a minimum cluster size of 20 points, ultimately filtering effective dendrite clusters from multiple connected sub-regions. The density threshold setting requires combining the metallographic image of the actual solidification structure of the cast billet, statistically analyzing the projected density of the dendrite junction region in the macroscopic image, and calculating the minimum guaranteed density value using a vector sampling interval of 4 pixels.

[0067] Furthermore, the spatial geometric center refers to the location point formed by the average of the x-coordinates and y-coordinates of all points in the planar point set; the intersection point coordinate set refers to the set of coordinate pairs of the geometric centers of all effective dendritic intersection clusters. When calculating the spatial geometric center, the arithmetic mean method is used: summing the x-coordinates of all points within the cluster and dividing by the number of points, and summing the y-coordinates and dividing by the number of points. For example, an effective dendritic intersection cluster contains 30 candidate points, with a total x-coordinate of 17460 and a total y-coordinate of 15300. The calculated geometric center x-coordinate is 17460 / 30 = 582, and the y-coordinate is 15300 / 30 = 510. This point (582, 510) is determined as one of the final intersection points, and the complete intersection point coordinate set includes three points: (582, 510), (618, 505), and (600, 490).

[0068] It should be noted that the preset divergence threshold, connectivity threshold, and density threshold are all pre-calibrated based on offline statistical analysis of a large number of historical qualified billet samples. Specifically, low-magnification images of continuously cast square billets under normal operating conditions of no less than 50 heats and their corresponding vector distribution fields are collected and analyzed. The local divergence distribution, spatial connectivity characteristics, and density of intersecting clusters in the normal dendrite growth region are statistically analyzed, and their statistical distributions, such as the 90th percentile and average distance, are taken as the benchmarks for the above-mentioned thresholds. In practical applications, the benchmark values ​​can be finely adjusted, for example, by ±20%, according to different steel grades, cross-sectional dimensions, or process windows, to adapt to specific production conditions. This calibration and adjustment process is a routine operation for those skilled in the art based on statistical common sense and process knowledge.

[0069] In one implementation, based on the set of intersection point coordinates, a topographic feature matrix is ​​constructed for the corresponding locations, and the convergence intensity is calculated based on the topographic feature matrix, including:

[0070] Based on the set of intersection point coordinates, a morphological feature matrix is ​​constructed for the corresponding positions of the intersection points to obtain the morphological feature matrix of the corner region.

[0071] The volume change trend of the corner region's topographic feature matrix is ​​calculated to obtain the radial contraction gradient field.

[0072] Based on the radial contraction gradient field, the peak value of the gradient modulus is identified to obtain the set of sector contraction boundary radii.

[0073] The discreteness of the set of sector contraction boundary radii is calculated to obtain the asymmetric contraction characteristic coefficient of the corner region. Based on the asymmetric contraction characteristic coefficient of the corner region, the convergence intensity is determined to obtain the convergence intensity characterizing the physical bonding tightness.

[0074] It should be noted that when constructing the morphological feature matrix corresponding to the intersection point, a local sampling window with a side length of 80 pixels needs to be constructed centered on the intersection point coordinates. Crystal microscopic image data within this window is then captured to obtain the corner region image, generating a corner region morphological feature matrix to characterize the transition morphology between the dendrite arm tip and the central convergence region. The corner region morphological feature matrix refers to a two-dimensional data array of grayscale values ​​arranged spatially within this window. For example, at the intersection point (582, 510), the local sampling window covers a row range of 542 to 622 and a column range of 470 to 550. The captured image data shows that the dendrite arm tip exhibits a centripetal contraction morphology, and this 80×80 matrix contains a total of 6400 grayscale values.

[0075] Furthermore, calculating the volume change trend of the morphological feature matrix in the corner region requires performing four-way differential operations in the horizontal, vertical, and 45-degree diagonal directions in the corner region to calculate the gray-level gradient magnitude, thereby constructing a radial shrinkage gradient field reflecting the local shrinkage trend of the material volume. The radial shrinkage gradient field refers to the vector distribution of the rate of change of gray-level values ​​from the outer corner towards the center; a larger gradient modulus indicates more significant shrinkage deformation. Multi-directional differential operations refer to the mathematical operations of calculating the derivative of the image's gray-level space along different angles. For example, a four-directional Sobel operator is used for differential calculation, specifically calculating the gray-level gradients in the 0°, 45°, 90°, and 135° directions, and synthesizing the gradient magnitude matrix. For instance, the 80×80 morphological feature matrix is ​​convolved with 3×3 differential templates in each of the four directions, and the gradient magnitude is calculated pixel-by-pixel, outputting an 80×80 radial shrinkage gradient field. In particular, in the annular region 35 pixels from the intersection point, the radial contraction gradient field shows a gradient amplitude of 0.72, indicating that the centripetal deformation caused by solidification contraction of the dendrites at this location is the most significant, and the gradient amplitude increases from the periphery to the center.

[0076] In this embodiment, when identifying the gradient modulus peak, the corner region needs to be divided into eight 45-degree sector sub-regions. Within each sector, the radially contracting gradient field is scanned to locate the gradient modulus peak position. The radial distance from this peak point to the intersection point is recorded, forming a set of sector contraction boundary radii containing eight radius values. A sector sub-region refers to a sector-shaped spatial partition with the intersection point as its vertex, covering a 45-degree angle. The gradient modulus peak refers to the maximum gradient magnitude within the sector. The sector contraction boundary radius refers to the Euclidean distance from the gradient peak position to the intersection point. For example, a polar coordinate system is established with the intersection point as the center. 45 rays are emitted outward from the intersection point with a step size of 1 pixel. Each ray corresponds to a sector centerline. The maximum gradient point is searched within a range of ±22.5 degrees. The distance between this point and the intersection point is calculated, and a set of sector contraction boundary radii containing eight values ​​is output. For example, the shrinkage boundary radii of the 8 sectors were measured to be 28, 31, 34, 29, 36, 33, 30, and 27 pixels, respectively, corresponding to the directions of 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°. The maximum value of 36 pixels is located in the 180° sector, and the minimum value of 27 pixels is located in the 315° sector.

[0077] Specifically, the asymmetric shrinkage characteristic coefficient of the corner region is the ratio of the standard deviation to the mean. This coefficient ranges from 0 to 1; a larger value indicates a more significant difference in shrinkage in different directions at the corner. Convergence intensity is a comprehensive index quantifying the physical bonding tightness of dendrite arms in the intersection region, ranging from 0 to 1. When calculating the dispersion of the set of sector shrinkage boundary radii, the standard deviation is used to calculate the dispersion, and an exponential mapping function is applied. Specifically, the arithmetic mean μ and standard deviation σ of the eight radius values ​​are first calculated. The asymmetric shrinkage characteristic coefficient is characterized by the coefficient of variation, i.e., σ / μ. When determining the convergence intensity, a negative exponential function is used to map the asymmetric shrinkage characteristic coefficient of the corner region to the convergence intensity S. The asymmetric shrinkage characteristic coefficient σ / μ is substituted into the exponential function, and the exponential operation is performed to output the convergence intensity S, such as convergence intensity S = exp(-3 × σ / μ). A higher convergence intensity indicates a tighter physical bonding of dendrite arms in the intersection region and a stronger inter-crystal overlap. Understandably, by quantifying the asymmetric shrinkage characteristics and correlating them with the convergence intensity, the differences in the bonding quality of dendritic structures at different junctions can be clearly revealed, providing a reliable basis for subsequent judgment on the existence of potential weak bonding regions.

[0078] In step S14, if the convergence intensity exceeds a preset intensity threshold, a rhombus segregation candidate region is extracted to obtain a rhombus segregation candidate region image. A gray-level co-occurrence matrix is ​​constructed based on the rhombus segregation candidate region image, the feature values ​​of the gray-level co-occurrence matrix are extracted, and a mapping calculation is performed to obtain a segregation severity index.

[0079] It is worth noting that this step applies a conditional threshold judgment to the convergence intensity output in step S13. When the intensity value exceeds the preset intensity threshold, it triggers the cropping of the candidate rhomboid segregation region image. Then, it extracts texture disorder features through the gray-level co-occurrence matrix, maps the degree of segregation, and fuses the region area information to calculate a comprehensive segregation severity index. Thus, a logical correlation is established between convergence intensity and rhomboid segregation risk, realizing a progressive detection from structural stability assessment to compositional segregation defect quantification. The preset intensity threshold is used to determine whether the convergence intensity reaches the standard for triggering rhomboid segregation quantification analysis. Its setting is based on the global statistical distribution characteristics of convergence intensity, the actual overlap quality level classification of dendrite arms, and the structural stability requirements of the continuous casting process. Here, a large-sample statistical quantile method is used, specifically requiring the collection of no less than 100 sets of continuous casting billet production data to calculate the convergence intensity of all intersection points. The baseline preset intensity threshold is statistically calculated to be 0.73.

[0080] Furthermore, the rhombic segregation candidate region image refers to the original grayscale image data of a rhombic geometric region constructed along the main direction of the dendrite arm, with the four vertices located at the ends of the dendrite arm and the side length being 1.2 times the distance from the intersection point to the end of the arm. For example, at the intersection point (582, 510), the convergence intensity S = 0.754 is calculated, which is greater than the preset threshold of 0.75. The judgment condition is met, and the system performs the subsequent rhombic segregation candidate region truncation operation. The distance to the right edge of the billet is calculated to be 68 pixels, and the side length of the rhombus is set to 1.2 × 68 = 82 pixels. The coordinates of the four vertices of the rhombus are (582, 510) ± (41, 0) and (582, 510) ± (0, 41), respectively. The size of the truncated rhombic region image is 82 × 82 pixels, which is the rhombic segregation candidate region image.

[0081] In one implementation, a gray-level co-occurrence matrix is ​​constructed based on the rhomboid segregation candidate region image. Feature values ​​of the gray-level co-occurrence matrix are extracted and mapped to obtain a segregation severity index, including:

[0082] Based on the rhomboid segregation candidate region image, a matrix is ​​constructed to obtain a gray-level co-occurrence matrix, and feature values ​​are extracted based on the gray-level co-occurrence matrix to obtain comprehensive texture disorder feature values.

[0083] The degree of segregation is mapped to the comprehensive texture disorder feature value to obtain the rhomboid segregation degree value;

[0084] The severity of the segregation is calculated by combining the rhomboid segregation candidate region image with the rhomboid segregation degree value, and the segregation severity index is obtained.

[0085] It should be noted that the gray-level co-occurrence matrix refers to the probability matrix that a pixel with gray value i and a pixel with gray value j in an image appear simultaneously under a specific spatial relationship; the comprehensive texture disorder feature value refers to the scalar value that quantifies the mapping of texture non-uniformity to the 0 to 1 interval by fusing four statistical parameters: contrast, correlation, energy, and inverse difference moment. When constructing the gray-level co-occurrence matrix for the rhomboid segregation candidate region image, a uniform quantization method is first used to linearly map the gray values ​​from 0 to 255 in the rhomboid segregation candidate region image to 32 quantization levels from 0 to 31, reducing the order of magnitude. Then, a standard parameter with a fixed spacing d=1 pixel and four directions θ∈{0°,45°,90°,135°} is used to define the spatial relationship. Next, a 32×32 zero matrix is ​​initialized, and all pixel pairs in the image are traversed. For each pair of pixels, the gray level (i,j) is incremented by 1 for the corresponding matrix element. After the statistics are completed, all elements of the matrix are divided by the total number of pixel pairs to achieve normalization. Finally, the matrix is ​​symmetric to obtain the gray-level co-occurrence matrix. Specifically, the asymmetric co-occurrence matrix is ​​converted into a symmetric matrix by adding the matrix to its transpose and then dividing by 2 to eliminate the directional influence, so that the subsequent texture feature extraction is not affected by the scanning direction and enhances the feature stability.

[0086] Furthermore, the extraction of comprehensive texture disorder features refers to the process of extracting four statistical parameters—contrast, correlation, energy, and inverse moment—from the normalized gray-level co-occurrence matrices in four directions, and then fusing them into a comprehensive texture disorder feature value F. Here, contrast reflects texture sharpness, correlation reflects the degree of gray-level linear dependence, energy reflects texture uniformity, and inverse moment reflects texture homogeneity. In one example, the four statistical parameters are weighted and summed according to preset weights to obtain a single comprehensive texture disorder feature value F. This value summarizes multi-dimensional texture information and is directly used for mapping the degree of segregation. The weight vector during fusion is set to [0.3, 0.2, 0.2, 0.3], corresponding to [contrast, correlation, energy, inverse moment], respectively. This is determined based on the sensitivity analysis of the parameters to segregated textures, specifically obtained through large-sample statistical analysis.

[0087] In this embodiment, the rhombus segregation degree value refers to a dimensionless index that quantifies the severity of enrichment of central components, with a value range of 0-1. When performing segregation degree mapping, the comprehensive texture disorder feature value is input into a preset segregation degree mapping model and converted into a rhombus segregation degree value in the range of 0 to 1, realizing a quantitative mapping from texture features to the severity of chemical component segregation. For a simple example, the segregation degree mapping model here adopts a three-segment linear mapping function. When the feature value F∈[0,0.35], the segregation degree P=0.5×F; when F∈(0.35,0.55], P=0.175+2×(F-0.35); when F>0.55, P=0.575+0.5×(F-0.55). The calculated comprehensive texture disorder feature value F is substituted into the corresponding interval formula to output the segregation degree P, where the comprehensive texture disorder feature value has been weighted and normalized as described above, and its value range is [0,1].

[0088] It should be noted that the segmentation points and coefficients of the three-segment linear mapping function were obtained by performing correlation regression analysis on the comprehensive texture disorder feature values ​​and the local carbon segregation index obtained from offline sampling chemical analysis. First, a batch of cast billet samples covering different degrees of segregation were prepared, and their low-magnification images and the chemical composition of drill cuttings at the corresponding locations were acquired simultaneously to calculate the texture feature values ​​and segregation index. Second, a piecewise linear regression algorithm was used to fit the texture feature value and segregation index data pairs, aiming to minimize the root mean square error, and automatically determining the optimal segmentation points and linear coefficients for each interval. The final obtained mapping relationship allows the visual texture features to be reliably mapped to the physical degree of segregation. Those skilled in the art can follow the same experimental and data analysis process to calibrate the model for their specific production line.

[0089] Specifically, the segregation severity calculation combines the area proportion W (the percentage of pixels in this area out of the total pixels in the billet cross-section) of the rhomboid segregation candidate region image with the segregation degree value P to calculate the final segregation severity index. This index integrates both segregation intensity and influence range. The segregation severity index is a quantitative index that comprehensively evaluates the impact of rhomboid segregation on the overall quality of the billet; the area proportion refers to the percentage of pixels occupied by the rhomboid candidate region in the observation field; the calculation method involves multiplying the segregation degree value P by the square root of the area proportion W to obtain the segregation severity index.

[0090] In step S15, according to the preset physical quantity mapping rules, the concentration distribution scalar field is transformed for the segregation severity index to obtain the concentration distribution scalar field, and the solidification state is characterized according to the concentration distribution scalar field to obtain the overall solidification front distribution.

[0091] It is worth noting that this step converts the dimensionless segregation severity index into a scalar field with concentration physical meaning by pre-setting physical mapping rules. Then, it combines the heat flux vector field to construct a solute-thermal coupling driving field. Through time stepping and three-dimensional surface interpolation, the overall solidification front distribution is reconstructed, realizing the leap from segregation defect quantification to solidification interface morphology inversion, and revealing the distortion mechanism of component segregation on the solidification front advancement path.

[0092] Here, the concentration distribution scalar field refers to the percentage increment of the solute concentration relative to the average value at each element in the two-dimensional matrix; the physical quantity mapping rule refers to the predefined piecewise linear function relationship between segregation severity and concentration increment. In this embodiment, the physical quantity mapping rule adopts a three-segment linear mapping function. When the segregation severity index S∈[0,0.08], the concentration increment C=0.8%×(S / 0.08); when S∈(0.08,0.15], C=0.8%+1.7%×(S-0.08) / 0.07; when S>0.15, C=2.5%+1.5%×(S-0.15) / 0.10. Based on the piecewise function coefficients fitted according to the calibration experimental data, the function mapping is performed point by point on the segregation severity index matrix, outputting a concentration distribution scalar field of the same size as the grid of the billet section.

[0093] In one implementation, solidification state is characterized based on the concentration distribution scalar field to obtain the overall solidification front distribution, including:

[0094] Obtain the heat flux vector field;

[0095] The migration gradient of the concentration distribution scalar field is calculated to obtain the solute migration gradient vector field, and the solute migration gradient vector field is coupled with the heat flux vector field to obtain the composite driving vector field.

[0096] The solidification front is reconstructed in three dimensions based on the composite driving vector field to obtain the overall solidification front distribution.

[0097] It is worth noting that the heat flux vector field refers to the magnitude and direction distribution of the heat flux density at each point on the cross-section of the billet, specifically measured in real time by an array of thermocouples arranged on the surface of the billet. The heat flux vector field provides the direction and intensity of the driving force for heat transfer during the solidification process, serving as the reference input for subsequent coupling calculations.

[0098] In this embodiment, the solute migration gradient vector field refers to the magnitude and direction distribution of the solute concentration gradient at each point, obtained through migration gradient calculation. When calculating the migration gradient, the Sobel gradient operator is used, specifically employing a 3×3 convolution kernel to calculate the first-order partial derivatives of the concentration distribution scalar field in the x and y directions, synthesizing the gradient vector. For example, at point (580, 510), the concentration distribution value is 2.06%, its right neighbor has a concentration of 2.31%, and its left neighbor has a concentration of 2.01%. The gradient in the x direction is calculated. C / x = (2.31 - 2.01) / 2 = 0.15% / pixel; upper neighbor density 2.18%, lower neighbor density 1.94%, gradient in the y direction. C / y=(2.18-1.94) / 2=0.12% / pixel, solute migration gradient vector g=(0.15,0.12)% / pixel, the direction is towards the direction of the fastest increase in concentration, and the magnitude is 0.192% / pixel; perform horizontal and vertical convolution operations on the 200×200 concentration distribution matrix to output a solute migration gradient vector field of the same size.

[0099] Furthermore, the composite driving vector field refers to the combined driving force distribution after coupling the solute gradient and heat flux vectors, which is obtained through coupling calculations of the solute migration gradient vector field and the heat flux vector field. A fixed weight ratio is used for coupling calculations, with the coupling weights typically set to prioritize the solute gradient. This is the result obtained through fitting analysis of large sample data; for example, the heat flux vector weight is 0.3, and the solute migration gradient weight is 0.7, thus highlighting the dominant influence of component segregation on the solidification front morphology. For example, at point (580, 510), the solute gradient g = (0.15, 0.12), the heat flux vector q = (360, 240), and the normalized q_norm = (0.83, 0.56), the composite driving vector D characterizing the relative driving intensity and direction is calculated as D = 0.7 × (0.15, 0.12) + 0.3 × (0.83, 0.56) = (0.105, 0.084) + (0.249, 0.168) = (0.354, 0.252). This vector is biased towards the solute-rich region, with a modulus of 0.435, driving the solidification interface to indent towards the solute-rich region.

[0100] It should be noted that the coupling weights are determined through an inverse problem optimization method. Based on known solidification structures, such as typical samples of dendrite spacing and final segregation morphology, the two weight parameters are optimized (the sum of which is 1) with the goal of maximizing the agreement between the solidification front reconstruction results and metallographic observation results. The optimization results of a large number of cases show that the reconstruction accuracy is best when the solute migration gradient weight is in the range of 0.65~0.75 and the heat flow vector weight is in the range of 0.25~0.35. The embodiment of this invention takes the middle value. This weight allocation reflects that in the later stage of solidification of continuously cast billets, the influence of solute redistribution on the compositional supercooling on the interface morphology is often stronger than that of simple heat flow transport.

[0101] Specifically, the overall solidification front distribution refers to the complete spatial morphology of the solid-liquid interface on the cross-section of a continuously cast billet, driven by the solute migration gradient and thermal flux coupling. It manifests as a continuous, smooth three-dimensional surface, where the coordinates of each spatial point correspond to the actual advancement position of the solidification interface at a specific moment, obtained through three-dimensional reconstruction of the solidification front. In one example, discrete advancement with a time step of 0.1 s and cubic spline surface interpolation are used to achieve the three-dimensional reconstruction of the solidification front. By accumulating each displacement step to 100 steps, a set of interface displacement points within 10 seconds is obtained, and then bicubic spline interpolation is performed on the point set. For example, at point (580, 510), the composite driving vector magnitude is 0.435 units / second, the step size is 0.1 seconds, and the single-step displacement is 0.0435 units. After accumulating 100 steps, the solidification front at this point advances 4.35 units towards the liquid core side. The central region has a significant depression due to the large driving vector magnitude. The reconstructed overall solidification front distribution shows that the central segregation zone front lags behind the corner by about 12 mm. The depression depth is positively correlated with the segregation severity index.

[0102] It should be noted that the time step and total advance time of 10 seconds are set based on the typical cooling rate and interface advance speed of the continuously cast billet at the end of solidification. For ordinary carbon steel billets with cross-sectional dimensions of 150mm×150mm~200mm×200mm, the advance speed of the mushy zone at the end of solidification is usually on the order of 0.1-1mm / s. Choosing a time step of 0.1 seconds ensures that the displacement in a single step is within the spatial grid resolution, meeting the requirements of numerical stability. Choosing a total simulation time of 10 seconds is sufficient to cover the critical stage from the mushy zone to complete solidification, thereby completely reconstructing the final solidification front morphology. Those skilled in the art can adjust the above time parameters proportionally according to the billet cross-sectional size and casting speed.

[0103] In step S16, based on the overall solidification front distribution, a matching judgment is made between the front advancement velocity and the heat flux distribution. If the front advancement velocity and the heat flux distribution do not match, the vector distribution field is adjusted to obtain an optimized intersection point, including:

[0104] Obtain the local heat flux intensity value;

[0105] The normal propagation rate is calculated based on the overall solidification front distribution, and the normal propagation rate is matrix-mapped with the local heat flux intensity value to obtain the velocity-heat flux coupling matrix.

[0106] Calculate the covariance of the velocity-thermal-fluid coupling matrix. If the covariance is lower than a preset consistency threshold, it is marked as a mismatch feature region.

[0107] Based on the mismatch feature region, a vector field superposition correction is performed to obtain a corrected composite vector field;

[0108] Based on the modified composite vector field, the solidification interface reconstruction deduction is performed to obtain the optimized intersection point.

[0109] It is worth noting that this step identifies heat-mass transport mismatch regions by comparing the spatial distribution consistency between the normal advance rate of the solidification front and the local heat flux intensity values. A deviation correction tensor is generated and superimposed onto the original vector field. The coordinates of the intersection point are then optimized through solidification interface reconstruction and deduction, correcting the prediction error of the solidification interface caused by component segregation and compensating for the deviation between the heat flux-driven and actual solidification rates. The local heat flux intensity value refers to the heat flow rate per unit area of ​​a micro-element of the billet cross-section, expressed in W / m². The heat flux density value at each point on the billet cross-section is obtained. Heat flux intensity characterizes the intensity of heat transfer from the inside of the billet to the outside and is the fundamental driving force of the solidification process; it needs to be used as a baseline input for subsequent coupling analysis.

[0110] In this embodiment, the normal advance rate refers to the speed at which the solidification front moves in the direction perpendicular to the interface, measured in mm / s. In the central region of a low-carbon steel continuous casting billet, the normal advance rate of the normal dendrite front is approximately 0.45 mm / s, while it can drop to 0.18 mm / s near the severely segregated zone, indicating a significant local solidification lag. The velocity-thermal-fluidity coupling matrix refers to the frequency distribution matrix formed by dividing the advance rate and heat flux intensity into two-dimensional grids according to numerical ranges and counting the number of points in each grid. Specifically, the local heat flux intensity q corresponding to the front node is extracted, and the heat flux intensity range [0, 1000] kW / m² is divided into 10 equal intervals, and the normal advance rate range [0, 0.6] mm / s is divided into 12 equal intervals, forming a 10×12 coupling matrix. The normal advance rate and heat flux intensity of each grid node are encoded into intervals, and the number of points falling into each interval combination is counted, outputting a 10×12 dimensional velocity-thermal-fluidity coupling matrix. In this matrix, rows represent different normal rate intervals, and columns represent heat flux intensity levels. For example, there are 123 points with a propulsion speed of 0.4 mm / s and a heat flux intensity of 850 W / m², which corresponds to 123 elements in the 8th row and 3rd column of the velocity-heat flux coupling matrix.

[0111] Furthermore, covariance refers to the linear correlation measure between the row variables (propelling rate) and column variables (heat flux intensity) of the coupling matrix, obtained by calculating the Pearson correlation coefficient between the distributions of the normal propulsion rate and the local heat flux intensity. The mismatch characteristic zone refers to the area of ​​the cast billet where the propulsion rate and heat flux intensity are severely disconnected. The consistency threshold is the critical covariance value used to determine whether heat flux and solidification rate are compatible; it is set based on statistical analysis of large sample data distributions and is here taken as 0.65. For example, in the central segregation zone, the calculated covariance of the coupling matrix is ​​0.51, which is lower than the threshold of 0.65. The system determines that there is a severe mismatch in this area and marks the mismatch characteristic zone covering a 40mm range on both sides of the cast billet centerline.

[0112] In one example, vector field superposition correction refers to a compensation operation that adds the correction tensor to the original vector field point by point. The correction tensor is constructed using the gradient difference method, specifically calculating the propulsion rate gradient and heat flux gradient within the mismatch region, such as the propulsion rate gradient (0.025, -0.012)s. -1 The heat flux gradient is (0.018, -0.008)s. -1 Calculate the corrected tensor (0.025, -0.012) - 0.5 × (0.018, -0.008) = (0.016, -0.008)s. -1 After being superimposed on the original vector field, the output is a modified composite vector field. The vector direction at this point is deflected by 12° toward the center of the billet, and the modulus value increases by 18%.

[0113] Specifically, after obtaining the corrected composite vector field, the solidification interface time integration deduction needs to be performed again to recalculate the convergence position of the growth vector and obtain the corrected intersection point coordinates, which are closer to the actual intersection position of the solidification front. Reconstruction deduction refers to the process of recalculating the solidification interface displacement based on the corrected vector field; the optimized intersection point refers to the more accurate dendrite growth intersection position coordinates after deviation correction. The reconstruction deduction uses the same time step and cumulative number of steps as step S15, but the input vector field is replaced with the corrected composite vector field. 100 time integration steps are performed on each grid node of the corrected vector field, and the accumulated displacement yields a new set of spatial coordinate points. Convergence regions with growth vector magnitudes less than 0.01 are detected, the geometric center of these regions is calculated, and the optimized intersection point coordinates are output. For example, the original intersection point coordinates were (582, 510), which deviated from the geometric center of the billet (600, 500) by 18 mm. After vector field correction and reconstruction, the new intersection point coordinates were (594, 502), with the deviation reduced to within 6 mm and the positioning accuracy improved by 67%. This optimized coordinate was used as the final intersection point input in step S17.

[0114] In step S17, interference data is extracted based on the optimized intersection point, and a quantitative assessment of rhombus transformation is performed based on the interference data to obtain the final quantitative assessment result of rhombus transformation.

[0115] It is worth noting that the interference data mentioned in this embodiment specifically refers to the multi-physics field data affecting the microstructure at the end of solidification, which is collected in real time by the online monitoring system of the continuous casting process or calculated in real time by the process model in the area near the optimized intersection point. This includes temperature fluctuation data and solute gradient data. The temperature fluctuation data is obtained by collecting the historical temperature curves of the surface and core of the billet at the end of solidification at the intersection point through a micro thermocouple array embedded in specific locations in the crystallizer and the secondary cooling zone, and calculating its short-term fluctuation amplitude and frequency. The solute gradient data is based on the concentration distribution scalar field generated in step S15, and the concentration gradient vector value of the area around the optimized intersection point is directly extracted. These data together constitute the input for evaluating the thermo-mechanical-solute coupling effect.

[0116] In one implementation, a quantitative assessment of rhombus variation is performed based on the interference data to obtain the final quantitative assessment result of rhombus variation, including:

[0117] Based on the interference data, calculate the thermal stress tensor and the cumulative plastic strain value;

[0118] Based on the thermal stress tensor, a lattice mismatch feature map is generated to obtain the geometric deviation corresponding to the lattice mismatch feature map.

[0119] By combining the geometric deviation and the cumulative plastic strain value for multidimensional mapping evaluation, the final quantitative evaluation result of the rhombic deformation is obtained.

[0120] It is worth noting that the thermal stress tensor refers to the three-dimensional stress state of a micro-element caused by uneven temperature and solute distribution; the cumulative plastic strain value refers to the accumulation of irreversible deformation of the material during thermal cycling, characterizing the degree of damage. In this embodiment, the thermal stress tensor and cumulative plastic strain value are calculated using a thermo-elastic-plastic incremental constitutive model. The specific model includes thermal expansion terms, phase transformation volume change terms, and high-temperature creep terms, and the material parameters consider the influence of carbon equivalent. By substituting the transient temperature fluctuation value and solute concentration gradient value from the interference factors into the thermo-elastic-plastic incremental constitutive equation, the stress increment and plastic strain increment are solved using the finite difference method, and the thermal stress tensor and cumulative plastic strain value are obtained iteratively. In particular, the thermal stress tensor is mainly generated by the thermal expansion incompatibility driven by the temperature gradient, and the cumulative plastic strain value often exceeds 0.008 in the severely segregated region, indicating that the material has undergone irreversible deformation.

[0121] Furthermore, the lattice mismatch feature map refers to a two-dimensional distribution map showing the degree of deviation of a crystal lattice from an ideal lattice. The generation of the lattice mismatch feature map employs a principal strain-lattice deviation linear mapping method. Specifically, the stress tensor is decomposed into a strain tensor, the deviation between the actual strain and the ideal solidification strain is calculated, and this deviation is quantized using the Frobenius norm to obtain the geometric deviation of the lattice. The ideal solidification strain is set to 0. The geometric deviation value refers to the statistical average of the deviation in the lattice distortion region, reflecting the intensity of lattice distortion; strain and deviation are calculated point-by-point for a 200×200 stress tensor field, and the geometric deviation is output.

[0122] It is worth noting that when the thermal stress tensor exhibits a clear alternating tensile-compressive distribution, the high mismatch regions in the lattice mismatch characteristic spectrum are concentrated in bands on both sides of the segregation band. The geometric deviation values ​​are typically between 0.012 and 0.028; the larger the value, the more significant the crystal lattice distortion, leading to inhomogeneity in the subsequent phase transformation structure. Understandably, this spectrum visually reflects the distorting effect of thermal stress on the microlattice, providing a visual basis for the microscopic mechanism of rhombic transformation formation.

[0123] In one example, multidimensional mapping assessment refers to a mathematical method that maps multiple physical quantities to a unified assessment index through weighted combination. The final quantitative assessment result of rhombic deformation refers to the comprehensive risk level value characterizing the macroscopic deformation and microscopic damage caused by the combined effect of rhombic segregation and thermal stress in the cast billet. The higher the value, the more severe the rhombic deformation risk. A two-dimensional assessment plane is constructed by weighting and fusing the geometric deviation value and the cumulative plastic strain value, and mapped to a comprehensive rhombic deformation assessment index in the 0-1 interval. This index directly corresponds to the severity level of the rhombic deformation defect, resulting in the final quantitative assessment result. Specifically, a linear weighted fusion model is used, weighting the geometric deviation and cumulative plastic strain with weights of 0.6 and 0.4, respectively. Substituting the geometric deviation value and the cumulative plastic strain value into the weighting formula, the assessment index in the 0-1 interval is output. For example, when the geometric deviation is 0.021 and the cumulative plastic strain value is 0.0095, the rhombic deformation assessment index reaches 0.76. The weight configuration is determined based on regression analysis of offline metallographic inspection data.

[0124] In summary, this invention discloses an online intelligent detection method based on low-magnification detection technology. Through a comprehensive analysis and optimization method targeting the solidification process of continuously cast billets, it solves the billet quality problems caused by dendrite growth boundaries, rhombic segregation degree, and uneven distribution of the solidification front. Furthermore, by acquiring macroscopic images of the continuously cast billets, it extracts dendrite growth boundaries, determines the dendrite growth direction vector, and constructs a global vector distribution field, intuitively depicting the directionality and spatial distribution of dendrite growth. It locates potential intersection points and extracts convergence intensity, quantifying the convergence characteristics of dendrites, thus solving the problem of existing technologies' inability to accurately depict the directionality of dendrite growth and its convergence characteristics. This invention also extracts rhombic segregation candidate regions by intercepting convergence intensity exceeding a preset threshold, focusing on high-risk segregation regions, quantifying segregation severity indicators, and converting them into a concentration distribution scalar field based on physical quantity mapping rules to depict the solidification state. This transforms visual features into quantifiable physical parameters, solving the problem of existing technologies' inability to quantify rhombic segregation defects and achieving accurate quantification of the severity of rhombic segregation.

[0125] Reference Figure 2 The second embodiment of the present invention provides an online intelligent detection system based on low-magnification detection technology, comprising:

[0126] The boundary data acquisition module is used to acquire a macroscopic image of the continuously cast billet and extract the dendrite growth boundary based on the macroscopic image of the continuously cast billet to obtain the boundary contour.

[0127] The vector distribution construction module is used to extract the discrete coordinate points of the boundary contour and calculate the dendrite growth direction vector. Based on the dendrite growth direction vector and the discrete coordinate points, a set of directed vectors is generated and spatially mapped to obtain the vector distribution field.

[0128] The convergence intensity extraction module is used to screen out effective dendritic convergence clusters based on the vector distribution field, calculate potential convergence points based on the effective dendritic convergence clusters, obtain a set of convergence point coordinates, construct a morphological feature matrix for the corresponding position based on the set of convergence point coordinates, and calculate the convergence intensity based on the morphological feature matrix.

[0129] The segregation degree quantification module is used to extract a diamond-shaped segregation candidate region if the convergence intensity exceeds a preset intensity threshold, obtain a diamond-shaped segregation candidate region image, construct a gray-level co-occurrence matrix based on the diamond-shaped segregation candidate region image, extract the feature values ​​of the gray-level co-occurrence matrix, and perform mapping calculation to obtain a segregation severity index.

[0130] The front distribution reconstruction module is used to perform a concentration distribution scalar field transformation on the segregation severity index according to a preset physical quantity mapping rule to obtain a concentration distribution scalar field, and to characterize the solidification state based on the concentration distribution scalar field to obtain the overall solidification front distribution.

[0131] The leading edge matching optimization module is used to determine the matching between the leading edge advance velocity and the heat flux distribution based on the overall solidification leading edge distribution. If the leading edge advance velocity and the heat flux distribution do not match, the vector distribution field is adjusted to obtain an optimized intersection point.

[0132] The rhombus transformation quantitative assessment module is used to extract interference data based on the optimized intersection point, and to perform rhombus transformation quantitative assessment based on the interference data to obtain the final rhombus transformation quantitative assessment result.

[0133] It should be noted that the online intelligent detection system based on low-magnification detection technology provided in this embodiment of the invention is used to execute all the process steps of the online intelligent detection method based on low-magnification detection technology in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0134] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0135] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that 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 for those skilled in the art.

Claims

1. An online intelligent detection method based on low-magnification detection technology, characterized in that, include: A macroscopic image of a continuously cast square billet is acquired, and dendrite growth boundaries are extracted based on the macroscopic image of the continuously cast square billet to obtain the boundary contour; Extract the discrete coordinate points of the boundary contour and calculate the dendrite growth direction vector. Generate a set of directed vectors based on the dendrite growth direction vector and the discrete coordinate points, and perform spatial mapping to obtain the vector distribution field. Effective dendrite junction clusters are selected based on the vector distribution field, potential junction points are calculated based on the effective dendrite junction clusters, a set of junction point coordinates is obtained, and a morphological feature matrix is ​​constructed based on the set of junction point coordinates. The convergence intensity is calculated based on the morphological feature matrix. If the convergence intensity exceeds a preset intensity threshold, a rhombus segregation candidate region is extracted to obtain a rhombus segregation candidate region image. A gray-level co-occurrence matrix is ​​constructed based on the rhombus segregation candidate region image. The feature values ​​of the gray-level co-occurrence matrix are extracted and mapped to obtain a segregation severity index. According to the preset physical quantity mapping rules, the concentration distribution scalar field is transformed for the segregation severity index to obtain the concentration distribution scalar field, and the solidification state is characterized according to the concentration distribution scalar field to obtain the overall solidification front distribution. Based on the overall solidification front distribution, a matching judgment is made between the front advancement velocity and the heat flux distribution. If the front advancement velocity and the heat flux distribution do not match, the vector distribution field is adjusted to obtain an optimized intersection point. Interference data is extracted based on the optimized intersection point, and a quantitative assessment of rhombus transformation is performed based on the interference data to obtain the final quantitative assessment result of rhombus transformation.

2. The online intelligent detection method based on low-magnification detection technology according to claim 1, characterized in that, The method extracts dendrite growth boundaries based on the macroscopic image of the continuously cast billet to obtain boundary contours, including: The macroscopic image of the continuously cast billet is converted to grayscale to obtain grayscale image data; Edge detection is performed on the grayscale image data to obtain boundary edge data; Boundary continuity analysis is performed on the boundary edge data. If boundary breaks exist, boundary break repair is performed to obtain the boundary contour of the dendrite growth boundary.

3. The online intelligent detection method based on low-magnification detection technology according to claim 1, characterized in that, The process of extracting discrete coordinate points of the boundary contour and calculating dendrite growth direction vectors, generating a set of directed vectors based on the dendrite growth direction vectors and the discrete coordinate points, and performing spatial mapping to obtain a vector distribution field includes: Traverse the coordinate sequence of the boundary contour, extract discrete coordinates, and obtain discrete coordinate points; The normal direction is calculated based on the discrete coordinate points to obtain the dendrite growth direction vector extending from the boundary to the liquid core side; By combining the discrete coordinate points and the dendrite growth direction vector, a set of directed vectors is generated, and the set of directed vectors is spatially mapped to obtain the vector distribution field characterizing the macroscopic dendrite growth trend.

4. The online intelligent detection method based on low-magnification detection technology according to claim 1, characterized in that, The process involves selecting effective dendrite junction clusters based on the vector distribution field, calculating potential junction points based on the effective dendrite junction clusters, and obtaining a set of junction point coordinates, including: Calculate the local divergence of the vector distribution field. If the local divergence is greater than a preset divergence threshold, mark it as a high-density aggregation candidate point to obtain a high-density aggregation candidate point set. Based on the set of high-density aggregated candidate points, spatial connectivity is constructed to obtain connected sub-regions, and density feature analysis is performed on the connected sub-regions to obtain highly cohesive effective dendritic junction clusters; Calculate the spatial geometric center of the effective dendrite cluster to obtain the set of intersection point coordinates that characterize the macroscopic dendrite growth intersection position.

5. The online intelligent detection method based on low-magnification detection technology according to claim 1, characterized in that, The step of constructing a topographic feature matrix for the corresponding location based on the set of intersection point coordinates, and calculating the convergence intensity based on the topographic feature matrix, includes: Based on the set of intersection point coordinates, a morphological feature matrix is ​​constructed for the corresponding positions of the intersection points to obtain the morphological feature matrix of the corner region. The volume change trend of the corner region's topographic feature matrix is ​​calculated to obtain the radial contraction gradient field. Based on the radial contraction gradient field, the peak value of the gradient modulus is identified to obtain the set of sector contraction boundary radii. The discreteness of the set of sector contraction boundary radii is calculated to obtain the asymmetric contraction characteristic coefficient of the corner region. Based on the asymmetric contraction characteristic coefficient of the corner region, the convergence intensity is determined to obtain the convergence intensity characterizing the physical bonding tightness.

6. The online intelligent detection method based on low-magnification detection technology according to claim 1, characterized in that, The process of constructing a gray-level co-occurrence matrix based on the rhomboid segregation candidate region image, extracting the feature values ​​of the gray-level co-occurrence matrix, and performing mapping calculations to obtain a segregation severity index includes: Based on the rhomboid segregation candidate region image, a matrix is ​​constructed to obtain a gray-level co-occurrence matrix, and feature values ​​are extracted based on the gray-level co-occurrence matrix to obtain comprehensive texture disorder feature values. The degree of segregation is mapped to the comprehensive texture disorder feature value to obtain the rhomboid segregation degree value; The severity of the segregation is calculated by combining the rhomboid segregation candidate region image with the rhomboid segregation degree value, and the segregation severity index is obtained.

7. The online intelligent detection method based on low-magnification detection technology according to claim 1, characterized in that, The method characterizes the solidification state based on the concentration distribution scalar field to obtain the overall solidification front distribution, including: Obtain the heat flux vector field; The migration gradient of the concentration distribution scalar field is calculated to obtain the solute migration gradient vector field, and the solute migration gradient vector field is coupled with the heat flux vector field to obtain the composite driving vector field. The solidification front is reconstructed in three dimensions based on the composite driving vector field to obtain the overall solidification front distribution.

8. The online intelligent detection method based on low-magnification detection technology according to claim 1, characterized in that, The process involves matching the leading edge velocity with the heat flux distribution based on the overall solidification front distribution. If the leading edge velocity and heat flux distribution do not match, the vector distribution field is adjusted to obtain an optimized intersection point, including: Obtain the local heat flux intensity value; The normal propagation rate is calculated based on the overall solidification front distribution, and the normal propagation rate is matrix-mapped with the local heat flux intensity value to obtain the velocity-heat flux coupling matrix. Calculate the covariance of the velocity-thermal-fluid coupling matrix. If the covariance is lower than a preset consistency threshold, it is marked as a mismatch feature region. Based on the mismatch feature region, a vector field superposition correction is performed to obtain a corrected composite vector field; Based on the modified composite vector field, the solidification interface reconstruction deduction is performed to obtain the optimized intersection point.

9. The online intelligent detection method based on low-magnification detection technology according to claim 1, characterized in that, The process involves performing a quantitative assessment of rhombic variation based on the interference data to obtain the final quantitative assessment result of rhombic variation, including: Based on the interference data, calculate the thermal stress tensor and the cumulative plastic strain value; Based on the thermal stress tensor, a lattice mismatch feature map is generated to obtain the geometric deviation corresponding to the lattice mismatch feature map. By combining the geometric deviation and the cumulative plastic strain value for multidimensional mapping evaluation, the final quantitative evaluation result of the rhombic deformation is obtained.

10. An online intelligent detection system based on low-magnification detection technology, characterized in that, include: The boundary data acquisition module is used to acquire a macroscopic image of the continuously cast billet and extract the dendrite growth boundary based on the macroscopic image of the continuously cast billet to obtain the boundary contour. The vector distribution construction module is used to extract the discrete coordinate points of the boundary contour and calculate the dendrite growth direction vector. Based on the dendrite growth direction vector and the discrete coordinate points, a set of directed vectors is generated and spatially mapped to obtain the vector distribution field. The convergence intensity extraction module is used to screen out effective dendritic convergence clusters based on the vector distribution field, calculate potential convergence points based on the effective dendritic convergence clusters, obtain a set of convergence point coordinates, construct a morphological feature matrix for the corresponding position based on the set of convergence point coordinates, and calculate the convergence intensity based on the morphological feature matrix. The segregation degree quantification module is used to extract a diamond-shaped segregation candidate region if the convergence intensity exceeds a preset intensity threshold, obtain a diamond-shaped segregation candidate region image, construct a gray-level co-occurrence matrix based on the diamond-shaped segregation candidate region image, extract the feature values ​​of the gray-level co-occurrence matrix, and perform mapping calculation to obtain a segregation severity index. The front distribution reconstruction module is used to perform a concentration distribution scalar field transformation on the segregation severity index according to a preset physical quantity mapping rule to obtain a concentration distribution scalar field, and to characterize the solidification state based on the concentration distribution scalar field to obtain the overall solidification front distribution. The leading edge matching optimization module is used to determine the matching between the leading edge advance velocity and the heat flux distribution based on the overall solidification leading edge distribution. If the leading edge advance velocity and the heat flux distribution do not match, the vector distribution field is adjusted to obtain an optimized intersection point. The rhombus transformation quantitative assessment module is used to extract interference data based on the optimized intersection point, and to perform rhombus transformation quantitative assessment based on the interference data to obtain the final rhombus transformation quantitative assessment result.

Images