Visual representation method of distribution characteristics for judging manufacturing quality of tungsten carbide powder

Abstract

The application discloses a distribution feature visual representation method for judging the quality of tungsten carbide powder, and particularly relates to the technical field of tungsten carbide powder quality detection. A multi-directional illumination image of tungsten carbide powder after vibration spreading is acquired to determine a particle group region. A particle group edge gray transition zone is extracted, and a non-stable dark area formed by particle shielding is removed. A continuous distribution unit and an edge attachment sequence are formed according to the fitting relationship of the stable dark transition zone. A distribution feature map is constructed according to the repeated position of the edge attachment sequence in different particle group regions, and is matched with a qualified powder reference distribution feature map. The application can identify the local aggregation and diffusion features of segregation powder along the edge of the particle group, and reduce the influence of shielding, dark spots and spreading fluctuations on the judgment result.

Description

Technical Field

[0001] This invention relates to the field of tungsten carbide powder quality inspection technology, and specifically to a visual characterization method for the distribution characteristics of tungsten carbide powder manufacturing quality assessment. Background Technology

[0002] Before tungsten carbide powder enters the compaction stage after being transported in a closed gas flow system, a small amount of free carbon segregated powder adheres to the edges of normal particle clusters, forming small and sparsely distributed dark transition zones. This anomaly typically does not significantly alter the average particle size, bulk density, or overall grayscale, but it can easily cause localized porosity expansion during subsequent sintering, leading to scattered edge chipping in thin-walled cemented carbide components. Existing visual inspection methods mostly rely on particle area, average color, or the number of agglomerates for discrimination, making it difficult to distinguish between normal occlusion shadows and the continuous transition distribution formed by segregated powder, especially prone to misjudgment when there are slight changes in powder thickness. Therefore, a quality discrimination method capable of characterizing the spatial distribution relationship of these small dark regions is needed. Summary of the Invention

[0003] The purpose of this invention is to provide a visual characterization method for the distribution characteristics of tungsten carbide powder manufacturing quality assessment, in order to address the shortcomings of the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a visual characterization method for the distribution characteristics of tungsten carbide powder manufacturing quality assessment, comprising: Acquire multi-directional illumination images of the tungsten carbide powder surface after vibration spreading, and determine the particle cluster region based on the consistency of brightness changes between images in each direction; A grayscale transition band continuously distributed from the inside to the outside is extracted around the particle cluster region. Based on the retention degree of the grayscale transition band under different lighting directions, unstable dark areas formed by particle occlusion are removed to obtain a stable dark transition region. Based on the adhesion relationship between the stable dark transition zone and the edge of the particle cluster, the stable dark transition zone is divided into several continuous distribution units, and an edge attachment sequence is generated according to the interval variation between adjacent continuous distribution units. Based on the repeated occurrence positions of the edge attachment sequence in different particle cluster regions, a distribution feature map is constructed to characterize the local aggregation state of segregated powder. The distribution feature map is matched with the reference distribution feature map corresponding to qualified tungsten carbide powder, and the manufacturing quality of the tungsten carbide powder to be judged is determined based on the matching difference.

[0005] Preferably, defining the particle cluster region includes: Multiple illumination units are activated at circumferential intervals along the surface of tungsten carbide powder, and overlapping illumination is retained during the switching of adjacent illumination units. Unidirectional illumination images and overlapping illumination images are obtained respectively. Dark boundary pixels and bright boundary pixels are searched along the illumination direction in the overlapping illumination image. Dark boundary pixels and bright boundary pixels with spacing conforming to the scale of particle protrusion edges are connected into local migration line segments. Based on the proximity of the midpoints and the continuity of the directions of the local migration lines in adjacent overlapping illumination images, the local migration lines are connected into a continuous path that returns around the same local center, and the area enclosed by the continuous path is determined as the candidate particle cluster region.

[0006] Preferably, determining the particle cluster region further includes: Sampling points are selected along a continuous path in the candidate particle cluster region. The brightness rise and fall order of the outer, edge and inner positions of the sampling points is read in each unidirectional illumination image, and the forward order, reverse order and invalid order that do not form a clear rise and fall relationship are marked respectively. Local areas that can form a corresponding reversal of the forward and reverse order under opposite lighting directions and maintain the continuous transmission of edge positions under adjacent lighting directions are marked as stable return regions. By merging stable return regions that are adjacent to each other along a continuous path and have the same tangential direction, and removing discrete dark spots that are only retained under a single illumination direction or whose area changes too much under adjacent illumination directions, the particle cluster region is obtained.

[0007] Preferably, the extraction of the grayscale transition band includes: Edge positions are selected at intervals along the continuous path of the particle cluster region in order of arc length, and radial sampling bands are established along the outward normal direction pointing to the outside of the particle cluster region from each edge position, so that the radial sampling bands extend from the inside of the particle cluster region to the outside of the particle cluster region. Each radial sampling band is divided into multiple radial sampling layers in order from the inside out, and the regions in which the brightness of multiple consecutive radial sampling layers decreases as a dark segment are defined. Based on the projection overlap relationship and the proximity relationship of the outer edge positions between dark segments in adjacent radial sampling bands, multiple dark segments are sequentially connected to form a gray-scale transition band distributed around the particle cluster region.

[0008] Preferably, a stable dark transition region is obtained by: The grayscale transition band is mapped to the unidirectional lighting image corresponding to the opposite lighting direction. Under the condition that the position of the continuous path remains unchanged, the grayscale transition band that is still attached to the same continuous path under the opposite lighting direction and whose outer edge position is restricted is retained to obtain the initial dark transition area. The movement sequence of the initial dark transition area under adjacent lighting directions is traced along the arc length direction of the continuous path. The initial dark transition area that continues to migrate towards the backlight side and detaches from the initial attachment position as the lighting direction changes is marked as the particle occlusion dark area. Remove the dark areas obscured by particles from the initial dark transition region, and merge the remaining areas according to the overlapping relationship on the continuous path to obtain a stable dark transition region.

[0009] Preferably, dividing the continuous distribution units includes: The continuous path of the particle cluster region is unfolded according to the arc length coordinate, and sampling positions are set at intervals along the continuous path. At each sampling position, the distance between the inner edge of the stable dark transition region and the continuous path, the radial width of the stable dark transition region, and the folding direction of the outer edge of the stable dark transition region are read along the outward normal direction. Adjacent sampling positions with continuous fit, consistent outer edge folding direction, and continuous radial width variation are assigned to the same initial distribution segment. For gaps between adjacent initial distribution segments that are not covered by a stable dark transition area, the radial width variation trend on both sides of the gap is traced back along the continuous path. When the radial width variation trend can continue to each other and the gap does not undergo a sharp turn, the adjacent initial distribution segments are merged into a continuous distribution unit.

[0010] Preferably, generating the edge attachment sequence includes: Record the bonding length of each continuous distribution unit and the edge spacing between adjacent continuous distribution units according to the clockwise arrangement order of each continuous distribution unit on the continuous path. Based on the change of the current edge spacing relative to the previous edge spacing, the edge spacing is sequentially encoded as narrowing, expanding or maintaining, and based on the proportion of the bonding length of each continuous distribution unit to the total arc length of the continuous path, the continuous distribution units are sequentially divided into short bonding level, medium bonding level or long bonding level. The encoding results of the bonding length level and the corresponding edge interval are recorded alternately in the order of arrangement to obtain the edge attachment sequence that represents the attachment rhythm of the stable dark transition zone along the edge of the particle cluster.

[0011] Preferably, constructing a segregation transmission chain includes: The edge attachment sequences corresponding to different particle cluster regions are cyclically expanded. Starting from any continuous distribution unit, the bonding length level and edge spacing coding results are compared segment by segment. The sequence segments that continuously appear with the same bonding length level and whose edge spacing coding change direction can be maintained are marked as repeated attachment segments. Map the midpoint of the continuous distribution unit corresponding to the starting position of the repeated attachment segment to the powder spreading surface to obtain the repeated attachment position, and search for other adjacent repeated attachment positions along the reverse conveying side starting from any repeated attachment position. A segregation transfer chain is formed by sequentially connecting multiple repeated attachment positions that are within the range of the equivalent diameter of the particle cluster region, whose angles deviate from the reverse conveying direction are kept within the range of oblique attachment, and whose angles change continuously between adjacent connections.

[0012] Preferably, constructing the distribution feature map includes: Taking any repeated attachment position in the segregation transfer chain as the center, count the number of repeated attachment positions belonging to the same segregation transfer chain in the neighborhood range associated with the equivalent diameter of the particle cluster region, and read the blank range on both sides where no repeated attachment segments appear along the normal direction of the segregation transfer chain. Based on the intrasample arrangement of the number of repeated attachment sites and the blank areas on both sides, the repeated attachment sites are divided into local aggregation core segments and peripheral diffusion segments. The powder spreading surface is divided into grids according to the equivalent diameter of the particle cluster region. In each grid, the weighted distribution intensity corresponding to the local aggregation core segment, the distribution intensity corresponding to the peripheral diffusion segment, the average offset direction of the effective step connection, and the normalized blank range are recorded to obtain the distribution feature map.

[0013] Preferably, determining the manufacturing quality of the tungsten carbide powder to be judged includes: Using the local agglomeration core segment as the matching starting point, the local agglomeration core segment in the distribution feature map corresponding to the tungsten carbide powder to be identified is translated step by step to the candidate position in the reference distribution feature map, and the corresponding peripheral diffusion segment is rotated synchronously around the translated local agglomeration core segment to keep the step-like offset direction of the segregation transmission chain aligned. The blank areas on both sides are searched in reverse along the aligned segregation transmission chain. The areas where the local aggregation core segments are close in position and the blank areas on both sides shrink relative to the reference distribution feature map are marked as restricted aggregation areas. The order in which the peripheral diffusion segments extend from the local aggregation core segments to both ends is preserved. By comparing the restricted aggregation area, the extension order of the peripheral diffusion segment, and the number of particle clusters spanned by the segregation transfer chain with the corresponding stratification markers in the reference distribution feature map, it is determined that the tungsten carbide powder to be judged has segregation adhesion anomalies when local aggregation differences and chain transfer differences occur continuously in the same area and the extension order of the peripheral diffusion segment remains.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention differs from existing technologies that identify powder anomalies solely based on dark area, average grayscale, or single-illumination images. It employs a combination of multi-directional illumination and short-term overlapping acquisition of adjacent illumination, enabling particle cluster edges to form traceable dark-to-bright boundary migration paths as illumination direction changes. Furthermore, by utilizing the reversal of brightness increase / decrease sequences under opposite illumination directions, interference from shallow scratches, random dark spots, and uneven local reflections is eliminated. Thus, even with slight fluctuations in powder thickness and local particle overlap, the true particle cluster region can be stably obtained, providing a reliable boundary for subsequent segregation and adhesion identification.

[0015] This invention further differs from existing technologies that treat abnormal dark areas as isolated patches. It extracts a continuously decreasing grayscale transition band along the edge of the particle cluster, and eliminates particle-occupied dark areas by using opposite illumination retention relationships and backlight migration sequences. Subsequently, the remaining stable dark transition areas are converted into continuous distribution units, edge attachment sequences, and segregation transfer chains across particle cluster regions. This invention can preserve the oblique, discontinuous, but recurring spatial attachment relationship of free carbon segregated powder along the edge of the particle cluster, distinguishing between locally aggregated core segments and peripheral diffusion segments, thereby reducing the misleading influence of occlusion, vibrational spreading bands, and occasional dark residues on manufacturing quality assessment. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the visual characterization method for distribution characteristics used in the quality judgment of tungsten carbide powder manufacturing according to the present invention.

[0018] Figure 2 This is a flowchart of the method for determining the manufacturing quality of tungsten carbide powder to be judged according to the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0020] Example 1, please refer to Figure 1As shown in this embodiment, the visual characterization method for distribution characteristics used in tungsten carbide powder manufacturing quality assessment includes: Multi-directional illumination images of the surface of tungsten carbide powder after vibration spreading were acquired, and the particle cluster regions were determined based on the consistency of brightness changes between the images in each direction.

[0021] In this embodiment, after the tungsten carbide powder has completed vibratory spreading, the planar stage carrying the powder is kept stationary. The powder spreading thickness is preferably 0.4 mm to 0.8 mm, allowing single-layer particles, locally stacked particles, and dark adhesion areas formed by free carbon segregation to simultaneously enter the imaging range. An area array camera is positioned above the stage, with the angle between the camera's optical axis and the normal to the powder surface not exceeding 3 degrees. The camera's focal plane coincides with the powder surface, and the actual length corresponding to a single pixel is preferably 2 micrometers to 6 micrometers.

[0022] Eight illumination units are spaced circumferentially along the powder surface, with an azimuth angle of 45 degrees between adjacent units. Each illumination unit uses a strip-shaped low-angle incident light source, with the angle between the incident light and the powder surface preferably between 18 and 25 degrees. The low-angle incident light causes the protruding parts of the particle clusters to form a dark boundary on the shaded side and a bright boundary on the illuminated side. The illumination units are activated sequentially, and the camera acquires one unidirectional illumination image after each illumination state stabilizes. During the switching between adjacent illumination units, a short period of overlapping illumination is maintained, preferably 15% to 25% of the duration of the unidirectional illumination. The camera acquires one overlapping illumination image at the midpoint of the overlapping illumination duration. Thus, each round of acquisition yields eight unidirectional illumination images and eight overlapping illumination images.

[0023] Let the grayscale value of the k-th unidirectional illumination image at pixel position (x, y) be Ik(x, y), and the grayscale value of the overlapping illumination image obtained when the k-th illumination unit and the (k+1)-th illumination unit are both activated be Jk(x, y). Here, k ranges from 1 to 8; when k is 8, k+1 corresponds to the first illumination unit. To reduce the impact of slight brightness fluctuations on the powder surface, local normalization is first performed on each image. A square neighborhood with a side length of 21 pixels is selected centered at pixel position (x, y). The mean grayscale value Mk(x, y) and the standard deviation Sk(x, y) of the neighborhood grayscale are calculated. The normalized grayscale value is obtained according to the following formula: Gk(x, y) = [Ik(x, y) - Mk(x, y)] ÷ [Sk(x, y) + 1].

[0024] The normalized gray values ​​Hk(x, y) for the overlapping illumination images are obtained in the same way. The denominator is increased by 1 to avoid numerical instability when the standard deviation of the neighborhood gray values ​​is close to 0.

[0025] For each overlapping illumination image, grayscale changes are searched along the illumination direction and in the opposite direction. Using the illumination direction unit vector as dk, the directional difference is calculated at pixel position p: Dk(p) = Hk(p + 2dk) - Hk(p - 2dk). When Dk(p) is less than the negative decision threshold, p is recorded as a dark boundary pixel; when Dk(p) is greater than the positive decision threshold, p is recorded as a bright boundary pixel. The absolute values ​​of the positive and negative decision thresholds are preferably the 82nd percentile of the absolute values ​​of all directional differences in the current overlapping illumination image.

[0026] It should be noted that using the 82nd percentile can automatically adjust the boundary screening scale according to changes in powder batch, light source attenuation, and spreading thickness, avoiding the omission of dark boundaries or excessive expansion of bright boundaries caused by using only a fixed grayscale difference.

[0027] A search area with a radius of 12 pixels is established around each dark boundary pixel, and the nearest bright boundary pixel is found along the illumination direction. If the distance between the dark boundary pixel and the bright boundary pixel is between 3 and 12 pixels, they are connected as a local migration line segment. For local migration lines in adjacent overlapping illumination images, the distance between the midpoints of the line segments and the angle between the line segments are calculated. If the midpoint distance is no greater than 8 pixels and the angle between the line segments is no greater than 35 degrees, the two local migration lines are connected end to end. After processing 8 overlapping illumination images in sequence, a continuous path for the migration from the dark boundary to the bright boundary is obtained.

[0028] Whether a continuous path returns around the same local center is determined by the ratio of the distance between the path endpoints to the cumulative path length. Let the continuous path contain m local migration segments, with the starting point of the first local migration segment being P1, the ending point of the mth local migration segment being Pm, and the cumulative path length being L. The return ratio is calculated using the following formula: .

[0029] When R is not greater than 0.28 and the continuous path covers no less than 5 different lighting directions, it is determined that the continuous path returns around the same local center.

[0030] The value of 0.28 comes from the powder spreading test: normal single particle edges or linear scratches usually only show migration segments in 2 to 4 illumination directions, with a return ratio greater than 0.40; particle clusters formed by multiple particles stacked on top of each other can form closed or nearly closed paths in at least 5 illumination directions, with a return ratio concentrated between 0.10 and 0.26. The area enclosed by continuous paths that meet the return conditions is recorded as the candidate particle cluster area.

[0031] Within the candidate particle cluster region, the edge brightness increase / decrease order in the unidirectional illumination image is further extracted. For any sampling point q on a continuous path, along the normal direction passing through q and pointing towards the local center, the normalized gray values ​​corresponding to the four pixels outside the path, the path position, and the four pixels inside the path are selected sequentially and denoted as outer gray value, edge gray value, and inner gray value, respectively. When the illumination direction is towards sampling point q, the illuminated side usually exhibits an order of increasing gray value from the outer side to the edge and then decreasing gray value towards the inside; when the illumination direction is reversed, the same position usually exhibits an order of decreasing gray value from the outer side to the edge and then increasing gray value towards the inside. The aforementioned two types of orders are encoded as forward order and reverse order, respectively.

[0032] Let Ck(q) be the order code of sampling point q in the k-th unidirectional illumination image. A forward order is denoted as 1, a reverse order as -1, and 0 when no clear ascending / descending relationship exists. The opposite illumination direction corresponds to the (k+4)-th unidirectional illumination image. For each sampling point, calculate the reversal score: F(q) satisfies Ck(q) × Ck+4(q), which is the number of -1 opposite illumination direction pairs ÷ 4.

[0033] Next, calculate the continuous transmission score: T(q) equals the number of direction pairs in adjacent lighting directions that satisfy the order code not being 0 and the edge position movement distance not exceeding 5 pixels ÷ 8.

[0034] When F(q) is not less than 0.75 and T(q) is not less than 0.625, the local region corresponding to the sampling point q is marked as the stable return region.

[0035] A value of F(q) not less than 0.75 indicates that at least three sets of opposite illumination directions can form a corresponding reversal; a value of T(q) not less than 0.625 indicates that at least five sets of adjacent illumination directions can maintain continuous transmission. This combination can eliminate random dark spots on the powder surface. Random dark spots are usually caused by tiny depressions, dust, or uneven local reflection. Their positions lack a continuous movement relationship when the illumination direction changes, and it is difficult to form a stable positive and negative sequence correspondence.

[0036] Adjacent stable return regions are merged. Adjacency is determined by the shortest distance between region edges; regions with a shortest edge distance of no more than 6 pixels are considered mergingable. The merging direction is constrained by continuous paths: merging is performed only when the tangential angle between the continuous paths corresponding to two stable return regions is no greater than 30 degrees. This process avoids incorrectly connecting adjacent but different particle clusters of local protrusions.

[0037] After merging, the retention of each merged region in the eight unidirectional illumination images is checked. If a merged region appears only in one illumination direction, or if the corresponding area change in adjacent illumination directions exceeds 65%, the merged region is identified as a discrete dark spot and removed. The area change is calculated by dividing the absolute value of the area difference between adjacent unidirectional illumination images by the larger area value. The retained merged regions are designated as particle cluster regions.

[0038] In a set of tungsten carbide powder samples with particle sizes ranging from 0.6 μm to 1.2 μm, when directly identifying dark areas using a single illumination image, shallow scratches and local occlusions on the powder surface are easily identified as particle clusters. However, using the aforementioned method, the migration lines corresponding to the shallow scratches cannot return in at least five illumination directions, and the dark spots formed by local occlusions cannot simultaneously meet the requirements for inversion score and continuous transfer score. Therefore, discrete dark spots can be removed while preserving the true particle cluster edges. The obtained particle cluster regions can be further used to extract continuously distributed grayscale transition bands from the inside out, providing a boundary basis for subsequent stable dark transition region identification.

[0039] A grayscale transition band continuously distributed from the inside to the outside is extracted around the particle cluster region. Based on the degree of retention of the grayscale transition band under different lighting directions, unstable dark areas formed by particle occlusion are removed to obtain a stable dark transition region.

[0040] After the particle cluster region is determined, grayscale transition bands are extracted based on the continuous paths corresponding to the particle cluster region. The continuous paths are formed by connecting the local migration line segments that migrate from the aforementioned dark boundary to the light boundary, and are distributed around the particle cluster region. To avoid local burrs at the edges of the particle cluster region affecting the radial grayscale change analysis, equidistant sampling is first performed according to the arc length order of the continuous paths. The arc length interval between adjacent edge positions on the continuous path is preferably 4 pixels. When the arc length interval is less than 3 pixels, adjacent radial sampling bands contain too many duplicate pixels; when the arc length interval is greater than 6 pixels, smaller stable dark transition areas are easily missed.

[0041] Let Qi be the position of the i-th edge on the continuous path. Along the arc length of the continuous path, take path segments of 2 pixels in length before and after Qi, using the direction from the beginning to the end of each path segment as the tangential direction. Denote the outward normal direction as Ni, which is perpendicular to the tangential direction and points from the inside of the particle cluster to the outside of the particle cluster. Establish a radial sampling band centered on Qi along Ni. The width of the radial sampling band in the tangential direction is preferably 5 pixels, and it extends from 3 pixels inside the particle cluster to 18 pixels outside the particle cluster in the outward normal direction. This setup can simultaneously cover the edge of the particle cluster, the segregated powder attached to the edge of the particle cluster, and the normally spread powder outside the segregated powder.

[0042] For the k-th unidirectional illumination image, the radial sampling band is divided into several radial sampling layers in an inward-outward order, with each radial sampling layer having a width of 1 pixel. The representative gray value of the r-th radial sampling layer in the i-th radial sampling band is calculated according to the following formula: Ak,i,r = Sum of all normalized gray values ​​in the radial sampling layer ÷ Number of pixels in the radial sampling layer. Here, r ranges from -3 to 18; a negative r indicates the location is inside a particle cluster region, and a positive r indicates the location is outside a particle cluster region. The normalized gray values ​​follow the local normalization processing results of the aforementioned unidirectional illumination image.

[0043] To identify continuously distributed dark segments from the inside out, the grayscale decrease between adjacent radial sampling layers is calculated: Bk,i,r = Ak,i,r⁻¹ - Ak,i,r; when Bk,i,r is positive, it indicates that the grayscale decreases from the inside out. For any set of continuous radial sampling layers, the number of radial sampling layers with positive grayscale decreases is counted, and the decrease preservation ratio is calculated: Ck,i = number of radial sampling layers with positive grayscale decreases ÷ total number of continuous radial sampling layers.

[0044] When the number of consecutive radial sampling layers is between 5 and 14, the decrease retention ratio is not less than 0.70, and the decrease in grayscale value of the outermost segment relative to the grayscale value of the innermost segment is not less than 0.35, the region corresponding to this group of consecutive radial sampling layers is recorded as a dark segment. The value of 0.35 is expressed using normalized grayscale values, which can reduce the impact of changes in illumination intensity. The decrease retention ratio is preferably 0.70 because when free carbon segregated powder adheres to the edge of the particle cluster, the dark attachment area is not completely uniform, allowing for a slight grayscale rebound in a small number of radial sampling layers; the dark areas formed by the highlights on the particle surface and random noise are usually difficult to maintain a continuous decrease relationship in more than 5 radial sampling layers.

[0045] Dark segments in adjacent radial sampling bands are connected by an overlapping relationship. Let Ei be the outer edge position of the dark segment in the i-th radial sampling band, and Ei+1 be the outer edge position of the dark segment in the (i+1)-th radial sampling band. Project the two dark segments onto the outer region of the continuous path, and calculate the ratio of their overlapping pixel area to the area of ​​the smaller dark segment: D = overlapping pixel area ÷ smaller dark segment area.

[0046] When D is not less than 0.40 and the distance between Ei and Ei+1 is not greater than 4 pixels, the two dark segments are concatenated. The concatenation operation is performed sequentially along a continuous path to obtain a grayscale transition band distributed around the grain cluster region. The grayscale transition band must contain at least 4 adjacent radial sampling bands. This requirement is used to exclude random grayscale fluctuations that occupy only a small number of edge positions.

[0047] After the grayscale transition bands are formed, each grayscale transition band is mapped to a unidirectional illumination image corresponding to the opposite illumination direction. The k-th unidirectional illumination image and the (k+4)-th unidirectional illumination image are in opposite illumination directions. During the mapping process, the continuous path position is preserved, and the dark segment is recalculated within the corresponding radial sampling band. Let Uk be the number of radial sampling bands covered by the grayscale transition band in the k-th unidirectional illumination image, and Vk be the number of radial sampling bands in the (k+4)-th unidirectional illumination image where the dark segment is re-obtained on the same continuous path. The preservation ratio is calculated according to the following formula: Pk = Vk ÷ Uk. At the same time, the swing distance of the outer edge position under the opposite illumination direction is calculated. For each radial sampling band covered by the grayscale transition band, the outer edge position in the k-th and (k+4)-th unidirectional illumination images is recorded respectively, the distance between the two outer edge positions is calculated, and then the average distance of all radial sampling bands is taken to obtain the average swing distance Wk.

[0048] When at least three sets of opposite lighting directions meet the retention ratio of not less than 0.65 and the average swing distance is not greater than 3 pixels, the corresponding grayscale transition zone is retained as the initial dark transition area.

[0049] A retention ratio of 0.65 allows for minor gaps at the edges of segregated powder due to localized particle reflection; an average swing distance of 3 pixels is used to limit the range of positional variation at the outer edge of the grayscale transition zone. Dark areas formed by particle occlusion typically move to the other side of the particle cluster region as the illumination direction reverses, making it difficult to maintain a high retention ratio on the same continuous path.

[0050] For the initial dark transition area, it is also necessary to track its movement sequence along the continuous path under adjacent lighting directions. The continuous path is unfolded clockwise into arc-length coordinates. For the initial dark transition area in the k-th unidirectional lighting image, the average value of the centers of all dark segments on the arc-length coordinates is calculated and denoted as Lk. The movement between adjacent lighting directions is calculated using the following formula: Mk = Lk + 1 - Lk. When crossing the starting point of the continuous path, a circular correction is performed on the movement to ensure that the absolute value of the movement does not exceed half of the total arc length of the continuous path. Then, based on the orientation and local tangential direction of the k-th lighting unit, the movement direction of the backlight side along the continuous path is determined. If Mk is in the same direction as the movement of the backlight side, it is recorded as one backlight migration; if Mk is in the opposite direction to the movement of the backlight side, it is recorded as one non-backlight migration.

[0051] In eight unidirectional illumination images, an initial dark transition area is marked as a particle-occupied dark area when it exhibits at least five consecutive backlight migrations, and the distance between the outer edge position after at least two migrations and the initial attachment position is greater than six pixels. Five consecutive backlight migrations indicate that the dark area continuously moves with the illumination direction; an outer edge distance greater than six pixels indicates that the dark area has detached from its original attachment position. Segregated powder stably attached to the edge of the particle cluster region does not continuously migrate towards the backlight side with the illumination direction, and therefore can be distinguished from the particle-occupied dark area.

[0052] After marking the dark areas obscured by grains, remove these areas from the initial dark transition region. For the remaining regions, merge them according to their overlapping relationships along the continuous path. Merging is performed when the interval between adjacent remaining regions along the continuous path is no more than 3 pixels, and the ratio of the overlapping pixel area of ​​the two regions in adjacent radial sampling bands to the area of ​​the smaller remaining region is no less than 0.35. After merging, a stable dark transition region is obtained.

[0053] In the spread samples of tungsten carbide powder after closed-loop gas flow transport, two types of dark regions often exist simultaneously at the edges of particle clusters. The first type of dark region is formed by particle stacking and obstruction; after changing the illumination direction, it migrates along the edge of the particle cluster towards the shaded side. The second type of dark region is formed by the adhesion of a small amount of free carbon segregated powder; after changing the illumination direction, it remains near the same continuous path, only exhibiting a small range of outer edge oscillation. After the aforementioned processing, the first type of dark region is removed due to continuous backlight migration, while the second type of dark region is retained due to a high retention rate under multiple sets of opposite illumination directions. The stable dark transition region can serve as input for subsequent continuous distribution unit division and edge attachment sequence construction, ensuring that slight changes in powder spread thickness, local particle obstruction, and changes in illumination direction no longer directly interfere with the quality judgment of segregated powder.

[0054] Based on the adhesion relationship between the stable dark transition zone and the edge of the particle cluster, the stable dark transition zone is divided into several continuous distribution units, and an edge attachment sequence is generated according to the interval variation between adjacent continuous distribution units.

[0055] After obtaining the stable dark transition region, arc length coordinates are established along the continuous path corresponding to the particle cluster region. The starting point of the continuous path is selected as the path pixel with the smallest x-coordinate and y-coordinate, and the path pixels are recorded sequentially in a clockwise direction. The total arc length of the continuous path is denoted as L, and the arc length coordinate corresponding to any edge position on the path is denoted as s. For closed continuous paths, s extends from 0 to L; for approximately closed continuous paths, the portion with a distance of no more than 6 pixels between the beginning and end is filled with a straight line before the unfolding operation is performed. The filled portion is only used to determine the arrangement order and does not participate in the area calculation of the stable dark transition region.

[0056] Sampling positions are established along the continuous path at 2-pixel arc length intervals. For each sampling position, a stable dark transition region is searched along the outward normal direction. The pixel with the smallest distance between the continuous path and the inner edge of the stable dark transition region is designated as the inner edge corresponding point, and the pixel with the largest distance from the continuous path along the same outward normal direction is designated as the outer edge corresponding point. The distance between the inner edge corresponding point and the continuous path is denoted as 'a', and the distance between the outer edge corresponding point and the inner edge corresponding point is denoted as 'b'. Distance 'a' represents the degree of fit, and distance 'b' represents the radial width.

[0057] When 'a' is no more than 2 pixels, the current sampling position is recorded as the matching position; when 'a' is greater than 2 pixels but no more than 4 pixels, and the two sampling positions before and after the current sampling position are both matching positions, the current sampling position is supplemented and recorded as the matching position. This supplementary recording method can preserve the short-distance inner edge offset caused by slight indentations at the particle cluster edges, while avoiding misidentifying discrete dark areas far from the particle cluster edges as matching areas.

[0058] For consecutively arranged contact positions, calculate the folding direction of the outer edge of the stable dark transition zone. Let Ei be the outer edge corresponding point of the i-th sampling position, Ei-1 be the outer edge corresponding point of the (i-1)-th sampling position, and Ei+1 be the outer edge corresponding point of the (i+1)-th sampling position. The vector pointing from Ei-1 to Ei is denoted as u, and the vector pointing from Ei to Ei+1 is denoted as v. The outer edge turning amount is calculated according to the following formula: Qi = (lateral component of u) / (vertical component of v) - (vertical component of u) / (lateral component of v).

[0059] When Qi is greater than 0, the i-th sampling position is recorded as an outward foldback; When Qi is less than 0, the i-th sampling position is recorded as an inward foldback; When Qi equals 0, the foldback direction of the previous sampling position is used.

[0060] To reduce directional jitter caused by fluctuations in a single pixel, the Qi values ​​at any three consecutive sampling positions are summed, and the sign of the summation result determines the foldback direction of the intermediate sampling position.

[0061] Simultaneously, the radial width change is calculated. Let the radial width corresponding to the i-th sampling position be bi, then the radial width change between adjacent sampling positions is calculated according to the following formula: Gi = bi + 1 - bi.

[0062] The local width variation is obtained by averaging the radial width changes at five consecutive sampling locations. A local width variation greater than one pixel is considered gradual widening; a local width variation less than -1 pixel is considered gradual narrowing; and a local width variation between -1 and 1 pixel is considered width maintenance. One pixel corresponds to the allowable surface roughness undulations between adjacent sampling locations. This value is preferably determined based on the 75th percentile of the absolute value of the radial width variation in a qualified tungsten carbide powder sample; when the 75th percentile is less than one pixel, one pixel is still used to avoid over-segmentation caused by imaging noise.

[0063] Check the mating positions sequentially along the arc length coordinate. When adjacent sampling positions are both mating positions, the outer edge folding direction is consistent, and the local width change state is consistent, or only switches between width maintenance and gradual widening, or only switches between width maintenance and gradual narrowing, the adjacent sampling positions are assigned to the same initial distribution segment.

[0064] When the outer edge folding direction changes, or the local width change state changes directly from gradually widening to gradually narrowing, or from gradually narrowing to gradually widening, the current initial distribution segment ends, and a new initial distribution segment is established from the next sampling position.

[0065] The initial distribution segment must cover at least four consecutive sampling locations, with an arc length of no less than eight pixels along the corresponding continuous path. Regions covering fewer than four consecutive sampling locations are prone to formation due to localized reflections from the powder surface, edge burrs, or single particle attachments, and will not participate in the subsequent edge attachment sequence construction.

[0066] After the initial distribution segments are formed, reverse tracing is performed on the gaps between adjacent initial distribution segments that are not covered by the stable dark transition region. Let the arc length coordinate corresponding to the end of the previous initial distribution segment be s1, and the arc length coordinate corresponding to the beginning of the next initial distribution segment be s2. The gap length is calculated according to the following formula: E = s2 - s1.

[0067] For adjacent initial distribution segments that cross the starting point of a continuous path, the gap length is calculated according to the following formula: E = L - s1 + s2.

[0068] When the gap length is no more than 10 pixels, the radial widths of the three sampling positions before the end of the previous initial distribution segment and the radial widths of the three sampling positions after the beginning of the next initial distribution segment are extracted. The radial width variation trend on the front side is calculated by subtracting the average radial width of the first three sampling positions from the radial width at the end, and the radial width variation trend on the back side is calculated by subtracting the radial width of the beginning from the average radial width of the last three sampling positions. If the radial width variation trends on both sides are both positive, both are both negative, or both absolute values ​​are no more than 1 pixel, it is considered that the radial width variations on both sides of the gap position can continue.

[0069] Further calculate the curvature of the continuous path corresponding to the gap location. Let H be the angle between the tangential directions at the beginning and end of the gap location. When H is not greater than 25 degrees, the gap length is not greater than 10 pixels, and the radial width changes on both sides of the gap location can continue, adjacent initial distribution segments are merged. The value of 25 degrees is used to limit the path direction change at the gap location and avoid erroneous merging across the sharp turning points of the particle cluster edges. The value of 10 pixels is preferably determined according to the 90th percentile of the width of the local edge depression in normally spread tungsten carbide powder; under the imaging conditions of a particle size of 0.6 micrometers to 1.2 micrometers and a single pixel corresponding to 4 micrometers, the 90th percentile is 9 pixels, so 10 pixels is taken as the maximum allowable gap length.

[0070] After reverse tracing and merging, continuous distribution units are obtained. Each continuous distribution unit consists of one or more initial distribution segments and corresponds to a continuous arc length range on the continuous path. For the j-th continuous distribution unit, the coordinates of the first arc length are denoted as Sj, and the coordinates of the last arc length are denoted as Tj. The bonding length is calculated according to the following formula: Kj = Tj - Sj. For continuous distribution units that cross the starting point of the continuous path, the bonding length is calculated according to the following formula: Kj = L - Sj + Tj.

[0071] The continuous distribution units are arranged in a clockwise direction. Let the coordinate of the arc length of the end of the j-th continuous distribution unit be Tj, and the coordinate of the arc length of the beginning of the (j+1)-th continuous distribution unit be Sj+1. Then, the edge interval between adjacent continuous distribution units is calculated as follows: Ij = Sj+1 - Tj. The edge interval between the end and the beginning of the arrangement is calculated as follows: In = L - Tn + S1. Where n is the number of continuous distribution units.

[0072] To characterize the rhythm of edge interval changes, adjacent edge intervals are compared. The change ratio of the j-th edge interval relative to the previous edge interval is calculated using the following formula: Vj = (Ij - Ij - 1) ÷ Ij - 1. When Vj is less than -0.20, the j-th edge interval is encoded as "narrowing"; when Vj is greater than 0.20, the j-th edge interval is encoded as "expanding"; when Vj is between -0.20 and 0.20, the j-th edge interval is encoded as "maintaining". The value of 0.20 is preferably determined by adding one interquartile range to the median of the absolute values ​​of the change ratios of adjacent edge intervals in the same batch of qualified tungsten carbide powder samples; when the calculation result is between 0.16 and 0.24, it is uniformly taken as 0.20 to ensure consistency in coding scale between different batches.

[0073] The bonding length level is determined based on the total arc length of the continuous path. The bonding length percentage of the j-th continuous distribution unit is calculated as follows: Uj = Kj ÷ L. When Uj is less than 0.06, the bonding length level is recorded as "short"; when Uj is not less than 0.06 and less than 0.14, the bonding length level is recorded as "medium"; when Uj is not less than 0.14, the bonding length level is recorded as "long". Using a percentage instead of a fixed pixel length reduces the impact of differences in particle cluster size.

[0074] The edge attachment sequence is constructed according to the arrangement order of consecutive distribution units. First, the adhesion length level of the first consecutive distribution unit is recorded, then the sequential code of the first edge interval relative to the previous edge interval is recorded, and then the adhesion length level and the sequential code of the corresponding edge interval of subsequent consecutive distribution units are recorded alternately. For a particle cluster region containing 4 consecutive distribution units, the edge attachment sequence can be represented as: "short-narrow-medium-maintain-long-expand-medium-narrow".

[0075] This edge attachment sequence does not directly depend on the absolute area of ​​the stable dark transition zone, but rather retains the interval rhythm and adhesion length differences formed when the stable dark transition zone adheres along the edge of the particle cluster. In tungsten carbide powder samples transported in a closed gas flow, normal particle shading residues typically present as small, dispersed, and discontinuously spaced dark regions; when free carbon segregated powder adheres to the edge of the particle cluster, continuous distribution units often exhibit a pattern of continuous narrowing followed by expansion, accompanied by medium or long-level adhesion lengths. By tracing back and merging the initial distribution segments on both sides of the local depression, it is possible to avoid splitting the same segregated attachment region into multiple isolated parts. By alternating the edge interval variations with adhesion length levels, a comparable edge attachment sequence can be provided for subsequently constructing distribution feature maps based on the recurring locations between different particle cluster regions.

[0076] Based on the recurrence of the edge attachment sequence in different particle cluster regions, a distribution feature map is constructed to characterize the local aggregation state of segregated powder.

[0077] After obtaining the edge adhesion sequence corresponding to each particle cluster region, a planar coordinate system is established for the powder spreading surface. The powder conveying direction is taken as the positive transverse direction, and the direction perpendicular to the powder conveying direction is taken as the positive longitudinal direction. The geometric center of each particle cluster region is defined as the center of the region enclosed by a continuous path. The x-coordinate of the geometric center is equal to the sum of the x-coordinates of all pixels within the region divided by the number of pixels within the region, and the y-coordinate of the geometric center is equal to the sum of the y-coordinates of all pixels within the region divided by the number of pixels within the region.

[0078] Edge attachment sequences are arranged alternately by adhesion length levels and sequential codes. Adhesion length levels include "short," "medium," and "long," while sequential codes include "narrowing," "expanding," and "maintaining." Since the continuous paths are closed or nearly closed, edge attachment sequences do not have a fixed starting point. To avoid misidentifying the same attachment rhythm as different sequences due to different starting positions, each edge attachment sequence is cyclically expanded according to the number of consecutive distribution units. During cyclic expansion, the end of the sequence is reconnected to the beginning, ensuring that any consecutive distribution unit can serve as a comparison starting point.

[0079] Any two particle cluster regions are denoted as particle cluster region a and particle cluster region b, respectively. Any consecutive distribution unit in the edge attachment sequence of particle cluster region a is selected as the starting position, and any consecutive distribution unit in the edge attachment sequence of particle cluster region b is selected as the corresponding starting position. Starting from the starting position, at least three consecutive distribution units are sequentially extracted in a clockwise direction, along with their corresponding adhesion length levels and sequential codes between adjacent consecutive distribution units, forming the sequence segment to be compared.

[0080] Suppose the sequence segments to be compared contain m adhesion length levels and m sequential codes. The number of times the a-th particle cluster region and the b-th particle cluster region have identical adhesion at the current starting position is denoted as N1, and the number of times the sequential codes match is denoted as N2. The degree of segment correspondence is calculated according to the following formula: R = (N1 + N2) ÷ (2 × m). Wherein, identical adhesion length levels are counted as 1 instance of identical adhesion; correspondences between "narrowing" and "narrowing," "expanding" and "expanding," or "maintaining" and "maintaining" are counted as 1 instance of sequential code consistency. Correspondences between "maintaining" and "narrowing," or between "maintaining" and "expanding," are not counted as consistency. This processing can preserve the directional changes in edge attachment rhythm and will not confuse slight fluctuations with continuous segregation attachment in the same state.

[0081] When m is not less than 3, R is not less than 0.83, and there are at least two consecutive identical adhesion length grades, the current sequence segment to be compared is marked as a duplicated attachment segment. A value of 0.83 indicates that at least five of the six comparison positions in a segment are consistent. This value is preferably determined jointly based on qualified tungsten carbide powder samples and tungsten carbide powder samples with free carbon segregation: in qualified samples, the correlation degree of segments formed by random dark residues is typically below 0.67; when segregated powder adheres along the edges of multiple particle clusters, the correlation degree is typically between 0.83 and 1.00. Therefore, using 0.83 can reduce misjudgments caused by random overlap.

[0082] For each repetitive attachment segment, its initial continuous distribution unit on the continuous path is recorded. The midpoint arc length coordinate of the initial continuous distribution unit is denoted as s. Based on the continuous path arc length coordinates, the midpoint of the initial continuous distribution unit is mapped back to the powder spreading surface to obtain the repetitive attachment position P. The repetitive attachment position P is not the center of the particle cluster, but rather the starting position where the stable dark transition zone forms a repetitive attachment rhythm at the edge of the particle cluster. Using the repetitive attachment position P can preserve the bias direction of the segregated powder at the edge of the particle cluster, avoiding the loss of spatial features caused by using only the center of the particle cluster.

[0083] After marking the repeated attachment segments, a reverse search order is established based on the spatial positions between particle cluster regions. The unit vector corresponding to the powder conveying direction is denoted as e. For any repeated attachment position Pi, other repeated attachment positions Pj located on the reverse conveying side are searched around Pi. The reverse conveying side is determined by the following formula: the inner product of (Pj - Pi) and e is less than 0.

[0084] A dot product less than 0 indicates that Pj is located on the reverse conveying side of Pi. For multiple repeated attachment positions that meet the reverse conveying side condition, the position with the smallest distance from Pi is selected as the next search position. The distance between Pi and Pj is calculated using the Euclidean distance formula, which will not be elaborated here.

[0085] To limit the search range between adjacent particle clusters, the median of the equivalent diameters of all particle cluster regions is calculated, denoted as D. The equivalent diameter of a particle cluster region is equal to the square root of the region's area divided by pi, multiplied by 2. When d is between 0.8D and 4D, the connection between Pi and Pj is preserved. When the distance is less than 0.8D, two repeated attachment sites usually belong to particle cluster regions that are in contact or almost overlapping, which is not conducive to characterizing cross-particle cluster transfer; when the distance is greater than 4D, the connection easily crosses unrelated particle cluster regions, weakening the ability to identify local aggregation states.

[0086] For the retained connections, calculate the angle between the connection direction and the powder conveying direction. Let this angle be denoted as α, and calculate it using the following formula: ;in, This is the current reattachment location. This is the next repeat attachment location found along the reverse conveying side.

[0087] The connection vector is Pj-Pi. Since the reverse lookup is performed along the reverse conveying side, α is between 90 and 180 degrees. For ease of comparison, the angle deviating from the reverse conveying direction is denoted as β, calculated as: β = 180 degrees - α. When β is between 12 and 38 degrees, the connection between Pi and Pj is considered an effective stepped connection. This angle range is used to preserve the oblique adhesion traces formed by segregated powder along the powder flow edge after closed-air conveying. When the angle is less than 12 degrees, the connection is close to a straight line arrangement in the conveying direction, easily originating from strips formed by vibration spreading; when the angle is greater than 38 degrees, the connection is close to a lateral dispersion, easily originating from random particle cluster distribution.

[0088] Starting from any repeated attachment position, continuously search for effective stepped connections along the reverse conveying side. When at least three repeated attachment positions are obtained consecutively, and the absolute value of the β change of adjacent effective stepped connections is no greater than 10 degrees, connect these repeated attachment positions to form a segregation transfer chain. The value of 10 degrees is preferably determined based on the change in the edge direction of the particle clusters in the same spread sample. The angular change of random connections in a normal spread sample is usually greater than 18 degrees; after segregated powder is affected by the airflow conveying direction, the stepped displacement formed between adjacent particle cluster regions usually has a relatively stable oblique transfer relationship, and the angular change is mostly concentrated between 3 and 9 degrees.

[0089] After the segregation transfer chain is established, the offset order of repeated attachment positions within the chain is checked. Let the same segregation transfer chain sequentially contain P1, P2, and so on, up to Pn. The lateral offset of adjacent repeated attachment positions is denoted as Δx, and the longitudinal offset as Δy, calculated using the following formulas: Δx = xj + 1 - xj; Δy = yj + 1 - yj. Under reverse lookup conditions, Δx should be less than 0. For the same segregation transfer chain, if all adjacent connections have positive Δy values, or if all adjacent connections have negative Δy values, then the repeated attachment positions within the chain are considered to have a unilateral step-like offset. If the sign of Δy changes repeatedly within the chain, the position where the sign changes is used as the splitting point, splitting the original segregation transfer chain into two separate segregation transfer chains. This process avoids incorrectly connecting repeated attachment segments that span different segregation directions.

[0090] Further distinguish between the local aggregation core segment and the peripheral diffusion segment. For any repeating attachment site Pi in the segregation transport chain, establish a circular area with a radius of 3D centered on Pi, and count the number of repeating attachment sites belonging to the same segregation transport chain within the circular area, denoted as ci. The local chain density is calculated according to the following formula: Simultaneously, other repeating attachment sites are searched for on both sides along the normal direction of the segregation transfer chain. The normal direction is perpendicular to the current effective step connection direction. The distances from the current segregation transfer chain to the nearest repeating attachment sites on both sides are denoted as g1 and g2, respectively. The width of the blank range on both sides is calculated according to the following formula: G = g1 + g2. When no other repeating attachment sites appear on one side within a 6D range, the distance on that side is recorded as 6D. Using 6D can limit the search range and avoid the influence of distant irrelevant particle clusters on the local analysis of the current segregation transfer chain.

[0091] For all segregation transport chains in the same powder spread sample, the density (ρi) is ranked, and the 70th percentile is used as the high-density cutoff, and the 40th percentile as the low-density cutoff. For any repeated attachment site, if ρi is not less than the high-density cutoff and G is not greater than 5D, the repeated attachment site is classified as the local aggregation core segment. If ρi is between the low-density cutoff and the high-density cutoff, or if G is greater than 5D but not greater than 9D, the repeated attachment site is classified as the peripheral diffusion segment. Locations with ρi below the low-density cutoff and G greater than 9D are not included in the stratification marking.

[0092] Using in-sample percentiles instead of fixed density values ​​allows for adaptation to variations in powder spreading area, particle count, and magnification. The localized core segment corresponds to a segregation region where repeated attachment sites are relatively concentrated and the surrounding blank areas are narrow; the peripheral diffusion segment corresponds to a diffusion edge where the number of repeated attachment sites decreases and the surrounding blank areas increase. Both are arranged continuously along the segregation transfer chain, reflecting the gradual extension of segregated powder from localized aggregation sites to the edges of surrounding particle clusters.

[0093] A distribution feature map is constructed on the powder-spread surface. The distribution feature map uses the same planar coordinates as the original image and is divided into square grids with side length D. Let the grid in row u and column v be Euv. The number of repeated attachment sites of local aggregation core segments falling into Euv is counted and denoted as Cuv; the number of repeated attachment sites of peripheral diffusion segments falling into Euv is counted and denoted as Fuv. The grid aggregation value is calculated according to the following formula: Zuv = 2 × Cuv + Fuv.

[0094] The local clustering core segment is weighted with 2, while the peripheral diffusion segment is weighted with 1. This setting allows the local clustering core segment to have higher expression intensity in the distribution feature map, while preserving the spatial extension relationship of the peripheral diffusion segment.

[0095] The average offset direction in each grid is further recorded. If the grid Euv contains q effective step connections, and the deviation angle between the r-th effective step connection and the reverse transport direction is βr, then the average offset angle of the grid is calculated according to the following formula: Auv = (β1 + β2 + ... + βq) ÷ q. When q equals 0, Auv is recorded as 0. The distribution characteristic map is represented by two layers of data. The first layer records Zuv, which is used to characterize the distribution intensity of the local aggregation core segment and the peripheral diffusion segment; the second layer records Auv, which is used to characterize the step-like offset direction of the segregation transmission chain. To retain the blank range on both sides, a third layer of data is established. In the third layer, G÷D is recorded in the grid corresponding to the local aggregation core segment, and G÷D is also recorded in the grid corresponding to the peripheral diffusion segment. Grids without segregation transmission chains are recorded as 0.

[0096] In tungsten carbide powder samples transported in a closed gas flow, a small amount of free carbon segregated powder tends to first adhere to the edge of the particle cluster downstream of the transport direction. Then, influenced by subsequent vibration and spreading, it forms oblique, intermittent, but rhythmically repeating attachment traces between adjacent particle cluster regions. When observing a single particle cluster region, this trace only appears as part of a stable, dark transition zone, making it difficult to distinguish from occasional attached powder. After the aforementioned processing, the recurring edge attachment sequence fragments in multiple particle cluster regions can be connected to form a segregation transfer chain. The step-like offset direction, the number of repeated attachment positions within the chain, and the blank areas on both sides can be used to distinguish between the local aggregation core segment and the peripheral diffusion segment.

[0097] In the same powder spreading sample, when a local agglomeration core segment continuously spans at least three grids, and the peripheral diffusion segments are distributed along the same segregation transfer chain at both ends of the local agglomeration core segment, the distribution feature map can exhibit a spatial structure extending from high agglomeration value regions to low agglomeration value regions. This spatial structure is distinct from isolated grids formed by random dark residues, and also from the approximately linear distribution formed by vibratory spreading strips. Therefore, the distribution feature map can serve as input for subsequent matching with the reference distribution feature map of qualified tungsten carbide powder, providing a visual representation with spatial transfer relationships for manufacturing quality assessment.

[0098] Please see Figure 2 As shown, the distribution feature map is matched with the reference distribution feature map corresponding to the qualified tungsten carbide powder, and the manufacturing quality of the tungsten carbide powder to be judged is determined based on the matching difference.

[0099] After constructing the distribution feature map, the distribution feature map corresponding to the tungsten carbide powder to be identified is matched with the reference distribution feature map corresponding to the qualified tungsten carbide powder. The distribution feature map continues to use the powder delivery direction as the positive lateral direction, and records the local aggregation core segment, peripheral diffusion segment, segregation transfer chain, blank areas on both sides, and average offset angle according to the aforementioned grid division method. The reference distribution feature map is constructed using no less than 20 qualified tungsten carbide powder samples. Each qualified sample undergoes the same vibration spreading, multi-directional illumination image acquisition, particle cluster region extraction, stable dark transition zone identification, continuous distribution unit division, and edge attachment sequence analysis process.

[0100] For each qualified sample, the grid edge length is uniformly converted to the median D of the equivalent diameter of the particle cluster region. At the same grid location, the cluster value, average offset angle, and blank range of each qualified sample are arranged separately. The cluster value and blank range are taken from the middle position of the arrangement as reference data. The average offset angle is merged according to the pointing relationship of each direction in the plane to avoid incorrectly separating data with similar directions but located on opposite sides of the angle starting point after using ordinary arithmetic averaging. This yields the reference cluster value, reference average offset angle, and reference blank range.

[0101] To mitigate the impact of magnification and particle size differences on the blanking ranges on both sides, the width of the blanking range where no repeated attached segments appear on either side of the segregation transport chain is divided by D to obtain the normalized blanking range: ;in, This represents the width of the blank space on both sides of the grid in row u and column v. This represents the normalized blank range corresponding to the grid. The normalized blank range in the reference distribution feature map is obtained in the same way.

[0102] During matching, each local clustering core segment in the distribution feature map to be discriminated is first extracted, and the coordinates of the repeated attachment positions contained in the local clustering core segment are averaged to obtain the center position of the core segment. The local clustering core segments in the reference distribution feature map are processed in the same way. Based on the center position of the core segment, the local clustering core segments in the distribution feature map to be discriminated are sequentially translated to the candidate positions in the reference distribution feature map. The preferred translation distance is 0.5D, and the preferred horizontal and vertical search ranges are both 2D outside the center position of the core segment. This setting can accommodate small misalignments generated during vibration spreading, while avoiding local clustering core segments crossing too many irrelevant grids.

[0103] After each local aggregation core segment is translated, the peripheral diffusion segment connected to that local aggregation core segment is processed. First, the difference between the average offset angle of the segregation transfer chain to be determined and the average offset angle of the reference segregation transfer chain is calculated, and this difference is used as the rotation angle of the peripheral diffusion segment. ;in, This represents the average offset angle of the segregation transmission chain to be determined. θ represents the average offset angle of the reference segregation transmission chain, and θ represents the rotation angle that the outer diffusion section needs to be adjusted.

[0104] The outer diffusion segment rotates synchronously around the center of the translated core segment. Let any position in the outer diffusion segment be (x, y), and the center of the local aggregation core segment be (x, y). The translation is (Δx, Δy), and the rotated coordinates are obtained according to the following formula: The method of translating the local agglomeration core segment and simultaneously rotating the peripheral diffusion segment preserves the original order of the segregation transmission chain, which extends stepwise from the local agglomeration core segment to the peripheral diffusion segment. If all meshes are rotated uniformly, the irregular arrangement of the particle clusters themselves can easily introduce additional misalignment; if only translation is performed, a slight deviation in the powder transport direction can cause the peripheral diffusion segment to fail to correspond to the reference position. The aforementioned method can correct the directional offset of the peripheral diffusion segment without altering the internal relationships of the local agglomeration core segment.

[0105] After each translation and rotation, the differences in cluster values, average offset angles, and normalized blank ranges within the matching mesh are statistically analyzed. The alignment degree is calculated using the following formula: ;in, , and These represent the clustered value, average offset angle, and normalized blank range in the distribution feature map to be discriminated, respectively. , and These represent the corresponding data in the reference distribution feature map; N represents the number of grids participating in the matching. The summation range includes the current local cluster core segment and the surrounding diffusion segment covering the grid connected to this local cluster core segment. When the angle difference is greater than 180 degrees, 360 degrees is used to subtract the original angle difference. The translation position and rotation angle with the largest Q are selected as the alignment result of the current local cluster core segment.

[0106] After alignment, the blank areas on both sides of the local clustering core segment are searched in reverse along the segregation transfer chain. For all grids covered by the local clustering core segment, the normalized average value of the blank area is calculated. The average value in the distribution feature map to be discriminated is recorded as the current blank area, and the average value in the reference distribution feature map is recorded as the reference blank area. The current blank area is subtracted from the reference blank area, and then divided by the reference blank area plus 1 to obtain the blank area shrinkage ratio.

[0107] When the distance between the centers of localized agglomeration core segments is no greater than 1.5D, and the shrinkage ratio of the blank area is no less than 0.25, the current localized agglomeration core segment is marked as a restricted agglomeration area. 1.5D is used to accommodate slight misalignment between adjacent particle clusters. 0.25 indicates that the blank areas on both sides are reduced by at least 25% relative to the qualified sample. This judgment threshold is preferably determined based on the 95th percentile of the shrinkage ratio of the blank area in the qualified sample. When the 95th percentile is between 0.20 and 0.28, 0.25 is used. This setting method can reduce the impact of variations in spreading density between different batches.

[0108] After marking the confined aggregation region, the extension order of the peripheral diffusion segments is preserved. The peripheral diffusion segment coverage mesh is read sequentially from the local aggregation core segment outwards along the segregation transfer chain. For adjacent peripheral diffusion segment meshes, the movement direction along the powder conveying direction and the movement direction perpendicular to the powder conveying direction are recorded respectively. If the adjacent movement direction in the distribution feature map to be judged is consistent with the corresponding movement direction in the reference distribution feature map, it is recorded as one extension retention. The extension retention count is divided by the total number of adjacent movements to obtain the peripheral diffusion segment extension retention ratio. When the peripheral diffusion segment extension retention ratio is not less than 0.70, it is considered that the peripheral diffusion segment extension order has not disappeared during the alignment process. 0.70 indicates that at least 70% of the peripheral diffusion segments still maintain the original stepped extension direction, which can exclude accidental correspondence caused only by local mesh overlap.

[0109] Subsequently, the restricted aggregation region, the order of extension of the peripheral diffusion segment, and the number of particle clusters traversed by the segregation transport chain were compared layer by layer with the corresponding stratification markers in the reference distribution characteristic map. Differences in local aggregation, diffusion extension, and chain transport were calculated uniformly using the following formula: ; ; ;in, Indicates local clustering differences; Indicates the difference in diffusion and extension; The difference in chain-like transmission is represented by q and q′, which represent the number of grids covered by the outer diffusion segment to be determined and the reference outer diffusion segment, respectively; RE represents the extension retention ratio of the outer diffusion segment; n and n′ represent the number of particle cluster regions crossed by the segregation transmission chain to be determined and the reference segregation transmission chain, respectively. represents the deviation angle of the valid step connection to be judged and the reference valid step connection relative to the reverse conveying direction, respectively; t represents the number of valid step connections participating in the angle comparison.

[0110] The thresholds for determining local aggregation differences, diffusion extension differences, and chain transmission differences were all obtained through cross-comparison of qualified samples. Each qualified sample was used sequentially as the sample to be judged, and the remaining qualified samples were used to establish a reference distribution characteristic map. The three types of difference data were statistically analyzed. The 95th percentile of the local aggregation difference data, the 95th percentile of the diffusion extension difference data, and the 95th percentile of the chain transmission difference data were used as the thresholds for chain transmission differences. Under the implementation conditions of a particle size of 0.6 μm to 1.2 μm and a powder spreading thickness of 0.4 mm to 0.8 mm, the preferred thresholds for local aggregation differences are 0.35, diffusion extension differences are 0.30, and chain transmission differences are 0.32.

[0111] Inspect adjacent grids along the powder conveying direction. If three or more consecutive adjacent grids within the same area simultaneously meet the criteria of local aggregation difference not less than 0.35, chain transmission difference not less than 0.32, and peripheral diffusion section extension retention ratio not less than 0.70, the area is identified as an abnormal segregation adhesion area. Three or more consecutive adjacent grids are used to exclude occasional dark adhesion near single particle clusters. If the abnormal segregation adhesion area further meets the criterion of diffusion extension difference not less than 0.30, the tungsten carbide powder to be judged is determined to have abnormal segregation adhesion, and the manufacturing quality is deemed unqualified.

[0112] In tungsten carbide powder spread samples after closed-loop airflow transport, while normal particle obstruction residues may produce high aggregation values ​​in localized areas, they typically do not simultaneously exhibit the combined characteristics of continuously narrowing blank areas on both sides, stable preservation of the extension order of the peripheral diffusion section, and the segregation transfer chain spanning multiple particle cluster regions. Although the banded dark marks formed by vibration spreading can extend along the transport direction, their stepped offset direction cannot be continuously maintained after the synchronous rotation of the peripheral diffusion section. When a small amount of free carbon segregated powder adheres to the edge of a particle cluster, the blank area near the local aggregation core section will significantly narrow, the number of particle cluster regions spanned by the segregation transfer chain will increase, and the peripheral diffusion section will be continuously distributed along an oblique stepped position.

[0113] Through the aforementioned matching process, the offset of the spreading position and slight deviation of the powder conveying direction can be eliminated first. Then, the manufacturing quality is judged by utilizing the differences in the confined aggregation area, the extension order of the peripheral diffusion section, and the chain transfer. This process does not only compare the area of ​​the dark region, but also preserves the spatial extension relationship formed by the segregated powder between multiple particle clusters, thereby reducing the interference of local occlusion, particle cluster size changes, and slight fluctuations in spreading thickness on the judgment results.

[0114] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A visual characterization method for distribution characteristics used in the quality assessment of tungsten carbide powder manufacturing, characterized in that, include: Acquire multi-directional illumination images of the tungsten carbide powder surface after vibration spreading, and determine the particle cluster region based on the consistency of brightness changes between images in each direction; A grayscale transition band continuously distributed from the inside to the outside is extracted around the particle cluster region. Based on the retention degree of the grayscale transition band under different lighting directions, unstable dark areas formed by particle occlusion are removed to obtain a stable dark transition region. Based on the adhesion relationship between the stable dark transition zone and the edge of the particle cluster, the stable dark transition zone is divided into several continuous distribution units, and an edge attachment sequence is generated according to the interval variation between adjacent continuous distribution units. Based on the repeated occurrence positions of the edge attachment sequence in different particle cluster regions, a distribution feature map is constructed to characterize the local aggregation state of segregated powder. The distribution feature map is matched with the reference distribution feature map corresponding to qualified tungsten carbide powder, and the manufacturing quality of the tungsten carbide powder to be judged is determined based on the matching difference.

2. The visual characterization method for distribution characteristics in tungsten carbide powder manufacturing quality assessment according to claim 1, characterized in that, Determining the particle cluster region includes: Multiple illumination units are activated at circumferential intervals along the surface of tungsten carbide powder, and overlapping illumination is retained during the switching of adjacent illumination units. Unidirectional illumination images and overlapping illumination images are obtained respectively. Dark boundary pixels and bright boundary pixels are searched along the illumination direction in the overlapping illumination image. Dark boundary pixels and bright boundary pixels with spacing conforming to the scale of particle protrusion edges are connected into local migration line segments. Based on the proximity of the midpoints and the continuity of the directions of the local migration lines in adjacent overlapping illumination images, the local migration lines are connected into a continuous path that returns around the same local center, and the area enclosed by the continuous path is determined as the candidate particle cluster region.

3. The visual characterization method for distribution characteristics in tungsten carbide powder manufacturing quality judgment according to claim 2, characterized in that, Determining the particle cluster region also includes: Sampling points are selected along a continuous path in the candidate particle cluster region. The brightness rise and fall order of the outer, edge and inner positions of the sampling points is read in each unidirectional illumination image, and the forward order, reverse order and invalid order that do not form a clear rise and fall relationship are marked respectively. Local areas that can form a corresponding reversal of the forward and reverse order under opposite lighting directions and maintain the continuous transmission of edge positions under adjacent lighting directions are marked as stable return regions. By merging stable return regions that are adjacent to each other along a continuous path and have the same tangential direction, and removing discrete dark spots that are only retained under a single illumination direction or whose area changes too much under adjacent illumination directions, the particle cluster region is obtained.

4. The visual characterization method for distribution characteristics in tungsten carbide powder manufacturing quality judgment according to claim 3, characterized in that, Extracting the grayscale transition band includes: Edge positions are selected at intervals along the continuous path of the particle cluster region in order of arc length, and radial sampling bands are established along the outward normal direction pointing to the outside of the particle cluster region from each edge position, so that the radial sampling bands extend from the inside of the particle cluster region to the outside of the particle cluster region. Each radial sampling band is divided into multiple radial sampling layers in order from the inside out, and the regions in which the brightness of multiple consecutive radial sampling layers decreases as a dark segment are defined. Based on the projection overlap relationship and the proximity relationship of the outer edge positions between dark segments in adjacent radial sampling bands, multiple dark segments are sequentially connected to form a gray-scale transition band distributed around the particle cluster region.

5. The visual characterization method for distribution characteristics in tungsten carbide powder manufacturing quality judgment according to claim 4, characterized in that, A stable dark transition region is obtained including: The grayscale transition band is mapped to the unidirectional lighting image corresponding to the opposite lighting direction. Under the condition that the position of the continuous path remains unchanged, the grayscale transition band that is still attached to the same continuous path under the opposite lighting direction and whose outer edge position is restricted is retained to obtain the initial dark transition area. The movement sequence of the initial dark transition area under adjacent lighting directions is traced along the arc length direction of the continuous path. The initial dark transition area that continues to migrate towards the backlight side and detaches from the initial attachment position as the lighting direction changes is marked as the particle occlusion dark area. Remove the dark areas obscured by particles from the initial dark transition region, and merge the remaining areas according to the overlapping relationship on the continuous path to obtain a stable dark transition region.

6. The visual characterization method for distribution characteristics in tungsten carbide powder manufacturing quality judgment according to claim 1, characterized in that, Dividing a continuous distribution unit includes: The continuous path of the particle cluster region is unfolded according to the arc length coordinate, and sampling positions are set at intervals along the continuous path. At each sampling position, the distance between the inner edge of the stable dark transition region and the continuous path, the radial width of the stable dark transition region, and the folding direction of the outer edge of the stable dark transition region are read along the outward normal direction. Adjacent sampling positions with continuous fit, consistent outer edge folding direction, and continuous radial width variation are assigned to the same initial distribution segment. For gaps between adjacent initial distribution segments that are not covered by a stable dark transition area, the radial width variation trend on both sides of the gap is traced back along the continuous path. When the radial width variation trend can continue to each other and the gap does not undergo a sharp turn, the adjacent initial distribution segments are merged into a continuous distribution unit.

7. The visual characterization method for distribution characteristics in tungsten carbide powder manufacturing quality judgment according to claim 6, characterized in that, The generation of edge attachment sequences includes: Record the bonding length of each continuous distribution unit and the edge spacing between adjacent continuous distribution units according to the clockwise arrangement order of each continuous distribution unit on the continuous path. Based on the change of the current edge spacing relative to the previous edge spacing, the edge spacing is sequentially encoded as narrowing, expanding or maintaining, and based on the proportion of the bonding length of each continuous distribution unit to the total arc length of the continuous path, the continuous distribution units are sequentially divided into short bonding level, medium bonding level or long bonding level. The encoding results of the bonding length level and the corresponding edge interval are recorded alternately in the order of arrangement to obtain the edge attachment sequence that represents the attachment rhythm of the stable dark transition zone along the edge of the particle cluster.

8. The visual characterization method for distribution characteristics in tungsten carbide powder manufacturing quality judgment according to claim 7, characterized in that, Constructing a segregation propagation chain includes: The edge attachment sequences corresponding to different particle cluster regions are cyclically expanded. Starting from any continuous distribution unit, the bonding length level and edge spacing coding results are compared segment by segment. The sequence segments that continuously appear with the same bonding length level and whose edge spacing coding change direction can be maintained are marked as repeated attachment segments. Map the midpoint of the continuous distribution unit corresponding to the starting position of the repeated attachment segment to the powder spreading surface to obtain the repeated attachment position, and search for other adjacent repeated attachment positions along the reverse conveying side starting from any repeated attachment position. A segregation transfer chain is formed by sequentially connecting multiple repeated attachment positions that are within the range of the equivalent diameter of the particle cluster region, whose angles deviate from the reverse conveying direction are kept within the range of oblique attachment, and whose angles change continuously between adjacent connections.

9. The visual characterization method for distribution characteristics in tungsten carbide powder manufacturing quality assessment according to claim 1, characterized in that, Constructing the distribution feature map includes: Taking any repeated attachment position in the segregation transfer chain as the center, count the number of repeated attachment positions belonging to the same segregation transfer chain in the neighborhood range associated with the equivalent diameter of the particle cluster region, and read the blank range on both sides where no repeated attachment segments appear along the normal direction of the segregation transfer chain. Based on the intrasample arrangement of the number of repeated attachment sites and the blank areas on both sides, the repeated attachment sites are divided into local aggregation core segments and peripheral diffusion segments. The powder spreading surface is divided into grids according to the equivalent diameter of the particle cluster region. In each grid, the weighted distribution intensity corresponding to the local aggregation core segment, the distribution intensity corresponding to the peripheral diffusion segment, the average offset direction of the effective step connection, and the normalized blank range are recorded to obtain the distribution feature map.

10. The visual characterization method for distribution characteristics in tungsten carbide powder manufacturing quality judgment according to claim 9, characterized in that, Determining the manufacturing quality of the tungsten carbide powder to be judged includes: Using the local agglomeration core segment as the matching starting point, the local agglomeration core segment in the distribution feature map corresponding to the tungsten carbide powder to be identified is translated step by step to the candidate position in the reference distribution feature map, and the corresponding peripheral diffusion segment is rotated synchronously around the translated local agglomeration core segment to keep the step-like offset direction of the segregation transmission chain aligned. The blank areas on both sides are searched in reverse along the aligned segregation transmission chain. The areas where the local aggregation core segments are close in position and the blank areas on both sides shrink relative to the reference distribution feature map are marked as restricted aggregation areas. The order in which the peripheral diffusion segments extend from the local aggregation core segments to both ends is preserved. By comparing the restricted aggregation area, the extension order of the peripheral diffusion segment, and the number of particle clusters spanned by the segregation transfer chain with the corresponding stratification markers in the reference distribution feature map, it is determined that the tungsten carbide powder to be judged has segregation adhesion anomalies when local aggregation differences and chain transfer differences occur continuously in the same area and the extension order of the peripheral diffusion segment remains.

Images