Casting joint line identification method based on point cloud

Abstract

The invention provides a casting joint line identification method based on point cloud, and belongs to the crossing field of three-dimensional industrial detection and intelligent manufacturing, and the method comprises the steps: carrying out the preprocessing of the obtained original three-dimensional point cloud data of a casting; forming a primary candidate set by point clouds of which the curvature change rate is greater than a curvature threshold value and which meet a normal mutation condition in the voxel three-dimensional point cloud data; screening connected domains from the primary candidate set by adopting a three-dimensional DBSCAN clustering algorithm, and setting a length-width ratio threshold value and a length threshold value of the connected domains to remove unqualified connected domains to obtain candidate connected domains; performing expansion and corrosion operation on each candidate connected domain, and performing topological constraint verification on the obtained result to obtain a consistent connected domain; and performing three-level geometric verification on the consistent connected domain in sequence. According to the invention, the joint line identification precision can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the cross - field of three - dimensional industrial inspection and intelligent manufacturing, and particularly relates to a method for identifying the mold - closing line of castings based on point cloud. Background Technique

[0002] The casting industry is one of the basic pillars of the manufacturing industry and is widely used in fields such as automobiles, aerospace, and energy equipment. As an inevitable product of the casting process, the detection accuracy of the mold - closing line directly affects the sealing performance, assembly accuracy, and fatigue life of castings. Taking the engine cylinder block of an automobile as an example, a position deviation of the mold - closing line exceeding 0.3 mm will cause the sealing between the cylinder block and the cylinder head to fail, leading to the risk of oil leakage, and the related quality recall cost can reach millions of dollars per batch.

[0003] For the detection of the mold - closing line, in the existing technology, either manual detection or a two - dimensional vision detection system is used. The former has the defects of low efficiency and high missed - detection rate, while the latter has the defect of high misjudgment rate. Three - dimensional digital detection is used for mold - closing line detection, but the existing point - cloud processing algorithms lag far behind the development of hardware. When the existing algorithms analyze the surface of castings, they are easily interfered by stray noise points. For example, when calculating the surface normal, a small number of abnormal points will cause the overall direction judgment to be incorrect, and the error will expand like a snowball. Although the improved moving least - squares method can alleviate some problems, its fixed calculation range parameters are difficult to cope with the difference in point - cloud density in different regions - it may overly blur details in dense regions and cannot accurately capture features in sparse regions. For such fine edge features as the mold - closing line, the existing methods use a unified curvature threshold screening (such as setting 0.05 as the standard), but the actual castings have significant material differences: the surface of cast iron is rough, while that of aluminum alloy is relatively smooth. This "one - size - fits - all" threshold setting will lead to a high missed - detection rate for thin - wall castings. More intractably, when the mold - closing line intersects or bifurcates, conventional clustering algorithms are prone to misjudging it as multiple discontinuous segments, and the deep - learning scheme relying on massive data training not only takes a long time and effort for annotation, but also the recognition effect will drop significantly after replacing the scanning device. In addition, traditional breakpoint repair operations will damage the original shape like an eraser, and the uniform dilation and erosion operations will distort the width of the mold - closing line. Summary of the Invention

[0004] The purpose of the present invention is to propose a method for identifying the mold - closing line of castings based on point cloud, which can effectively improve the accuracy of mold - closing line recognition.

[0005] The present invention is realized through the following technical solutions: A method for identifying the mold - closing line of castings based on point cloud, comprising the following steps: Step S1: Preprocess the acquired original 3D point cloud data of the casting. The preprocessing includes obtaining the normal vectors of each point cloud in the original 3D point cloud data, suppressing noise in the original 3D point cloud data to obtain denoised 3D point cloud data, and performing voxel filtering on the denoised 3D point cloud data to obtain voxel 3D point cloud data; Step S2: Compose a primary candidate set from the point clouds in the voxel 3D point cloud data whose curvature change rate is greater than the curvature threshold and satisfy the normal mutation condition. The normal mutation condition is related to the normal angle between adjacent points in the point cloud neighborhood; Step S3: Use a 3D DBSCAN clustering algorithm to screen connected domains from the primary candidate set, and set the aspect ratio threshold and length threshold of the connected domains to eliminate unqualified connected domains, obtaining candidate connected domains. Among them, the neighborhood radius in the 3D DBSCAN clustering algorithm is dynamically adjusted according to the curvature change rate; Step S4: Perform dilation and erosion operations on each candidate connected domain, and perform topological constraint verification on the obtained results to obtain consistent connected domains. Among them, in the dilation operation, the expansion direction is dominated by the normal direction of the point clouds in the candidate connected domain. The topological constraint verification is used to verify whether the shape fidelity condition and density condition are satisfied; Step S5: Perform three-level geometric verification on the consistent connected domains in sequence. The first-level geometric verification is used to retain the consistent connected domains with continuous angles. The second-level geometric verification is used to retain the consistent connected domains whose width and height differences both meet the set thresholds, and obtain the mold closing line trajectory based on this consistent connected domain. The third-level geometric verification is used to judge whether the length of the mold closing line trajectory is within the error range of the standard mold closing line length, and eliminate the mold closing line trajectories outside the error range.

[0006] Further, in the step S1, for the point clouds in the original 3D point cloud data p i , use the moving least squares method to fit the local surface, and obtain the normal vector of the point cloud according to the gradient direction of the local surface n i .

[0007] Further, in the step S1, the noise suppression of the original 3D point cloud data includes: Step S11: For the point clouds in the original 3D point cloud data p i , when the total weight of the neighborhood of the point cloud W i meets , it is then eliminated, thus obtaining point cloud data P Gauss , where , , q j is the point cloud pi within the neighborhood the j th point cloud n i and n j respectively represent the normal vectors of the point cloud p i and the point cloud q j ; is the standard deviation of the spacing of the point clouds within the neighborhood of the point cloud p i ; is the set tolerance threshold for normal vector deviation α is the proportionality coefficient k is the number of neighborhood points; Step S12: For the point cloud P Gauss in the point cloud data p j , if it meets , then this point cloud is far-field noise and is removed. If it meets and the normal deviation angle is greater than the set threshold, it is near-field noise and is removed. In this way, the denoised three-dimensional point cloud data P filter is obtained, where μ is the mean position of the point cloud data P Gauss ; is the global standard deviation of the point cloud data P Gauss ; μ j is the local mean of the neighborhood of the point cloud p j ; is the standard deviation of the spacing of the point clouds within the neighborhood of the point cloud p j .

[0008] Furthermore, in the step S1, the voxel side length is used to perform voxel filtering on the denoised three-dimensional point cloud data to obtain the voxel three-dimensional point cloud data P voxel , where r base is the average spacing of the point clouds in the denoised three-dimensional point cloud data is the curvature change rate of the point clouds in the denoised three-dimensional point cloud data is the set first curvature threshold.

[0009] Furthermore, in the step S2, for the point cloud p i' in the voxel three-dimensional point cloud data, if it meetsR i' > R t , then it satisfies the normal mutation condition, where, , k is the number of neighborhood point clouds, is the point cloud p i' and the point cloud within its neighborhood p j' of the normal vector angle, is the indicator function, which returns 1 when the condition is met and 0 otherwise, is the normal vector angle threshold, R t = 0.3, is the point cloud p i' of the normal vector, is the point cloud p j' of the normal vector.

[0010] Furthermore, in the step S3, the number of neighborhood point clouds of the three-dimensional DBSCAN clustering algorithm is expressed as , where, is the set base radius, is the set second curvature threshold; when the aspect ratio of the connected domain is greater than the aspect ratio threshold or the length is less than the length threshold, it is determined that the connected domain is an unqualified connected domain.

[0011] Furthermore, in the step S4, according to the formula the candidate connected domain R i in the point cloud p i'' is dilated to obtain a new point cloud p i''_new , the shape fidelity condition is expressed as , the density condition is expressed as , where, α' is the dilation step size, is the point cloud p i'' of the normal vector, represents finding the new point cloud p i''_new and the point cloud p i'' the distance between, l c is the set distance threshold, ρ final is the candidate connected domain R i the density of the new connected domain point cloud obtained after dilation and erosion operations,ρ original is a candidate connected region R i the density of the point cloud, m ∈(0,1).

[0012] Further, in the step S5, the primary geometric verification is performed through the formula where, , s j is the j th edge point of the consistent connected region, is the set angle threshold.

[0013] Further, in the step S5, the secondary geometric verification specifically includes the following steps: Step S51: Receive the continuous edge path obtained after the primary geometric verification . For any point cloud in this continuous edge path, construct a local coordinate system , where, is the tangent vector of the point cloud , is the normal vector of the point cloud , is the binormal vector of the point cloud , N is the number of point clouds in the continuous edge path; Step S52: Project each neighboring point cloud in the neighborhood of the point cloud q k onto the normal vector to obtain . Then the neighborhood width of the point cloud is . If this width meets the set width threshold range, it is determined that the width at this point cloud is qualified; otherwise, it is marked as an abnormal point; Step S53: Extract the height value sequence of all point clouds on the continuous edge path. Calculate the height change between adjacent point clouds according to the formula and calculate the overall height fluctuation range according to the formula . If the height change is less than the set height threshold and the overall height fluctuation range is less than the set height threshold, it is determined that the height difference of this continuous edge path is qualified; otherwise, it is marked as a defective area and removed, where, is the j'' th value in the height value sequence.

[0014] Further, in step S5, the two-way dynamic Dijkstra algorithm is used to obtain the mold closing line trajectory.

[0015] The present invention has the following beneficial effects: 1. First, the present invention preprocesses the original three-dimensional point cloud data of the casting, screens the primary candidate set from the preprocessed data according to the curvature change rate and the normal mutation condition, then uses the three-dimensional DBSCAN clustering algorithm to screen the connected domains from the primary candidate set, and sets the aspect ratio threshold and length threshold of the connected domains to eliminate unqualified connected domains to obtain candidate connected domains. Then, dilation and erosion operations are performed on each candidate connected domain, and the obtained results are verified by topological constraints to obtain consistent connected domains. Finally, three-level geometric verification is sequentially performed on the consistent connected domains to obtain the mold closing line trajectory that meets the design requirements, thereby effectively improving the recognition accuracy of the mold closing line. In the preprocessing stage, noise and valid points can be effectively distinguished. When screening connected domains, the neighborhood radius is dynamically adjusted according to the curvature change rate, and the aspect ratio threshold and length threshold of the connected domains are set, which can solve the problem of over-segmentation of sharp edges and enhance the screening ability of connected domains in complex structures. In the dilation and erosion operations, shape fidelity conditions and density conditions are set, which can effectively fill the holes in the point cloud and smooth the surface, and ensure the stability of the geometric structure. By using three-level geometric verification, the continuity of the mold closing line can be optimized, so that it is superior to traditional technologies in noise suppression, feature retention, clustering robustness, and repair accuracy, and is applicable to high-precision industrial inspection scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described in detail below with reference to the accompanying drawings.

[0017] Figure 1 is a flowchart of the present invention.

[0018] Figure 2 is a schematic diagram of the normal vector of the original three-dimensional point cloud data of the present invention.

[0019] Figure 3 is a schematic diagram of the present invention for removing the point cloud that does not meet the curvature threshold and the normal mutation condition.

[0020] Figure 4 is a schematic diagram of the present invention for removing unqualified connected domains.

[0021] Figure 5 is a schematic diagram of the consistent connected domain of the present invention after dilation and erosion.

[0022] Figure 6 is a schematic diagram of the consistent connected domain of the present invention where both the width difference and the height difference meet the threshold.

[0023] Figure 7 is a schematic diagram of the standard mold closing line of the present invention.

[0024] Figure 8 Schematic diagram of the mold clamping line trajectory of the present invention. Specific embodiments

[0025] As Figure 1 shown, the method for identifying the mold clamping line of a casting based on point cloud includes the following steps: Step S1: Preprocess the acquired original three-dimensional point cloud data of the casting. The preprocessing includes obtaining the normal vectors of each point cloud in the original three-dimensional point cloud data, suppressing noise in the original three-dimensional point cloud data to obtain denoised three-dimensional point cloud data, and performing voxel filtering on the denoised three-dimensional point cloud data to obtain voxel three-dimensional point cloud data; Specifically, it includes the following steps: Step S10: For the th i point cloud in the original three-dimensional point cloud data, use the moving least squares method to fit the local surface, and obtain the normal vector of this point cloud according to the gradient direction of the local surface n i ; The moving least squares fitting is specifically to use a second-order polynomial to fit the neighborhood points, and solve the parameters of the second-order polynomial by the weighted least squares method. The local surface of the i th point cloud fitted is expressed as , set the polynomial order to 2, the search radius , is the average point cloud spacing, is the Gaussian kernel function, h is the Gaussian kernel width parameter, is the point cloud p i 's k neighborhood points, k is the neighborhood radius, which is dynamically adjusted according to the point cloud density, usually taking k = 10 - 30, represents the square of the distance between the point p i in the point cloud and the neighborhood point , used to measure spatial proximity; Using the quadratic surface fitting as described above can improve the normal estimation accuracy of complex surfaces. The normal vector is obtained through the gradient direction of the fitted local surface, which can ensure the adaptability to complex surfaces; the point cloud normal vector is as Figure 2 shown; Step S11: For the point cloud p i in the original three-dimensional point cloud data, when the total neighborhood weight W i satisfies It is removed when the condition is met, and thus the point cloud data is obtained. P Gauss , where , is the dynamic weight assignment formula, q j is the point cloud p i in the th j point cloud in the neighborhood, n i and n j respectively represent the normal vectors of the point cloud p i and the point cloud q j , is the standard deviation of the point cloud spacing (i.e., local density) in the neighborhood of the point cloud p i , is the set normal vector deviation tolerance threshold, usually taking a value of 15°, α is the proportionality coefficient ( α = 0.3 - 0.5), k is the number of neighborhood points; is the spatial weight form of the Gaussian kernel function, is the normal weight form of the Gaussian kernel function, w ij By imposing double constraints on the spatial distance and normal deviation, it is possible to effectively distinguish noise points from valid points, thereby improving the filtering accuracy; Step S12: For the point cloud P Gauss in the point cloud data p j , if it meets , then this point cloud is far - field noise and is removed. If it meets and the normal deviation angle is greater than the set threshold (25°), then it is near - field noise and is removed. In this way, the denoised three - dimensional point cloud data P filter is obtained, where μ is the mean position of the point cloud data P Gauss , is the global standard deviation of the point cloud data P Gauss , μ j is the local mean of the neighborhood of the point cloud p j , is the point cloud p ​j The standard deviation of the point cloud spacing within the neighborhood, and the normal deviation angle is the point cloud p j The normal angle between the point cloud and the neighboring points in the k-neighborhood point set.

[0026] Mean position Calculated by the following formula: , is the point cloud p j coordinates, n is the point cloud data P Gauss in the number of point clouds, is the point cloud p j The distance between and the mean position; This can effectively distinguish different types of noise points and adopt different processing methods according to their characteristics, thereby enhancing the filtering effect; Step S13: Use the voxel side length Perform voxel filtering on the denoised three-dimensional point cloud data to obtain voxel three-dimensional point cloud data P voxel , where r base is the average spacing of the point clouds in the denoised three-dimensional point cloud data, is the curvature change rate of the point clouds in the denoised three-dimensional point cloud data, and are the curvatures in the minimum principal direction and the maximum principal direction respectively, is the set first curvature threshold; This can achieve downsampling with a larger voxel size for flat areas to improve efficiency, retain the original resolution for high-curvature areas to maintain details, and avoid excessive smoothing of details by traditional voxel filtering.

[0027] Step S2: The point clouds in the voxel three-dimensional point cloud data with a curvature change rate greater than the curvature threshold and satisfying the normal mutation condition form a primary candidate set, and the normal mutation condition is related to the normal angle between adjacent points in the point cloud neighborhood; Specifically, for the point cloud p i' in the voxel three-dimensional point cloud data, if it meets R i' > R t , then it satisfies the normal mutation condition, where , k is the number of neighboring point clouds, is the point cloud p i' and its neighboring point clouds p j'Normal angle Is an indicator function that returns 1 when the condition is met and 0 otherwise Is the normal angle threshold Represents the dot product Is the point cloud p i' Normal vector of Is the point cloud p j' Normal vector of In this embodiment, the curvature threshold is set to 0.8 R t = 0.3, and the point cloud that does not meet the curvature threshold and the normal mutation condition is removed as Figure 3 Shown

[0028] Step S3: Use the 3D DBSCAN clustering algorithm to screen connected regions from the primary candidate set, and set the aspect ratio threshold and length threshold of the connected region to remove unqualified connected regions, obtaining candidate connected regions. Among them, the neighborhood radius in the 3D DBSCAN clustering algorithm is dynamically adjusted according to the curvature change rate Specifically, the number of neighborhood point clouds in the 3D DBSCAN clustering algorithm is expressed as Among them Is the set base radius Is the set second curvature threshold; when the aspect ratio of the connected region is greater than the aspect ratio threshold or the length is less than the length threshold, the connected region is determined to be an unqualified connected region. In this embodiment, the aspect ratio threshold is set to 5:1. Connected regions with too large an aspect ratio are mostly noise or scanning errors (such as flying points), and the length threshold is set to 10 mm. Connected regions with too small a length are usually caused by point cloud noise or insufficient local sampling and do not conform to the actual object size

[0029] Removing unqualified connected regions as Figure 4 Shown, significantly improving the clustering accuracy of sharp edges. Combining the topological constraints of the aspect ratio and length, effectively removing non-physical connected regions, taking into account both the robustness and computational efficiency of the algorithm

[0030] Step S4: Perform dilation and erosion operations on each candidate connected region, and perform topological constraint verification on the obtained results to obtain consistent connected regions. Among them, in the dilation operation, the expansion direction is dominated by the normal direction of the point cloud in the candidate connected region, and the topological constraint verification is used to verify whether the shape fidelity condition and density condition are satisfied Specifically, according to the formula Perform dilation operation on the point cloud R i In the candidate connected region p i'' To obtain a new point cloud p i''_new, the shape fidelity condition (i.e., the anisotropic diffusion strategy) is expressed as , the density condition is expressed as , where α' = 0.1 mm is the dilation step size, is the point cloud p i'' 's normal vector, represents finding the new point cloud p i''_new and the point cloud p i'' 's distance, l c = 0.3 mm is the set distance threshold, ρ final is the candidate connected domain R i 's density of the new connected domain point cloud obtained after dilation and erosion operations, ρ original is the candidate connected domain R i 's point cloud density, m ∈ (0, 1).

[0031] Performing the morphological closing operation of first dilating and then eroding on each candidate connected domain as above can fill the holes in the point cloud or voxel model, connect the broken regions, and smooth the surface geometric features. The dilation operation expands along the normal direction and introduces an anisotropic diffusion strategy to control the diffusion range; the erosion operation is based on the density preservation criterion and dynamically adjusts the threshold to avoid excessive damage to the original structure. The consistent connected domain obtained after dilation and erosion is as Figure 5 shown.

[0032] Step S5: Perform three-level geometric verification on the consistent connected domain in sequence. The first-level geometric verification is used to retain the consistent connected domain with continuous angles, the second-level geometric verification is used to retain the consistent connected domain whose width and height difference both meet the set threshold, and based on this consistent connected domain, the mold joint line trajectory is obtained. The third-level geometric verification is used to determine whether the length of the obtained mold joint line trajectory is within the error range compared with the standard mold joint line length, and eliminate the mold joint line trajectories outside the error range; Specifically, it includes the following steps: Step S50: The first-level geometric verification is carried out through the formula . If it meets the condition, it is considered that the current consistent connected domain is continuous and enters the second-level geometric verification, where , s j is the j th edge point of the consistent connected domain obtained in step S4, is the set angle threshold.

[0033] Step S51: Receive the continuous edge path of the consistent connected region obtained after the first-level geometric verification. For any point cloud in this continuous edge path , construct a local coordinate system . Among them, is the tangent vector of the point cloud p j' . If j'' = 1 or j'' = N , then . N is the number of point clouds in the continuous edge path, is the normal vector of the point cloud , which can be estimated from the neighborhood points through principal component analysis (PCA). is the binormal vector of the point cloud ; Step S52: Project each neighboring point cloud in the neighborhood of the point cloud q k onto the normal vector to obtain . Then the neighborhood width of the point cloud is . If this width meets the set width threshold range, it is determined that the width at this point cloud is qualified; otherwise, it is marked as an abnormal point. Among them, the neighborhood radius is taken as 2 - 3 mm, and the width threshold range is [1.5 mm, 4 mm]; Step S53: Extract the height value sequence of all point clouds on the continuous edge path. Calculate the height change between adjacent point clouds according to the formula , and calculate the overall height fluctuation range according to the formula . If the height change is less than the set height threshold and the overall height fluctuation range is less than the set height threshold, it is determined that the height difference of this continuous edge path is qualified; otherwise, it is marked as a defective area and removed. Among them, , is the j'' th value in the height value sequence, and the height threshold is 0.2 mm ; The consistent connected regions where both the width difference and the height difference meet the thresholds are as Figure 6 shown.

[0034] In obtaining the mold - closing line trajectory based on the consistent connected region, the bidirectional dynamic Dijkstra algorithm is used to obtain the mold - closing line trajectory. In this algorithm, not only the geometric distance is considered, but also the curvature change and the signal - to - noise ratio are introduced. The formula is , ; Geometric distance: Indicates the geometric distance of each segment of the path, α = 0.6 is the weight coefficient; Curvature change: Indicates the curvature change of each segment of the path, β = 0.3 is the weight coefficient. The curvature change reflects the smoothness of the path. A large curvature change may result in an unnatural or difficult-to-implement path; Signal-to-noise ratio: Indicates the reciprocal of the signal-to-noise ratio, γ = 0.1 is the weight coefficient, which reflects the reliability of the path. A lower signal-to-noise ratio may mean that the path is unreliable under certain conditions.

[0035] In the three-level geometric verification, as Figure 7 shown, the standard mold closing line is obtained according to the CAD model of the casting. According to the formula the length of the mold closing line trajectory is obtained. According to the formula it is judged whether it is qualified to ensure that the manufacturing meets the design requirements, is the three-dimensional coordinate of the i point on the mold closing line trajectory, N is the total number of points on the mold closing line trajectory, is the length of the standard mold closing line, .

[0036] The above is only a preferred embodiment of the present invention, and thus cannot limit the scope of implementation of the present invention. That is, equivalent changes and modifications made according to the scope of the patent application of the present invention and the content of the specification should still fall within the scope covered by the patent of the present invention.

Claims

1. A method for identifying the mold closing line of a casting based on point cloud, characterized in that: It includes the following steps: Step S1: Preprocess the obtained original 3D point cloud data of the casting. The preprocessing includes obtaining the normal vectors of each point cloud in the original 3D point cloud data, suppressing noise in the original 3D point cloud data to obtain denoised 3D point cloud data, and performing voxel filtering on the denoised 3D point cloud data to obtain voxel 3D point cloud data; Step S2: Combine the point clouds in the voxel 3D point cloud data with a curvature change rate greater than the curvature threshold and satisfying the normal mutation condition to form a primary candidate set. The normal mutation condition is related to the angle between the normal vectors of adjacent points in the point cloud neighborhood; Step S3: Use the 3D DBSCAN clustering algorithm to screen connected domains from the primary candidate set, and set the aspect ratio threshold and length threshold of the connected domains to eliminate unqualified connected domains, obtaining candidate connected domains. Among them, the neighborhood radius in the 3D DBSCAN clustering algorithm is dynamically adjusted according to the curvature change rate; Step S4: Perform dilation and erosion operations on each candidate connected domain, and perform topological constraint verification on the obtained results to obtain consistent connected domains. Among them, in the dilation operation, the expansion direction is dominated by the normal direction of the point cloud in the candidate connected domain, and the topological constraint verification is used to verify whether the shape fidelity condition and density condition are met; Step S5: Perform three-level geometric verification on the consistent connected domains in sequence. The first-level geometric verification is used to retain the consistent connected domains with continuous angles, the second-level geometric verification is used to retain the consistent connected domains with the width and height difference both meeting the set thresholds, and obtain the mold closing line trajectory based on this consistent connected domain. The third-level geometric verification is used to determine whether the length of the mold closing line trajectory is within the error range of the standard mold closing line length, and eliminate the mold closing line trajectories outside the error range.

2. The method for identifying the mold closing line of a casting based on point cloud according to claim 1, wherein: In the step S1, for the point cloud in the original three-dimensional point cloud data p i , the local surface is fitted by the moving least squares method, and the normal vector of the point cloud is obtained according to the gradient direction of the local surface n i .

3. The method for identifying the mold closing line of a casting based on point cloud according to claim 2, wherein: In the said step S1, suppressing noise in the original 3D point cloud data includes: Step S11: For the point cloud in the original three-dimensional point cloud data p i , when the total weight of the neighborhood of this point cloud W i meets , it is removed, and in this way, the point cloud data P Gauss is obtained. Among them, , , q j is the p i nth point cloud within the neighborhood of the point cloud . j , n i and n j respectively represent the normal vectors of the point cloud p i and the point cloud q j . is the standard deviation of the point cloud spacing within the neighborhood of the point cloud p i . is the set normal vector deviation tolerance threshold, α is the proportionality coefficient, k is the number of neighborhood points; Step S12: For the point cloud data P Gauss in the point cloud p j , if it meets , then this point cloud is far-field noise and is removed. If it meets and the normal deviation angle is greater than the set threshold, it is near-field noise and is removed. In this way, the denoised three-dimensional point cloud data P filter is obtained, where μ is the mean position of the point cloud data P Gauss , is the global standard deviation of the point cloud data P Gauss , μ j is the local mean of the neighborhood of the point cloud p j , is the standard deviation of the point cloud spacing within the neighborhood of the point cloud p j .

4. A method for identifying the mold closing line of a casting based on point cloud according to claim 3, characterized in that: In the step S1, a voxel side length is used to perform voxel filtering on the denoised three-dimensional point cloud data to obtain voxel three-dimensional point cloud data P voxel , where r base is the average spacing of the point cloud in the denoised three-dimensional point cloud data, is the curvature change rate of the point cloud in the denoised three-dimensional point cloud data, is the set first curvature threshold.

5. The method for identifying the mold closing line of a casting based on point cloud according to claim 4, characterized in that: In step S2, for the point cloud in the voxel three-dimensional point cloud data p i' , if it meets R i' > R t , then it satisfies the normal mutation condition, where , k is the number of neighboring point clouds, is the point cloud p i' and the point clouds within its neighborhood p j' 's normal angle, is the indicator function, which returns 1 when the condition is met, otherwise returns 0, is the normal angle threshold, R t = 0.3, is the normal vector of the point cloud p i' , is the normal vector of the point cloud p j' .

6. The method for identifying the mold closing line of a casting based on point cloud according to claim 5, wherein: In the step S3, the number of neighboring point clouds of the three-dimensional DBSCAN clustering algorithm is expressed as , where is the set basic radius, is the set second curvature threshold; when the aspect ratio of the connected domain is greater than the aspect ratio threshold or the length is less than the length threshold, it is determined that the connected domain is an unqualified connected domain.

7. A method for identifying the mold closing line of a casting based on point cloud according to claim 6, characterized in that: In the step S4, according to the formula the candidate connected region R i the point cloud in p i'' is dilated to obtain a new point cloud p i''_new , the shape fidelity condition is expressed as , the density condition is expressed as , where α' is the dilation step size, is the normal vector of the point cloud p i'' , represents the distance between the new point cloud p i''_new and the point cloud p i'' , l c is the set distance threshold, ρ final is the density of the new connected region point cloud obtained after the candidate connected region R i undergoes dilation and erosion operations, ρ original is the density of the point cloud of the candidate connected region R i , m ∈(0,1).

8. A method for identifying the mold closing line of a casting based on point cloud according to claim 7, characterized in that: In the step S5, the primary geometric verification is performed through the formula , where , s j is the j th edge point of the consistency connected region, is the set angle threshold.

9. The method for identifying the mold closing line of a casting based on point cloud according to claim 8, wherein: In the said step S5, the second-level geometric verification specifically includes the following steps: Step S51: Receive the continuous edge path obtained after first-level geometric verification , for any point cloud in this continuous edge path , construct a local coordinate system , where is the tangent vector of the point cloud , is the normal vector of the point cloud , is the binormal vector of the point cloud , N is the number of point clouds in the continuous edge path; Step S52: Project each neighboring point cloud within the neighborhood of the point cloud q k onto the normal vector to obtain . Then, the neighborhood width of the point cloud is . If this width meets the set width threshold range, it is determined that the width at this point cloud is qualified; otherwise, it is marked as an abnormal point. Step S53: Extract the height value sequence of all point clouds on the continuous edge path , and calculate the height change between adjacent point clouds according to the formula , and calculate the overall height fluctuation range according to the formula . If the height change is less than the set height threshold and the overall height fluctuation range is less than the set height threshold, it is determined that the height difference of this continuous edge path is qualified; otherwise, it is marked as a defective area and removed. Among them, is the j'' th value in the height value sequence.

10. A method for identifying the mold closing line of a casting based on point cloud according to claim 9, characterized in that: In the said step S5, the bidirectional dynamic Dijkstra algorithm is used to obtain the mold closing line trajectory.

Images