Casting joint line polishing method based on vision

Abstract

The invention provides a casting joint line polishing method based on vision, and belongs to the crossing field of mechanical manufacturing automation and intelligent manufacturing, and the casting joint line polishing method comprises the following steps: obtaining template surface point cloud of a standard casting, carrying out reverse modeling on the template surface point cloud to obtain first template point cloud, and obtaining first measurement point cloud of a casting to be polished; performing rough matching and fine matching on the first template point cloud in sequence, and converting the original joint line model into a new joint line model; points, closest to track points of the template track, in the new joint line model are obtained, for each closest point, the point, having the largest vertical distance from the plane where the new joint line model is located, in the neighborhood of the closest point is obtained, and all the points having the largest distances form a path planning point cloud; and traversing all planning points in the path planning point cloud according to the machining direction, selecting a machining mode according to the machining allowance corresponding to each planning point, and generating a final machining track. The grinding precision, the grinding quality and the grinding efficiency of the joint line can be improved, and the grinding device is also applicable to complex curved surfaces.

Description

Technical Field

[0001] The present invention belongs to the cross - field of mechanical manufacturing automation and intelligent manufacturing, and particularly relates to a vision - based grinding method for the mold - closing line of castings. Background Art

[0002] As a core basic component in modern industrial fields such as automobiles, aerospace, and shipbuilding, the surface quality of castings directly affects the assembly accuracy and service performance. However, due to the inherent burrs, mold - closing line (parting surface) defects, and surface roughness in the casting process, castings must undergo precision grinding for post - treatment. Traditional manual grinding relies on the operator's experience, with problems such as low efficiency, poor consistency, and safety hazards, and it is difficult to meet the requirements of large - scale production in modern industry.

[0003] Although existing automated grinding systems partially replace manual labor, they still face multiple technical bottlenecks. First, most systems use two - dimensional vision positioning. Limited by the principle of planar imaging, they cannot accurately reconstruct the three - dimensional morphology of complex curved surfaces, resulting in positioning errors often exceeding 0.2 mm, making it difficult to adapt to the grinding requirements of high - precision components. Second, the modular design of the system leads to data islands. Links such as scanning, recognition, and path planning operate independently and lack real - time collaboration capabilities, making it difficult to cope with the 0.1 - 0.5 mm - level warping deformation of castings caused by material shrinkage or mold wear. Third, traditional five - axis robots rely on preset trajectories to perform grinding tasks and cannot dynamically perceive the actual geometric features (such as sudden curvature changes, grooves, protrusions, etc.) of castings, resulting in deviations between the tool posture and the surface normal of the curved surface, affecting the surface treatment quality. In addition, the problem of insufficient adaptive ability for complex curved surfaces is particularly prominent. In the fixed - trajectory mode, the system cannot adjust the grinding parameters in real time according to the local morphology, resulting in frequent over - cutting or under - machining phenomena in the edge area. Summary of the Invention

[0004] The purpose of the present invention is to propose a vision - based grinding method for the mold - closing line of castings, which can improve the grinding accuracy, grinding quality, and grinding efficiency, and is also applicable to complex curved surfaces.

[0005] The present invention is realized through the following technical solutions: A vision - based grinding method for the mold - closing line of castings, comprising the following steps: Step S1: Obtain the template surface point cloud of a standard casting whose mold - closing line has been ground, perform reverse modeling on the template surface point cloud to obtain the first template point cloud, and obtain the measured surface point cloud of the casting to be ground as the first measured point cloud; Step S2: Coarsely match and finely match the first template point cloud successively according to the feature relationship between the first template point cloud and the first measurement point cloud. Convert the original mold closing line model into a new mold closing line model according to the transformation relationship corresponding to the coarse match and the fine match. Among them, the original mold closing line model includes the teaching trajectory and its neighborhood point cloud. The teaching trajectory is formed in the first template point cloud through manual teaching. The new mold closing line model includes the template trajectory converted from the teaching trajectory. Step S3: Respectively obtain the points closest to each trajectory point of the template trajectory in the new mold closing line model. For each point closest to the distance, obtain the point with the largest vertical distance from its neighborhood to the plane where the new mold closing line model is located. The path planning point cloud is formed by all the points with the largest distance. And use the vertical distance corresponding to each point in the machining trajectory as the machining allowance of this point. Step S4: Traverse all the planned points in the path planning point cloud according to the machining direction. Select the machining method according to the machining allowance corresponding to each planned point and generate the final machining trajectory. The machining methods include one-time milling, two-time milling or multiple-time milling. One-time milling adopts the fast feed mode, two-time milling adopts rough machining and finish machining in sequence, and multiple-time milling adopts layer machining.

[0006] Further, in the step S1, the standard casting is set on the programmable turntable. Rotate the programmable turntable at a set angle. After the i th rotation, obtain the original point cloud of the standard casting , and according to the formula fuse all the collected original point clouds to obtain the template surface point cloud , where represents the product, represents the i th rotation corresponding rotation matrix, represents the i th rotation angle, T ( d i ) represents the translation matrix, is located in the turntable coordinate system.

[0007] Further, in the step S1, the reverse modeling specifically includes the following steps: Step S11: Construct the Poisson equation , and use the adaptive octree grid discretization strategy to solve the Poisson equation to obtain the implicit function , where is the Laplace operator, v is the normal vector field obtained by extending the normal vector set of the template surface point cloud to the three-dimensional space through the moving least squares method, is the gradient operator; Step S12. Initial grid extraction: Dynamically adjust the grid size according to the point cloud density of the template surface points. The size of the extracted initial grid is expressed as , where x represents the normal vector of any point in the three-dimensional space, d max represents the set maximum grid size, β represents the proportionality coefficient, ρ (x) represents the point cloud density; Step S13. Non-uniform grid subdivision: Dynamically divide the subdivision levels according to the subdivision criterion and perform Loop subdivision on the area. The number of subdivision times is limited to , where represents the local curvature of the grid vertex, G represents the gradient modulus length of the grid, represents the maximum local curvature of all vertices in the grid, G max represents the maximum gradient modulus length of all vertices in the grid, represents the set subdivision threshold, represents the preset maximum fixed-point density, d current represents the current vertex density; Step S14. Adaptive Laplacian smoothing: Iteratively update the position of the subdivided grid vertex p i to obtain the first template point cloud. The formula for the k (i + 1)-th iteration is , where λ is the step size for controlling the smoothing intensity, represents the neighborhood of p i , p j represents the point in the neighborhood, w ij represents the adaptive weight.

[0008] Furthermore, step S2 specifically includes the following steps: Step S21. Based on the feature points or feature descriptors of the extracted first template point cloud M and the first measurement point cloud T , calculate the first transformation matrix H 1 = { R 1, t 1} through singular value decomposition. According to the formula , perform rough matching on the first template point cloud M to obtain the second template point cloud ; Step S22. From the template point cloud Identify the original mold closing line model M line The corresponding point cloud set, remove the part in the first measured point cloud T corresponding to this point cloud set to obtain the second measured point cloud T 1. Use the iterative closest point algorithm to finely match the second template point cloud with the second measured point cloud T 1 to obtain the third template point cloud, and generate the second transformation matrix H 2 = { R 2, t 2}; Step S23. According to the formula Convert the original mold closing line model M line into the intermediate mold closing line model M lineSrc , and perform statistical filtering on the outlier points in the intermediate mold closing line model M lineSrc to obtain the new mold closing line model , where R 1, R 2 respectively represent the rotation matrices in the rough matching and fine matching processes, t 1, t 2 respectively represent the translation matrices in the rough matching and fine matching processes.

[0009] Furthermore, obtaining the machining allowance in step S3 includes the following steps: Step S31. According to the formula Obtain the point in the new mold closing line model m that is closest to the p m th trajectory point of the template trajectory q m_nearest ; Step S32. The equation of the plane where the new mold closing line model is located is , and the perpendicular distance from the point q m_nearest in the neighborhood of the point q near = ( q x , q y , q z ) to this plane is , and the projection point q near of the point q proj on this plane is , and select the point with the largest perpendicular distance in the neighborhood as the highest pointq max , the path planning point cloud is formed by each highest point, and the highest point q max and its corresponding projection point q proj The distance between them is used as the machining allowance of this point , where is the normal vector of the plane where the new mold closing line model is located, is the reference point on this plane.

[0010] Furthermore, in the step S4, the specific method of selecting the machining method according to the machining allowance corresponding to each trajectory point includes: If , select one-time milling; if , select two-time milling; if , select multiple-time milling, where T 1 and T 2 are user-adjustable parameters, and the selection ranges are 2-4mm and 4-6mm respectively.

[0011] Furthermore, in the step S4, the layer machining is to decompose the machining task into multiple layers, and each layer is gradually milled according to a predetermined thickness h , and the cooling time is arranged after each milling. The machining parameters for each layer are , where F 3l represents the feed rate, S 3l represents the spindle speed, l represents the number of machining layers, △ F represents the feed rate adjustment amount, △ S represents the spindle speed adjustment amount.

[0012] Furthermore, in the step S4, generating the final machining trajectory specifically includes: Step S41, initialize the set for storing the finally generated paragraphs and the set for representing the currently processed paragraph ; Step S42, traverse all the planning points in the path planning point cloud according to the machining direction. If the i 'th planning point Tr i' has the same machining allowance as the previous planning point Tr i'-1 , then add the planning point Tr i'-1 to the set S, that is, S = S ∪ { Tr i'-1}; If the current set S is empty or the current planning pointTr i' The machining allowance is different from the previous planning point Tr i'-1 If the machining allowance is different, the current set S is incorporated into the set Segs, and the planning point Tr i'-1 is added to the new set S, that is, Segs = Segs ∪ {S}, S = { Tr i'-1}; After the traversal, if the current segment set S is not empty, the set S is added to the set Segs, and the final machining trajectory is the final set Segs.

[0013] Furthermore, in step S23, the intermediate mold closing line model M lineSrc is statistically filtered for outliers to obtain a new mold closing line model Specifically, it includes: Step S231: For the M lineSrc th point in the intermediate mold closing line model m ', according to the formula calculate the average distance from this point to its k ' nearest neighbors , and according to the formulas and calculate the global average distance and the standard deviation , where is the neighborhood formed by k ' nearest neighbors, is the k 'th neighbor, represents the distance calculation, M represents the intermediate mold closing line model M lineSrc ; Step S232: Points that satisfy are outliers, filter the outliers to obtain a new mold closing line model , where is the multiple threshold of the average distance.

[0014] Furthermore, in step S1, after the i th rotation, the binocular 3D camera is used to collect the point cloud of the standard casting as , and using the formula transform the point cloud in the camera coordinate system to the turntable coordinate system to obtain the original point cloud , where is the known turntable calibration matrix.

[0015] The present invention has the following beneficial effects: 1. First, the present invention obtains the template surface point cloud of the standard casting whose mold joint line has been polished, performs reverse modeling on the template surface point cloud to obtain the first template point cloud, obtains the measured surface point cloud of the casting to be polished as the first measurement point cloud, then performs rough matching and fine matching on the first template point cloud in sequence according to the feature relationship between the first template point cloud and the first measurement point cloud, converts the original mold joint line model into a new mold joint line model according to the transformation relationship corresponding to the rough matching and fine matching, then respectively obtains the points in the new mold joint line model that are closest to each trajectory point of the template trajectory, for each closest point, obtains the point with the largest vertical distance from the plane where the new mold joint line model is located within its neighborhood, forms a path planning point cloud from all the points with the largest distances, and takes the vertical distance corresponding to each point in the machining trajectory as the machining allowance of this point. Finally, traverses all the planned points in the path planning point cloud according to the machining direction, selects the machining method according to the machining allowance corresponding to each planned point, and generates the final machining trajectory, which can more accurately reconstruct the three-dimensional morphology of the casting surface, thereby reducing the positioning error, having strong real-time collaboration ability, being able to cope with the warpage deformation of 0.1 - 0.5 mm level caused by material shrinkage or mold wear of the casting, paying attention to the machining allowance to dynamically perceive the actual geometric features of the casting, avoiding the deviation of the tool posture from the surface normal matching, and finally realizing the improvement of the mold joint line polishing accuracy, polishing quality and polishing efficiency of the casting. 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 first template point cloud for reverse modeling of the present invention.

[0019] Figure 3 is a schematic diagram of the second template point cloud of the present invention.

[0020] Figure 4 is a schematic diagram of the second measurement point cloud of the present invention.

[0021] Figure 5 is a schematic diagram of the third template point cloud of the present invention.

[0022] Figure 6 is a machining flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] As Figure 1 shown, the vision-based mold joint line polishing method for castings includes the following steps: Step S1: Obtain the template surface point cloud of the standard casting after the mold closing line has been polished. Perform reverse modeling on the template surface point cloud to obtain the first template point cloud. Obtain the measured surface point cloud of the casting to be polished as the first measurement point cloud. In this embodiment, the casting is set on the turntable of the programmable turntable. A binocular 3D camera is used to photograph the casting, and a working five-axis robot is used to polish the mold closing line of the casting. The calibration of the turntable of the programmable turntable uses existing technologies, such as the calibration method disclosed in Patent CN202411983146.2. The eye-to-hand calibration of the five-axis robot and the turntable uses the Eye-to-Hand eye-outside-hand calibration mode, and equations are established through a checkerboard calibration board to solve the hand-eye matrix , where represents the homogeneous transformation matrix from the end-effector coordinate system of the robot to the base coordinate system, which is a known quantity and can be calculated in real time through the forward kinematics model of the robot. represents the homogeneous transformation matrix from the camera coordinate system to the end-effector coordinate system of the robot, which is an unknown quantity. represents the homogeneous transformation matrix from the camera coordinate system to the base coordinate system of the robot, which is a known quantity and can be calculated from the images of the checkerboard calibration board in multiple poses and the robot pose data.

[0024] The standard casting is set on the programmable turntable. Each time, the programmable turntable is rotated by a set angle . After the i th rotation, the binocular 3D camera is used to collect the point cloud of the standard casting as . Using the formula , the point cloud in the camera coordinate system is transformed to the turntable coordinate system to obtain the original point cloud , where is the known turntable calibration matrix. According to the formula , multi-view fusion is performed on all the collected original point clouds to obtain the template surface point cloud , where represents the product of consecutive multiplications, transforming each piece of point cloud to the position of the first piece of point cloud. represents the rotation matrix corresponding to the i th rotation. represents the i th rotation angle. T ( d i ) represents the translation matrix (i.e., the translation compensation amount) corresponding to the i th rotation, and is in the turntable coordinate system.

[0025] The process of obtaining the first measurement point cloud is the same as the process of obtaining the first template point cloud.

[0026] Since there may be slight pose errors in the data obtained from multi-view fusion, in order to reduce the errors, the point cloud on the template surface obtained by fusion is precisely registered using the neighborhood matching algorithm. The number of iterations of the neighborhood matching algorithm is set to 300 times, the convergence threshold is 0.01 mm, and the error in the overlapping area of the point cloud is less than 0.1 mm.

[0027] Inverse modeling is performed using improved Poisson reconstruction, which can accelerate the inverse reconstruction efficiency and improve the reconstruction effect. The specific steps are as follows: Step S11: Construct the Poisson equation , and solve the Poisson equation using the adaptive octree grid discretization strategy to obtain the implicit function , where is the Laplace operator, v is the normal vector field obtained by extending the normal vector set of the template surface point cloud to the three-dimensional space using the moving least squares method, is the gradient operator; The adaptive octree grid discretization strategy for solving the Poisson equation is a prior art. In this embodiment, the maximum depth is set to 10 layers, and the point cloud density threshold is 0.1 points / mm²; the grid is only encrypted near the point cloud to reduce waste of computing resources; the normal vector confidence weight is also introduced to optimize the robustness of the solution in the noise area, P represents the set of template surface point clouds, represents the distance from the spatial point x to the nearest point cloud point, is the adjustment parameter; Step S12: Use the improved Marching Cubes algorithm to extract the initial triangular mesh from the implicit function , and enable high-precision interpolation in the area where the gradient magnitude is greater than the threshold. Specifically, for the initial mesh extraction: the size of the initial mesh extracted is dynamically adjusted according to the point cloud density of the template surface point cloud, and is expressed as , and the phenomenon of holes or fragmentation is avoided through the topology preservation criterion, where x represents the normal vector of any point in the three-dimensional space, d max represents the set maximum mesh size, β represents the proportionality coefficient, ρ (x) represents the point cloud density; Step S13: Non-uniform grid subdivision: Dynamically divide the subdivision levels according to the subdivision criterion , where the larger the value, the higher the required subdivision level. Perform Loop subdivision on the area, and the number of subdivision times is limited to , where represents the local curvature of the grid vertex, GRepresents the gradient magnitude of the mesh, represents the maximum local curvature of all vertices in the mesh, G max represents the maximum gradient magnitude of all vertices in the mesh, represents the set subdivision threshold, and its specific value is dynamically adjusted according to the mesh complexity, represents the preset maximum fixed-point density, d current represents the current vertex density; Loop subdivision is an existing recursive subdivision surface algorithm for triangular meshes. The core goal is to generate a continuous and smooth surface by iteratively subdividing the mesh and adjusting the vertex positions, while preserving the topological structure and features of the original mesh.

[0028] Step S14, Adaptive Laplacian smoothing: For the vertex p of the subdivided mesh i Iteratively update the position to obtain the first template point cloud. The k (i + 1)-th iteration formula is , and the vertices participating in the smoothing satisfy the normal angle . After smoothing, the sharp edge angle deviation ≤ 1.5°, which is better than 3.2° of the traditional method. Among them, is the step size for controlling the smoothing intensity and needs to be adjusted according to the mesh stability, represents the neighborhood of p i , p j represents the point in the neighborhood, represents the adaptive weight, and respectively represent the local curvatures of the mesh vertices p i and p j , is the curvature difference tolerance threshold, controls the weight decay speed. The greater the curvature difference, the smaller the weight.

[0029] The first template point cloud obtained by reverse modeling is as Figure 2 shown, which significantly improves the reconstruction quality and computational efficiency of complex geometries while maintaining sharp features.

[0030] Step S2, According to the feature relationship between the first template point cloud and the first measurement point cloud, perform rough matching and fine matching on the first template point cloud in sequence. According to the transformation relationship corresponding to the rough matching and fine matching, convert the original mold closing line model into a new mold closing line model. Among them, the original mold closing line model includes the teaching trajectory and its neighborhood point cloud, the teaching trajectory is formed by manual teaching in the first template point cloud, and the new mold closing line model includes the template trajectory converted from the teaching trajectory; Specifically, it includes the following steps: Step S21, Based on the extracted first template point cloudM and the first measured point cloud T of the feature points or feature descriptors, calculate the first transformation matrix through singular value decomposition H 1 = { R 1, t 1}, according to the formula perform rough matching on the first template point cloud M to obtain the second template point cloud ; Feature points refer to the points in the point cloud with significant geometric or statistical characteristics (such as corner points, edge points, high curvature points), and feature descriptors refer to the vectors describing the geometric attributes of the local neighborhood of the feature points (such as normal distribution, curvature histogram). In this embodiment, by using the FPFH (Fast Point Feature Histograms) algorithm, feature descriptors are generated by calculating the geometric statistics of the local neighborhood of the point cloud (such as normal angle, distance distribution). This process is prior art, and the process of calculating the first transformation matrix through singular value decomposition is also prior art; The second template point cloud is as shown in Figure 3 ; Rough matching aligns the first template point cloud M and the first measured point cloud T as much as possible on a large scale, laying a foundation for subsequent finer processing; Step S22, identify the point cloud set corresponding to the original mold closing line model from the template point cloud M line , remove the part of the first measured point cloud T corresponding to this point cloud set to obtain the second measured point cloud as shown in Figure 4 1, and use the existing iterative closest point algorithm to perform fine matching on the second template point cloud T and the second measured point cloud 1 to obtain the third template point cloud as shown in T . A second transformation matrix Figure 5 2 = { H 2, R 2, t 2} is generated during this process. Through fine matching, the local consistency between models can be significantly improved; Step S23, according to the formula convert the original mold closing line model M line to the intermediate mold closing line model M lineSrc , and perform statistical filtering on the outlier points in the intermediate mold closing line model M lineSrc to obtain the new mold closing line model , where R 1, R 2 respectively represent the rotation matrices in the rough matching and fine matching processest 1. t 2 represents the translation matrix of the coarse matching and fine matching process respectively; Among them, the middle parting line model M lineSrc The outliers in the matrix are statistically filtered to obtain a new parting line model Specifically include: Step S231: middle parting line model M lineSrc The m ' points, according to the formula Calculate the point to k The average distance of the nearest neighbors , and according to the formula and Calculate the global average distance and standard deviation ,in, Reason k 'The neighborhood formed by the nearest neighbors, For the k 'A neighbor, Indicates the distance. Represents the middle parting line model M lineSrc the number of midpoints; Step S232: Satisfy The points are outliers, and the outliers are filtered to obtain the new parting line model ,in, The threshold is a multiple of the distance from the mean.

[0031] Step S3, respectively obtaining the points in the new parting line model that are closest to each track point of the template track, and for each point with the closest distance, obtaining the point in its neighborhood with the largest vertical distance to the plane where the new parting line model is located, forming a path planning point cloud from all the points with the largest distance, and taking the vertical distance corresponding to each point in the processing track as the processing allowance of the point; Obtaining machining allowance specifically includes the following steps: Step S31: According to the formula Get New Parting Line Model The first m Track points p m The nearest point q m_nearest , Express request p m and q m The Euclidean distance between Step S32. The equation of the plane where the new mold closing line model lies is , and the perpendicular distance from the point q m_nearest within the neighborhood of the point q near = ( q x , q y , q z ) to this plane is , and the projection point q near of the point q proj on this plane is . Select the point with the maximum perpendicular distance within the neighborhood of the point q m_nearest as the highest point q max . From the new mold closing line model , form a path planning point cloud with the highest points of the points closest to each trajectory point of the template trajectory respectively. After converting the path planning point cloud from the turntable coordinate system to the camera coordinate system and then to the robot coordinate system, the distance between the highest point q max and its corresponding projection point q proj m_nearest is used as the machining allowance of this point. Among them, is the normal vector of the plane where the new mold closing line model lies, is the reference point on this plane, and the radius of the neighborhood of the point q m_nearest i' is 5 - 10 mm; The process of obtaining the equation of the plane where the new mold closing line model lies is prior art; Step S4. Traverse all the planning points in the path planning point cloud according to the machining direction, select the machining method according to the machining allowance corresponding to each planning point, and generate the final machining trajectory. The machining methods include one - time milling, two - time milling, or multiple - time milling. One - time milling adopts the fast - feed mode, two - time milling adopts rough machining and finish machining in sequence, and multiple - time milling adopts layer - by - layer machining; Specifically, generating the final machining trajectory includes the following steps: Step S41. Initialize the set for storing the finally generated paragraphs and the set for representing the currently processed paragraph; i Step S42. Traverse all the planning points in the path planning point cloud according to the machining direction. If the machining allowance of the Tr 'th planning point Tr i' is different from that of the previous planning point Tri'-1 If the machining allowance is the same, the planned point Tr i'-1 is added to the set S, that is, S = S ∪ { Tr i'-1}; if the current set S is empty or the machining allowance of the current planned point Tr i' is different from that of the previous planned point Tr i'-1 , the current set S is incorporated into the set Segs and the planned point Tr i'-1 is added to the new set S, that is, Segs = Segs ∪ {S}, S = { Tr i'-1}; after the traversal, if the current segment set S is not empty, the set S is added to the set Segs, and the final machining trajectory is the final set Segs. The final machining trajectory is converted into G-code to accurately guide the five-axis robot to perform the predetermined machining task.

[0032] For the machining method: If , one milling is selected, and the rapid feed mode is adopted. The feed speed F v = 1000 mm / min, and the spindle speed S n = 8000 rpm. The path is planned directly using the rapid feed mode or the shortest path algorithm. This ensures an efficient and direct machining process, reduces unnecessary tool movement, and improves machining efficiency; If , two milling operations are selected, which are divided into rough machining and finish machining steps. First, rough machining is performed to remove most of the material, and then it is adjusted to the finish machining mode to complete the final forming. This can ensure machining accuracy and surface quality while reducing workpiece stress. Among them, the rough machining parameters are: feed speed F 2a v = 800 mm / min, and the spindle speed S 2a n = 6000 rpm; the finish machining parameters: F 2b v = 1200 mm / min, S 2b n = 10000 rpm; If , multiple milling operations are selected, and the layer machining strategy is adopted. Layer machining decomposes the machining task into multiple layers, and each layer is milled step by step according to the predetermined thickness h t = 1 mm. A cooling time is arranged after each milling. The machining parameters for each layer are , where F 3l represents the feed speed,S 3l represents the spindle speed, l represents the number of machining layers, F base is the initial feed rate, S base is the initial spindle speed, △ F represents the feed rate adjustment amount, △ S represents the spindle speed adjustment amount.

[0033] Among them, T 1 and T 2 are user-adjustable parameters, and the default values are 3 mm and 5 mm respectively.

[0034] Before generating the G-code, it is first necessary to further optimize the machining trajectory, such as removing redundant paths and eliminating repeated or unnecessary movements; considering tool radius compensation and adjusting the path according to the tool size to ensure that no collision or undercut occurs, etc.

[0035] For paragraphs with different machining allowances, corresponding G-code blocks are generated. Set the corresponding machining parameters (such as feed rate, spindle speed) at the beginning of each paragraph, and restore the default values or prepare the parameters for the next paragraph at the end.

[0036] The specific machining process is as Figure 6 shown.

[0037] As mentioned above, it is only a preferred embodiment of the present invention, so the scope of implementation of the present invention cannot be limited thereby. 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 vision-based method for grinding the mold closing line of castings, characterized in that: It includes the following steps: Step S1: Obtain the template surface point cloud of the standard casting with the mold closing line polished, perform reverse modeling on the template surface point cloud to obtain the first template point cloud, and obtain the measured surface point cloud of the casting to be polished as the first measurement point cloud; Step S2: Coarsely match and finely match the first template point cloud in sequence according to the feature relationship between the first template point cloud and the first measurement point cloud, and convert the original mold closing line model into a new mold closing line model according to the transformation relationship corresponding to the coarse match and the fine match. Among them, the original mold closing line model includes the teaching trajectory and its neighborhood point cloud, the teaching trajectory is formed in the first template point cloud through manual teaching, and the new mold closing line model includes the template trajectory converted from the teaching trajectory; Step S3: Respectively obtain the points in the new mold closing line model that are closest to each trajectory point of the template trajectory. For each point closest in distance, obtain the point with the largest vertical distance from the plane where the new mold closing line model is located within its neighborhood. The path planning point cloud is formed by all the points with the largest distance, and the vertical distance corresponding to each point in the machining trajectory is used as the machining allowance of that point; Step S4: Traverse all the planning points in the path planning point cloud according to the machining direction, select the machining method according to the machining allowance corresponding to each planning point and generate the final machining trajectory. The machining methods include one-time milling, two-time milling or multiple-time milling. One-time milling adopts the fast feed mode, two-time milling adopts rough machining and finish machining in sequence, and multiple-time milling adopts layer-by-layer machining.

2. The method for grinding the mold clamping line of a casting based on vision according to claim 1, wherein: In the step S1, the standard casting is arranged on the programmable turntable, and the programmable turntable is rotated at a set angle. After the i th rotation, the original point cloud of the standard casting is obtained . According to the formula , all the collected original point clouds are fused to obtain the surface point cloud of the template . Among them, represents the continuous multiplication, represents the rotation matrix corresponding to the i th rotation, represents the angle of the i th rotation, T ( d i ) represents the translation matrix, is in the turntable coordinate system.

3. The method for grinding the mold clamping line of a casting based on vision according to claim 2, wherein: In the step S1, the reverse modeling specifically includes the following steps: Step S11: Construct the Poisson equation , and use the adaptive octree grid discretization strategy to solve the Poisson equation to obtain the implicit function , where is the Laplace operator, v is the normal vector field obtained by extending the normal vector set of the template surface point cloud to the three-dimensional space through the moving least squares method, is the gradient operator; Step S12, Initial grid extraction: Dynamically adjust the grid size according to the point cloud density of the template surface points. The size of the initially extracted grid is expressed as , where x represents the normal vector of any point in the three-dimensional space, d max represents the set maximum grid size, β represents the proportionality coefficient, ρ (x) represents the point cloud density; Step S13, non-uniform grid subdivision: According to the subdivision criterion dynamically divide the subdivision levels, and perform Loop subdivision on the region, and the number of subdivision times is limited to , where represents the local curvature of the grid vertex, G represents the gradient norm of the grid, represents the maximum local curvature of all vertices in the grid, G max represents the maximum gradient norm of all vertices in the grid, represents the set subdivision threshold, represents the preset maximum fixed-point density, d current represents the current vertex density; Step S14, Adaptive Laplacian Smoothing: For the mesh vertex p after subdivision i Iteratively update the position to obtain the first template point cloud. The k iteration formula for the (i + 1)-th iteration is , where λ is the step size for controlling the smoothing intensity, denotes the neighborhood of p i , p j denotes the points in the neighborhood, w ij denotes the adaptive weight.

4. A vision-based grinding method for the mold closing line of a casting according to claim 3, characterized in that: The step S2 specifically includes the following steps: Step S21, based on the extracted first template point cloud M and the first measurement point cloud T feature points or feature descriptors, calculate the first transformation matrix H 1 = { R 1, t 1} according to the formula to perform rough matching on the first template point cloud M to obtain the second template point cloud ; Step S22: Identify the original mold closing line model from the template point cloud M line and the corresponding point cloud set, and remove the part in the first measurement point cloud T corresponding to this point cloud set to obtain the second measurement point cloud T 1. Use the iterative closest point algorithm to perform fine matching between the second template point cloud and the second measurement point cloud T 1 to obtain the third template point cloud, and generate the second transformation matrix H 2 = { R 2, t 2}; Step S23. According to the formula transform the original mold - closing line model M line into an intermediate mold - closing line model M lineSrc , and perform statistical filtering on the outlier points in the intermediate mold - closing line model M lineSrc to obtain a new mold - closing line model , where R 1. R 2 respectively represent the rotation matrices in the rough matching and fine matching processes, t 1. t 2 respectively represent the translation matrices in the rough matching and fine matching processes.

5. A vision-based method for grinding the mold closing line of a casting according to claim 4, characterized in that: In the step S3, obtaining the machining allowance includes the following steps: Step S31. According to the formula obtain the new mold closing line model and the m th trajectory point p m of the template trajectory that is the closest q m_nearest ; Step S32. The equation of the plane where the new die - joint line model lies is , and the perpendicular distance from the point q m_nearest in the neighborhood of the point q near =( q x , q y , q z ) to this plane is . The projection point q near of the point q proj on this plane is . Select the point with the maximum perpendicular distance in the neighborhood as the highest point q max . The path - planning point cloud is formed by each highest point. The distance between the highest point q max and its corresponding projection point q proj is used as the machining allowance of this point. Among them, is the normal vector of the plane where the new die - joint line model lies, and is the reference point on this plane.

6. A vision-based grinding method for the mold clamping line of castings according to claim 5, characterized in that: In the step S4, selecting the machining method according to the machining allowance corresponding to each trajectory point specifically includes: If , select single milling; if , select double milling; if , select multiple milling, where T 1 and T 2 are user-adjustable parameters, and the selection ranges are 2 - 4 mm and 4 - 6 mm respectively.

7. A vision-based grinding method for the mold closing line of castings according to claim 6, characterized in that: In the step S4, the layer machining is to decompose the machining task into multiple layers, and each layer is gradually milled according to a predetermined thickness. After each milling, a cooling time is arranged, and the machining parameters for each layer are h , where , F 3l represents the feed rate, S 3l represents the spindle speed, l represents the number of machining layers, △ F represents the feed rate adjustment amount, △ S represents the spindle speed adjustment amount.

8. A vision-based grinding method for the mold clamping line of a casting according to claim 7, characterized in that: In the step S4, generating the final machining trajectory specifically includes: Step S41: Initialize a set for storing the finally generated paragraphs and a set for representing the paragraphs currently being processed ; Step S42: Traverse all the planned points in the path planning point cloud according to the machining direction. If the machining allowance of the i 'th planned point Tr i' is the same as the machining allowance of the previous planned point Tr i'-1 , add the planned point Tr i'-1 to the set S, i.e., S = S ∪ { Tr i'-1}; if the current set S is empty or the machining allowance of the current planned point Tr i' is different from the machining allowance of the previous planned point Tr i'-1 , incorporate the current set S into the set Segs and add the planned point Tr i'-1 to the new set S, i.e., Segs = Segs ∪ {S}, S = { Tr i'-1}; after the traversal, if the current segment set S is not empty, add the set S to the set Segs, and the final machining trajectory is the final set Segs.

9. A vision-based grinding method for the mold closing line of a casting according to any one of claims 4 to 8, characterized in that: In the step S23, the intermediate mold clamping line model M lineSrc is statistically filtered for outliers to obtain a new mold clamping line model Specifically, it includes: Step S231: For the i-th point in the intermediate mold clamping line model M lineSrc , calculate the average distance from this point to its m ' nearest neighbors according to the formula . And calculate the global average distance k and the standard deviation respectively according to the formulas and , where is the neighborhood formed by the ' nearest neighbors, is the k ' neighbor, represents the i-th k neighbor, represents calculating the distance, M represents the intermediate mold clamping line model M lineSrc , and n is the number of points in the middle point; Step S232: Points that satisfy are outliers. Filter the outliers to obtain a new mold closing line model , where is the multiple threshold of the average distance.

10. A vision-based grinding method for the mold closing line of a casting according to any one of claims 2 to 8, characterized in that: In the step S1, after the i -th rotation, the binocular 3D camera is used to collect the point cloud of the standard casting as . Using the formula , the point cloud in the camera coordinate system is transformed into the turntable coordinate system to obtain the original point cloud , where is the known calibration matrix of the turntable.

Images