A point cloud-based path planning method for robot column-nail roller surfacing welding

Abstract

The present invention relates to the field of intelligent welding manufacturing technology, in particular to a point cloud-based robot column nail roller surfacing path planning method. The surfacing path is projected onto the column nail roller cylinder center axis twice to complete the order of the centroid of the ceramic column. The upper and lower edges of the two adjacent ceramic columns on the same column are the end point and the starting point of the segmented welding path respectively. The small segments of the welding path are connected in sequence to finally form an automatic welding trajectory. The key features of the weld are segmented and extracted by a deep learning algorithm. The characteristic that the ceramic column protrudes from the weld surface in a stepped shape is used. The model senses the height size of the data and improves the segmentation accuracy. The column nail roller center axis is obtained by fitting a three-dimensional circle in space by column and then fitting a straight line from the center of each circle. The centroid is sorted by two auxiliary projections, and the planning of the surfacing path is finally realized. When the work content changes, there is no need for teaching, which reduces the workload of the welding operator. It has a wide range of applications and improves the stability of the surfacing quality.

Description

Technical Field

[0001] The present invention relates to the field of intelligent welding manufacturing technology, and in particular to a point cloud-based robot column nail roller surfacing path planning method. Background Art

[0002] The roller press is a highly efficient grinding device that plays a key role in the processing of ores, cement, and metals. It is an energy-saving grinding device and an intermediate device in cement grinding and metal ore crushing. It is a key device that achieves more crushing and less grinding, significantly reducing energy consumption. It comprises a feeder, an articulated frame, a movable roller, a fixed roller, a roller body, and a roller sleeve. The articulated frame is equipped with two roller mechanisms with studded roller surfaces: one is a fixed roller, and the other is a movable roller that can reciprocate horizontally. Several carbide studs are evenly distributed on the roller surface of the roller body.

[0003] Currently, traditional overlay welding methods utilize a combination of semi-automatic and manual methods. Semi-automatic welding refers to the use of welding robots. Currently, most robots used in industrial production are teach-and-play and offline programmable. Teach-and-play robots require a welder to teach them once, allowing the robot to accurately repeat each taught step over and over again. Compared to manual welding, they offer unparalleled advantages in completing highly repeatable and high-intensity welding tasks. However, the level of automation employed by this welding method remains low. Changes in the workload (e.g., welding stud rollers of different sizes) require the welder to teach them again. Furthermore, since teach-and-play robots lack the ability to send or receive feedback for real-time adjustments, errors in the fabrication or assembly of the weldment after the robot's teaching can lead to unacceptable deviations between the weld trajectory and the taught trajectory during re-creation, resulting in poor weld quality and further increased production costs. Furthermore, teach-and-play robots rely heavily on manual mapping, making them difficult to meet requirements for complex tasks (e.g., those with a large number of ceramic stud rollers).

[0004] Manual welding, on the other hand, is labor-intensive and often involves harsh working environments. During the welding process, welders must constantly monitor the position of the welding torch and the weld seam, adjusting the torch's position and angle to align it with the weld seam. Obviously, the quality of traditional welding often depends on the operator's experience and skill, resulting in inconsistent weld quality from one worker to the next, making it difficult to guarantee consistent and reliable welding quality.

[0005] In addition, welding workshops inevitably produce welding smoke, harmful gases, noise, light pollution and other factors that endanger the health of welding operators. Simple protective measures can only reduce the harm to operators to a certain extent, but cannot completely eliminate the damage caused by pollution.

[0006] Therefore, the combination of semi-automatic and manual methods used in traditional cladding welding has problems such as long teaching time, heavy workload, long production cycle, poor adaptability to changing scenarios, high welder costs, difficulty in recruiting welders, and poor stability of cladding quality. Summary of the Invention

[0007] The technical problem to be solved by the present invention is: in order to solve the problems existing in the existing technology in the above-mentioned background technology, an improved method for improving model performance and segmentation accuracy is provided, in which the central axis of the column nail roller is obtained by first fitting a three-dimensional circle in space by column and then fitting a straight line from the center of each circle; and the sorting of the center of mass is completed through two auxiliary projections, and finally a point cloud-based robot column nail roller welding path planning method is realized.

[0008] The technical solution adopted by the present invention to solve the technical problem is: a point cloud-based robot column nail roller surfacing path planning method, comprising the following steps:

[0009] Step 1: Image acquisition: Use a depth camera to capture images of the stud-roller weldment and characterize them using point cloud data;

[0010] Step 2: Extract weldment features. Segment the column-nail-roller weldment collected in step 1 to extract the ceramic column.

[0011] Step 2.1: De-noise the captured image using a combination of straight-through filtering and Gaussian filtering. Specifically, the straight-through filtering removes noise far from the stud-roller weldment, and the Gaussian filtering removes outliers on the stud-roller weldment surface.

[0012] Step 2.2: Use deep learning to segment the denoised point cloud image and extract the ceramic columns;

[0013] Step 3: Build-up welding path planning, specifically: the build-up welding path is projected onto the central axis of the column roller cylinder twice to complete the order of the center of mass of the ceramic column. The upper and lower edges of the two adjacent ceramic columns in the same column are the end and starting points of the segmented welding path respectively. The small segments of the welding path are connected in sequence to finally form an automatic welding trajectory.

[0014] Step 3.1: Project the centroid of the ceramic column onto the central axis of the column spike roller. The centroids in the same column will be clustered together. Cluster the projections on the central axis. The total number of categories is the number of columns. Compare the coordinates of the centroid projections in each category to obtain the column ranking.

[0015] Step 3.2: Project the centroid of the ceramic column onto a plane perpendicular to the central axis of the column spike roller. The centroids in the same row will be clustered together. Cluster the projections. The total number of categories is the number of rows. Sorting the clustered results can also obtain the row ranking.

[0016] Step 3.3: The segmented paths can be planned by sorting the rows in each column.

[0017] Furthermore, the method of fitting the central axis of the column spike roller in step 3 specifically includes the following steps:

[0018] Step 3.1.1: The centroid coordinates of two adjacent points on the same row are spaced by ∆Y. The centroid coordinates are clustered based on the Y value and the threshold value to complete the rough classification of the ceramic pillars.

[0019] Step 3.1.2: Fit a three-dimensional circle to each point in each column;

[0020] Step 3.1.3: Fit a three-dimensional circle to each column;

[0021] Step 3.1.4: Compare the radius of the circle fitted in step 3.1.3 with the actual radius, eliminate the circles with larger errors, and fit a straight line to the centers of the remaining circles. The resulting straight line is the central axis of the column roller cylinder.

[0022] Furthermore, step 2.2 specifically includes the following steps:

[0023] Step 2.2.1: Use CloudCompare software (CloudCompare is a 3D point cloud (mesh) editing and processing software) to annotate the filtered point cloud file, select the ceramic pillars and label them;

[0024] Step 2.2.2: Use the PointNet++ model to train and segment the model. Specifically, based on the PointNet++ model, a training strategy and data enhancement method suitable for ceramic column segmentation are designed. The difference between the z-axis coordinate of each point in the point cloud and the z-axis coordinate of the lowest point in the point cloud is calculated and spliced ​​as a new column of features.

[0025] Furthermore, step 3.1.2 specifically includes the following steps:

[0026] Step 3.1.2.1: Use the least squares method to fit the plane of each column and calculate the normal vector to the plane. Specifically, assume that the coordinates of the center of the circle are C, and take any two points as P1 and P2. Let the vector connecting P1 and P2 be vector1, and the vector formed by the coordinates of the center of the circle and the midpoint of the line connecting P1 and P2 be vector2. If the plane equation vector1·vector2=0 is satisfied, find the minimum value of vector1·vector2.

[0027] Step 3.1.2.2: Use the conditional extreme value solution method to solve the coordinates of the circle center;

[0028] Step 3.1.2.3: Calculate the radius based on the average distance from the center coordinates found in step 3.1.2.2 to each point on the circle.

[0029] Furthermore, in step 3, the order of rows and columns of the ceramic columns after segmentation by two auxiliary projections is sorted. Specifically, projection sorting is used, and welding is performed by column. The points in each sorted column are traversed in turn, and the points in the column are sorted according to the size of the Y coordinate. The two adjacent points are the starting point and end point of a small segment path of the column.

[0030] Furthermore, step 3.4 is included: performing an intersection operation on the two points on the adjacent ceramic columns taken out each time with the centroid clusters on the previous auxiliary surface projection, and the points existing in the column projection cluster and the row projection cluster are actual welding feature points.

[0031] Furthermore, it also includes:

[0032] The coordinates of the centroid of the ceramic column are offset. Specifically, the direction vector of two adjacent centroid points is calculated and offset along this direction to obtain the actual welding point.

[0033] Calculate the vertical direction of the ceramic column by projecting the center of mass of the ceramic column and a point on the central axis of the column nail roller cylinder onto the projection plane. Calculate the direction vector of the two projection points as the vertical direction of the ceramic column. Offset the actual welding point along the direction vector to the path of each welding gun entry and exit. Connect each welding segment path to form a planned path.

[0034] Furthermore, step 4 is also included: verifying whether the path planning method meets the welding accuracy requirements.

[0035] Furthermore, step 4 includes the following steps:

[0036] Step 4.1: The coordinates of the starting point and end point of each segmented path are obtained by offsetting the center of mass of the ceramic column. The length of each segmented path can be calculated based on the coordinates of the two ends of the path;

[0037] Step 4.2: Obtain the coordinates of the welding points of each segment path through manual teaching, and calculate the length of the segment welding path in the teaching mode;

[0038] Step 4.3: Compare the planned path obtained after offsetting the center of mass of the ceramic column obtained in step 4.1 with the taught path obtained by manual teaching in step 4.2. If it is within the error accuracy range, it is qualified; if it is not within the error accuracy range, it is unqualified.

[0039] Furthermore, step 1: image acquisition, specifically: the depth camera is installed at the end of the robotic arm, and the camera and robotic arm need to be calibrated by hand and eye to improve the accuracy of image acquisition; since the size of the column nail roller weldment is much larger than the camera field of view, the image captured by a single shot is a local area of ​​the column nail roller weldment, and it is also necessary to splice the images of different areas of the column nail roller weldment captured by multiple shots to complete the three-dimensional reconstruction of the column nail roller weldment surface. The reconstructed column nail roller weldment surface is the welding surface to be processed.

[0040] The beneficial effects of the present invention are as follows: 1) the key features of the weld are segmented and extracted through a deep learning algorithm, and the characteristic of the ceramic column protruding from the weld surface in a stepped shape is utilized to find appropriate data enhancement methods, so that the model can sense the height size of the data, thereby improving the model performance and the segmentation accuracy, laying the foundation for welding path planning;

[0041] 2) To address the large errors that occur during the fitting process of incomplete cylinders with small curvature, a method is proposed to first fit a three-dimensional spatial circle by column and then fit a straight line from the center of each circle to obtain the central axis of the pin roller. The centroids are then sorted through two auxiliary projections, ultimately achieving the planning of the surfacing path.

[0042] 3) Verify that the path planning method meets the welding accuracy requirements, which is of certain significance for realizing welding automation and improving production efficiency;

[0043] 4) It has a high degree of automation. When the work content changes, there is no need for teaching, which reduces the workload of welding operators. It has a wide range of applications and is suitable for column nail rollers of different sizes and specifications and column nail rollers with different numbers of ceramic columns. It has low welding costs, low operating difficulty, and improves the stability of cladding quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention will be further described below with reference to the accompanying drawings and examples.

[0045] Figure 1 It is a structural schematic diagram of the present invention;

[0046] Figure 2 It is a schematic diagram of the centroid of the ceramic column of the present invention projected onto the centroid of the same row on the central axis of the column nail roller cylinder;

[0047] Figure 3 It is a schematic diagram of the centroid of the ceramic column of the present invention projected onto the centroid of the same row on the central axis of the column nail roller cylinder;

[0048] Figure 4 This is the effect diagram after cylinder fitting of two incomplete cylinder curvatures in the prior art of the present invention;

[0049] Figure 5 This is a schematic diagram of the present invention fitting a three-dimensional circle to each point on each column to obtain the central axis of the cylindrical spike roller;

[0050] Figure 6 It is a schematic diagram of the planned path finally formed by the present invention;

[0051] Figure 7 This is a comparison diagram of the planned path and the taught path of the present invention;

[0052] In the figure: 1. Substrate, 2. Ceramic pillars. DETAILED DESCRIPTION

[0053] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0054] like Figures 1 to 7 A point cloud-based robot column nail roller surfacing path planning method is shown. The structure of the column nail roller is as follows: Figure 1 As shown, the base material of the substrate 1 is cast steel, the height h of the ceramic column 2 is 5-6 mm, and the height h is actually a variable and can be automatically identified. The center-of-mass distance e between two adjacent ceramic columns 2 is 36 mm, and the center-of-mass distance e is actually a variable and can be automatically identified. It is necessary to use a surfacing method to fill the gaps around the ceramic columns. The ceramic columns 2 do not melt during the surfacing.

[0055] First, a depth camera is used to capture images of the stud-roller weldment. The depth camera is installed at the end of the robotic arm. In order to improve the accuracy of image acquisition, the camera and robotic arm need to be calibrated by hand and eye. Since the size of the stud-roller weldment is much larger than the camera's field of view, the image captured by a single shot is a local area of ​​the stud-roller weldment. Therefore, the images of different areas of the stud-roller weldment captured by multiple shots are stitched together to obtain a three-dimensional stitched image. Specifically, since the point cloud obtained by each data acquisition is generated in the camera coordinate system, the camera coordinate system will change with the movement of the camera shooting position. The point cloud images in different camera coordinate systems cannot be directly stitched together to form a complete welding surface. Therefore, the point cloud coordinate system needs to be moved to a fixed reference coordinate system to automatically stitch the point cloud and complete the three-dimensional reconstruction of the welding surface. Since the robotic arm is fixed on a horizontal plane and its base remains fixed during movement, the coordinate transformation matrix obtained from the hand-eye calibration can be used to transform the point cloud coordinates in the camera coordinate system to the robotic arm base coordinate system to complete the point cloud splicing, thereby completing the 3D reconstruction of the stud-roller weldment surface. The reconstructed stud-roller weldment surface is the welding surface to be processed and is represented by point cloud data.

[0056] Then, the surface of the characterized column-nail-roller weldment is segmented to extract the ceramic columns. However, due to the influence of the industrial environment, a large amount of noise is generated during the image acquisition process. Therefore, the collected image is first subjected to noise reduction processing. A combined noise reduction method combining straight-through filtering and Gaussian filtering is used to remove noise far away from the column-nail-roller weldment. Then, Gaussian filtering is used to remove outliers on the surface of the column-nail-roller weldment.

[0057] Secondly, a deep learning method is used to segment and extract ceramic pillars from the denoised point cloud images. This algorithm has good stability. First, the filtered point cloud file is annotated using CloudCompare software (CloudCompare is a three-dimensional point cloud (mesh) editing and processing software). The ceramic pillars are selected and labeled. Then, the PointNet++ model is used (PointNet++ first divides the point cloud set into a number of overlapping local areas through distance measurement, and then extracts local features from the neighborhood, and then extracts higher-level features from these local features. This process is repeated until the global features of the entire point cloud set are obtained) to train and segment it.

[0058] Based on the PointNet++ model, a set of data augmentation methods suitable for ceramic pillar segmentation was designed. Specifically, the data samples were expanded through random rotation and random dithering. Then, point cloud resampling and label smoothing were used to improve the model segmentation accuracy. In addition, considering that ceramic pillar 2 protrudes from the surface of substrate 1 in a stepped shape and has a significant height characteristic, the height appending method was further used for data augmentation. Then, the difference between the z-axis coordinate of each point in the point cloud and the z-axis coordinate of the lowest point in the point cloud was calculated and spliced ​​as a new column of features. This approach aims to make deep learning aware of the height dimension of the data, allowing the model to learn the feature differences between different point sets, laying the foundation for path planning.

[0059] Finally, the surfacing path is projected onto the central axis of the column roller cylinder twice to complete the row and column order of the center of mass of the ceramic columns. The upper and lower edges of the two adjacent ceramic columns in the same column are the end and starting points of the segmented welding path respectively. The small welding paths are connected in sequence to finally form an automatic welding trajectory.

[0060] Since the density of the ceramic columns is uniform and the shape is symmetrical, their centroids are located at the geometric center, and each centroid represents a ceramic column. When the centroids of the ceramic columns are projected onto the central axis of the column nail roller, the centroids in the same column will gather into a cluster, such as Figure 2 As shown, the projections on the central axis are clustered, and the total number of categories is the number of columns. The column sorting can be obtained by comparing the coordinates of the centroid projections of each category. The centroid of the ceramic column is projected onto a plane perpendicular to the central axis of the column nail roller cylinder, and the centroids in the same row will be clustered into a cluster, as shown in Figure 3 As shown, the projections are clustered, and the total number of categories is the number of rows. The row sorting can also be obtained by sorting the clustered results; each column can be used to plan the segmented path according to the sorted rows.

[0061] In the traditional point cloud algorithm library PCL, for cylindrical objects such as stud rollers, the SACMODEL_CYLINGER model provided by the Sample_consensus module is used to fit the cylinder. This method can obtain a point on the cylinder axis, the axis direction vector, and the cylinder radius. Due to the limitations of the camera field of view, each fitting is performed on the incomplete cylinder after 3D reconstruction. When the curvature of the incomplete cylinder in the camera field of view is large, a better fitting effect can be obtained, such as Figure 4 (b). However, if the curvature of the incomplete cylinder is small, it will cause a large error, such as Figure 4 (a) A large error will cause the centroid projections of different rows or columns to gather together, making it impossible to accurately divide the rows and columns, ultimately affecting the planning of the surfacing path.

[0062] In order to prevent the calculation of the central axis of the column spike roller from being affected by the curvature of the incomplete cylinder and to improve the stability of the algorithm, a method for fitting the central axis of the column spike roller cylinder is proposed:

[0063] First, the centroid coordinate interval between two adjacent points on the same row is ∆Y, which is 36 mm. Therefore, the centroid coordinates are clustered based on the Y value and 5 as the threshold to complete the rough classification of ceramic columns.

[0064] Next, fit a three-dimensional circle in space to each point on each column. First, use the least squares method to fit the plane where each column is located, and calculate the normal vector of the plane. Assuming that all points are on the spatial circle, the central axis of the line connecting any two points must pass through the center of the circle. Assuming that the coordinates of the center of the circle are C, and any two points are P1 and P2, let the vector vector1 of the line connecting P1 and P2 be vector1, and the vector formed by the coordinates of the midpoint of the line connecting the center of the circle and P1 and P2 be vector2, then the plane equation vector1·vector2=0 is satisfied. Since the number of points on each column is greater than 3, the equation constructed by this constraint is an overdetermined equation (an overdetermined system of equations refers to a system of equations in which the number of equations is greater than the number of unknowns. For the system of equations Ra=y, R is an n×m matrix. If the R column is full rank and n>m, then the system of equations has no exact solution. In this case, the system of equations is called overdetermined. equations), and since the center of the circle is on the plane fitted by the least squares method, the center of the circle must also satisfy the above plane equations. Finally, the minimum value of vector1·vector2 is solved under the constraints of the plane equations. The coordinates of the center of the circle can be solved through the conditional extreme value (the conditional extreme value is the extreme value under certain additional conditions. The conditional maximum point and the conditional minimum point are collectively referred to as the conditional extreme value points. The function values ​​of the conditional maximum point and the conditional minimum point are the conditional extreme values ​​of the function f(x) under the constraint condition gj(x)). The radius is calculated based on the average distance from the center coordinates to each point on the circle.

[0065] like Figure 5 As shown in (a), a three-dimensional circle is fitted to each column; Figure 5 As shown in (b), the radius of the fitted circle is compared with the actual radius, and the circles with larger errors are eliminated; Figure 5 As shown in (c), the remaining center of the circle is fitted with a straight line, and the resulting straight line is the central axis of the column roller cylinder.

[0066] Projection sorting is used. Since welding is performed in columns, the points in each sorted column are traversed in turn, and the points in the column are sorted according to the size of the Y coordinate. The two adjacent points are the starting point and the end point of a small segment path in the column.

[0067] However, considering that the noise points located above or below the column will also cluster together with the points of the column on the projection line during projection, it is necessary to intersect the two points on the adjacent ceramic columns 2 taken out each time with the centroid clusters on the previous auxiliary surface projection. At the same time, the points existing in the column projection cluster and the row projection cluster are the actual welding feature points.

[0068] Since the starting point and the end point are described by the centroid of the ceramic column, and the actual welding starting point and the end point should be located at the edge of the ceramic column, the coordinates of the centroid of the ceramic column are offset. By calculating the direction vector of the two adjacent centroid points, the actual welding point can be obtained by offsetting along this direction.

[0069] At the same time, the ceramic column 2 has a height. Each time the welding gun is welded along the path, it needs to enter in a direction perpendicular to the ceramic column 2 and leave in a vertical direction after welding each small section. Therefore, it is also necessary to calculate the vertical direction of the ceramic column 2, project the center of mass of the ceramic column and a point on the central axis of the column nail roller cylinder onto the projection plane, calculate the direction vector of the two projection points as the vertical direction of the ceramic column 2, offset the actual welding point along the direction vector to the path of each entry and exit of the welding gun, connect each welding segment path, and finally form a completed planning path, as shown in the figure. Figure 6 shown.

[0070] After completing the path planning, verify the accuracy of the path planning by performing the following experiments:

[0071] First, the coordinates of the starting point and the end point of each segment path are obtained by offsetting the center of mass of the ceramic column. According to the coordinates of the two ends of the path, the length of each segment path can be calculated, such as Figure 7 The length of the planned path shown by the solid line;

[0072] Secondly, the coordinates of the welding points of each segment path are obtained through manual teaching, and the length of the segment welding path in the teaching mode is calculated, such as Figure 7 The length of the teaching path shown by the dotted line;

[0073] Then, the planned path length obtained after the center of mass of the ceramic column is offset is compared with the taught path length obtained by manual teaching. Figure 7 In the figure, the vertical axis represents the length of the segmented welding path, and the horizontal axis represents the segmented path number, totaling 106 segments. It can be seen that the maximum error between the two is 0.98mm, the minimum error is 0.49mm, and the average error is 0.746mm. Furthermore, considering the camera's inherent accuracy of ±0.5mm and camera shooting errors, if the error between the planned welding path and the actual path is within the error accuracy range, the path is considered acceptable, which is of certain significance for achieving welding automation and improving production efficiency. If it is not within the error accuracy range, the path is considered unqualified.

[0074] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A point cloud-based robot column nail roller surfacing path planning method, characterized by: The steps include: Step 1: Image acquisition: Use a depth camera to capture images of the stud-roller weldment and characterize them using point cloud data; Step 2: Extract weldment features. Segment the column-nail-roller weldment collected in step 1 to extract the ceramic column. Step 2.1: De-noise the captured image using a combination of straight-through filtering and Gaussian filtering. Specifically, the straight-through filtering removes noise far from the stud-roller weldment, and the Gaussian filtering removes outliers on the stud-roller weldment surface. Step 2.2: Use deep learning to segment the denoised point cloud image and extract the ceramic columns; Step 3: Build-up welding path planning: The build-up welding path is projected onto the central axis of the column roller cylinder twice to complete the order of the center of mass of the ceramic columns. The upper and lower edges of the two adjacent ceramic columns in the same column are the end and starting points of the segmented welding path respectively. The small welding paths are connected in sequence to finally form an automatic welding trajectory. Step 3.1: Project the centroid of the ceramic column onto the central axis of the column spike roller. The centroids in the same column will be clustered together. Cluster the projections on the central axis. The total number of categories is the number of columns. Compare the coordinates of the centroid projections in each category to obtain the column ranking. Step 3.2: Project the centroid of the ceramic column onto a plane perpendicular to the central axis of the column spike roller. The centroids in the same row will be clustered together. Cluster the projections. The total number of categories is the number of rows. Sorting the clustered results can also obtain the row ranking. Step 3.3: The segmented paths can be planned by sorting the rows in each column.

2. The point cloud-based robot stud roller surfacing path planning method according to claim 1 is characterized in that: The method for fitting the central axis of the column spike roller in step 3 specifically includes the following steps: Step 3.1.1: The centroid coordinates of two adjacent points on the same row are spaced by ∆Y. The centroid coordinates are clustered based on the Y value and the threshold value to complete the rough classification of the ceramic pillars. Step 3.1.2: Fit a three-dimensional circle to each point in each column; Step 3.1.3: Fit a three-dimensional circle to each column; Step 3.1.4: Compare the radius of the circle fitted in step 3.1.3 with the actual radius, eliminate the circles with larger errors, and fit a straight line to the centers of the remaining circles. The resulting straight line is the central axis of the column roller cylinder.

3. The point cloud-based robot stud roller surfacing path planning method according to claim 1 is characterized by: The step 2.2 specifically includes the following steps: Step 2.2.1: Use CloudCompare software to annotate the filtered point cloud file, select the ceramic column and label it; Step 2.2.2: Use the PointNet++ model to train and segment the model. Specifically, based on the PointNet++ model, a training strategy and data enhancement method suitable for ceramic column segmentation are designed. The difference between the z-axis coordinate of each point in the point cloud and the z-axis coordinate of the lowest point in the point cloud is calculated and spliced ​​as a new column of features.

4. The point cloud-based robot stud roller surfacing path planning method according to claim 2 is characterized in that: The step 3.1.2 specifically includes the following steps: Step 3.1.2.1: Use the least squares method to fit the plane of each column and calculate the normal vector to the plane. Specifically, assume that the coordinates of the center of the circle are C, and take any two points as P1 and P2. Let the vector connecting P1 and P2 be vector1, and the vector formed by the coordinates of the center of the circle and the midpoint of the line connecting P1 and P2 be vector2. If the plane equation vector1·vector2=0 is satisfied, find the minimum value of vector1·vector2. Step 3.1.2.2: Use the conditional extreme value solution method to solve the coordinates of the circle center; Step 3.1.2.3: Calculate the radius based on the average distance from the center coordinates found in step 3.1.2.2 to each point on the circle.

5. The point cloud-based robot stud roller surfacing path planning method according to claim 1 is characterized in that: In the step 3, the order of rows and columns of the ceramic columns after segmentation is sorted by two auxiliary projections. Specifically, projection sorting is used, and welding is performed by column. The points in each sorted column are traversed in turn, and the points in the column are sorted according to the size of the Y coordinate. The two adjacent points are the starting point and end point of a small segment path of the column.

6. The point cloud-based robot stud roller surfacing path planning method according to claim 1, characterized in that: The method also includes step 3.4: performing an intersection operation on the two points on the adjacent ceramic columns taken out each time and the centroid clusters on the previous auxiliary surface projection, and the points existing in the column projection cluster and the row projection cluster are actual welding feature points.

7. The point cloud-based robot stud roller surfacing path planning method according to claim 6, characterized in that: Also includes: The coordinates of the centroid of the ceramic column are offset. Specifically, the direction vector of two adjacent centroid points is calculated and offset along this direction to obtain the actual welding point. Calculate the vertical direction of the ceramic column by projecting the center of mass of the ceramic column and a point on the central axis of the column nail roller cylinder onto the projection plane. Calculate the direction vector of the two projection points as the vertical direction of the ceramic column. Offset the actual welding point along the direction vector to the path of each welding gun entry and exit. Connect each welding segment path to form a planned path.

8. The point cloud-based robot stud roller surfacing path planning method according to claim 7, characterized in that: It also includes step 4: verifying whether the path planning method meets the welding accuracy requirements.

9. The point cloud-based robot stud roller surfacing path planning method according to claim 8, characterized in that: Described step 4 comprises the following steps: Step 4.1: The coordinates of the starting point and end point of each segmented path are obtained by offsetting the center of mass of the ceramic column. The length of each segmented path can be calculated based on the coordinates of the two ends of the path; Step 4.2: Obtain the coordinates of the welding points of each segment path through manual teaching, and calculate the length of the segment welding path in the teaching mode; Step 4.3: Compare the planned path obtained after offsetting the center of mass of the ceramic column obtained in step 4.1 with the taught path obtained by manual teaching in step 4.

2. If it is within the error accuracy range, it is qualified; if it is not within the error accuracy range, it is unqualified.

10. The point cloud-based robot stud roller surfacing path planning method according to claim 1, characterized in that: The step 1: image acquisition, specifically: a depth camera is installed at the end of the robotic arm, and in order to improve the accuracy of image acquisition, the camera and the robotic arm need to be calibrated by hand and eye; since the size of the stud roller weldment is much larger than the camera field of view, the image captured by a single shot is a local area of ​​the stud roller weldment, and it is also necessary to splice the images of different areas of the stud roller weldment captured by multiple shots to complete the three-dimensional reconstruction of the stud roller weldment surface. The reconstructed stud roller weldment surface is the welding surface to be processed.

Images