A method and system for extracting full-path welds of complex components based on multi-dimensional features and topological reasoning

Abstract

The application discloses a kind of complex component full-path weld extraction method and system based on multi-dimensional feature and topological reasoning, belong to robot welding vision technical field.Aiming at the problem that box type, I-beam and other components are not complete due to structural obstruction, prone to misjudgment of rolling fillet and other pseudo-weld, the application collects three-dimensional point cloud and two-dimensional image to extract initial geometric intersection line;Fusion three-dimensional gap depth mutation and two-dimensional point solid weld texture features, true and false discrimination and weighted screening are carried out on initial intersection line;Based on real weld and intersection, a space topological graph with node degree attribute is constructed;Using node degree constraint and orthogonal priori, collinear extension and orthogonal complete reasoning are carried out on suspended breakpoint, and complete welding path is reconstructed.The application can effectively eliminate the interference of pseudo-weld, realize accurate reconstruction of full path under limited field of view, and improve the robustness and efficiency of welding automation system.

Description

Technical Field

[0001] This invention belongs to the field of robot welding vision technology, specifically relating to a method and system for extracting full-path welds of complex components based on multi-dimensional features and topological reasoning. It is applicable to weld recognition, false weld removal, and full-path welding trajectory reconstruction of orthogonal steel structure components such as box-type and I-beams under local field of view. Background Technology

[0002] With the advancement of intelligent manufacturing, robotic automated welding has been widely applied in fields such as heavy machinery and steel structures. When automatically welding complex components with multiple intersecting surfaces, such as box joints, H-beams, and I-beams, obtaining accurate and complete full-path weld trajectories is a prerequisite for achieving intelligent welding.

[0003] Existing vision-guided robots typically locate welds by extracting the intersection lines of surfaces. However, two serious problems exist in real-world complex components: First, due to severe occlusion caused by the component's own structure, a single camera shot often only captures a portion of the weld within a local field of view, failing to form a complete path; second, the components contain numerous smooth rounded corners inherent to rolled steel sections, sheet metal bends, and machined edges, and these geometric intersection lines are easily misidentified as real welds by purely 3D point cloud algorithms.

[0004] Traditional welding path acquisition methods primarily rely on manual teaching or multiple roaming scans performed by controlling robots to frequently change their observation postures and stitch together local field-of-view data. These methods not only significantly reduce welding cycle time and production efficiency, and increase the cost of manual intervention, but also only achieve simple splicing of geometric intersections, failing to fundamentally distinguish between genuine welds and pseudo-weld features such as rolled fillets and sheet metal bends, thus making it difficult to eliminate pseudo-weld interference and the risk of misidentification. Furthermore, the multiple-scanning and teaching modes heavily depend on on-site operational experience, exhibiting poor flexibility and adaptability, and cannot meet the demands of high-efficiency, high-stability, and high-robust intelligent automated welding of diverse, small-batch, complex components under complex working conditions. Summary of the Invention

[0005] Existing technologies suffer from two major technical shortcomings: First, the smooth transition lines of numerous rolled steel fillets, sheet metal bends, and machined edges on the surface of complex components are easily misidentified as real welds by pure geometric point cloud algorithms, leading to safety risks such as miswelding and collisions. Second, the structural features of large steel components severely obscure the view, and conventional vision systems can only acquire limited local field-of-view data, failing to continuously extract and reconstruct the complete welding path. This invention provides a system and method for extracting welds from complex components based on multidimensional features and topological reasoning. It aims to achieve accurate identification of genuine and false welds through the fusion of multidimensional physical features, eliminating interference from false welds at the source. Furthermore, based on spatial topological graphs and mathematical vector analytical models, it performs precise logical reasoning and path completion for welds in obscured blind spots within a limited local field of view, achieving complete reconstruction of the entire weld path.

[0006] To achieve the above objectives, the technical solution of this invention is implemented as follows: On one hand, this invention provides a complex component weld extraction system based on multi-dimensional features and topological reasoning, comprising: a 3D vision camera, used to acquire a three-dimensional point cloud model and two-dimensional image data of the complex component to be welded, and to extract the initial geometric intersection line within the field of view; and a feature identification and anti-spoofing module, used to perform authenticity identification and probability calculation on the initial geometric intersection line based on multi-dimensional physical features, eliminate false welds, and obtain a set of true weld line segments. Specifically, this module uses the following multi-dimensional physical feature fusion formula for calculation: in, For the first The probability score of an initial geometric intersection line being identified as a real weld, when If the value exceeds the preset threshold, it is confirmed as a genuine weld segment. These are the weighting coefficients for the three-dimensional contour features; For the first The abrupt change value of the three-dimensional profile depth at the initial geometric intersection line is used to characterize the assembly gap feature; These are the weighting coefficients for the features of the two-dimensional image; For the first The grayscale variation feature value of the two-dimensional image at the initial geometric intersection line is used to characterize the spot welding or light and shadow tomography features.

[0007] The topology graph construction module is used to construct a spatial topology graph with node degree attributes based on the set of real weld seam segments and their intersecting endpoints in three-dimensional space.

[0008] The logical reasoning and completion module is used to perform mathematical analytical reasoning on abnormal nodes in the spatial topology graph based on orthogonal prior knowledge of box-type or I-beam components and node degree constraint rules, thereby completing unknown weld seam segments located in blind spots. The spatial topology reasoning process based on node degree is as follows: Figure 3 shown. Specifically: (1) For suspended abnormal nodes with a degree of 1 caused by local occlusion, the collinear extension formula is adopted. Extend a ray to the known physical boundary and calculate the intersection point of the extension. The coordinates; (2) For undetermined outlier nodes with a degree of 2 due to missing intersecting edges, the vector cross product formula is used. Calculate the third orthogonal direction vector perpendicular to the two known sides. And along this direction, a hidden weld seam is generated in the blind spot of the field of view.

[0009] The robot is used to obtain the full-path 3D welding trajectory based on the completed spatial topology map and control the welding torch to perform automatic welding operations.

[0010] On the other hand, this invention provides a method for extracting weld seams from complex components based on multidimensional features and topological reasoning, such as... Figure 5 As shown, it includes the following steps: Acquire the 3D point cloud model and 2D image data of the complex component to be welded, and extract the initial geometric intersection line within the field of view; Based on the multi-dimensional physical feature fusion formula, the weighted score of the three-dimensional depth mutation value and the two-dimensional gray value of each intersection line is calculated to perform multi-dimensional feature authenticity identification. Smooth transition intersection lines with scores below the preset threshold are discarded as fake welds to obtain a set of real weld line segments. Construct a spatial topology graph by using real weld seam segments as edges and their spatial intersections as nodes, and assign the degree of each node. Blind zone weld reasoning, based on orthogonal prior knowledge of components, completes the suspended breakpoints with degree 1 using the spatial vector extension formula, and completes the undetermined nodes with degree 2 by calculating the third-dimensional orthogonal direction using the vector cross product formula. By utilizing the structural symmetry properties to perform parallel or symmetric mapping, a complete full-path 3D welding trajectory is generated, which is then sent to the robot to complete automated welding.

[0011] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages.

[0012] 1. Achieving highly robust "false weld identification" completely eliminates interference from fake welds. Existing intersection line extraction or plane fitting algorithms are prone to misidentifying rolled fillets and bent edges of components as welds. This invention overcomes the limitations of pure geometric extraction and establishes a multi-dimensional physical feature fusion model. By using "assembly gap (depth abrupt change)" and "spot welding trace (grayscale abrupt change)" in the actual welding process as mathematical evaluation indicators, continuous smooth pseudo edges are directly filtered out from the bottom layer of the algorithm, which greatly reduces the system's misidentification rate and collision risk.

[0013] 2. Breaking through the visual limitations of "what you see is what you get," this invention achieves precise analytical reconstruction of blind-spot welds. Addressing the industry pain point of severe occlusion in large steel structures, this invention innovatively introduces the concepts of spatial topology graphs and node degree conservation from graph theory. It utilizes the collinear extension formula... With the cross product formula Even with only a partially incomplete field of view on one side, the robot automatically "reasoned" and filled in the blind welds hidden on the back side or inside corner through rigorous algebraic spatial geometry calculations. This significantly reduced the roaming scanning actions performed by the robot to find the full path and greatly improved the production cycle time.

[0014] 3. Possesses exceptional flexibility and generalization capabilities without the need for teaching. The logical reasoning mechanism of this invention is based on orthogonal prior knowledge of the underlying physical space of box-type and I-beam components, rather than relying on precise CAD template matching of specific dimensions or deep learning networks driven by massive amounts of data. Therefore, this system does not require retraining or teaching programming for each type of component to be welded, truly achieving highly flexible intelligent extraction for complex components in small batches of various types. Attached Figure Description

[0015] Figure 1 This is a block diagram of a complex component weld extraction system in an embodiment of the present invention; Figure 2 This is a comparative schematic diagram of multi-dimensional feature identification of genuine and counterfeit welds in an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatial topology reasoning process based on node degree in an embodiment of the present invention; Figure 4 This is a schematic diagram of a complex component full-path weld extraction system based on multi-dimensional features and topological reasoning in an embodiment of the present invention; Figure 5 This is a flowchart of the intelligent weld seam extraction method in an embodiment of the present invention; Among them, 1 is the host computer, 2 is the robot controller, 3 is the six-axis industrial robot, 4 is the 3D vision camera, 5 is the industrial welding machine, and 6 is the complex steel structure component. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] Example 1: A system for extracting weld seams from complex components based on multidimensional features and topological reasoning.

[0018] This embodiment provides a system for extracting weld seams from complex components based on multi-dimensional features and topological reasoning. For example... Figure 4 As shown, the hardware system used includes: a six-axis industrial robot 3, which is an ABB IRB1410 model, with a repeatability of ±0.05mm and a load capacity of 5kg, suitable for medium and light-duty welding operations; a robot controller 2, which is an IRC5 industrial robot controller that is compatible with the ABB IRB1410 six-axis industrial robot; an industrial welding machine 5, which is a Megmeet CM350 digital inverter welding machine, supporting a current adjustment range of 30A~350A and a voltage adjustment range of 10V~38V; and a 3D vision camera 4, which is a Gocator 2400 series three-dimensional intelligent sensor, with the following technical parameters: X-direction resolution of 0.12mm~0.36mm, Z-direction linearity of ±0.05mm, field of view of 80mm~200mm, and a built-in red laser projector (wavelength 660nm), which can simultaneously output three-dimensional point cloud and two-dimensional grayscale images. The host computer 1 uses an industrial-grade embedded controller, configured with an Intel Core i7-9700E processor, a high-performance discrete graphics card, 16GB of DDR4 memory, and a 512GB solid-state drive. This hardware is interconnected with the fieldbus via Gigabit Ethernet and centrally managed by the host computer to achieve automatic extraction of welding paths and robot control.

[0019] During system operation, the complex steel structural component 6 to be welded is placed on the workbench. The 3D vision camera 4 acquires local 3D point cloud and 2D images of the component using an "eye-in-hand" method and transmits them to the host computer 1 via gigabit network cable. The host computer 1 is equipped with core algorithm processing software, mainly including a feature identification and anti-spoofing module, a topology graph construction module, and a logical reasoning completion module.

[0020] Feature identification and anti-spoofing module: First, all initial geometric intersection lines within the field of view are extracted through point cloud normal vector calculation and region growing algorithm. Then, this module establishes a local bounding box along each intersection line and extracts the maximum 3D depth gradient value within the bounding box as the depth abrupt change value. (A physical quantity characterizing assembly gaps), and extract the maximum gray-level gradient value within the corresponding two-dimensional image region as the gray-level change value. (Physical quantities characterizing spot weld marks). Calculation formula using multidimensional physical characteristics: Calculate the confidence score for each intersection line to be a true weld. Among them, and These are normalized weighting coefficients preset based on engineering experience. The values ​​of the weighting coefficients are determined based on engineering experience and experimental calibration. Specifically, 50 samples each of real welds and spurious welds are collected, and their values ​​are calculated separately. and The value is determined by using a grid search method to find the weight combination that maximizes the identification accuracy. , When the score is... If the score exceeds the preset identification threshold, the intersection line is confirmed as a genuine weld; otherwise, it is judged as a continuous, smooth rolled fillet or bent pseudo-weld and removed from memory. The identification threshold is set as follows: 10 known genuine welds and 10 known pseudo-welds are selected, and their Si scores are calculated respectively. The score range for genuine welds is 0.75-0.95, and the score range for pseudo-welds is 0.15-0.45. Based on this, the identification threshold is set to 0.7.

[0021] The topology graph construction module instantiates the selected and retained real weld seam segments as "edges" in the graph data structure, and instantiates the endpoints and intersections of the segments in three-dimensional space as "vertices". The module traverses all nodes and calculates their "degree", that is, counts the number of valid edges connected to the node.

[0022] Logical reasoning and completion module: This is the core of the system. For welds that are truncated due to camera field of view (nodal degree of 1, i.e., suspended breakpoints), the module calls the collinear extension formula. The system extracts the physical boundary bounding box plane of the entire component and calculates the extension distance scalar using a mathematical analytical method of finding the intersection of rays and planes. Thus, the coordinates of the new intersection point within the blind zone are obtained. .

[0023] For interior angles of intersecting edges that are missing due to occlusion (node ​​degree is 2, i.e., undetermined right-angle nodes), the module extracts the unit direction vectors of the two known perpendicular welds. and Call the vector cross product formula Calculate the third orthogonal direction vector perpendicular to the two known sides. Combined with the internal spatial vector constraints of the point cloud bounding box, determine the unique positive and negative directions of the cross product result, and then generate the hidden weld seam in the blind spot along this direction.

[0024] Example 2: A method for extracting weld seams from complex components based on multidimensional features and topological reasoning.

[0025] The structural block diagram of the complex component weld extraction system based on multidimensional features and topological reasoning is as follows: Figure 1As shown, the system mainly includes a data acquisition layer, an algorithm processing layer, and an execution control layer. The data acquisition layer includes a 3D vision camera 4, used to acquire 3D point clouds and RGB images. The algorithm processing layer includes a feature identification and anti-spoofing module, a topology graph construction module, and a logical reasoning completion module, which completes the authenticity judgment, topology construction, and blind zone weld seam reasoning. The execution control layer consists of a host computer 1, a robot controller 2, a six-axis industrial robot 3, and an industrial welding machine 5, which completes automated welding according to the 3D welding trajectory. This embodiment describes in detail the specific steps and procedures for accurately extracting blind zone weld seams using the above system, specifically including the following steps.

[0026] Step S1: Multimodal data acquisition and initial intersection line extraction. The six-axis industrial robot 3 moves to the preset observation point and controls the 3D vision camera 4 to perform a single local shot of the component to be welded, simultaneously acquiring the point cloud model and two-dimensional texture image. The point cloud is subjected to pass-through filtering and voxel downsampling to remove environmental noise. The RANSAC (Random Sample Consensus) algorithm is used to fit the local plane, and the intersection lines of the local plane are extracted as the initial geometric intersection line set.

[0027] Step S2: Perform multi-dimensional feature identification of genuine and fake welds. For each intersection line in the set, the system automatically samples the tubes along the line. For genuine assembly welds, due to the splicing gap, their three-dimensional contour depth profile will exhibit a "V-shaped" high-frequency abrupt change. The peak value of this abrupt change is extracted. Meanwhile, real weld seams are often accompanied by artificially applied tack weld protrusions or light and shadow breaks caused by differences in the reflectivity of the steel plate material. The peak values ​​of these image grayscale abrupt changes are extracted as... Substitute into the formula After calculation, the actual weld seam Extremely high; and the smooth rounded corners inherent in rolled steel sections... Approaching 0, Uniform variation results in extremely low scores, leading the system to automatically eliminate false welds. Taking a candidate intersection line at the junction of the flange and web of an H-beam as an example, a local bounding box is established along this intersection line, the three-dimensional depth gradient is extracted, and the depth abrupt change value is measured. Simultaneously extract the gray-level gradient of the corresponding two-dimensional image region and measure the gray-level change value. Take the weighting coefficients. , Substituting into the formula, we get The score is greater than the preset threshold of 0.7, therefore it is determined to be a genuine weld. The score was measured at a rolled fillet on the same component. , Calculated If the value is below the threshold, it will be rejected.

[0028] Step S3: Spatial Topology Mapping and Degree Analysis. Import the confirmed actual weld seam segments into the graph theory algorithm, defining the two ends of each segment as nodes. To eliminate point cloud errors, merge neighboring nodes whose spatial Euclidean distance is less than the tolerance threshold (e.g., 2mm) to achieve convergence. Calculate the degree of each node after merging. In ideal box-shaped or I-beam orthogonal members, the degree of a legal assembly intersection should be strictly constrained to 2 (planar right-angle intersection) or 3 (spatial trihedral intersection).

[0029] Step S4: Dimensional reduction analysis and targeted reasoning of blind zone welds. Traverse the topology graph and trigger the reasoning mechanism for abnormal nodes as follows.

[0030] (1) Suspended breakpoint reasoning: If the node degree is 1, it indicates that the camera only captured half of the weld. The system obtains the coordinates of the suspended breakpoint. and the direction vector of the known half-section weld. Establish the ray parametric equations. Combine the pre-obtained geometric plane equations of the component's base plate or flange with the system's parameters, and solve for the intersection points of the ray and the plane, thus obtaining the extension distance. .

[0031] Substitute the parameters into the formula Obtain the complete boundary points within the blind zone. Complete the targeted completion of the occluded and broken lines. The reasoning process is as follows: Figure 3 (b) ① is shown.

[0032] (2) Orthogonal interior angle reasoning: If the degree measure of an interior right-angle node is 2, it indicates that the third weld perpendicularly downwards along the web is blocked. The system extracts the direction vectors of the known two sides. and Call the vector cross product formula The spatial orthogonal direction is obtained. Hidden welds are generated along this direction until they intersect with the plane of the component's base plate. The system follows... A virtual projection line is generated within the blind zone and solidified as a third hidden weld, thus upgrading the degree of the anomalous node from 2 to a legitimate 3. The reasoning process is as follows: Figure 3 (b) As shown in ②. Suppose that at a certain node to be determined, the direction vectors of the two actual welds are known as: (along the positive X-axis direction), (Along the positive Y-axis). The direction of the third side is calculated using the vector cross product formula: Based on the internal spatial vector constraints of the point cloud bounding box, and given that the web of the component extends in the negative Z-axis direction, the opposite direction is taken. .

[0033] Step S5: Global Mapping and Trajectory Output Based on Component Symmetry. After completing the local map reasoning under a single-sided view, the system utilizes the inherent central axis symmetry of complex steel structural components (such as H-beams) to mirror the extracted single-sided complete topology map along the component's central symmetry plane, directly deriving the weld topology completely located on the invisible back side. Finally, the global closed-loop topology map undergoes coordinate system transformation (from the camera coordinate system to the robot's base coordinate system) to generate a smooth, continuous 3D welding trajectory, which is then sent to the robot controller for automated welding.

[0034] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A system for extracting weld seams from complex components based on multidimensional features and topological reasoning, characterized in that, include: 3D vision camera, feature identification and anti-spoofing module, topology graph construction module, logical reasoning completion module, and robot.

2. The complex component weld extraction system according to claim 1, characterized in that, In the feature identification and despiking module, the three-dimensional depth features of the weld are compared. Genuine and fake welds exhibit obvious assembly gaps and abrupt depth changes, while fake welds show a smooth transition characteristic formed by rolling fillets. A comparison of their two-dimensional grayscale features reveals that genuine welds exhibit abrupt grayscale changes, either from spot welding or light and shadow breaks, while fake welds show uniform grayscale changes and no obvious texture abrupt changes. The comparison of genuine and fake welds is shown in Figure 2. The formula for authenticity identification and probability calculation is as follows: in, For the first The probability score of an initial geometric intersection line being identified as a real weld, when If the value exceeds the preset threshold, it is confirmed as a genuine weld segment. These are the weighting coefficients for the three-dimensional contour features; For the first The abrupt change value of the three-dimensional profile depth at the initial geometric intersection line is used to characterize the assembly gap feature; These are the weighting coefficients for the features of the two-dimensional image; For the first The grayscale variation feature value of the two-dimensional image at the initial geometric intersection line is used to characterize the spot welding or light and shadow tomography features.

3. The complex component weld extraction system according to claim 1, characterized in that, In the logical reasoning completion module, abnormal nodes in a suspended state are completed using a collinear extension formula, which is as follows: in, To provide the spatial three-dimensional coordinates of the new nodes that intersect with the physical boundary of the component after completion; The spatial three-dimensional coordinates of the suspended node extracted visually; It is the scalar distance of the straight line extension from the suspended node to the known physical boundary plane of the component; Let be the spatial unit direction vector of the actual weld line segment connected to the suspended node. This formula is based on the analytic geometry theory of ray-plane intersection: in orthogonal steel structures (box-type, I-beam), the weld line must extend to the physical boundary of the component.

4. The complex component weld extraction system according to claim 1, characterized in that, In the logical reasoning completion module, the missing intersecting edges of abnormal nodes are completed using the vector cross product formula, which is as follows: in, The spatial direction vector of the hidden weld seam located in the blind spot of the field of vision is inferred. The spatial direction vector of the first known real weld line segment connected to the abnormal node; Let be the spatial direction vector of the second known real weld line segment connected to the abnormal node. This formula is based on the orthogonal basis theorem in spatial analytic geometry: in a Cartesian coordinate system, given two mutually perpendicular direction vectors, their cross product uniquely determines the third orthogonal direction.

5. A method for extracting weld seams from complex components based on multidimensional features and topological reasoning, characterized in that, Includes the following steps: Acquire the 3D point cloud and 2D image of the component, and extract the initial geometric intersection line; The authenticity of the weld is identified by multi-dimensional physical feature fusion; a topological graph is constructed, the degree of the nodes is counted and constrained to 2 or 3; based on orthogonal priors, the collinear extension of the degree 1 breakpoints and the vector cross product of the degree 2 nodes are used to complete the blind weld seam; a complete welding trajectory is generated and the welding gun is controlled to complete the automated welding.

6. The method for extracting weld seams from complex components according to claim 5, characterized in that, The method for identifying the authenticity of the initial geometric intersection based on the multi-dimensional physical feature fusion formula is as follows: Multimodal data acquisition is performed to obtain the three-dimensional point cloud and two-dimensional image of the component; the depth abrupt change value representing the assembly gap and the grayscale change value representing the spot weld are extracted at the initial geometric intersection; the depth abrupt change value and the grayscale change value are multiplied by their respective weighting coefficients and summed to obtain the probability score that the initial geometric intersection is a real weld; initial geometric intersections with a probability score lower than a preset threshold are marked as fake welds caused by rolling fillets or bending and are discarded.

7. The method for extracting weld seams from complex components according to claim 5, characterized in that, The process of constructing a spatial topology graph based on the set of real weld seam segments and calculating the degree of each node includes: defining the real weld seam segments as edges of the topology graph, defining the intersections of the real weld seam segments in three-dimensional space as nodes of the topology graph; counting the number of edges connected to each node and defining them as the degree of that node; in box-shaped or I-beam orthogonal members, the degree constraint of a reasonable node is 2 or 3.

8. The method for extracting weld seams from complex components according to claim 7, characterized in that, During the mathematical logic reasoning process for nodes with abnormal degrees: identify nodes with a degree of 1 in the spatial topology graph and mark them as suspended breakpoints caused by occlusion; Using the collinear extension formula, a ray is extended to the physical boundary plane inside the component, the coordinates of the extension intersection point are calculated, and the extended segment is added to the topology as the inferred hidden weld.

9. The method for extracting weld seams from complex components according to claim 7, characterized in that, In the process of mathematical logic reasoning for nodes with abnormal degrees: identify nodes with a degree of 2 in the spatial topology graph and mark them as undetermined nodes; extract the direction vectors of the two known real weld line segments connected to the undetermined node; calculate the third orthogonal direction vector using the vector cross product formula, and generate the third hidden weld along this orthogonal direction, so that the degree of the undetermined node is upgraded to 3.

10. The method for extracting weld seams from complex components according to claim 5, characterized in that, The method also includes, based on the structural symmetry properties of the component, performing translational projection or symmetrical mapping of the identified front weld features along the central axis of the component, and inferring the trajectory of the hidden weld on the back side of the complete blind zone.

Images