Welding pose track generation method for fillet weld of workpiece, medium and equipment

Abstract

The invention relates to a welding pose track generation method of a workpiece fillet weld, a medium and equipment. The method comprises the following steps: acquiring a two-dimensional workpiece weld point; performing three-dimensional reconstruction on the two-dimensional workpiece welding seam points to obtain a three-dimensional workpiece welding seam point cloud; a plurality of geometric equations in the three-dimensional workpiece welding seam point cloud are extracted, track regeneration is conducted according to the geometric equations, and an initial welding track and key welding seam point coordinates are obtained; and attitude mapping is conducted according to the prior welding attitude and the key welding seam point coordinates, interpolation calculation is conducted according to the initial welding track, and a final welding pose track is obtained. The problems that in the prior art, the actual space pose of a workpiece cannot be sensed online, the three-dimensional contour of a fillet weld cannot be automatically recognized, and the motion trail of a robot cannot be generated in real time are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of welding automation, in particular to a welding pose trajectory generation method for a workpiece fillet weld, a medium and equipment. BACKGROUND

[0002] During the ship assembly stage, a large number of lifting lugs are temporarily or permanently fixed to the outer plate, bulkhead and reinforcing structure to provide lifting and fixing points for subsequent section turning, docking, closing and mooring test. The lifting lug is usually made of high-strength steel plate with a thickness of 20-60 mm after numerical control cutting and bevel processing, and two symmetrical fillet welds are formed between the root and the ship body. The fillet weld not only bears the tensile and shear in the plane of the lifting lug, but also bears the out-of-plane bending moment generated by lifting impact, and has high requirements for penetration, weld formation and internal quality. The welding quality is directly related to the structural strength and safety of the whole ship.

[0003] At present, domestic and foreign shipyards mainly use manual welding or semi-automatic welding methods for the above-mentioned fillet welds, but manual welding requires operators to adjust current, voltage, wire angle and swing amplitude in real time according to the assembly gap, which has high dependence on personnel experience, large welding quality fluctuation, and is prone to defects such as undercut, slag inclusion, incomplete fusion and root incomplete penetration. At the same time, the welding position is mostly overhead welding or vertical welding, and the working environment is accompanied by high temperature, strong ultraviolet radiation and high concentration of metal dust, which has high labor intensity and high occupational health risk. The semi-automatic welding method needs to record the welding gun pose through manual "point-by-point teaching" method, and the fixed path is formed and then circulates, which is tedious and inefficient, and lacks real-time perception and recognition ability of the workpiece. For example, in the face of common lifting lug installation deviation, structural deformation or assembly error in actual production, the traditional teaching path cannot be dynamically adjusted, resulting in inconsistency between the welding trajectory and the actual weld position.

[0004] Therefore, how to improve the problems that the prior art cannot online perceive the actual spatial pose of the workpiece, automatically recognize the three-dimensional contour of the fillet weld, and generate the robot motion trajectory in real time is a technical problem to be solved in the field. SUMMARY

[0005] Based on this, the present application aims to provide a welding pose trajectory generation method for a workpiece fillet weld, a medium and equipment to solve at least one technical problem mentioned in the background.

[0006] In a first aspect, the present application provides a welding pose trajectory generation method for a workpiece fillet weld, comprising: acquiring two-dimensional workpiece weld points; three-dimensionally reconstructing the two-dimensional workpiece weld points to obtain three-dimensional workpiece weld point clouds; Extracting several geometric equations in the three-dimensional workpiece weld point cloud, regenerating the trajectory according to the geometric equations, and obtaining the initial welding trajectory and the key weld point coordinates; According to the prior welding posture and the key weld point coordinates, the posture mapping is performed, and the interpolation calculation is performed according to the initial welding trajectory, so as to obtain the final welding posture trajectory.

[0007] Further, the step of acquiring the two-dimensional workpiece weld point comprises: A weld point recognition model is constructed and trained, taking the original weld image as the input and the weld point probability map as the output; The original weld image of the workpiece surface is collected and input into the trained weld point recognition model to obtain the weld point probability map; The points with a confidence lower than a set threshold in the weld point probability map are removed to obtain several connected domains; The peak value of each connected domain is detected to obtain the point with the highest confidence in each connected domain as the two-dimensional workpiece weld point.

[0008] Further, the specific structure of the weld point recognition model comprises: An input layer is used to receive the original weld image; A feature extraction layer is used to extract features from the original weld image to obtain a shallow feature map, a middle feature map and a deep feature map; A feature fusion layer is used to sequentially fuse the shallow feature map, the middle feature map and the deep feature map to obtain a global fusion feature map; An output layer is used to generate the confidence of each weld point according to the global fusion feature map and output the weld point probability map.

[0009] Further, the feature extraction layer comprises a shallow convolution block, a middle convolution block and a deep convolution block connected in sequence; The shallow convolution block is used to extract local features from the original weld image to obtain a shallow feature map; The middle convolution block is used to further extract local features from the shallow feature map to obtain a middle feature map; The deep convolution block is used to further extract local features from the middle feature map to obtain a deep feature map; The feature fusion layer comprises a multi-scale sampling block, a hierarchical sampling block and a global sampling block connected in sequence; The multi-scale sampling block is connected with the outputs of the shallow convolution block and the deep convolution block, and is used to fuse the shallow feature map and the deep feature map to combine the high-resolution details of the shallow feature map and the overall structure cognition of the deep feature map, so as to obtain a multi-scale fusion feature map; The hierarchical sampling block is connected with the outputs of the middle layer convolution block and the multi-scale sampling block, and is used for fusing the middle layer feature map and the multi-scale fusion feature map to further combine local features of the middle layer feature map and improve classification accuracy, so as to obtain a hierarchical fusion feature map. The global sampling block is connected with the output of the hierarchical sampling block, and is used for upsampling and pooling operations on the hierarchical fusion feature map to expand a network receptive field and enhance feature expression capability, so as to obtain a global fusion feature map.

[0010] Further, the step of performing three-dimensional reconstruction on the two-dimensional workpiece weld points to obtain a three-dimensional workpiece weld point cloud includes: S21: performing three-dimensional reconstruction on the two-dimensional workpiece weld points according to prior camera calibration parameters and a laser light plane geometric model to obtain an initial three-dimensional workpiece weld point cloud; S22: fitting a current plane equation from a plurality of points in the initial three-dimensional workpiece weld point cloud; S23: obtaining distances between each point in the initial three-dimensional workpiece weld point cloud and the current plane equation to determine current inliers and count the number of current inliers; S24: determining whether a set iteration condition is reached, if not, returning to step S22, and if yes, selecting a current plane equation corresponding to the largest number of current inliers as a weld fitting equation according to the number of inliers counted by each plane equation; S25: calculating perpendicular distances between each point in the initial three-dimensional workpiece weld point cloud and the weld fitting equation, and determining whether the distance is less than a set threshold, if not, removing the corresponding point cloud to obtain an optimized three-dimensional workpiece weld point cloud.

[0011] Further, the step of extracting a plurality of geometric equations from the three-dimensional workpiece weld point cloud and performing trajectory regeneration according to the geometric equations to obtain an initial welding trajectory and key weld point coordinates includes: projecting the three-dimensional workpiece weld point cloud to the finally determined plane equation to obtain a projected point cloud; obtaining a plurality of geometric equations according to the projected point cloud, and generating an initial welding trajectory by interpolation calculation according to parameters of each equation; obtaining intersection points of adjacent geometric equations to obtain key weld point coordinates.

[0012] Further, the step of obtaining a plurality of geometric equations from the projected point cloud and generating an initial welding trajectory by interpolation calculation according to parameters of each equation includes: S321: performing circular fitting on the projected point cloud to obtain a circular equation, and removing inliers of the circular equation in the projected point cloud to obtain a first remaining projected point cloud; S322: determining whether the number of the first remaining projected point cloud is greater than a first set threshold, if not, performing step S325; S323: If so, then perform line fitting on the first residual projection point cloud to obtain several line equations, and remove the points inside each line equation in the residual projection point cloud to obtain the second residual projection point cloud. S324: Determine whether the number of the second remaining projection point cloud is greater than the second set threshold. If not, proceed to step S325. If yes, perform point cloud clustering on the second remaining projection point cloud to obtain several clustered point clouds. Obtain the linear equations of adjacent points in each clustered point cloud. Based on the intersection and endpoints of the adjacent linear equations, obtain the arc equation parameters of each clustered point cloud. S325: Perform interpolation calculations based on the parameters of each geometric equation to generate the initial welding trajectory.

[0013] Furthermore, the steps of performing attitude mapping based on the prior welding posture and the coordinates of key weld points, and interpolating based on the initial welding trajectory to obtain the final welding pose trajectory, include: Using any critical weld point as the calibration point, and based on the coordinates of the corresponding prior critical weld point, the translation matrix is ​​obtained; Obtain the prior welding posture corresponding to each key point, and perform posture mapping based on the coordinates of each key weld point to obtain the current welding posture of each key point. The current welding pose of each key point is obtained by transforming the current welding posture of each key point using the translation matrix; Pose interpolation calculations are performed between adjacent key points to obtain the final welding pose trajectory.

[0014] Secondly, this application also provides a computer storage medium storing executable program code; the executable program code is used to execute the method for generating the welding pose trajectory of the fillet weld of the workpiece as described in any one of the first aspects.

[0015] Thirdly, this application also provides a terminal device, including a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to execute the method for generating the welding pose trajectory of the fillet weld of the workpiece as described in any one of the first aspects.

[0016] This invention provides a method, medium, and device for generating welding pose trajectories for fillet welds on workpieces. It acquires two-dimensional weld points on the workpiece to accurately reflect the complex geometric features of the weld area, providing foundational data for subsequent steps. Then, it performs three-dimensional reconstruction of the two-dimensional weld points to obtain a three-dimensional weld point cloud, improving data quality and significantly enhancing the accuracy and stability of subsequent trajectory planning and attitude estimation. This also provides a reliable reference for subsequent geometric primitive extraction. Furthermore, by extracting several geometric equations from the three-dimensional weld point cloud, the trajectory is regenerated based on these equations to obtain the initial welding trajectory and the coordinates of key weld points. The complex weld point cloud is abstracted into geometric primitives such as lines, arcs, and spline curves, achieving semantic and structural representation of the weld. This method uses geometric primitives to represent trajectories, facilitating understanding and processing. The trajectories generated based on geometric primitives inherently possess continuity and smoothness, avoiding jitter or unevenness caused by directly connecting discrete points, thus improving welding quality. Furthermore, it identifies key feature points such as the start, end, inflection, and midpoint of the weld seam using geometric primitives, providing crucial references for subsequent attitude planning. Finally, attitude mapping is performed based on prior welding attitudes and the coordinates of key weld seam points, and interpolation calculations are performed based on the initial welding trajectory to obtain the final welding pose trajectory. First, prior welding attitudes corresponding to each key point are introduced to ensure that the generated welding attitude conforms to process specifications, avoiding abrupt changes or unreachability. Then, the initial trajectory is optimized using an interpolation algorithm to achieve a smooth transition in the welding attitude, resulting in the final welding pose trajectory. This method solves the problems of existing technologies that cannot perceive the actual spatial pose of the workpiece online, automatically identify the 3D contour of fillet welds, and generate robot motion trajectories in real time. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for generating the welding pose trajectory of a workpiece fillet weld according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the weld point recognition model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the lifting lug workpiece according to an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the three-dimensional workpiece weld point cloud and the initial weld trajectory in an embodiment of the present invention. Figure 5 This is a schematic diagram of a key weld point in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that if the embodiments of the present invention involve directional indications, such as up, down, left, right, front, back, etc., these directional indications are only used to explain the relative positional relationships and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly. Furthermore, if the embodiments of the present invention involve descriptions such as "first," "second," "S1," "S2," "step one," "step two," etc., these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance, or implicitly indicating the number of technical features indicated or the order of method execution. Those skilled in the art will understand that anything that does not violate the inventive concept and is within the scope of the present invention should be included in the protection scope of the present invention.

[0020] like Figure 1 As shown, the present invention provides a method for generating the welding pose trajectory of a workpiece fillet weld: S1: Obtain the weld seam points of the two-dimensional workpiece; Specifically, but not limited to, placing the workpiece at a set position, scanning the workpiece surface to obtain an original weld image, and inputting the original weld image into a trained weld point recognition model to obtain two-dimensional workpiece weld points. The specific structure of this weld point recognition model is not limited here.

[0021] Preferably, when scanning the workpiece surface to obtain the original weld image, background edges, scratches, and reflections may also be misidentified as "welds" due to their reflectivity, causing the robot to deviate from the actual bevel during welding, resulting in problems such as undercut and incomplete penetration. Therefore, the step of obtaining the two-dimensional workpiece weld point may optionally include: S11: Construct and train a weld point recognition model that takes the original weld image as input and the weld point probability map as output; Specifically, this can be optional, but not limited to, acquiring several existing raw weld images and labeling the weld point probabilities in these images to obtain a training dataset. The weld point recognition model can employ any existing neural network model. Using the labeled raw weld images, a training dataset is constructed. Then, based on the training dataset, an initial neural network model is input to train its parameters, resulting in a trained neural network model, which serves as the weld point recognition model. It is noteworthy that this step is preparatory; it can be pre-built and trained, only needing to be completed before using the weld point recognition model for analysis.

[0022] Preferably, the specific structure of the weld point recognition model is as follows: Figure 2 As shown, the options include: The input layer is used to receive the raw weld images; The feature extraction layer is used to extract features from the original weld image, resulting in shallow feature maps, medium feature maps, and deep feature maps. The feature fusion layer is used to sequentially fuse shallow feature maps, mid-level feature maps, and deep feature maps to obtain a global fused feature map; The output layer is used to generate the confidence level of each weld point based on the global fused feature map and output the weld point probability map.

[0023] Specifically, the input layer receives the original weld image to provide a data foundation for subsequent steps. Then, a feature extraction layer extracts features from the original weld image, resulting in shallow, medium, and deep feature maps. The shallow feature map preserves fine spatial location information and edge details, improving the accuracy of weld boundary positioning. The medium feature map extracts structural features of the weld (such as ripples and weld width) to balance semantic information and spatial resolution, improving robustness to weld morphology changes. Finally, the deep feature map captures global semantics and the overall weld pattern, enhancing the receptive field and anti-interference capabilities, thereby suppressing the effects of noise such as illumination and smoke. By simultaneously acquiring detailed, structural, and global information, the system adapts to different weld morphologies. The diversity of scales and the complementarity of features at different levels can significantly improve the defect detection rate and false alarm suppression capability. Then, a feature fusion layer sequentially fuses shallow, mid-level, and deep feature maps to obtain a global fused feature map. By spatially aligning and semantically negotiating features at different scales (shallow, mid-level, and deep), each pixel incorporates a triple perspective of "macro-context + meso-structure + micro-detail." Thus, when a shallow feature map identifies a weld edge but a deep feature map identifies it as background texture, the fusion layer automatically calibrates through a cross-scale voting mechanism, significantly reducing the false detection rate. Finally, a prior activation function is introduced into the output layer to generate the confidence score of each weld point based on the global fused feature map, outputting a weld point probability map. Preferably, the activation function used is the Sigmoid function.

[0024] Preferably, the specific structure of the feature extraction layer includes: a shallow convolutional block, a middle convolutional block, and a deep convolutional block connected in sequence; Shallow convolutional blocks are used to extract local features from the original weld seam image to obtain a shallow feature map. The intermediate convolutional block is used to further extract local features from the shallow feature map to obtain the intermediate feature map; Deep convolutional blocks are used to further extract local features from the intermediate feature maps to obtain deep feature maps.

[0025] Specifically, shallow convolutional blocks can be used to extract local features from the original weld image to obtain a shallow feature map, which extracts fine features such as edges and textures, improves the sensitivity to small changes in weld defects, and retains accurate spatial location information, which is beneficial for accurately locating weld points. Then, a middle-layer convolutional block is used to further extract local features from the shallow feature map to obtain a middle-layer feature map, which captures more complex local patterns (such as weld shape and structure) on the basis of shallow features, in order to balance local details and regional context information. Finally, a deep convolutional block is used to further extract local features from the middle-layer feature map to obtain a deep feature map, which obtains high-level semantics such as the overall structure and orientation of the weld, in order to identify complex defect patterns and nonlinear features, and enhances adaptability to changes in illumination and contrast.

[0026] Preferably, the specific structure of the feature fusion layer includes: a multi-scale sampling block, a hierarchical sampling block, and a global sampling block connected in sequence; Multi-scale sampling blocks, connected to the outputs of shallow and deep convolutional blocks, are used to fuse shallow and deep feature maps to combine the high-resolution details of shallow feature maps with the overall structural understanding of deep feature maps, resulting in a multi-scale fused feature map. Hierarchical sampling blocks, connected to the outputs of intermediate convolutional blocks and multi-scale sampling blocks, are used to fuse intermediate feature maps and multi-scale fused feature maps to further combine local features of intermediate feature maps, improve classification accuracy, and obtain hierarchical fused feature maps. The global sampling block, connected to the output of the hierarchical sampling block, is used to perform upsampling and pooling operations on the hierarchical fused feature map to expand the network's receptive field and enhance its feature representation ability, thus obtaining the global fused feature map.

[0027] Specifically, shallow and deep feature maps can be fused using multi-scale sampling blocks. This combines the high spatial resolution (precise location) of shallow features with the high semantic information (overall structure) of deep features, thereby effectively distinguishing genuine welds from visually similar interference such as metal scratches, oxide scale, and reflections by utilizing the overall cognition of deep features, resulting in a multi-scale fused feature map. Then, mid-level and multi-scale fused feature maps are fused using hierarchical sampling blocks. Based on the multi-scale fusion, local patterns of mid-level features (such as weld width and continuous shape) are further injected to provide richer local context, making weld point classification decisions more accurate, resulting in a hierarchical fused feature map. Finally, the hierarchical fused feature map is upsampled and pooled using global sampling blocks to capture the global direction and topology of the weld, so that the integrated global feature map has the triple expression capabilities of detail, local, and global, resulting in a global fused feature map, which is more adaptable to changes in image quality and local occlusion.

[0028] S12: Collect the original weld seam image on the surface of the workpiece and input it into the trained weld seam point recognition model to obtain the weld seam point probability map; Specifically, a probability threshold can be set, but is not limited to, and then the workpiece surface is scanned to acquire the original weld image. Since the deep learning model can learn complex weld features and has good recognition ability for welds of different types, under different lighting conditions and with different surface conditions, the original weld image is input into the weld point recognition model trained in step S11 to obtain the weld point probability map. The confidence of each point is obtained through the weld point probability map, and the weld points are automatically identified through the trained deep learning model to improve detection efficiency and robustness.

[0029] S13: Remove points with confidence levels below a set threshold from the weld point probability graph to obtain several connected components; S14: Perform peak detection on each connected component to obtain the point with the highest confidence in each connected component, which is the weld point of the two-dimensional workpiece.

[0030] Specifically, the probability map of weld points can be filtered based on a confidence threshold, but is not limited to this. By eliminating low-confidence points that are misjudged by the model, several connected regions composed of high-confidence points are obtained to reduce noise interference and improve the accuracy of subsequent processing. At the same time, high-confidence weld points are retained, so that subsequent processing is more focused on connected regions composed of high-confidence points, reducing the amount of computation. Then, peak detection is performed in each connected region to accurately find the point with the highest built-in confidence in that region, which is used as the final determined 2D workpiece weld point, improving positioning accuracy. Retaining only one peak point in each connected region can avoid redundancy caused by multiple neighboring points, simplifying the subsequent 3D reconstruction process and reducing the amount of computation.

[0031] S2: Perform three-dimensional reconstruction of the weld seam points of the two-dimensional workpiece to obtain a three-dimensional weld seam point cloud of the workpiece; Specifically, due to the complex welding environment, images are often affected by various factors such as weld slag adhesion, high-temperature arc light, strong light interference, and side plate reflection, resulting in noise, light spots, occlusion and other interference phenomena in the image, which in turn cause false detection and missed detection in the weld point recognition process. Therefore, it is possible to perform three-dimensional reconstruction on the weld points of the two-dimensional workpiece to obtain a three-dimensional workpiece weld point cloud.

[0032] Preferably, the step of performing three-dimensional reconstruction of the weld seam points of a two-dimensional workpiece to obtain a three-dimensional weld seam point cloud may include: S21: Based on the prior camera calibration parameters and the laser light plane geometric model, the weld seam points of the two-dimensional workpiece are reconstructed in three dimensions to obtain the initial three-dimensional workpiece weld seam point cloud. Specifically, the weld seam points of a two-dimensional workpiece can be reconstructed into three dimensions based on the prior camera calibration parameters and laser light plane geometric model obtained in advance by those skilled in the art. By elevating the weld seam points in the two-dimensional image to three-dimensional space, their true three-dimensional coordinate information is obtained, providing basic data for subsequent steps such as robot guidance and welding trajectory generation, and obtaining the initial three-dimensional workpiece weld seam point cloud.

[0033] For example, based on the prior camera calibration parameters and the laser light plane geometric model, the steps for three-dimensional reconstruction of the weld points of a two-dimensional workpiece can be as shown in Equation 2-1: 2-1 Where (u, v) are the pixel coordinates of the laser beam plane geometric model. Determined by the camera's internal structure, these are the camera's intrinsic parameters, while A, B, and C are the laser light plane parameters. These are the three-dimensional coordinates converted to the camera coordinate system.

[0034] S22: Select several points from the initial three-dimensional workpiece weld point cloud and fit them to obtain the current plane equation; Specifically, it is possible, but not limited to, to select a number of points from the point cloud of the weld seam of the three-dimensional workpiece based on planar features and perform planar fitting to obtain the current planar equation.

[0035] For example, the initial plane equations can be optionally set according to the plane characteristics and the equations they need to satisfy, denoted as: And randomly select three non-collinear points from the point cloud. The normal vector of the plane It can be done The cross product of the two direction vectors is calculated, and then... Substituting these values ​​into the initial plane equation, we obtain the plane parameter D, thus yielding the current plane equation. It is worth noting that the equations this plane equation must satisfy and the parameters used are for illustrative purposes only and are not intended to be limiting.

[0036] S23: Obtain the distance between each point in the initial 3D workpiece weld point cloud and the current plane equation to determine the current interior point and count the current interior point; S24: Determine whether the set iteration conditions have been met. If not, return to step S22. If yes, select the current plane equation corresponding to the maximum number of current internal points as the weld fitting equation based on the number of internal points counted for each plane equation. Specifically, it is optional, but not limited to, setting a first distance threshold and using existing distance calculation formulas to calculate the distance between each point in the initial 3D workpiece weld seam point cloud and the current plane equation. It is then determined whether the distance is greater than the first distance threshold. If it is not greater, the point is an interior point of the current plane equation, and the number of current interior points is counted. After traversing each point in the 3D workpiece weld seam point cloud, it is determined whether the set iteration condition has been met. If yes, it means that the number of samples obtained is insufficient, and it is necessary to return to step S22. If no, it means that the number of plane equations obtained is sufficient to obtain the final weld fitting equation. The more interior points of the plane equation, the higher the overlap between the corresponding plane equation and the plane where the weld is located. Therefore, the current plane equation corresponding to the maximum number of current interior points is selected as the weld fitting equation.

[0037] Preferably, the iteration conditions can be set, but are not limited to reaching the maximum number of iterations or the iteration time, in which case the iteration process will end early.

[0038] S25: Calculate the vertical distance between each point in the initial 3D workpiece weld point cloud and the final determined weld fitting equation, and determine whether the distance is less than the set threshold. If it is not less than the threshold, remove the corresponding point cloud to obtain the optimized 3D workpiece weld point cloud.

[0039] Specifically, a second distance threshold can be set, and the vertical distance between each point in the initial 3D workpiece weld point cloud and the finally determined weld fitting equation can be calculated. It is then determined whether the distance is less than the second distance threshold. If it is not less than the threshold, it means that the distance between the point and the weld is large, which is an outlier point. The corresponding point cloud needs to be removed to obtain the optimized 3D workpiece weld point cloud, thereby removing outliers, noise points, etc., and reducing the computational complexity of subsequent steps.

[0040] S3: Extract several geometric equations from the point cloud of the weld seam of the 3D workpiece, regenerate the trajectory based on the geometric equations, and obtain the initial welding trajectory and the coordinates of the key weld seam points; Specifically, based on the point cloud of the weld seam of the 3D workpiece, any existing geometric equation extraction algorithm (such as straight line, circular arc, spline fitting) can be used to identify the basic shape of the weld seam (such as straight seam, fillet weld, bevel, V-shaped seam, curved weld, etc.). Then, based on the parameters of these equations, a continuous and smooth welding trajectory can be regenerated, and the coordinates of key weld seam points on the welding trajectory can be extracted to provide a data foundation for subsequent steps.

[0041] Preferably, the steps of extracting several geometric equations from the 3D workpiece weld point cloud, regenerating the trajectory based on the geometric equations, and obtaining the initial welding trajectory and the coordinates of key weld points may include: S31: Project the point cloud of the weld seam of the three-dimensional workpiece onto the finally determined plane equation to obtain the projected point cloud; S32: Obtain several geometric equations based on the projected point cloud, and perform interpolation calculations based on the parameters of each equation to generate the initial welding trajectory; Specifically, such as Figure 4 As shown, the original weld point cloud often contains measurement errors, surface roughness, or slight deformation. By projecting it onto the finally determined plane equation, local disturbances can be effectively smoothed, making subsequent geometric extraction more stable. Projecting the 3D workpiece weld point cloud onto the finally determined plane equation achieves dimensionality reduction from 3D to 2D, which can significantly reduce the computational load of subsequent curve fitting and equation solving, improving algorithm efficiency, especially suitable for real-time welding systems. At the same time, the projected point cloud lies on the same plane, providing a unified reference coordinate system for subsequent geometric modeling (such as lines, arcs, splines, etc.), avoiding trajectory deviations caused by perspective or installation deviations. Then, by fitting geometric equations such as lines, arcs, ellipses, and splines, complex welds can be decomposed into several standard geometric segments, achieving structured expression and facilitating robot path planning. Simultaneously, by interpolating or subtracting equation parameters (such as start point, end point, radius, curvature), dense and smooth intermediate path points can be generated, obtaining the initial welding trajectory corresponding to each geometric equation, avoiding robot motion jitter and improving welding quality.

[0042] Preferably, since workpieces in the same field often share similar shapes, taking lifting lugs in the shipbuilding industry as an example, because they are widely used in ship lifting, transportation, and fixing, they typically employ circular or arc-shaped structures to distribute stress; that is, most are composed of circular, arc-shaped, and straight structures. Figure 3 As shown, the steps of obtaining several geometric equations from the projected point cloud and generating the initial welding trajectory by interpolation calculation based on the parameters of each equation may include: S321: Perform circular fitting on the projected point cloud to obtain the circular equation, and remove the points inside the circular equation in the projected point cloud to obtain the first remaining projected point cloud. S322: Determine whether the number of the first remaining projected point cloud is greater than the first set threshold. If not, proceed to step S325. Specifically, a first remaining quantity threshold can be set. Since the core function of the lifting lug workpiece is "lifting", it usually has lifting holes or circular lugs. Prioritize fitting a circle to match its structural characteristics to improve the fitting success rate. Then, robust fitting algorithms such as RANSAC can be used to identify and remove points inside the circle to obtain the first remaining projected point cloud, which provides a data basis for subsequent steps. Then, it is determined whether the number of the first remaining projected point cloud is greater than the first set threshold. If not, it means that the remaining point cloud may be discrete noise or boundary residuals, which do not have statistical significance. That is, the current workpiece only has one circular structure and no other structures. Therefore, it is only necessary to perform interpolation calculations based on the parameters of the circular equation to obtain the initial welding trajectory.

[0043] S323: If so, then perform line fitting on the first residual projection point cloud to obtain several line equations, and remove the points inside each line equation in the residual projection point cloud to obtain the second residual projection point cloud. S324: Determine whether the number of the second remaining projection point cloud is greater than the second set threshold. If not, proceed to step S325. If yes, perform point cloud clustering on the second remaining projection point cloud to obtain several clustered point clouds. Obtain the linear equations of adjacent points in each clustered point cloud. Based on the intersection and endpoints of the adjacent linear equations, obtain the arc equation parameters of each clustered point cloud. Specifically, a second remaining quantity threshold can be optionally set. When the number of remaining projected point clouds is greater than the first set threshold, it indicates that the current workpiece may still have other structures, and further fitting is required. At the same time, the lug structure is usually "circular hole + radial straight edge" (such as square lug plate, trapezoidal reinforcing plate, circular segmental composite plate and arched composite plate) and the arc structure that plays a transition role between the circular hole and the adjacent straight edge. Therefore, a straight line fitting can be performed on the first remaining projected point cloud to obtain several straight line equations, and the points in each straight line equation in the remaining projected point cloud are removed to obtain the second remaining projected point cloud. It is then determined whether the number of the second remaining projected point cloud is greater than the second set threshold. If not, it indicates that the remaining point cloud may be discrete noise or boundary residual points, which do not have statistical significance. At this time, all the geometric structures of the current workpiece are obtained. If so, point cloud clustering can be performed on the second remaining projected point cloud to obtain several clustered point clouds. Each clustered point cloud corresponds to an arc structure. The adjacent straight line equations of each clustered point cloud can be obtained to obtain the arc equation parameters of each clustered point cloud based on the intersection and endpoints of the adjacent straight line equations.

[0044] S325: Perform interpolation calculations based on the parameters of each geometric equation to generate the initial welding trajectory.

[0045] Specifically, interpolation calculations can be performed based on the parameters of each geometric equation to generate dense and ordered welding trajectory points, thereby reducing welding deviations and improving the stability of welding quality. At the same time, the ordered trajectory points facilitate subsequent trajectory optimization, welding speed planning, and other processing to obtain the initial welding trajectory.

[0046] S33: Obtain the intersection points of adjacent geometric equations to get the coordinates of the key weld points.

[0047] Specifically, since the connection points of the geometric equations are precisely the points of curvature discontinuity, if these points are ignored, the robot will force its way through them with a single speed / acceleration, causing problems such as overshoot, jitter, and sudden changes in weld width. Furthermore, robotic welding typically uses teach points or landmark points for path planning, and the intersection points of adjacent geometric equations serve as natural segment boundaries, ensuring that each segment uses the most suitable motion commands and process parameters. Therefore, the intersection points of adjacent geometric equations can be selected as critical weld points, thus obtaining the coordinates of the critical weld points as follows:Figure 5 As shown. Preferably, when the weld seam area on the current workpiece surface only contains a circular equation, the circular equation is downsampled to obtain several points as key weld seam points.

[0048] by Figure 3 Taking a circular composite plate as an example, a circular extraction method based on Random Sample Consensus (RANSAC) can be used to extract the circles from the projected point cloud, obtaining the parametric equations of the circles. Next, points within a certain range of the circle's distance fitting are filtered out, leaving points approximately located on straight lines. These points are then further processed to extract straight lines. Finally, the intersection of the extracted circles and straight lines determines two key points, A and B. Then, paths rotating clockwise around the Z-axis are simulated with directions from A to B and from B to A, respectively. The rotation angles are compared, and the longer path is identified as point A, thus completing the ordered identification of key points.

[0049] by Figure 3 Taking the arched cladding plate as an example, the projected point cloud can be fitted with a circle based on the Random Sample Consensus (RANSAC) method to obtain the parametric equation of the circle. Then, points within a certain range of the circumference are extracted to obtain the arc point cloud Circle1 located between points C and D, and this circle is removed from the original weld seam identification point cloud to obtain the remaining point cloud Left1.

[0050] Based on this, a unit vector is constructed by extracting the start and end points (points C and D) of Circle1. This vector is then used as the reference direction for line extraction. Vectors parallel to line CD are extracted to obtain line HG. After removing the point cloud near line HG, a new residual point cloud, Left2, is obtained.

[0051] Subsequently, the cross product of the HG line and the normal vector of the plane containing the circle is used to obtain the CI direction. Based on this direction, further line extraction is performed on the Left2 point cloud to obtain the CI line and the DF line. To distinguish between the two lines, since the point cloud is acquired sequentially according to the scan trajectory, it has a sequential nature in the original point cloud. By comparing the average index positions of the corresponding points of the CI line and the DF line in the original point cloud, the line with the smaller average index is closer to the scan start point and is identified as the CI line, while the other is the DF line.

[0052] Then, using the parametric equations of the circle and line obtained from the above information, the key points are determined. The intersection of line CI and line HG is point I, and the intersection of line DF and line HG is point F. Point C is obtained by finding the intersection of line CI and the circle, and point D is obtained by finding the intersection of line DF and the circle. Since the T-shaped plate HJ and EG segments are circular transitions with a fixed radius of 30mm, given points I and F, the positions of H, J, G, E, and the center of the circle can be deduced. Assuming the position of point I is (X, Y, Z), and the unit normal vector of line CI is L... IC The direction is from I to C, and the unit normal vector of line HG is L. HG The direction is H pointing to G. The position of point H can be found as (X). H Y H Z H The position of point J can be found as (X). J Y J Z J Find the center position of the arc segment HJ (X) HJC Y HJC Z HJC The specific formula is as follows: H = I + 30 * L HG J = I + 30 * L IC Reuse L HG Same as L IC Construct a vector pointing from I to the center point of HJ, and then use the radius of √2 to find the center of the HJ arc. Further, the arc trajectory of HJ segment can be generated, and the transition trajectory of EG arc can be generated in the same way.

[0053] by Figure 3 Taking the circular composite plate as an example, the RANSAC method can be used to fit the projected point cloud into a circle to obtain the parametric equation of the circle and generate the corresponding circular trajectory.

[0054] S4: Based on the prior welding posture and the coordinates of the key weld points, perform posture mapping and interpolation calculation based on the initial welding trajectory to obtain the final welding posture trajectory.

[0055] Specifically, a person skilled in the art can manually teach the key weld points of each workpiece to obtain the prior welding posture corresponding to each key weld point, and map the prior welding posture to the corresponding key weld point of the current workpiece to obtain the current welding posture of each key weld point of the current workpiece. Then, based on the initial welding trajectory obtained in step S3 and the current welding posture of each key weld point, interpolation calculation is performed to obtain the final welding posture trajectory.

[0056] Preferably, the steps of performing attitude mapping based on the prior welding posture and the coordinates of key weld points, and interpolating based on the initial welding trajectory to obtain the final welding posture trajectory, may include: S41: Using any critical weld point as the calibration point, and based on the coordinates of the corresponding prior critical weld point, obtain the translation matrix; S42: Obtain the prior welding posture of each prior key point, and perform posture mapping according to the translation matrix to obtain the current welding posture of each key point; Specifically, since the coordinates of the critical weld points on the current workpiece and the prior critical weld points obtained through teaching are both located in the robot's base coordinate system, the corresponding critical weld point coordinates can be obtained separately to calculate the translation matrix between them. Based on this translation matrix, the prior welding posture of each prior critical point is transformed to the current critical weld point coordinates, thus obtaining the current welding posture of each critical point. Through the translation matrix between the current critical point and the prior weld points, all predefined ideal welding postures are batch-automatically mapped onto the current workpiece, eliminating the need for individual manual teaching or complex on-site adjustments for each point. This significantly improves processing efficiency and ensures that the welding torch can weld in the designed posture during the actual welding process, thereby guaranteeing welding quality.

[0057] Preferably, when the workpiece contains a circular structure, the coordinates of the center of the circle can be selected as the calibration point to improve stability.

[0058] S43: Perform pose interpolation calculations based on the current welding posture of each adjacent key point to obtain the final welding pose trajectory.

[0059] Specifically, since welding usually requires the welding torch to move in a smooth and continuous path and posture, pose interpolation (such as linear interpolation, circular interpolation, or spline interpolation) can be performed between calibrated key points to generate a continuous, abrupt, and speed-controllable welding path. Through a reasonable interpolation algorithm, a better trajectory in robot kinematics can be planned, reducing vibration and overshoot, making the robot move more smoothly and quickly, thereby improving production efficiency.

[0060] This embodiment presents a method for generating the welding pose trajectory of a workpiece fillet weld according to the present invention. It obtains two-dimensional workpiece weld points to accurately reflect the complex geometric features of the weld area, providing basic data for subsequent steps. Then, it performs three-dimensional reconstruction of the two-dimensional workpiece weld points to obtain a three-dimensional workpiece weld point cloud, improving data quality and significantly enhancing the accuracy and stability of subsequent trajectory planning and attitude estimation. Simultaneously, it provides a reliable reference benchmark for subsequent geometric primitive extraction. Furthermore, it extracts several geometric equations from the three-dimensional workpiece weld point cloud and regenerates the trajectory based on these equations to obtain the initial welding trajectory and the coordinates of key weld points. This abstracts the complex weld point cloud into geometric primitives such as straight lines, arcs, and spline curves, realizing the semanticization and structural refinement of the weld. The structured representation facilitates understanding and processing. Furthermore, the trajectory generated based on geometric primitives inherently possesses continuity and smoothness, avoiding jitter or unevenness issues caused by directly connecting discrete points, thus improving welding quality. It also identifies key feature points such as the start, end, inflection, and midpoint of the weld seam based on geometric primitives, providing crucial references for subsequent attitude planning. Finally, attitude mapping is performed based on prior welding attitudes and the coordinates of key weld seam points, and interpolation calculations are performed based on the initial welding trajectory to obtain the final welding pose trajectory. First, the prior welding attitudes corresponding to each key point are introduced to ensure that the generated welding attitude conforms to process specifications, avoiding abrupt attitude changes or unreachability. Then, the initial trajectory is optimized through an interpolation algorithm to achieve a smooth transition of the welding attitude, resulting in the final welding pose trajectory. This solves the problems of existing technologies that cannot perceive the actual spatial pose of the workpiece online, automatically identify the 3D contour of fillet welds, and generate robot motion trajectories in real time.

[0061] On the other hand, the present invention also provides a computer storage medium storing executable program code; the executable program code is used to execute the above-mentioned method for generating the welding pose trajectory of any workpiece fillet weld.

[0062] On the other hand, the present invention also provides a terminal device, including a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to execute the above-mentioned method for generating the welding pose trajectory of any workpiece fillet weld.

[0063] For example, the program code can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the program code in the terminal device.

[0064] The terminal device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the terminal device may also include input / output devices, network access devices, buses, etc.

[0065] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0066] The memory can be an internal storage unit of the terminal device, such as a hard drive or RAM. The memory can also be an external storage device of the terminal device, such as a plug-in hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units of the terminal device. The memory is used to store the program code and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output.

[0067] The aforementioned computer storage medium and terminal device are created based on the aforementioned method for generating welding pose trajectories of workpiece fillet welds. Their technical functions and beneficial effects will not be elaborated here. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0068] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method of generating a welding pose trajectory of a corner joint of a workpiece, characterized by, The method comprises the steps of: acquiring two-dimensional workpiece weld points; three-dimensional reconstruction is performed on the two-dimensional workpiece weld points to obtain three-dimensional workpiece weld point clouds; extracting a plurality of geometric equations from the three-dimensional workpiece weld point clouds, regenerating a trajectory according to the geometric equations to obtain an initial welding trajectory and key weld point coordinates; mapping the welding pose according to the prior welding pose and the key weld point coordinates, and performing interpolation calculation according to the initial welding trajectory to obtain a final welding pose trajectory.

2. The method of claim 1, wherein, The step of acquiring two-dimensional workpiece weld points comprises: constructing and training a weld point recognition model taking an original weld image as input and a weld point probability map as output; collecting an original weld image of the workpiece surface and inputting the trained weld point recognition model to obtain a weld point probability map; removing points with a confidence lower than a set threshold in the weld point probability map to obtain a plurality of connected domains; performing peak value detection on each connected domain to obtain the point with the highest confidence in each connected domain as the two-dimensional workpiece weld point.

3. The method of claim 2, wherein, The specific structure of the weld point recognition model comprises: an input layer for receiving the original weld image; a feature extraction layer for extracting features from the original weld image to obtain shallow feature maps, middle feature maps and deep feature maps; a feature fusion layer for sequentially fusing the shallow feature maps, the middle feature maps and the deep feature maps to obtain a global fusion feature map; an output layer for generating the confidence of each weld point according to the global fusion feature map and outputting a weld point probability map.

4. The method of claim 3, wherein, The feature extraction layer comprises sequentially connected shallow convolution blocks, middle convolution blocks and deep convolution blocks; the shallow convolution blocks are used to extract local features from the original weld image to obtain shallow feature maps; the middle convolution blocks are used to further extract local features from the shallow feature maps to obtain middle feature maps; the deep convolution blocks are used to further extract local features from the middle feature maps to obtain deep feature maps; the feature fusion layer comprises sequentially connected multi-scale sampling blocks, hierarchical sampling blocks and global sampling blocks; the multi-scale sampling blocks are connected with the outputs of the shallow convolution blocks and the deep convolution blocks and are used to fuse the shallow feature maps and the deep feature maps to combine the high-resolution details of the shallow feature maps and the overall structural cognition of the deep feature maps to obtain multi-scale fusion feature maps; the hierarchical sampling blocks are connected with the outputs of the middle convolution blocks and the multi-scale sampling blocks and are used to fuse the middle feature maps and the multi-scale fusion feature maps to further combine the local features of the middle feature maps to improve the classification accuracy and obtain hierarchical fusion feature maps; the global sampling blocks are connected with the output of the hierarchical sampling blocks and are used to perform upsampling and pooling operations on the hierarchical fusion feature maps to expand the network receptive field and enhance the feature expression capability to obtain the global fusion feature map.

5. The method of claim 1, wherein, The step of three-dimensional reconstruction on the two-dimensional workpiece weld points to obtain three-dimensional workpiece weld point clouds comprises: S21: three-dimensional reconstruction is performed on the two-dimensional workpiece weld points according to the prior camera calibration parameters and the laser light plane geometric model to obtain an initial three-dimensional workpiece weld point cloud; S22: a plurality of points are selected from the initial three-dimensional workpiece weld point cloud to fit a current plane equation; S23: the distance between each point in the initial three-dimensional workpiece weld point cloud and the current plane equation is obtained to determine the current in-point and count the number of current in-points; S24: judging whether the set iteration condition is reached, if not, returning to step S22, if yes, selecting the current plane equation corresponding to the most current number of inliers as the weld fitting equation according to the number of inliers counted by each plane equation; S25: calculating the perpendicular distance between each point in the initial three-dimensional workpiece weld point cloud and the weld fitting equation, and judging whether the distance is less than the set threshold, if not, eliminating the corresponding point cloud to obtain the optimized three-dimensional workpiece weld point cloud.

6. The method of claim 1, wherein, The steps of extracting a plurality of geometric equations from the three-dimensional workpiece weld point cloud, generating an initial welding trajectory and key weld point coordinates according to the geometric equations, include: projecting the three-dimensional workpiece weld point cloud onto the finally determined plane equation to obtain a projected point cloud; obtaining a plurality of geometric equations according to the projected point cloud, and performing interpolation calculation according to the parameters of each equation to generate an initial welding trajectory; obtaining the intersection points of each adjacent geometric equation to obtain the key weld point coordinates.

7. The method of claim 5, wherein, The steps of obtaining a plurality of geometric equations according to the projected point cloud, and performing interpolation calculation according to the parameters of each equation to generate an initial welding trajectory, include: S321: performing circular fitting on the projected point cloud to obtain a circular equation, and eliminating the inliers in the projected point cloud within the circular equation to obtain a first remaining projected point cloud; S322: judging whether the number of the first remaining projected point cloud is greater than a first set threshold, if not, executing step S325; S323: if yes, performing straight line fitting on the first remaining projected point cloud to obtain a plurality of straight line equations, and eliminating the inliers in the remaining projected point cloud within each straight line equation to obtain a second remaining projected point cloud; S324: judging whether the number of the second remaining projected point cloud is greater than a second set threshold, if not, executing step S325, if yes, performing point cloud clustering on the second remaining projected point cloud to obtain a plurality of clustered point clouds, obtaining adjacent straight line equations of each clustered point cloud, and obtaining the circular arc equation parameters of each clustered point cloud according to the intersection and endpoints of the adjacent straight line equations; S325: performing interpolation calculation according to the parameters of each geometric equation to generate an initial welding trajectory.

8. The method of claim 1, wherein, The steps of performing pose mapping according to the prior welding pose and the key weld point coordinates, and performing interpolation calculation according to the initial welding trajectory to obtain a final welding pose trajectory, include: taking any key weld point as a calibration point, and obtaining a translation matrix according to the corresponding prior key weld point coordinates; obtaining the prior welding poses corresponding to each key point, and performing pose mapping according to the coordinates of each key weld point to obtain the current welding poses of each key point; converting the current welding poses of each key point according to the translation matrix to obtain the current welding poses of each key point; performing pose interpolation calculation between each adjacent key point to obtain a final welding pose trajectory.

9. A computer storage medium, characterized in that The storage has executable program code; the executable program code is used to execute the welding pose trajectory generation method of the workpiece fillet weld in any one of claims 1 to 8.

10. A terminal device, comprising: It includes a memory and a processor; the memory stores program code executable by the processor; the program code is used to execute the welding pose trajectory generation method of the workpiece fillet weld in any one of claims 1 to 8.

Images