Robot welding trajectory control method and system based on three-dimensional visual recognition

Abstract

The application discloses a robot welding track control method and system based on three-dimensional visual identification, which is applied to automatic welding production of automobile key parts. The method adopts a double-view field hierarchical acquisition architecture, first completes workpiece global registration through a first view field point cloud, solves pose deviation and corrects a global welding path to obtain a coarse positioning track; in a welding process, a second view field with higher resolution is used to collect weld seam point clouds in front of a welding gun in real time, instantaneous shape features of the weld seam are extracted, a target welding posture is matched with process parameters, a track correction amount is solved and a dynamic welding track is generated and executed synchronously. The application can improve welding track control precision and weld seam forming consistency, and adapt to strict requirements of automobile key part mass production welding.

Description

Technical Field

[0001] This application relates to the field of robotic welding technology, specifically to a robotic welding trajectory control method and system based on three-dimensional vision recognition. Background Technology

[0002] Currently, with the rapid development of the automotive industry, especially the new energy vehicle industry, the lightweight design and structural complexity of key automotive components (including body load-bearing structural parts, chassis suspension parts, power battery pack housings, drive motor housings, etc.) are continuously increasing. This places more stringent mass production requirements on the dimensional accuracy, forming consistency, structural strength, and airtightness of welding processes. In the field of welding technology, automated welding using industrial robots has become a core process in the production and processing of key automotive components. The precise and stable control of the welding trajectory is a crucial factor determining the welding quality, production efficiency, and overall vehicle safety performance of automotive parts.

[0003] Currently, the mainstream approach for robotic welding trajectory control of key automotive components is to combine offline programming with pre-scanning visual positioning. This involves acquiring overall point cloud data of the workpiece through pre-scanning, completing workpiece pose registration, and then correcting the offline-planned welding path. Some optimized solutions introduce online visual acquisition modules to correct the weld position in real time during the welding process. However, in these solutions, the pre-scanning global positioning and online local correction generally adopt a serial decoupled architecture, making it difficult to simultaneously achieve both high global positioning efficiency and accurate local weld identification. Furthermore, most existing online vision solutions only perform single-dimensional trajectory correction on the weld's spatial position, failing to link and adapt the weld's morphological features such as gaps, misalignments, and curvature with welding process parameters. This makes it difficult to reliably handle complex dynamic conditions in mass production, such as workpiece clamping deviations, bevel gap fluctuations, and welding thermal deformation.

[0004] For mass production welding of key automotive components, existing technologies still have many shortcomings in adaptability: On the one hand, many key automotive components have special structures such as complex curved surface welds, spatial lap welds, and variable curvature welds. Existing solutions have poor adaptability to the welding gun posture of curved welds, which easily leads to welding defects such as insufficient penetration and weld deviation due to posture deviation. On the other hand, the arc light and high-temperature spatter continuously generated during the mass production welding of automotive components can easily interfere with the stability of visual acquisition. Existing solutions cannot guarantee the quality of point cloud acquisition and the real-time performance of trajectory correction under continuous mass production conditions, resulting in a high failure rate of welding of key automotive components, which cannot meet the long-term stability and high reliability requirements of new energy vehicles for the welding quality of core components. Summary of the Invention

[0005] This application provides a robot welding trajectory control method and system based on three-dimensional vision recognition, which at least addresses the problems existing in the prior art.

[0006] A first aspect of this application provides a robot welding trajectory control method based on three-dimensional vision recognition. This method is executed based on a three-dimensional vision sensor mounted at the end effector of the robot in front of the welding torch, and includes the following steps:

[0007] S1. In response to the welding start command, control the three-dimensional vision sensor to collect the first field of view point cloud data of the key automotive component to be welded. The first field of view point cloud data includes the workpiece's preset reference feature area.

[0008] S2. Globally register the reference feature area with the preset 3D model of the workpiece, calculate the actual pose deviation of the workpiece in the robot base coordinate system, correct the preset global welding path, and obtain the coarse positioning trajectory.

[0009] S3. During the welding process of the robot along the coarse positioning trajectory, the three-dimensional vision sensor is controlled to collect the second field of view point cloud data of the weld area in front of the welding gun in real time. The resolution of the second field of view point cloud data is higher than that of the first field of view point cloud data.

[0010] S4. Based on the point cloud data of the second field of view, extract the instantaneous morphological feature parameters of the weld, including the weld centerline, weld gap, misalignment amount and weld direction curvature.

[0011] S5. Based on the instantaneous morphological characteristic parameters, the target welding posture and target welding process parameters corresponding to the current weld point are matched and obtained.

[0012] S6. Based on the target welding posture and the curvature of the weld direction, calculate the position offset and posture compensation of multiple trajectory correction nodes on the coarse positioning trajectory.

[0013] S7. The position offset, attitude compensation, and coarse positioning trajectory are fused to generate a dynamic welding trajectory. The robot is then controlled to perform welding synchronously according to the dynamic welding trajectory and the target process parameters.

[0014] In this embodiment, by constructing a welding trajectory control process with dual-field-of-view hierarchical acquisition, global-local two-level closed loop, and synchronous linkage between trajectory and process, the problems of workpiece clamping deviation and weld morphology fluctuation in the welding of key automotive components can be solved simultaneously, greatly improving the welding trajectory control accuracy and weld formation consistency, and adapting to the stringent requirements of mass production welding of key automotive components.

[0015] In some embodiments of this application, in step S1, the three-dimensional vision sensor is controlled to move sequentially to multiple preset global scanning poses to collect point cloud data of the first field of view. Each set of scanning poses is preset based on the structural features of the key automotive components to be welded.

[0016] In this embodiment, by using multiple sets of global scanning poses preset based on the structure of key automotive components, the occlusion of reference features by the complex structure of the components can be effectively avoided, ensuring the complete acquisition of the reference features required for global registration and improving the stability and accuracy of global pose calculation.

[0017] In some embodiments of this application, in step S2, a subset of point clouds corresponding to the reference feature region is first segmented from the point cloud data of the first field of view. After removing abnormal points in the subset of point clouds, a registration constraint range is set based on the prior information of the clamping and positioning of automotive parts. A constrained iterative nearest point algorithm is used to complete global registration, and the rotation matrix and translation vector corresponding to the pose deviation are calculated.

[0018] In this embodiment, the iterative nearest point registration algorithm based on prior constraints for automotive component clamping can effectively eliminate abnormal point cloud interference caused by workpiece surface splashes and oil stains, reduce the registration solution range, significantly improve the efficiency and accuracy of global pose registration, and avoid welding defects caused by clamping deviations.

[0019] In some embodiments of this application, in step S3, the arc intensity data of the welding pool is collected in real time, and the exposure parameters of the three-dimensional vision sensor and the laser projection intensity are dynamically adjusted according to the arc intensity. At the same time, the sampling frequency of the sensor is adjusted synchronously according to the robot's welding walking speed to ensure that the density of the second field of view point cloud data is constant.

[0020] In this embodiment, by adjusting the sensor acquisition parameters in conjunction with the arc intensity and welding walking speed, the interference of molten pool arc light and spatter on visual acquisition can be effectively offset, ensuring the constant density and stable quality of the second field of view point cloud data during continuous welding, and providing a reliable basis for accurate extraction of weld features.

[0021] In some embodiments of this application, in step S3, the acquisition distance in front of the welding torch is 8mm-25mm, and the acquisition distance is dynamically adjusted with the welding travel speed and arc intensity; when the welding travel speed increases or the arc intensity increases, the acquisition distance is increased accordingly.

[0022] In this embodiment, by dynamically adjusting the weld seam acquisition distance in front of the welding torch, the direct interference of molten pool arc light and high-temperature spatter on visual acquisition is further avoided while ensuring the real-time performance of trajectory correction, thereby improving the anti-interference capability of visual acquisition during the continuous mass production welding of automotive parts.

[0023] In some embodiments of this application, in step S5, the weld gap and misalignment are combined with the material, plate thickness and bevel type of the key automotive component to be welded and imported into a pre-trained welding process decision model, and the target process parameters including welding current, voltage, wire feed speed and travel speed are output. The model is trained based on the historical welding data and quality inspection results of the corresponding automotive component, and iteratively optimized after each workpiece welding is completed.

[0024] In this embodiment, by using a welding process decision model adapted to the attributes of key automotive components, the optimal process parameters corresponding to weld gap and misalignment can be accurately matched. At the same time, the model adaptability is continuously improved through iterative optimization of mass production data, which significantly reduces the defect rate of insufficient penetration and incomplete weld penetration in automotive component welding.

[0025] In some embodiments of this application, in step S6, trajectory correction nodes are dynamically selected based on the weld curvature; when the weld curvature is less than a preset threshold, nodes are selected at equal intervals along the coarse positioning trajectory; when the weld curvature is greater than or equal to the preset threshold, the node spacing is reduced as the curvature increases, ensuring that the correction node density of the curved weld segment is higher than that of the straight weld segment.

[0026] In this embodiment, the selection method of trajectory correction node with weld curvature adaptive balances the welding efficiency of straight welds with the trajectory correction accuracy of curved welds. This can effectively adapt to the welding requirements of complex curved surfaces and variable curvature welds of key automotive components, ensuring smooth transition of welding torch posture and uniformity of welding formation in curved weld segments.

[0027] In some embodiments of this application, in step S7, the position offset and attitude compensation of each trajectory correction node are first smoothed by spline interpolation to generate a continuous correction compensation curve; the correction compensation curve is verified based on the robot's kinematic constraints and joint speed limits, and after eliminating motion singularities and speed over-limit problems, it is fused with the coarse positioning trajectory to generate a dynamic welding trajectory.

[0028] In this embodiment, by using smooth interpolation and robot kinematic constraint verification, the problems of robot motion singularities and joint speed exceeding limits caused by trajectory correction can be effectively eliminated, ensuring that the welding torch travel speed is constant during the welding process, avoiding poor weld formation caused by robot vibration, and improving the surface quality and consistency of automotive parts welding.

[0029] In some embodiments of this application, in step S4, after extracting the instantaneous morphological feature parameters of the weld, the temperature data of the current weld pool is collected simultaneously, and the thermal deformation offset of the subsequent weld is predicted based on the pre-established weld pool temperature-welding thermal deformation mapping model; in step S6, when calculating the position offset, the thermal deformation offset is included in the calculation to pre-correct the subsequent welding trajectory.

[0030] In this embodiment, the weld offset caused by welding thermal deformation is pre-corrected by mapping the molten pool temperature with the welding thermal deformation. This can effectively solve the problem of weld deviation caused by thermal deformation in the welding of thin plates of key automotive components, and greatly improve the dimensional accuracy and pass rate of welding thin-walled automotive components.

[0031] A second aspect of this application provides a robot welding trajectory control system based on three-dimensional vision recognition. The system includes a robot body, a three-dimensional vision sensor, a global registration module, a feature extraction module, a process matching module, a trajectory correction module, and a robot control cabinet. The three-dimensional vision sensor and welding torch are mounted at the end of the robot body. The global registration module is used to complete workpiece pose calculation and coarse positioning trajectory generation. The feature extraction module is used to extract instantaneous morphological feature parameters of the weld. The process matching module is used to match the target welding posture and process parameters. The trajectory correction module is used to generate a dynamic welding trajectory. The robot control cabinet is used to control the robot to perform welding operations.

[0032] In this embodiment of the application, a robot welding trajectory control system is constructed, which covers the entire process control requirements for automated welding of key automotive components. It can be directly adapted to the industrial robot architecture of existing mass production welding lines for automotive components, and has strong engineering feasibility and application value. Attached Figure Description

[0033] Figure 1 A flowchart illustrating a robot welding trajectory control method based on three-dimensional vision recognition, provided as an embodiment of this application; Detailed Implementation

[0034] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0035] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1 The following is an explanation using specific examples.

[0036] In the automated welding production of key automotive components, industrial robots are core equipment. Currently, there are two main methods for controlling robot welding trajectories: one relies entirely on offline programming, where all welding paths are pre-planned in software. However, this method cannot handle deviations between the actual and theoretical weld positions caused by workpiece clamping, thermal deformation, etc. The other method introduces a vision system for guidance, but mostly employs a serial decoupled architecture of "global scanning and positioning followed by online local correction." In this architecture, global scanning typically uses low-resolution point cloud data to cover the entire workpiece, resulting in slow processing speeds. Furthermore, the positioning results are used to guide the entire welding process, and cannot respond promptly to new deviations (such as thermal deformation) that occur during subsequent welding. Online local correction often only focuses on the weld center position, failing to provide sufficient information to guide the synchronous adjustment of welding torch posture and welding processes for complex conditions such as weld gap changes, misalignment, and curvature variations in curved welds. Therefore, existing technologies struggle to consistently and stably guarantee welding quality in the complex scenarios of high-frequency, high-precision, and multi-condition mass production welding of automotive parts.

[0037] To address the aforementioned issues, this application provides an in-depth analysis of the sources of deviations and the visual guidance mechanism in the welding process of key automotive components. The inventors discovered that deviations during welding are not singular or static, but rather constitute a combination of "global deviations in workpiece clamping" and "local dynamic deviations caused by gaps, thermal deformation, etc., during welding." Global deviations are relatively fixed and can be corrected through a single, low-resolution scan with a large field of view, thus obtaining a rough but globally accurate guidance trajectory. Local dynamic deviations, such as fluctuations in bevel gaps and deformation of thin plates after heating, are instantaneous and require high precision, necessitating real-time sensing using high-resolution, small-field-of-view vision sensors before the welding torch arrives. More importantly, these local dynamic deviations not only affect the geometric position of the weld but also directly determine the filling state and forming quality of the weld pool. For example, excessive gaps can lead to burn-through, excessive misalignment can result in incomplete fusion, and abrupt changes in curvature at points in curved welds require instantaneous adjustment of the welding torch posture. Therefore, simple trajectory position correction is insufficient to solve all problems. It is necessary to deeply couple the instantaneous morphological characteristics of the weld (such as gap, misalignment, and curvature) with welding process parameters (such as current, voltage, and welding torch posture) to achieve integrated and coordinated control of trajectory, posture, and process.

[0038] Based on the above analysis, this application proposes an innovative technical solution. This solution constructs a dual-field-of-view, two-level closed-loop control architecture of global coarse positioning and local fine correction. Furthermore, it uses the weld morphology features extracted during the local fine correction stage as the basis for dynamically adjusting the welding posture and process parameters, thereby achieving multi-dimensional, end-to-end collaborative control from global to local, from position to posture, and from trajectory to process. Specifically, this application first utilizes low-resolution, large-field-of-view point cloud data to quickly calculate the global pose deviation of the workpiece, generating a global coarse positioning trajectory, thus solving the fundamental problem caused by workpiece clamping deviation. Subsequently, during the welding process, high-resolution, small-field-of-view point cloud data is used to capture the weld morphology in front of the welding torch in real-time and with precision, including the centerline, gap, misalignment, and curvature. These morphological parameters are used to dynamically match the optimal welding torch posture and welding process parameters, and simultaneously calculate the trajectory offset and posture compensation. Finally, these compensations are smoothly integrated into the coarse positioning trajectory to generate a dynamic welding trajectory that can adapt to global workpiece deviations and local dynamic changes, guiding the robot to complete the welding with the optimal posture and process. The entire solution organically integrates global static deviation compensation with local dynamic change response, forming a closed-loop, efficient, and precise welding trajectory control system, providing a brand-new approach to solving the problem of high-quality welding of key automotive components.

[0039] Please refer to Figure 1 , Figure 1 An embodiment of this application provides a robot welding trajectory control method based on three-dimensional vision recognition. This method is executed based on a three-dimensional vision sensor mounted at the end of the robot in front of the welding torch, and includes the following steps:

[0040] S1. In response to the welding start command, control the three-dimensional vision sensor to collect the first field of view point cloud data of the key automotive component to be welded. The first field of view point cloud data includes the workpiece's preset reference feature area.

[0041] S2. Globally register the reference feature area with the preset 3D model of the workpiece, calculate the actual pose deviation of the workpiece in the robot base coordinate system, correct the preset global welding path, and obtain the coarse positioning trajectory.

[0042] S3. During the welding process of the robot along the coarse positioning trajectory, the three-dimensional vision sensor is controlled to collect the second field of view point cloud data of the weld area in front of the welding gun in real time. The resolution of the second field of view point cloud data is higher than that of the first field of view point cloud data.

[0043] S4. Based on the point cloud data of the second field of view, extract the instantaneous morphological feature parameters of the weld, including the weld centerline, weld gap, misalignment amount and weld direction curvature.

[0044] S5. Based on the instantaneous morphological characteristic parameters, the target welding posture and target welding process parameters corresponding to the current weld point are matched and obtained.

[0045] S6. Based on the target welding posture and the curvature of the weld direction, calculate the position offset and posture compensation of multiple trajectory correction nodes on the coarse positioning trajectory.

[0046] S7. The position offset, attitude compensation, and coarse positioning trajectory are fused to generate a dynamic welding trajectory. The robot is then controlled to perform welding synchronously according to the dynamic welding trajectory and the target process parameters.

[0047] The execution component in this embodiment is a three-dimensional vision sensor, which is front-mounted on the welding torch at the robot's end effector and moves with the torch. This front-mounted installation allows the sensor to always be positioned in front of the welding torch's forward direction, enabling it to detect the area to be welded in advance. This sensor can be a device based on principles such as structured light, laser stripes, or binocular stereo vision, capable of acquiring three-dimensional point cloud data of the space surrounding the welding torch.

[0048] In this application, key automotive components refer to components that constitute the core functional structure of an automobile, have a decisive impact on the safety, reliability, and performance of the entire vehicle, and whose manufacturing precision and welding quality are directly related to the product qualification rate and service life. Specifically, these include, but are not limited to, body load-bearing structural components, chassis suspension components, power battery pack housings, drive motor housings, subframes, and various support brackets and connectors involving airtightness or structural strength. These components are usually made of high-strength steel, aluminum alloys, or lightweight composite materials, and have complex curved surfaces, variable curvature welds, thin plate overlaps, or spatially staggered structures, which impose stringent mass production requirements on dimensional stability, thermal deformation control, and weld formation consistency during the welding process.

[0049] This application defines two types of point cloud data for different purposes. The first type is first-field-of-view point cloud data, acquired before welding begins. It features a large field of view but relatively low resolution, and its purpose is to quickly acquire a complete point cloud containing pre-defined reference feature regions on the workpiece to identify the workpiece's actual position and orientation in space. These reference feature regions can be specially designed positioning holes or surfaces on the workpiece, or corners or edges with distinct contour features, such as positioning pin holes on a car battery tray or specific mounting surfaces on a car body side panel. The second type is second-field-of-view point cloud data, acquired in real-time during welding. It features high resolution but a smaller field of view, specifically designed for finely perceiving details in a small section of the weld seam in front of the welding torch. Its resolution is typically much higher than that of the first-field-of-view point cloud data, enabling clear differentiation of minute morphological changes in the weld seam, such as gaps or misalignments at the 0.1mm level.

[0050] In this embodiment, global registration refers to the process of matching and aligning the acquired first field-of-view point cloud data with the pre-stored 3D digital model of the workpiece. Using an algorithm (such as a constrained iterative nearest-point algorithm, or any existing algorithm), the deviation between the actual and theoretical poses of the workpiece can be calculated, typically represented by a rotation matrix and a translation vector. Based on this deviation, the originally preset global welding path can be corrected to obtain a coarse positioning trajectory that more closely matches the actual workpiece position. This trajectory provides a relatively accurate reference for subsequent fine welding.

[0051] During the welding process, the system extracts instantaneous morphological feature parameters of the weld from the second field of view point cloud data. These parameters mainly include: the weld centerline, i.e., the geometric centerline of the weld trajectory; the weld gap, referring to the size of the gap at the butt joint or lap joint of the welded workpieces; the misalignment, referring to the height difference caused by misalignment between two welded workpieces at the butt joint; and the weld curvature, i.e., the degree of curvature of the weld centerline in a plane or space. These parameters collectively describe the real-time geometric state of the weld. To cope with the complex working conditions in the mass production welding of automotive parts, this application introduces an adaptive adjustment mechanism when acquiring the second field of view point cloud data. The system acquires the arc light intensity emitted by the weld pool in real time and dynamically adjusts the exposure time of the 3D vision sensor and the laser projection intensity according to the changes in arc light intensity to avoid image saturation or insufficient contrast due to excessive arc light. At the same time, it also adjusts the sampling frequency of the sensor synchronously according to the robot's welding walking speed to ensure that the point cloud density acquired per unit length of weld remains constant at different speeds, thereby ensuring the stability and accuracy of weld feature extraction.

[0052] Furthermore, this application does not merely perform trajectory correction, but rather links weld morphology with process parameters. Specifically, the system inputs the extracted weld gap and misalignment, along with prior information such as the material, plate thickness, and bevel type of the workpiece to be welded, into a pre-trained welding process decision model. This model, trained with a large amount of historical welding data and corresponding quality inspection results, can intelligently recommend optimal target process parameters such as welding current, voltage, wire feed speed, and travel speed based on the input weld morphology. Moreover, after each workpiece welding is completed, the model can iteratively optimize based on the data from the welding process and the final quality results, making its recommended parameters increasingly accurate. When generating the final dynamic welding trajectory, this application introduces the concept of trajectory correction nodes. The system does not correct all trajectory points, but selectively selects some nodes on the coarse positioning trajectory for calculation. The node selection strategy is related to the weld curvature: when the weld curvature is small, nodes can be selected at equal intervals to improve computational efficiency; when the weld curvature is large, it means the weld is more curved, requiring a higher correction density. Therefore, the node spacing is reduced as the curvature increases to ensure smooth and accurate trajectory correction even in curved weld segments. For the calculated node position offsets and attitude compensations, the system first performs spline interpolation to smoothly transition them from discrete node values ​​to a continuous correction compensation curve. This avoids abrupt changes in the robot's motion trajectory. Subsequently, this curve is verified based on robot kinematic constraints (such as maximum joint velocity and acceleration) to eliminate potential motion singularities or velocity exceedances, ensuring that the generated dynamic welding trajectory is smoothly executable by the robot.

[0053] Furthermore, to address the common problem of welding thermal deformation in thin plate welding, this application proposes a pre-correction scheme. While extracting the instantaneous morphology of the weld, the system simultaneously acquires the temperature data of the weld pool and, based on a pre-established mapping model between the weld pool temperature and welding thermal deformation, predicts the potential offset of the weld due to thermal effects. When calculating the position offset of subsequent trajectory correction nodes, this predicted thermal deformation offset is also taken into account, thus achieving "early prediction and early correction" and preventing weld deviation caused by the welding torch failing to respond to thermal deformation in a timely manner.

[0054] In summary, the technical solution of this application achieves significant synergistic effects through the organic combination of the aforementioned features. First, the dual-field-of-view hierarchical acquisition strategy balances the efficiency of global positioning with the accuracy of local perception, laying a solid data foundation for the entire control process. Second, combining the coarse positioning trajectory generated by global registration with local correction based on real-time weld morphology forms a closed-loop control system from macro to micro and from static to dynamic, capable of simultaneously addressing clamping deviations and welding process fluctuations. Third, the deep coupling of weld morphology features with welding process parameters and welding torch posture changes the previous single control mode that only considered position and ignored morphology, achieving refined control of welding quality. Finally, optimization of node selection, trajectory smoothing, motion calibration, and thermal deformation pre-compensation ensures the stability, smoothness, and foresight of the entire control process, resulting in a highly accurate dynamic welding trajectory that is executed smoothly and reliably by the robot. These features support each other and work together, enabling this application to effectively solve the problem of unstable welding quality in complex working conditions of key automotive components, significantly improving the weld pass rate and consistency.

[0055] In some embodiments of this application, in step S1, the three-dimensional vision sensor is controlled to move sequentially to multiple preset global scanning poses to collect point cloud data of the first field of view. Each set of scanning poses is preset based on the structural features of the key automotive components to be welded.

[0056] In this embodiment, multiple global scanning poses refer to a series of different imaging positions and angles pre-set in the robot program to completely acquire the reference feature areas of the key automotive components to be welded. Since key automotive components typically have complex three-dimensional structures, such as battery trays with reinforcing ribs, motor housings with deep cavities, or multi-layered subframes, visual acquisition from a single perspective often cannot simultaneously cover all preset reference feature areas; some areas may be obscured by the component's own protruding structures. Therefore, multiple scanning poses can be pre-planned based on the component's three-dimensional digital model, allowing the robot, equipped with a three-dimensional vision sensor, to sequentially move to these poses and capture images of the workpiece from different directions and distances. Each scanning pose includes the sensor's position coordinates and orientation in space, set according to the distribution of reference feature areas on the component to be welded and the overall structural characteristics of the component. For example, for a subframe with a complex curved surface, scanning poses can be set at its front, side, and end to ensure that positioning holes or feature edges located on different planes can be clearly captured. Based on the pre-defined structural features of the key automotive components to be welded, this means that the planning of the aforementioned multiple sets of scanning poses is not arbitrary, but rather fully considers the geometry, size, and spatial layout of the reference features of the components. Specifically, during the offline programming stage, technicians import the 3D digital model of the component to be welded into simulation software to analyze its surface undulations, hollow structures, reinforcing rib orientation, and the location of the pre-designed positioning reference. Based on these structural characteristics, through simulation or teaching, several imaging poses are planned for the sensor to capture complete reference features without obstruction. These poses typically avoid areas of high curvature abrupt changes, deep cavities, or structures with inverted joints, selecting locations with a wide field of view and a small angle between the sensor and the normal direction of the reference features to obtain higher quality point cloud data with less distortion. This pre-defined process ensures that regardless of any clamping deviations in the actual incoming material, as long as the sensor sequentially captures data according to this pre-defined set of poses, it can capture sufficient information for global positioning from multiple angles.

[0057] This application utilizes the collaborative work of multiple pre-set poses to provide complete and high-quality input for subsequent global registration. If only a single scanning pose is used, for complex automotive critical components, occlusion can easily lead to missing point clouds in the reference feature region, making it difficult for subsequent global registration algorithms to find sufficient matching features. This affects the accuracy of pose deviation calculation and may even cause registration failure. However, by pre-setting multiple sets of structural feature-based scanning poses, the robot can guide the vision sensor to sequentially acquire data from multiple advantageous angles. After fusing the point cloud data from different perspectives, the entire picture of the reference feature region can be reconstructed, effectively avoiding the occlusion problem of a single viewpoint. Furthermore, since these scanning poses are pre-planned optimal acquisition positions based on the structural characteristics of the component, compared to real-time random angle searching, the time consumed in the global scanning stage is significantly shortened, improving production line cycle time. This pre-planned, orderly acquisition method lays a stable and reliable data foundation for the entire welding control system, enabling subsequent global registration based on this data to be completed more quickly and accurately.

[0058] In some embodiments of this application, in step S2, a subset of point clouds corresponding to the reference feature region is first segmented from the point cloud data of the first field of view. After removing abnormal points in the subset of point clouds, a registration constraint range is set based on the prior information of the clamping and positioning of automotive parts. A constrained iterative nearest point algorithm is used to complete global registration, and the rotation matrix and translation vector corresponding to the pose deviation are calculated.

[0059] In this embodiment, point cloud subset segmentation refers to extracting 3D data points belonging to the reference feature region (e.g., positioning holes, positioning surfaces, or specific geometric corners on the workpiece) from the entire first field-of-view point cloud data based on spatial location or geometric features, forming an independent point cloud set so that subsequent processing is only performed on this region, reducing computational load. Outlier removal involves using statistical filtering, radius filtering, and other methods to remove outliers caused by oil stains, splashes, or sensor noise on the workpiece surface after obtaining the point cloud subset. These outliers interfere with the accuracy of registration, and removal results in cleaner geometric features. Clamping and positioning prior information refers to the possible pose range of the workpiece or the geometric constraint relationship between reference features known in advance based on the workpiece's installation method on the fixture. For example, after a car battery tray is installed on the fixture, the angle between the axis direction of its positioning pin hole and the robot's base coordinate system is limited to a small fluctuation range. This information can be used as boundary conditions during registration. The constrained iterative closest point algorithm is an improved form of the iterative closest point algorithm. It incorporates constraints based on prior information about clamping and positioning into the traditional iterative process of finding the optimal rotation matrix and translation vector. For example, it limits the range of rotation angles or the length of the translation vector. This ensures that the registration result satisfies both the geometric matching degree between point clouds and the actual physical installation state of the workpiece, avoiding incorrect matching due to similar local features in the point clouds. The rotation matrix and translation vector are mathematical expressions describing the deviation between the actual pose of the workpiece and its theoretical pose. The rotation matrix represents the rotation angle around the three coordinate axes, and the translation vector represents the distance moved along the three coordinate axes. Together, they constitute a rigid transformation relationship from the theoretical numerical model coordinate system to the actual workpiece coordinate system.

[0060] In step S2, this application establishes a close collaborative relationship. First, by segmenting the point cloud into subsets, the processing scope is narrowed from the global point cloud to the reference feature region, directly reducing the data volume of subsequent registration algorithms and improving processing efficiency. The subsequent outlier removal ensures that the feature data used for registration is clean and reliable, reducing the negative impact of interference factors on registration accuracy from the source. Based on this, prior information from clamping and positioning is used to set registration constraints for the iterative nearest-point algorithm, essentially providing the algorithm with a solution boundary that conforms to physical reality. This effectively avoids the algorithm getting trapped in local optima due to repeated workpiece surface textures or incomplete point clouds, making the solved rotation matrix and translation vector more accurate and unique. These four steps are sequentially linked, from data refinement and quality purification to constraint solving, jointly ensuring the high efficiency and robustness of the workpiece's global pose deviation calculation, laying a solid foundation for the subsequent generation of accurate and reliable coarse positioning trajectories.

[0061] In some embodiments of this application, in step S3, the arc intensity data of the welding pool is collected in real time, and the exposure parameters of the three-dimensional vision sensor and the laser projection intensity are dynamically adjusted according to the arc intensity. At the same time, the sampling frequency of the sensor is adjusted synchronously according to the robot's welding walking speed to ensure that the density of the second field of view point cloud data is constant.

[0062] In this embodiment, arc intensity data refers to the quantified value of the intensity of the electric arc light generated by the molten pool during welding. This can be acquired in real time by integrating a photoelectric sensor into the 3D vision sensor or by using grayscale information collected by an image sensor. The arc intensity fluctuates with changes in welding current, voltage, and shielding gas state. Dynamically adjusting exposure parameters automatically adjusts the exposure time and gain value of the 3D vision sensor based on the real-time acquired arc intensity. For example, when the arc light intensifies, the exposure time is shortened to avoid overexposure; when the arc light weakens, the exposure time is appropriately extended to enhance details in dark areas, ensuring the sensor always operates within a suitable light-sensitive range. Laser projection intensity refers to the power of the structured light or laser stripes projected by the 3D vision sensor to acquire point cloud data. Synchronously adjusting the laser projection intensity according to the arc light intensity can increase the laser power to enhance the signal-to-noise ratio when the arc light is strong, ensuring the laser stripes remain clearly discernible against a strong arc light background. Robot welding walking speed refers to the speed at which the welding torch at the robot's end moves along the weld seam, typically determined by welding process parameters. In mass production welding, this speed may change due to process switching or adaptive adjustments. Synchronous adjustment of the sensor's sampling frequency refers to linking the data acquisition frequency of the 3D vision sensor with the welding walking speed. The sampling frequency is increased when the welding walking speed is high and decreased when the speed is low. The aim is to keep the number of point cloud frames acquired per unit weld length approximately the same. Maintaining constant point cloud data density means that, through the above adjustments, the final stitched weld area point cloud has a uniform sampling interval along the weld direction. For example, regardless of changes in welding speed, approximately a fixed number of point cloud data points can be obtained per millimeter of weld length, avoiding situations where the point cloud is sparse when the speed is high and redundant when the speed is low.

[0063] In step S3, this application establishes a mutually supportive stabilization mechanism. The coordinated adjustment of arc light intensity and laser projection intensity effectively creates an adaptive exposure and illumination system for the vision sensor in a dynamically changing, highly interfering environment. This ensures that the sensor can consistently acquire clear and stable original images during mass production welding operations characterized by arc flashes and frequent spatter. Simultaneously, the synchronized adjustment of welding travel speed and sampling frequency guarantees a uniform spatial distribution of the point cloud, preventing excessive density differences in different areas of the weld due to speed fluctuations. These two mechanisms work synergistically, ensuring the quality of point cloud acquisition by counteracting arc light interference and maintaining spatial distribution consistency through speed coordination. Together, they provide a high-quality, stable, and uniformly dense second field-of-view point cloud data foundation for subsequent weld morphology feature extraction, enabling the feature extraction algorithm to obtain reliable and consistent input during continuous welding.

[0064] In some embodiments of this application, in step S3, the acquisition distance in front of the welding torch is 8mm-25mm, and the acquisition distance is dynamically adjusted with the welding travel speed and arc intensity; when the welding travel speed increases or the arc intensity increases, the acquisition distance is increased accordingly.

[0065] In this embodiment, the acquisition distance refers to the spatial distance between the 3D vision sensor and the weld area to be acquired in front of the welding torch. It is typically measured along the direction of the welding torch's movement. This distance determines the environmental conditions the sensor is in during welding, as well as the field of view and resolution of the acquired weld area. Dynamic adjustment means that the acquisition distance is not fixed but automatically and adaptively adjusted according to the real-time changing parameters during welding, ensuring it remains within a range conducive to stable acquisition. Welding travel speed refers to the speed at which the robot moves the welding torch along the weld. When the welding travel speed increases, the molten pool area and spatter-affected area in front of the welding torch move towards the sensor more quickly. Appropriately increasing the acquisition distance allows the sensor more reaction time and avoidance space. Arc intensity refers to the strength of the electric arc light generated by the welding molten pool. When the arc intensity increases, the thermal radiation and light interference around the molten pool are more intense. Appropriately increasing the acquisition distance can reduce the direct radiation interference of the molten pool arc light on the sensor imaging, placing the sensor in a relatively favorable acquisition environment. 8mm-25mm is an optimal working range for acquisition distance. The lower limit of 8mm helps to ensure sufficient point cloud resolution to identify minute weld features, while the upper limit of 25mm balances the safe distance between the sensor and the molten pool with the effectiveness of the acquisition field of view.

[0066] This application establishes an adaptive acquisition distance control mechanism that adjusts according to welding conditions. Welding travel speed and arc intensity, as two key real-time operating condition indicators, affect the sensor's acquisition environment from two dimensions: physical space and optical interference, respectively. When the welding speed increases or the arc intensity intensifies, the threat to the sensor from high-temperature spatter and strong light interference from the molten pool increases accordingly. Dynamically increasing the acquisition distance at this time is equivalent to actively retracting the sensor to a position with relatively weaker interference, avoiding the risk of sensor damage from spatter or image saturation, while maintaining the geometric relationship of the welding torch and sensor in the front mounting. At the same time, the 8mm-25mm range setting provides clear boundary constraints for this dynamic adjustment, ensuring that the sensor is neither placed in the center of harsh operating conditions due to excessive distance, nor loses necessary point cloud resolution due to excessive distance. This method of actively adjusting the acquisition distance according to operating conditions complements the previously described adjustment of exposure parameters and sampling frequency, further enhancing the anti-interference capability and adaptability of the vision acquisition system in continuous mass production welding from the spatial distance dimension.

[0067] In some embodiments of this application, in step S5, the weld gap and misalignment are combined with the material, plate thickness and bevel type of the key automotive component to be welded and imported into a pre-trained welding process decision model, and the target process parameters including welding current, voltage, wire feed speed and travel speed are output. The model is trained based on the historical welding data and quality inspection results of the corresponding automotive component, and iteratively optimized after each workpiece welding is completed.

[0068] In this embodiment, weld gap refers to the width of the gap between two workpieces during welding, usually expressed in millimeters. For example, it is the opening distance between the two base materials on both sides of a butt weld. If the gap is too small, it is easy to cause incomplete penetration, while if the gap is too large, it may cause burn-through or weld beads. Misalignment refers to the height deviation caused by misalignment at the lap or butt joint of two workpieces. It is manifested as the degree to which one workpiece protrudes or recedes relative to the other. Excessive misalignment will affect the fusion quality of the weld and the joint strength. Material refers to the type of material of the workpieces to be welded. For example, high-strength steel, aluminum alloy, or magnesium alloy commonly used in key automotive components. Different materials have significant differences in thermal conductivity, melting point, and welding process window. Plate thickness refers to the thickness dimension of the workpieces to be welded. When welding plates of different thicknesses, it is necessary to adjust the heat input to balance the penetration of the thinner plate and the full fusion of the thicker plate. Bevel type refers to the geometry opened at the joint of the workpieces before welding, such as a V-shaped bevel, a U-shaped bevel, or a single-sided V-shaped bevel. Different bevel forms determine the filler amount of welding wire and the oscillation mode of the welding torch.

[0069] The pre-trained welding process decision model is an artificial intelligence model that has learned from a large amount of historical data. It can be an existing structure such as a neural network, random forest, or support vector machine. Its function is to map the optimal combination of welding process parameters based on the input weld morphology and workpiece attributes; this application does not limit this. Historical welding data refers to complete process parameters, weld morphology information, and corresponding process sensor data recorded from previous production of the same type of automotive parts. This data constitutes the basic samples for model training. Quality inspection results refer to evaluation indicators obtained from non-destructive testing, mechanical property testing, or dimensional measurement of the welded workpiece, such as weld penetration depth, porosity, joint tensile strength, or weld appearance quality. These results serve as labels to guide the model in learning the correspondence between excellent processes. Target process parameters are the specific welding settings output by the model for the current weld point, including welding current, voltage, wire feed speed, and travel speed. These parameters collectively determine the magnitude of the welding heat input and the filling behavior of the molten pool. Iterative optimization refers to using the actual weld morphology, process parameters, and final quality inspection results of each workpiece as new samples after welding, and then incrementally training or fine-tuning the model to improve the model's decision-making ability as production batches accumulate.

[0070] In step S5, this application establishes a complete closed loop from data input to decision output and then to self-evolution. Weld gap and misalignment reflect the immediate geometric state of the weld point, while material, plate thickness, and groove type represent the inherent properties of the workpiece itself. These five types of inputs together constitute the complete basis for welding process decisions, considering both real-time fluctuations and prior knowledge. The pre-trained welding process decision model maps this multi-dimensional input to target process parameters, achieving an intelligent leap from weld morphology to process setting. More importantly, the iterative optimization mechanism after completing each workpiece allows the model to continuously absorb successful experiences and lessons learned from the latest production processes. As the number of welded workpieces increases, the model's understanding of the matching relationship between different weld morphologies and process parameters becomes more accurate and stable. This continuous evolution capability means that process parameters no longer rely on fixed process specification tables but can dynamically adapt to complex factors such as workpiece consistency fluctuations and environmental changes during mass production, providing an adaptive technical foundation for maintaining consistent welding quality.

[0071] In some embodiments of this application, in step S6, trajectory correction nodes are dynamically selected based on the weld curvature; when the weld curvature is less than a preset threshold, nodes are selected at equal intervals along the coarse positioning trajectory; when the weld curvature is greater than or equal to the preset threshold, the node spacing is reduced as the curvature increases, ensuring that the correction node density of the curved weld segment is higher than that of the straight weld segment.

[0072] In this embodiment, the weld curvature refers to the degree of curvature of the weld centerline in space. Mathematically, it can be understood as the rate of change of the tangent angle at a certain point on the curve. The greater the curvature, the more abrupt the weld curvature. For example, the weld curvature at the corner of a car battery tray is much greater than the curvature of its straight planar segment. Trajectory correction nodes refer to several discrete positions selected on the coarse positioning trajectory. The system will specifically calculate the required position offset and attitude compensation at these points, and the trajectory correction between nodes is smoothly transitioned through interpolation. The preset threshold is a pre-set curvature critical value used to determine whether the degree of weld curvature reaches the point where the node selection strategy needs to be adjusted. This threshold can be set according to the robot's kinematic performance or welding process requirements. Equal-spacing node selection refers to selecting a correction node at fixed intervals along the coarse positioning trajectory in gentle areas with small weld curvature. This uniform distribution method is simple to calculate and can meet the correction requirements for straight or gently curved welds. Reducing the node spacing as curvature increases means that in curved areas where the weld has significant curvature, the spacing between adjacent correction nodes is reduced accordingly as the curvature value increases, resulting in a denser distribution of nodes at the bends. Correction node density refers to the number of correction nodes set per unit length of weld trajectory. The node density is higher in curved weld segments than in straight weld segments, indicating that the system performs trajectory correction with finer granularity in curved sections.

[0073] In step S6, this application achieves a balance between computational efficiency and correction accuracy. By introducing the weld curvature as an adaptive basis for node selection, the system can identify differences in weld geometry and adopt differentiated processing strategies. For straight or gently curved sections with small curvature, selecting nodes at equal intervals avoids excessive computational overhead, allowing resources to be concentrated on subsequent welding processes. For curved sections with large curvature, reducing the node spacing as curvature increases ensures sufficient correction nodes in areas where the weld direction changes rapidly, enabling the system to more precisely describe the curvature changes of the trajectory, thus providing denser control basis for welding torch attitude compensation. This method of dynamically adjusting the correction density based on weld geometry ensures that coarse positioning trajectory correction does not consume excessive computational power in simple areas, nor does it result in insufficient trajectory fitting accuracy in complex areas due to too few correction points, ultimately achieving better adaptability between resource utilization and control effect in the entire trajectory correction process.

[0074] In some embodiments of this application, in step S7, the position offset and attitude compensation of each trajectory correction node are first smoothed by spline interpolation to generate a continuous correction compensation curve; the correction compensation curve is verified based on the robot's kinematic constraints and joint speed limits, and after eliminating motion singularities and speed over-limit problems, it is fused with the coarse positioning trajectory to generate a dynamic welding trajectory.

[0075] In this embodiment, the position offset and attitude compensation of each trajectory correction node refer to the position correction value and attitude correction value calculated from the instantaneous morphological characteristics of the weld at discrete points selected on the coarse positioning trajectory. The former describes the displacement that the welding torch end needs to move, and the latter describes the pointing angle that the welding torch needs to adjust. Spline interpolation smoothing is a mathematical method that uses spline curves, such as cubic splines or B-splines, to fit the offset and compensation of discrete correction nodes into a continuous curve, making the correction amount exhibit a smooth change between nodes rather than a step change, thereby avoiding sudden turns or jitters in the robot's motion trajectory. The continuous correction compensation curve refers to the function curve that changes continuously along the entire weld direction after spline interpolation. It provides a corresponding correction value for any position point on the coarse positioning trajectory, making the correction process complete. The robot's kinematic constraints refer to the limitations determined by the robot's structure and motion laws, such as the rotation range of each joint, the reachable space of the end effector, and the mapping relationship between joint velocity and end effector velocity determined by the kinematic equations. Joint speed limits refer to the maximum permissible rotational speed of the motors driving each joint of the robot. This is a physical upper limit; trajectory planning exceeding this limit will prevent the robot from accurately tracking and executing it. Motion singularities refer to the phenomenon where, in certain special poses, the robot's degrees of freedom degenerate, causing the end effector to require infinitely high speeds for even small movements in a specific direction. Examples include wrist or shoulder singularities common in six-axis robots. Speed ​​over-limit problems occur when the calculated speed of a joint in the robot's motion trajectory generated based on the correction compensation curve exceeds the joint's physical maximum speed limit, preventing the robot from stably operating along the planned trajectory. The coarse positioning trajectory is the preliminary welding path obtained after correcting the overall workpiece pose deviation based on global registration; it serves as the baseline for generating dynamic welding trajectories. The dynamic welding trajectory is the actual motion trajectory generated by algebraically superimposing and fusing the smoothed and motion-calibrated correction compensation curve with the coarse positioning trajectory to control the robot in completing the welding operation.

[0076] In step S7, this application establishes a progressive relationship from discrete correction to continuous execution. Spline interpolation smoothing first transforms the correction values ​​of discrete nodes into a continuous curve, eliminating potential abrupt motion changes between nodes and ensuring geometric smoothness of the trajectory. Subsequently, robot kinematic constraints and joint velocity limits are introduced to verify this continuous curve, essentially setting an executability threshold at the physical level. By adjusting the local shape of the curve or eliminating infeasible motion segments, potential problems such as motion singularities and velocity exceedances are eliminated, ensuring that the corrected trajectory can be stably tracked by the robot. After these two steps, the corrected compensation curve is fused with the coarse positioning trajectory. The resulting dynamic welding trajectory includes compensation for global workpiece deviations and fine-tuning of the weld's local shape, while also satisfying the robot's own motion capability boundaries. This process—geometric smoothing, physical verification, and final fusion—ensures that the final dynamic welding trajectory adapts to the complex changes in the weld's spatial shape and meets the robot's execution capabilities in terms of motion characteristics, providing a reliable trajectory foundation for the smooth operation of the welding process.

[0077] In some embodiments of this application, in step S4, after extracting the instantaneous morphological feature parameters of the weld, the temperature data of the current weld pool is collected simultaneously, and the thermal deformation offset of the subsequent weld is predicted based on the pre-established weld pool temperature-welding thermal deformation mapping model; in step S6, when calculating the position offset, the thermal deformation offset is included in the calculation to pre-correct the subsequent welding trajectory.

[0078] In this embodiment, the molten pool temperature data refers to the real-time temperature value of the molten metal region during welding, which can be obtained through infrared thermal imagers, thermocouples, or spectral analysis. This data reflects the intensity of the current welding heat input and the heating state of the workpiece. The pre-established molten pool temperature-welding thermal deformation mapping model is a mathematical model built based on physical simulation or experimental data. It describes the correspondence between the molten pool temperature and the deformation of the subsequent weld area due to thermal expansion and contraction. For example, for thin aluminum alloy welding, the model can fit a function curve between the molten pool center temperature and the transverse shrinkage of the weld. The thermal deformation offset refers to the displacement that will occur in the unwelded area after welding, predicted by the mapping model based on the current molten pool temperature. This offset may manifest as a transverse drift of the weld centerline or a gap change caused by longitudinal shrinkage. The subsequent weld refers to the weld section in front of the current welding torch position that has not yet been welded but is about to be welded. Due to the heat conduction and cumulative effects of the welding heat input, the current molten pool temperature will affect the deformation trend of the subsequent area. Including the position offset in the calculation means that when determining the position offset of the trajectory correction node in step S6, the thermal deformation offset is included as an input and participates in the calculation along with other features. This ensures that the final correction result considers both the current weld morphology and the prediction of subsequent thermal deformation trends. Pre-correction refers to compensating for the predicted thermal deformation offset in the trajectory before welding reaches the area, allowing the welding torch to adapt to the upcoming shape changes of the workpiece during subsequent welding.

[0079] This application establishes a collaborative mechanism with feedforward control characteristics between steps S4 and S6. Step S4 simultaneously acquires the molten pool temperature while extracting the instantaneous weld morphology, effectively obtaining both the current weld geometry and the thermal state of the current welding process within a single acquisition cycle. Based on the prediction of the molten pool temperature-welding thermal deformation mapping model, this thermal state information is transformed into a prediction of subsequent weld deformation, realizing the extrapolation from the current state to future trends. When step S6 calculates the position offset, this predicted thermal deformation offset is combined with the current weld geometry characteristics, allowing the generated correction quantity to simultaneously perform both immediate compensation and forward-looking adjustment. This mechanism transforms welding thermal deformation, a physical phenomenon typically treated as a hysteresis error in traditional control methods, into a feedforward control quantity that can respond in advance. This helps reduce trajectory deviation caused by heat accumulation in conditions with significant thermal deformation, such as thin-plate welding and long-weld welding, thus expanding the adaptability of the welding process.

[0080] In some embodiments of this application, in step S3, the three-dimensional vision sensor adopts a laser stripe structured light sensor. When the laser stripe structured light sensor acquires the point cloud data of the second field of view, it simultaneously acquires the reflection spectrum distribution data of the weld area. Based on the reflection spectrum distribution data, it identifies the surface state of the current weld area, including the degree of oil stain coverage, oxide film thickness, and surface roughness. When it is identified that the degree of oil stain coverage exceeds a preset threshold or the oxide film thickness exceeds the allowable range of the process, when matching the target welding process parameters in step S5, it simultaneously outputs a weld surface pretreatment command. The pretreatment command includes increasing the laser cleaning power, adjusting the welding torch oscillation amplitude, or extending the arc preheating time. Before welding the current weld point, the robot control cabinet controls the corresponding actuator to complete the surface pretreatment operation.

[0081] In this embodiment, the laser stripe structured light sensor is a three-dimensional vision sensor. It projects one or more laser stripes onto the surface of the object being measured, captures deformed images of the laser stripes with a camera, and calculates the three-dimensional point cloud data of the object's surface using the principle of triangulation. Simultaneously, the laser stripe-irradiated area also carries the spectral reflectance information of the object's surface. Reflectance spectral distribution data refers to the light intensity distribution information of different wavelengths or gray levels collected by the sensor camera after the laser stripes irradiate the weld area. Different surface states have different characteristics in terms of laser reflection, absorption, and scattering. For example, oil stains weaken the reflected light intensity and reduce the uniformity of distribution, while oxide films may alter the polarization characteristics of reflected light or produce characteristic spectral absorption peaks. Surface state is a comprehensive indicator describing the cleanliness and chemical state of the surface to be welded in the weld area. The degree of oil stain coverage reflects the severity of contamination of the workpiece surface with cutting fluid, rust-preventive oil, and other pollutants. Oxide film thickness refers to the thickness of the oxide layer formed naturally or through heat treatment on the surface of materials such as aluminum alloys. Surface roughness represents the degree of undulation in the microscopic geometry of the workpiece surface. All three directly affect the arc stability, molten pool flow behavior, and weld formation quality during welding. The preset threshold is a pre-defined upper limit for the allowable degree of oil contamination or oxide film thickness. When the actual detected value exceeds this upper limit, it indicates that the current surface condition has exceeded the range that conventional welding processes can adapt to, requiring additional intervention. The weld surface pretreatment command is an action command issued by the system when it detects a poor surface condition, used to guide the actuator to perform surface treatment on the weld area before formal welding. Increasing laser cleaning power means increasing the output energy of the laser cleaning equipment to remove surface oil and oxide film through laser ablation or vaporization; adjusting the welding torch oscillation amplitude means increasing the oscillation width of the welding torch before formal welding or in the early stages of welding, using the mechanical brushing effect of the arc to remove surface contaminants; extending the arc preheating time means appropriately extending the residence time of the arc in the initial area after arc ignition, using the arc heat input to decompose oil or break down the oxide film. The corresponding actuator refers to the equipment unit required to complete the above pretreatment operations, which may include the laser cleaning head, the oscillation mechanism of the welding torch itself, and the timing control module of the welding power supply. These mechanisms are all coordinated and controlled by the robot control cabinet according to the pretreatment commands.

[0082] This application establishes a linkage mechanism between surface quality perception and process compensation between steps S3 and S5. While acquiring 3D point cloud data, the laser stripe structured light sensor simultaneously obtains reflectance spectral distribution data using a co-source laser, achieving synchronous acquisition of geometric and surface condition information without requiring additional sensors or extended acquisition time. When oil stains, oxide films, or roughness anomalies are identified through reflectance spectrum analysis, the system no longer passively welds based on the current surface condition. Instead, it actively outputs preprocessing instructions while matching process parameters in step S5. Increasing laser cleaning power removes contaminants at the source, while adjusting the welding torch oscillation amplitude and extending the arc preheating time enhances the adaptability to poor surfaces within the welding process itself. This method of pre-sensing surface condition and intervening before welding extends the starting point of the welding process from the default ideal surface conditions to active adaptation to actual surface conditions. This helps reduce defects such as welding porosity and incomplete fusion caused by fluctuations in workpiece surface cleanliness and improves the welding process's tolerance to residual contamination from previous processes.

[0083] In some embodiments of this application, in step S6, when calculating the attitude compensation amount of the trajectory correction node, spatial attitude parameters of the welding torch relative to the weld cross-section are introduced. The spatial attitude parameters include the welding torch tilt angle, the welding torch pointing angle, and the welding torch rotation angle. Based on the weld gap, misalignment amount, and weld direction curvature in the instantaneous morphological characteristic parameters of the weld, combined with the spatial attitude parameters of the welding torch relative to the weld cross-section, a welding torch attitude adaptation model is constructed. The welding torch attitude adaptation model takes the weld penetration uniformity and weld surface forming quality as optimization objectives, and outputs an attitude compensation amount that keeps the welding torch and the weld groove wall within a preset distance range and the welding wire extension length constant, so that the welding torch can maintain stable droplet transition in both curved weld sections and variable gap weld sections.

[0084] In this embodiment, the welding torch tilt angle refers to the angle between the welding torch axis and the normal to the weld surface, also known as the working angle. It determines the heating bias of the arc on both sides of the bevel. For example, in a fillet weld, an excessively large tilt angle can lead to insufficient penetration on one side and complete melting on the other. The welding torch pointing angle refers to the angle between the projection of the welding torch axis onto the weld's forward direction and the weld tangent, also known as the travel angle or forward tilt angle. It affects the flow direction of the molten pool and the weld formation coefficient. A larger forward tilt angle results in a wider and shallower weld, while a larger backward tilt angle results in a narrower and deeper weld. The welding torch spin angle refers to the angle at which the welding torch rotates around its own axis. For welding torches with curved nozzles or specific wire feeding mechanisms, the spin angle affects the wire extension direction and wire feeding stability, and needs to be controlled in complex spatial curve welds. The weld cross-section refers to a plane perpendicular to the weld centerline. On this cross-section, the geometry of the weld can be observed, such as the bevel shape, gap width, misalignment height, and the final molten pool filling profile. Weld penetration uniformity refers to the stability of the weld depth and width along the weld length. Uneven penetration can easily lead to localized incomplete penetration or burn-through defects. Weld surface formation quality refers to the appearance of the weld surface after welding, including the height of the reinforcement, the uniformity of the scale pattern, the consistency of the weld width, and the presence or absence of defects such as undercut and weld beads. The welding torch posture adaptation model is a computational model built based on geometric relationships and welding process principles. Its inputs are the weld morphology characteristics and the current posture of the welding torch, and its output is the optimized posture compensation amount. The preset distance range refers to the safe interval that should be maintained between the tip of the welding torch nozzle or contact tip and the weld groove wall. If the interval is too small, it can easily lead to a short circuit between the contact tip and the workpiece. If the interval is too large, it may cause poor shielding gas coverage or a decrease in arc stability. The wire extension length refers to the length of the welding wire from the tip of the contact tip to the arc ignition point. Maintaining a constant wire extension length helps to maintain stable welding current and consistent droplet transfer. Stable droplet transfer refers to the transition of molten metal droplets from the tip of the welding wire into the molten pool with uniform size and frequency, without large droplet splashing or short-circuit transfer abnormalities. This is an important prerequisite for ensuring the quality of weld formation.

[0085] In step S6, this application constructs a complete control link from three-dimensional attitude parameters to welding stability. By introducing three spatial attitude parameters—welding torch tilt angle, pointing angle, and rotation angle—the control dimension of the welding torch is extended from traditional planar positioning to attitude freedom within the six degrees of freedom of space, providing sufficient descriptive capability for precise control of complex welds. The welding torch attitude adaptation model uses weld penetration uniformity and surface forming quality as optimization objectives, meaning the model pursues not only the ability of the welding torch to pass through the weld space but also the optimization of welding quality. The model outputs an attitude compensation amount that maintains a preset distance between the welding torch and the groove wall while keeping the wire extension length constant. These two constraints ensure the rationality of attitude adjustment from the dimensions of safe distance and process stability, respectively. Finally, in curved weld sections and variable gap weld sections, this attitude optimization enables stable droplet transfer, directly addressing the technical challenge of poor attitude adaptation when welding complex curved surfaces and variable curvature welds in key automotive components. By finely controlling the pointing of the welding torch in three-dimensional space, the arc heat input is always applied to the most needed position on the groove.

[0086] In some embodiments of this application, after the welding of a single key automotive component is completed, complete welding quality data of the workpiece is collected. The welding quality data includes weld formation images, welding process electrical parameter time-series records, and post-weld three-dimensional scanning point clouds. The welding quality data is associated and stored with the weld instantaneous morphological feature parameter sequence, target welding process parameter sequence, and dynamic welding trajectory recorded during the welding process to construct a welding quality traceability sample library. Based on the welding quality traceability sample library, the offline retraining of the welding process decision model is periodically triggered, and welding quality scores are introduced to weight the samples during the retraining process, so that the model is given higher decision sensitivity to the weld morphological feature regions corresponding to high-frequency defects.

[0087] In this embodiment, welding quality data refers to a multi-dimensional information set collected after the workpiece welding is completed, used to evaluate the quality of the welding result. This data constitutes the basis for judging whether the welding process is successful. Weld formation images are post-weld photographs of the weld appearance taken with an industrial camera, which can intuitively reflect the weld reinforcement height, width, uniformity of the fish-scale pattern, and visible defects such as undercut, weld beads, and surface porosity. Welding process electrical parameter time-series records refer to continuous waveform data showing the changes in parameters such as current, voltage, and wire feed speed output by the welding power source over time throughout the welding process. This data reflects the dynamic stability of the welding process. Post-weld 3D scanning point cloud data is point cloud data obtained by scanning the welded workpiece as a whole or partially using a 3D scanner. It is used to compare with the theoretical digital model of the workpiece to evaluate the welding deformation and the deviation between the actual weld formation size and the target size. Association storage refers to establishing a corresponding relationship between various types of data generated by the same workpiece before, during, and after welding, and organizing and storing them according to the workpiece's unique identifier, so that subsequent queries can completely trace back the workpiece's entire lifecycle data. The instantaneous morphological characteristic parameter sequence of the weld refers to the continuous record of parameters such as the weld centerline, gap, misalignment, and curvature extracted along the entire weld in step S4 and arranged sequentially over time or space. The target welding process parameter sequence refers to the time-series record of the set values ​​of welding current, voltage, wire feed speed, and travel speed matched for each weld point or weld segment in step S5. The dynamic welding trajectory refers to the spatial motion path record generated in step S7, which is the record of the robot's actual welding operation. The welding quality traceability sample library is an accumulated data storage structure that pairs and stores the quality data of each workpiece with the corresponding process control data, forming a dataset that can be used for subsequent analysis and training. Periodic triggered offline retraining refers to retraining or fine-tuning the welding process decision model using new data from the sample library during production line downtime or low-load periods, according to a preset time period or cumulative sample quantity, so that the model can absorb the latest production experience. The welding quality score is an indicator for quantitatively evaluating the welding quality of each workpiece. It can be derived from weld formation image analysis, 3D scanning deviation calculation, and non-destructive testing results. The score reflects the quality of the workpiece's welding. Sample weighting refers to assigning different weight coefficients to different samples during model training. This allows samples with higher or lower quality scores to have a greater impact on model parameter updates, thus guiding the model to pay more attention to weld morphology feature areas prone to quality fluctuations. High-frequency defect weld morphology feature areas refer to the combination of morphological features corresponding to weld sections where welding defects repeatedly occur in actual production. For example, if porosity frequently appears at the corner of a battery tray in a certain vehicle model, the combination of the gap range and curvature value in this area constitutes a feature area where decision sensitivity needs to be improved.

[0088] This application constructs a complete closed-loop system from data acquisition to model self-optimization. Weld formation images, electrical parameter time-series records, and post-weld 3D scanning point clouds, acquired after welding, describe the true state of welding quality from three dimensions: appearance, process stability, and dimensional accuracy. These quality data are associated and stored with weld morphology features, process parameter settings, and dynamic trajectories during the welding process, making each workpiece a labeled sample carrying complete process and result information. Periodically triggered offline retraining uses the accumulated sample library to iteratively update the welding process decision model, while a sample weighting mechanism gives higher attention to samples that produce defects during training. This design ensures that the welding process decision model is not fixed after a one-time training but continuously learns from successful and failed cases in actual production as production batches progress, developing a more accurate process parameter matching capability for weld morphology areas prone to defects, thereby achieving continuous optimization and stable control of welding quality during long-term mass production.

[0089] A second aspect of this application provides a robot welding trajectory control system based on three-dimensional vision recognition. This system implements the aforementioned robot welding trajectory control method based on three-dimensional vision recognition, and includes: a robot body, a three-dimensional vision sensor, a global registration module, a feature extraction module, a process matching module, a trajectory correction module, and a robot control cabinet. The three-dimensional vision sensor and welding torch are pre-mounted at the end of the robot body. The global registration module is used to complete workpiece pose calculation and coarse positioning trajectory generation. The feature extraction module is used to extract instantaneous weld seam morphological feature parameters. The process matching module is used to match the target welding posture and process parameters. The trajectory correction module is used to generate a dynamic welding trajectory. The robot control cabinet is used to control the robot to perform welding operations.

[0090] In the above embodiments, the system's constituent modules form a well-defined overall architecture with clear division of labor and data flow. The 3D vision sensor, as the perception front end, provides raw data input to the system; the global registration module completes pre-welding global positioning and generates a reference trajectory; during welding, the feature extraction module obtains weld morphology information from the high-resolution point cloud, the process matching module determines the posture and parameters based on this information, and the trajectory correction module converts the decision results into trajectory compensation quantities and fuses them to generate the final trajectory; the robot control cabinet, as the execution end, uniformly schedules motion and process output. This modular division of labor allows global positioning, local perception, process decision-making, trajectory correction, and motion control to be relatively independent yet efficiently collaborative, which is beneficial for achieving stable and controllable automated operation processes in mass production welding scenarios.

[0091] A third aspect of this application provides an electronic device, including: one or more processors, one or more input devices, one or more output devices, and one or more memories. The processors, input devices, output devices, and memories communicate with each other via a communication bus. The memories store computer programs, including program instructions. The processors execute the program instructions stored in the memories. The processors are configured to invoke the program instructions to execute the aforementioned robot welding trajectory control method based on three-dimensional vision recognition.

[0092] It should be understood that, in the embodiments of this application, the processor may be a Central Processing Unit (CPU), or it may be 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 may be a microprocessor or any conventional processor.

[0093] Input devices may include touchpads, fingerprint sensors (for collecting the user's fingerprint information and fingerprint orientation information), microphones, etc., while output devices may include displays (LCDs, etc.), speakers, etc.

[0094] The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store information about the device type.

[0095] In specific implementations, the processor, input device, and output device described in the embodiments of this application can execute the implementation methods described in any embodiment of the robot welding trajectory control method based on three-dimensional vision recognition provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0096] In another embodiment of this application, an electronic device is provided. The electronic device stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the robot welding trajectory control method based on three-dimensional vision recognition described in the above embodiments. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in an electronic device, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0097] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0098] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0099] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0100] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0101] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0102] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0103] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered 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 robot welding trajectory control method based on three-dimensional vision recognition, characterized in that, The method is based on a three-dimensional vision sensor mounted at the front of the welding torch at the end of the robot and includes the following steps: S1. In response to the welding start command, control the three-dimensional vision sensor to collect the first field of view point cloud data of the key automotive component to be welded, the first field of view point cloud data includes the workpiece's preset reference feature area. S2. Globally register the reference feature region with the preset three-dimensional digital model of the workpiece, calculate the actual pose deviation of the workpiece in the robot base coordinate system, correct the preset global welding path, and obtain the coarse positioning trajectory. In step S2, the point cloud subset corresponding to the reference feature region is first segmented from the point cloud data of the first field of view. After removing the outliers in the point cloud subset, the registration constraint range is set based on the prior information of the clamping and positioning of the automotive parts. The global registration is completed by using the constrained iterative nearest point algorithm, and the rotation matrix and translation vector corresponding to the pose deviation are calculated. S3. During the welding process of the robot along the coarse positioning trajectory, the three-dimensional vision sensor is controlled to collect the second field of view point cloud data of the weld area in front of the welding gun in real time. The resolution of the second field of view point cloud data is higher than that of the first field of view point cloud data. S4. Based on the second field of view point cloud data, extract the instantaneous morphological feature parameters of the weld, including the weld centerline, weld gap, misalignment amount and weld direction curvature. In step S4, after extracting the instantaneous morphological feature parameters of the weld, the temperature data of the current weld pool is collected simultaneously. Based on the pre-established weld pool temperature-welding thermal deformation mapping model, the thermal deformation offset of the subsequent weld is predicted. S5. Based on the instantaneous morphological feature parameters, the target welding posture and target welding process parameters corresponding to the current weld point are matched and obtained; S6. Based on the target welding posture and weld curvature, calculate the position offset and posture compensation of multiple trajectory correction nodes on the coarse positioning trajectory. In step S6, when calculating the position offset, the thermal deformation offset is included in the calculation to pre-correct the subsequent welding trajectory; when calculating the attitude compensation of the trajectory correction node, the spatial attitude parameters of the welding torch relative to the weld cross-section are introduced. The spatial attitude parameters include the welding torch tilt angle, the welding torch pointing angle, and the welding torch rotation angle; based on the weld gap, misalignment, and weld curvature in the instantaneous morphological characteristic parameters of the weld, combined with the spatial attitude parameters of the welding torch relative to the weld cross-section, a welding torch attitude adaptation model is constructed; the welding torch attitude adaptation model takes the weld penetration uniformity and weld surface forming quality as optimization objectives, and outputs the attitude compensation amount that keeps the welding torch and the weld groove wall within a preset distance range and the welding wire extension length constant, so that the welding torch can maintain stable droplet transition in curved weld sections and variable gap weld sections; S7. The position offset, attitude compensation, and coarse positioning trajectory are fused to generate a dynamic welding trajectory, and the robot is controlled to perform welding synchronously according to the dynamic welding trajectory and the target process parameters.

2. The method according to claim 1, characterized in that, In step S1, the three-dimensional vision sensor is controlled to move sequentially to multiple preset global scanning poses to collect point cloud data of the first field of view. Each set of scanning poses is preset based on the structural features of the key automotive components to be welded.

3. The method according to claim 1, characterized in that, In step S3, the arc intensity data of the welding pool is collected in real time. The exposure parameters of the three-dimensional vision sensor and the laser projection intensity are dynamically adjusted according to the arc intensity. At the same time, the sampling frequency of the sensor is adjusted synchronously according to the robot's welding walking speed to ensure that the density of the point cloud data in the second field of view is constant.

4. The method according to claim 3, characterized in that, In step S3, the acquisition distance in front of the welding torch is 8mm-25mm, and the acquisition distance is dynamically adjusted according to the welding walking speed and arc intensity. When the welding walking speed or arc light intensity increases, the acquisition distance should be increased accordingly.

5. The method according to claim 1, characterized in that, In step S5, the weld gap and misalignment are combined with the material, plate thickness and bevel type of the key automotive component to be welded and imported into the pre-trained welding process decision model, which outputs target process parameters including welding current, voltage, wire feed speed and travel speed. The model is trained based on the historical welding data and quality inspection results of the corresponding automotive component and is iteratively optimized after each workpiece welding is completed.

6. The method according to claim 1, characterized in that, In step S6, trajectory correction nodes are dynamically selected based on the weld curvature. When the weld curvature is less than a preset threshold, nodes are selected at equal intervals along the coarse positioning trajectory. When the weld curvature is greater than or equal to the preset threshold, the node spacing is reduced as the curvature increases, ensuring that the correction node density of the curved weld segment is higher than that of the straight weld segment.

7. The method according to claim 1, characterized in that, In step S7, the position offset and attitude compensation of each trajectory correction node are first smoothed by spline interpolation to generate a continuous correction compensation curve. The correction compensation curve is then verified based on the robot's kinematic constraints and joint speed limits. After eliminating motion singularities and speed over-limit problems, it is fused with the coarse positioning trajectory to generate a dynamic welding trajectory.

8. A robot welding trajectory control system based on three-dimensional vision recognition, characterized in that, The system, used to perform the method as described in any one of claims 1 to 7, comprises a robot body, a three-dimensional vision sensor, a global registration module, a feature extraction module, a process matching module, a trajectory correction module, and a robot control cabinet; the three-dimensional vision sensor and the welding torch are pre-mounted at the end of the robot body; the global registration module is used to complete workpiece pose calculation and coarse positioning trajectory generation; the feature extraction module is used to extract instantaneous morphological feature parameters of the weld; the process matching module is used to match the target welding posture and process parameters; the trajectory correction module is used to generate a dynamic welding trajectory; and the robot control cabinet is used to control the robot to perform welding operations.

Images