Robot intelligent welding unit and method based on 3D visual guidance

Abstract

The invention discloses a robot intelligent welding unit and method based on 3D visual guidance, and the unit comprises an industrial robot arm, a welding gun, welding equipment, at least one 3D visual sensor and a visual processing controller. A three-dimensional reconstruction module, a self-adaptive path planning module, a real-time feedback control module and a safety monitoring module are integrated in the visual processing controller; a workpiece is scanned through the 3D vision sensor, a three-dimensional model is reconstructed, a welding seam is automatically recognized through the self-adaptive path planning module, an optimized welding path with embedded technological parameters is generated, and in the welding process, the real-time feedback control module dynamically corrects the robot track and adjusts the welding parameters based on a prediction-correction strategy, so that the welding quality is improved. And meanwhile, the safety monitoring module utilizes real-time point cloud to construct a virtual fence for collision protection, so that the whole-process intelligence from pre-welding sensing and autonomous planning to real-time correction and safety monitoring in welding is realized, and the welding precision, the process adaptability and the operation safety are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and automated welding technology, and in particular to a robotic intelligent welding unit and method based on 3D vision guidance. Background Technology

[0002] Current industrial robot welding operations typically employ teach-and-playback or offline programming modes. Operators manually guide the robot along a predetermined path, recording key points, or plan the path in software using a 3D model of the workpiece, and then the robot repeatedly executes this fixed trajectory. This programming method relies on the high consistency of workpiece dimensions, the precise repeatability of clamping positions, and the controllability of thermal deformation during welding. In actual mass production, operators pre-set process parameters such as the robot's movement speed, welding torch posture, welding current, and voltage, and maintain these parameters throughout the entire welding process.

[0003] First, in continuous welding operations, the workpiece undergoes nonlinear thermal deformation due to heat, causing the actual weld position to deviate from the originally planned path. Furthermore, due to factors such as incoming material processing errors, assembly tolerances in previous processes, and fixture wear, the weld positions and gaps of workpieces in the same batch are difficult to guarantee are completely consistent after clamping. When these deviations accumulate to a certain extent, the robotic welding torch, operating along a fixed trajectory, will deviate from the actual weld center, leading to welding defects such as weld misalignment, incomplete fusion, and burn-through, severely impacting welding quality.

[0004] Secondly, traditional welding systems lack the ability to sense the welding process status and cannot dynamically adjust process parameters according to changes in the molten pool and weld gap. For example, when the gap is too large in some areas, the original fixed wire feed speed and welding power cannot provide enough deposited metal, which can easily lead to insufficient filling. Conversely, when the gap is too small, excessive deposited metal can cause the weld reinforcement to exceed the standard. At the same time, complex structural parts often contain multiple welds with complex spatial orientations. Manually teaching the path is not only time-consuming and labor-intensive, but also makes it difficult to ensure the smoothness of complex curve trajectories and process consistency. Especially when there are significant individual differences in the workpieces, the pre-taught path often needs to be repeatedly adjusted, resulting in low production efficiency and high rework rates.

[0005] Therefore, in response to the problems mentioned above, this invention proposes a robotic intelligent welding unit and method based on 3D vision guidance. Summary of the Invention

[0006] To overcome the problems of traditional robotic welding systems that rely on fixed programming paths, making it difficult to cope with workpiece deformation, positional deviations, or complex welding environments, resulting in low welding accuracy, uneven welds, high rework rates, and even safety hazards, this invention proposes a 3D vision-guided intelligent robotic welding unit and method. By integrating 3D vision sensors, it achieves real-time 3D reconstruction of the workpiece and weld feature recognition. An adaptive path planning algorithm is used to generate an optimized welding path, and during the welding process, visual feedback is used to correct the robot's motion trajectory and dynamically adjust welding process parameters in real time. At the same time, a vision-based dynamic safety fence is constructed, thereby achieving intelligent welding operations with high precision, high adaptability, and high safety.

[0007] The technical solution of this invention is: a robotic intelligent welding unit based on 3D vision guidance, comprising: An industrial robot arm with a welding torch mounted at its end; Welding equipment, connected to a welding torch, used to provide welding power and welding wire; At least one 3D vision sensor is installed at a fixed position in the intelligent welding unit or at the end of an industrial robot arm to scan the workpiece and its surrounding environment before and during welding to acquire three-dimensional point cloud data. A vision processing controller, communicating with a 3D vision sensor and an industrial robot arm, includes: The 3D reconstruction module is used to receive and process 3D point cloud data to reconstruct a full-size 3D model of the welded workpiece. The adaptive path planning module has a built-in weld feature extraction algorithm, which is used to identify and extract the weld area from the full-size 3D model. Based on the geometric features of the weld area and the preset welding process database, it plans an optimized welding path in real time. The optimized welding path not only includes the robot's motion trajectory, but also the welding process parameter instructions that are dynamically associated with the motion trajectory. The real-time feedback control module is used to simultaneously receive real-time point cloud data from a 3D vision sensor during the process of the industrial robot arm guiding the welding torch to perform welding along the optimized welding path. By comparing the real-time point cloud data with the full-size 3D model or the optimized welding path, it calculates the deviation of the weld position and gap caused by workpiece thermal deformation or clamping error, and generates compensation instructions for real-time correction of the motion trajectory of the industrial robot arm and / or the process parameters of the welding equipment. The safety monitoring module is used to construct a dynamic virtual safety fence containing the industrial robot arm, welding torch, welding workpiece and surrounding equipment based on real-time environmental point cloud data acquired by 3D vision sensors throughout the welding process. When it is predicted that any part will enter the preset dangerous approach distance, it immediately sends a deceleration or emergency stop signal to the industrial robot arm.

[0008] Preferably, the 3D vision sensor is a laser profile scanner or a structured light 3D camera, which is installed at the end of an industrial robot arm and moves with the robot arm. It can scan the welding workpiece from multiple angles to obtain high-precision point cloud data, thereby eliminating scanning blind spots under a single viewpoint.

[0009] Preferably, the adaptive path planning module includes: The bevel feature recognition module is used to analyze point cloud data of the weld area and identify the weld type, bevel angle, blunt edge height and root gap. The process parameter matching module, based on the output of the bevel feature recognition module, retrieves and matches the optimal welding current, voltage, welding speed, and oscillation parameters from the preset welding process database, and embeds them into the instruction for optimizing the welding path.

[0010] Preferably, the real-time feedback control module is specifically used to: extract the weld point cloud in the area near the current molten pool in real time during the welding process, and compare it with the theoretical weld position in the optimized welding path; if the deviation in the lateral or height direction exceeds the preset threshold, send a position correction amount to the industrial robot arm; if the weld gap change is detected, send an instruction to the welding equipment to adjust the wire feeding speed or welding power.

[0011] Preferably, the safety monitoring module is also used to simulate the robot's movement trajectory in the no-load operation mode before welding begins, and to predict collision risks based on a dynamic virtual safety fence. If a potential collision risk is detected, an alarm is issued and the startup program is locked.

[0012] This invention proposes a robotic intelligent welding method based on 3D vision guidance, comprising the following steps: S1 controls the 3D vision sensor to scan the workpiece placed on the welding station, obtain its complete three-dimensional point cloud data, and reconstruct a high-precision three-dimensional model of the workpiece through the vision processing controller. S2, the vision processing controller extracts features from the 3D model of the workpiece, automatically identifies one or more weld seams to be welded, and autonomously generates an initial optimized welding path based on the geometric shape of the weld seam (including the 3D spatial orientation, bevel shape and gap size) and the preset welding process knowledge base. This path data not only includes the robot's motion trajectory points, but also embeds welding parameter instructions that match the path segment. Among the identified weld seams, based on multi-objective constraints such as shortest path, obstacle avoidance and robot reachability, a globally optimal welding sequence and robot transfer path are automatically planned to minimize the robot's idle travel time and reduce the overall heat input. S3, the industrial robot arm starts to move according to the optimized welding path, guiding the welding gun to weld. During this process, the 3D vision sensor collects point cloud data of the welding area and the front of the molten pool in real time. The vision processing controller compares the real-time point cloud data with the 3D model, calculates the deviation between the actual position of the weld and the planned path caused by factors such as thermal deformation, and immediately generates a trajectory correction command to send to the robot arm. At the same time, it adjusts the output power or wire feeding speed of the welding equipment in real time according to the change of the actual weld gap. Among them, the trajectory correction command is generated based on the predictive-correction control strategy. The vision processing controller first predicts the deviation amount at the next moment and performs feedforward compensation based on the deviation change trend in the recent period, and then performs feedback correction based on the actual deviation at the current moment, so as to achieve dynamic high-precision tracking of complex trajectories. The real-time adjustment of the output power or wire feeding speed of the welding equipment is specifically as follows: when an increase in weld gap is detected, the vision processing controller increases the amount of deposited metal by increasing the welding current or wire feeding speed; when a decrease in weld gap is detected, the welding current or wire feeding speed is reduced accordingly to prevent burn-through or excessive weld height, thereby achieving adaptive welding under variable gap conditions. S4. During the welding process, the safety monitoring module uses the dynamic environmental data acquired in real time by the 3D vision sensor to construct and update a dynamic three-dimensional scene that includes the moving robot, welding torch, workpiece and surrounding equipment. Through the spatial collision detection algorithm, it judges in real time whether there is a collision risk. Once the risk value exceeds the safety threshold, it immediately triggers an emergency stop to ensure the safety of human-machine collaborative operation. S5. During the welding process or after the welding of a single weld seam is completed, the 3D vision sensor scans the weld area again to obtain the point cloud data of the weld formation. The vision processing controller compares the point cloud data of the weld formation with the standard weld seam model to automatically evaluate the weld seam's reinforcement height, weld width, and surface forming quality, and generates a quality inspection report for real-time process adjustment or subsequent traceability.

[0013] The beneficial effects of this invention are: 1. This invention performs high-precision three-dimensional reconstruction of the welded workpiece through a three-dimensional reconstruction module, realizing accurate digital perception of the individual differences of each workpiece. It breaks away from the traditional welding system's over-reliance on idealized digital models and fixed clamping positions, providing a real and reliable geometric reference for subsequent adaptive welding.

[0014] 2. This invention automatically identifies weld types and extracts groove features through an adaptive path planning module, and intelligently matches process parameters by combining the built-in welding process database. This changes the traditional manual teaching and offline programming mode, enabling the welding system to make autonomous decisions for different workpieces.

[0015] 3. This invention uses a real-time feedback control module to dynamically track the changes in weld position and gap during the welding process, and achieves advance compensation based on a predictive-correction control strategy. This overcomes the tracking lag problem caused by system delay in traditional feedback control, and realizes high-precision real-time correction of dynamic offset caused by thermal deformation.

[0016] 4. This invention achieves adaptive molten metal quantity control under variable gap conditions by dynamically adjusting the welding current and wire feeding speed according to the real-time gap change through a real-time feedback control module. This solves the problem of weld penetration or insufficient filling that easily occurs when the gap changes abruptly in traditional fixed parameter welding, and significantly reduces the welding defect rate.

[0017] 5. This invention utilizes real-time point cloud data to construct a dynamic virtual safety fence through a safety monitoring module, enabling real-time collision prediction and active protection of moving robots, welding torches, and the surrounding environment. This overcomes the limitations of traditional safety solutions that rely on static protection through physical fences and area scanning, providing reliable protection for high-speed human-machine collaborative operations. Attached Figure Description

[0018] Figure 1 The diagram shown is a schematic representation of a system framework of the present invention. Figure 2 The diagram shown illustrates the working principle of the three-dimensional reconstruction module of this invention. Figure 3 The diagram shown illustrates the working principle of the adaptive path planning module of this invention. Figure 4 The diagram shown illustrates the working principle of the real-time feedback control module of this invention. Figure 5 The diagram shown is a simulation flow diagram of the safety monitoring module of the present invention; Figure 6 The diagram shown illustrates the working principle of the safety monitoring module of this invention. Figure 7 The diagram shown is a schematic representation of the overall process of this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides an embodiment: a robotic intelligent welding unit based on 3D vision guidance, comprising: The system comprises an industrial robot arm, a welding torch mounted at the end of the robot arm, welding equipment connected to the welding torch, at least one 3D vision sensor, and a vision processing controller. The industrial robot arm is a multi-degree-of-freedom articulated robot capable of flexibly moving the welding torch in three-dimensional space. The welding equipment includes a welding power source, wire feeding mechanism, and shielding gas source, providing the welding torch with the necessary electrical energy, welding wire, and shielding gas. Its operating parameters can be dynamically adjusted via external commands. The 3D vision sensor acquires three-dimensional point cloud data of the workpiece and its surrounding environment. Depending on the actual working conditions, it can be fixedly mounted on a support above or to the side of the welding unit to achieve overall scanning of the workpiece; alternatively, it can be mounted at the end of the industrial robot arm, moving with the robot arm to perform detailed scanning of complex workpieces from multiple angles, eliminating blind spots in single-view scanning and acquiring higher-precision local point cloud data. The vision processing controller integrates an image processing unit and control algorithm software, interacting with the industrial robot arm controller, welding equipment, and 3D vision sensor in real time via a communication bus.

[0021] In this embodiment, the vision processing controller will be described in detail: The vision processing controller is logically divided into multiple functional modules, including a 3D reconstruction module, an adaptive path planning module, a real-time feedback control module, and a safety monitoring module.

[0022] Please see Figure 2 In this embodiment, the three-dimensional reconstruction module will be described in detail: After the 3D vision sensor completes the scanning of the welded workpiece, it first transmits the raw point cloud data to the 3D reconstruction module. (The raw point cloud data usually contains a lot of redundant information and noise points caused by environmental interference. Specifically, in this invention, the redundant and interference information usually includes outlier noise points caused by arc light interference, false point clouds formed by spatter weld slag adhering to the workpiece surface, incomplete data areas caused by smoke and dust obstruction, and abnormal specular reflection points caused by highly reflective areas on the workpiece surface. Especially when the workpiece has deep holes, grooves, or narrow bevel structures, the sensor's incident angle is limited, which often leads to data loss or uneven density in these key areas. Directly using these data will affect the accuracy and efficiency of subsequent processing, so preprocessing is required.)

[0023] The first step in preprocessing is point cloud filtering. In the welding environment, strong arc light, spattering weld slag, and airborne fumes can interfere with the acquisition by the vision sensor, resulting in isolated outlier noise points in the point cloud. The 3D reconstruction module uses a statistical filtering algorithm to calculate the average distance between each point and its neighbors, identifying and removing points whose distance distribution significantly deviates from the overall mean as noise. Furthermore, because the workpiece surface may have reflective areas or deep holes, the acquired point cloud often contains localized data gaps, forming holes. The 3D reconstruction module uses interpolation algorithms to initially repair these small areas of missing data, ensuring the integrity of the point cloud data.

[0024] The principle of the statistical filtering algorithm includes: for each point in the point cloud, calculating the average distance between it and its k nearest neighbors, the distribution of which approximately follows a Gaussian distribution. Let a point in the point cloud... The set of k-neighbor points is N( If the average neighborhood distance is ), then The formula for calculation is: ; in, Point With neighboring points The Euclidean distance between them. The average distance over all points in the entire point cloud. Perform statistical analysis and calculate its mean μ and standard deviation σ: ; ; Where n is the total number of points in the point cloud. Based on the statistical characteristics of Gaussian distribution, points whose average distance exceeds the range of μ±λ·σ are identified as noise points, where λ is the set standard deviation multiple threshold.

[0025] To address the high interference characteristics of welding environments, the threshold value of λ is typically set between 1.0 and 3.0, and the value of k is typically set between 20 and 50. In a preferred embodiment of the present invention, λ is set to 1.5, and k is set to 30. This preferred value is based on the following: outlier noise points generated by arc spatter in the welding environment are usually sparsely distributed, with their average distance significantly deviating from the main distribution. A λ value of 1.5 effectively filters out outlier noise while avoiding misclassification of points with normal geometric features, such as bevel edges and corners, as noise, ensuring the complete preservation of detailed information in the weld area. A k value of 30 balances the reliability of local statistics with computational efficiency; a k value that is too small is easily affected by local fluctuations, leading to misclassification, while a k value that is too large increases computational overhead and is insensitive to minute features. Through the above parameter settings, the statistical filtering algorithm can accurately distinguish between real geometric features and false noise points under multiple interference environments such as welding fumes, spatter, and reflections, providing a high-quality point cloud data foundation for subsequent multi-view registration and surface reconstruction.

[0026] After preprocessing, the 3D reconstruction module will downsample the point cloud (a high-precision 3D vision sensor may generate millions or even tens of millions of points in a single scan, such a large amount of data will seriously affect the subsequent processing speed). Downsampling reduces the density of points in flat areas and retains enough points in edges and corners with drastic curvature changes, while maintaining the overall geometric features of the workpiece, through uniform sampling or curvature adaptive sampling. This significantly reduces the amount of data and improves processing efficiency while ensuring model accuracy.

[0027] Specifically, this embodiment uses a longitudinal beam structure of an automobile chassis as the test object, and designs a comparative experiment for verification. Three longitudinal beam workpieces with the same geometric features are selected as test samples. The workpieces are approximately 1200mm long and contain a total of eight welds, including fillet welds, butt welds, and T-welds, with moderate structural complexity. A laser contour scanner is used to perform a full-view scan of the workpiece to obtain the original point cloud data. The experiment is divided into three groups: In the control group, no downsampling was performed on the point cloud; the original point cloud was used directly for subsequent registration, reconstruction, and path planning.

[0028] Experimental group 1 used a uniform downsampling method, setting the target point spacing to 1.0 mm to uniformly reduce the point cloud density.

[0029] Experimental group 2 uses a curvature adaptive downsampling method, setting the target point spacing to 2.0 mm in flat areas and 0.5 mm in weld edges and corners with drastic curvature changes, thus preserving the point cloud density of the feature areas.

[0030] The experimental results are as follows:

[0031] From the table above, we can see that: Experimental groups 1 and 2 achieved data compression rates of 85.1% and 86.0%, respectively, reducing the original 8.5 million point cloud to approximately 1.2 million points. This significant compression greatly reduced the data size for subsequent processing, creating conditions for welding path planning with high real-time requirements.

[0032] The 3D reconstruction time for the control group was 47.3 seconds, while that for experimental groups 1 and 2 was reduced to 12.8 seconds and 13.1 seconds, respectively, representing an improvement of approximately 72%. Regarding path planning time, the control group took 23.6 seconds, while experimental groups 1 and 2 were reduced to 6.2 seconds and 6.5 seconds, respectively, representing an improvement of approximately 72.5%. Both experiments validated the significant improvement in processing efficiency achieved by downsampling.

[0033] While the uniform downsampling method in Experiment 1 significantly improved processing efficiency, the model accuracy deviation reached 0.18 mm, and the weld recognition accuracy dropped to 92.5%, with some weld edges becoming blurred. This is because the weld area is a critical part with drastic curvature changes. Uniform downsampling uses the same downsampling intensity in both flat and feature areas, resulting in insufficient point cloud density and loss of detail information in the feature areas.

[0034] Experimental group 2 maintained a high compression ratio of 86.0%, while the model accuracy deviation was only 0.06 mm, and the weld recognition accuracy reached 97.8%, basically on par with the control group. The weld edges were clear and complete, and the measurement errors of key geometric parameters such as bevel angle and root gap were controlled within 0.05 mm, fully meeting the accuracy requirements of welding path planning. This is because the curvature adaptive sampling strategy significantly reduces point cloud density in flat areas to improve efficiency, while retaining enough points in areas with abrupt curvature changes such as weld edges and corners, thus ensuring that key feature information is not lost.

[0035] For welded workpieces with complex structures, such as automotive chassis longitudinal beams, large box beams, multi-faceted pipe joints, and workpieces with deep cavities or complex stiffeners, a single scan often cannot obtain their complete three-dimensional information. It is necessary to scan from multiple angles. Multi-view point cloud registration technology precisely aligns the point cloud data collected from these different perspectives into the same coordinate system and stitches them together to form a complete workpiece point cloud.

[0036] When a 3D vision sensor is installed at the end of an industrial robot arm and scans as it moves, the 3D reconstruction module first uses the pose information of the robot's end effector at each scanning moment recorded by the robot's kinematic model to roughly transform the point cloud from different perspectives to the robot's base coordinate system, completing the initial registration. This step eliminates most of the coordinate differences caused by the robot's movement.

[0037] After initial registration, the module employs an iterative nearest-point algorithm for precise registration. This algorithm continuously searches for the closest point pairs between two point clouds and calculates the rotation matrix and translation vector that minimizes the distance between these point pairs. Then, it iterative optimization gradually reduces the registration error. In this embodiment, the 3D reconstruction module performs various optimizations on the iterative nearest-point algorithm. First, it pre-identifies points with significant geometric features (corner points, edge points) in the two point clouds through feature point extraction. Matching is performed only on these feature points, significantly improving the registration speed. Then, a dynamic threshold is introduced during the iteration process, gradually tightening the convergence condition as the registration accuracy improves, ensuring that the final registration result achieves sub-millimeter level accuracy.

[0038] Specifically, during the iteration of the nearest point algorithm, the module sets two thresholds: a distance threshold and a translation / rotation threshold. The distance threshold is used to determine whether a point pair is a valid matching point pair, and its initial value is set to 2.0 mm. This means that in the early stages of registration, as long as the distance between two points is less than 2.0 mm, they are considered valid matching point pairs and participate in the calculation of the transformation matrix. This relatively wide initial threshold allows the algorithm to establish a sufficient number of matching point pairs even if there is a large deviation in the initial registration, ensuring that the iteration can start smoothly.

[0039] As the number of iterations increases, the registration error gradually decreases, and the module automatically tightens the two thresholds. Specifically, after each iteration, the module uses the root mean square error (RMSE) of all current matching point pairs as a reference and sets the distance threshold for the next iteration to 1.5 times the current RMSE. When the RMSE decreases from the initial 1.5mm to 0.3mm, the distance threshold is tightened from approximately 2.2mm to approximately 0.45mm. This ensures that the algorithm retains only high-precision matching point pairs for subsequent calculations, eliminating erroneous matches caused by local geometric similarity. Simultaneously, the translation and rotation thresholds are also gradually tightened from the initial 0.5mm and 0.5° to 0.05mm and 0.05°. The algorithm considers convergence complete only when the changes in the transformation matrix calculated in two consecutive iterations are both less than this threshold.

[0040] The relatively wide initial threshold ensures robustness in the early stages of registration, avoiding the possibility of getting stuck in local optima due to insufficient matching point pairs caused by large initial pose deviations. As the iteration progresses, the threshold is gradually tightened, which ensures the high accuracy of the final registration result, keeping the stitching error between point clouds from different viewpoints within 0.1mm, achieving sub-millimeter level accuracy requirements.

[0041] After precise registration, the point cloud data from all perspectives are unified into the same coordinate system, forming a fused point cloud set containing complete geometric information of the workpiece (at this time, the overlapping areas from different perspectives may have excessively high point cloud density or slight misalignment, which requires subsequent processing for optimization).

[0042] After registration is completed, the 3D reconstruction module enters the point cloud fusion and surface reconstruction stage, transforming the discrete point cloud data into a continuous 3D mesh model that can be used for path planning.

[0043] In the point cloud fusion stage, the module first performs data fusion processing on the point clouds in overlapping areas (because the incident angle of the sensor is different when scanning from different perspectives, the position of the same physical point may have slight differences in multiple perspectives). The module uses a weighted average algorithm to assign different weights to each point according to the confidence level of each point relative to the incident angle of the sensor (the closer the incident angle is to perpendicular, the higher the weight), and calculates the optimal spatial position of the physical point. At the same time, for areas with excessive density, the module performs uniform sampling to remove redundant points, making the spatial distribution of the entire point cloud more uniform.

[0044] After point cloud fusion, a spatially scattered set of points is obtained, which needs to be transformed into a continuous triangular mesh model using a surface reconstruction algorithm. The 3D reconstruction module employs a surface reconstruction algorithm based on the Poisson equation. This algorithm can effectively handle point cloud data with noise and uneven sampling, generating smooth and detailed surfaces. The algorithm uses the normal vector information of the point cloud as a constraint, constructing an indicator function by solving a Poisson equation. The zero isosurface of this function is the reconstructed surface. In this embodiment, the module first estimates the normal vector direction of each point, ensuring that all normal vectors face the same direction. Then, it constructs an adaptive octree structure to partition the space, refining the mesh in dense point cloud regions and maintaining a larger mesh in sparse regions, thereby controlling the number of meshes while ensuring model accuracy. Finally, a continuous triangular mesh model is generated by extracting the isosurfaces of the octree nodes.

[0045] Specifically, the 3D reconstruction module employs a surface reconstruction algorithm based on the Poisson equation. This algorithm effectively handles point cloud data with noise and uneven sampling, generating smooth and detailed surfaces. The algorithm uses the normal vector information of the point cloud as constraints and constructs an indicator function χ by solving a Poisson equation. The zero isosurface of this function is the reconstructed surface.

[0046] Let the point cloud dataset be P = { , , …, }, each point Corresponding to a normal vector The normal vector is calculated using principal component analysis of the local neighborhood. The goal of the algorithm is to find an indicator function χ whose gradient... χ best approximates the vector field V defined by the point cloud normal vectors. This relationship can be expressed as: χ ≈ V; The vector field V is defined throughout the space by interpolating and diffusing the normal vectors of the discrete point cloud. To solve for the indicator function χ, we take the divergence on both sides of the above equation, transforming it into the form of a Poisson equation: Δχ = ·V; Where Δ is the Laplace operator, ·V is the divergence of the vector field V. The Poisson equation in three-dimensional space is: ; In the actual solution process, the algorithm first constructs an adaptive octree structure to partition the space. Let the depth of the octree be d, and each node correspond to a spatial region. In dense point cloud regions, the octree is refined to a deeper level, forming smaller mesh cells, thus preserving high resolution in areas rich in detail such as weld edges and bevel corners; in flat regions, the octree remains shallower, using larger mesh cells to reduce computational cost. Each node in the octree defines a set of basis functions. These basis functions are non-zero in the spatial region corresponding to the node, and the indicator function χ is approximated by a linear combination of the basis functions: ; in, Let be the nodal coefficients to be solved. Substituting the above expression into the Poisson equation, the coefficients are solved by minimizing the error function. This process is transformed into solving a sparse linear system of equations: A·x = b; Here, A is a symmetric positive definite sparse matrix with a dimension equal to the total number of nodes in the octree, and each element represents the Laplace operator interaction between basis functions; b is the right-hand side term, which is composed of the divergence integral of the vector field V. This system of linear equations is solved iteratively using the conjugate gradient method with multigrid preconditions, leveraging the multi-resolution characteristics of the octree to accelerate convergence.

[0047] After obtaining the indicator function χ, the zero isosurface of χ is found using an isosurface extraction algorithm (such as the moving cube algorithm). This involves identifying the location where the function value is zero in each octree node and connecting these locations to form a continuous triangular mesh. For characteristic areas such as bevels and chamfers on welded workpieces, the gradient of the indicator function changes drastically, and isosurface extraction can accurately capture these details. For the main plane of the workpiece, the indicator function changes gradually, resulting in a smooth and continuous mesh surface.

[0048] After surface reconstruction is complete, the 3D reconstruction module further enhances the generated mesh model by calculating the curvature of each vertex and the geometric changes in the local neighborhood. It identifies regions with significant curvature abrupt changes, typically corresponding to workpiece edges, corners, and weld locations. This module attaches these feature information as attributes to the mesh model, providing a direct basis for the subsequent adaptive path planning module to identify welds. Finally, the 3D reconstruction module outputs a high-precision, full-size 3D model of the workpiece, containing geometric information and feature annotations. This model accurately reflects the actual shape, size, and key weld feature areas of the workpiece.

[0049] Please see Figure 3 In this embodiment, the adaptive path planning module will be described in detail: After the adaptive path planning module is activated, it first performs a comprehensive analysis of the 3D mesh model of the workpiece generated by the 3D reconstruction module. The built-in weld feature extraction algorithm in this module locates potential weld areas by calculating the geometric properties of the model surface. Specifically, the algorithm traverses the triangular facets of the entire model, calculating the principal curvature and normal vector changes of each vertex. In typical welded workpieces, welds are usually located at the edge regions where two or more plates meet. These regions exhibit drastic curvature changes and discontinuous normal vector directions. The adaptive path planning module automatically filters out candidate regions that match the geometric characteristics of the weld by setting curvature thresholds and normal vector angle thresholds.

[0050] Specifically, the curvature threshold is used to distinguish the curvature abrupt changes between flat areas and weld areas on a workpiece surface. For most welded workpieces (such as automotive chassis longitudinal beams and engineering machinery structural components), the Gaussian curvature of the main surface is typically between 0.01 and 0.05 mm. -1 Within this range, the principal curvature at locations such as the weld groove edge and the root of the fillet weld can reach 0.5-2.0 mm due to drastic changes in geometry. -1 Therefore, the module sets the curvature threshold to 0.2 mm. -1 Specifically, all principal curvatures greater than 0.2 mm. -1 The vertices are marked as potential weld feature points. This threshold can effectively filter out minor curvature changes such as machining textures and slight scratches on the workpiece surface, while preserving the significant geometric features of the weld area.

[0051] The normal vector angle threshold is used to identify weld areas at the junctions of different sheet metals. On a continuous surface of a workpiece, the angle between the normal vectors of adjacent triangular facets is usually small, generally less than 15°. However, at the junctions of two sheet metals (e.g., T-joints, corner joints), the direction of the normal vector changes abruptly, and the angle can reach 60° to 120°. The module sets the normal vector angle threshold to 45°, meaning that when the angle between the normal vectors of adjacent facets exceeds 45°, the area is determined to be a potential weld boundary. This threshold can effectively distinguish weld areas from natural corners of the workpiece surface (e.g., at sheet metal bends, the angle between the normal vectors is usually between 30° and 45°), avoiding misidentification of non-weld structural features as welds.

[0052] The aforementioned thresholds are set based on statistical analysis of the geometric features of typical welded workpieces. During actual operation, the module will fine-tune the thresholds according to the workpiece type and scanning accuracy: for precision thin plates, the curvature threshold can be reduced to 0.15mm. -1 This allows for the capture of subtle weld features; for thick, heavy-duty structural components, the curvature threshold can be increased to 0.3 mm. -1 This is to avoid false detections caused by rough-machined surfaces.

[0053] For the selected candidate regions, this module employs a region growing algorithm for precise segmentation. Starting from the seed point with the largest curvature, the algorithm gradually expands to the surrounding neighborhood, merging continuous regions with similar geometric features into a complete weld region. During this process, the module also classifies and identifies welds based on their geometric morphology. For fillet welds formed by the T-shaped intersection of two plates, the point cloud distribution exhibits a distinct groove feature; for butt welds formed by the joining of plates, the surface shows continuous gaps or steps. By analyzing these features, the adaptive path planning module can accurately determine the type of each weld and assign a unique identifier to each weld, recording its spatial location, path length, and start and end point coordinates.

[0054] After the weld area is identified, the adaptive path planning module calls its internal bevel feature recognition module to perform refined geometric analysis on each weld. The bevel feature recognition module first extracts local point cloud slices of the weld area from the 3D model, and extracts a cross-sectional profile at regular intervals along the weld direction. For each cross-sectional profile, the bevel feature recognition module automatically identifies key geometric parameters by analyzing the distribution pattern of the point cloud.

[0055] Specifically, in the cross-sectional profile of a fillet weld, the bevel feature recognition module fits the surface planes of the two plates, calculates the included angle between these two planes to determine the welding angle, and measures the weld leg size and actual deposition space by detecting the depth and width of the point cloud depression near the intersection of the two planes. For butt welds, the bevel feature recognition module identifies the edges of the plates on both sides of the weld, accurately measures the minimum distance between the two edges as the root gap, and estimates the blunt edge height and bevel angle by analyzing the thickness direction of the plates at the edges and combining the back reflection information of the plates obtained during sensor scanning. For welds with backing materials, the bevel feature recognition module can also identify the position and outline of the backing material. All these geometric parameters are quantified and recorded as weld attribute data, and stored in association with the spatial location information of the weld.

[0056] Throughout the analysis process, the bevel feature recognition module intelligently handles point cloud data anomalies caused by local deformation or oxide scale. For areas with drastic gap changes, the bevel feature recognition module marks the location as a special process section and processes it separately during subsequent process parameter matching to ensure welding quality.

[0057] After obtaining the precise geometric parameters of each weld, the adaptive path planning module calls its internal process parameter matching module to make intelligent decisions on welding process parameters. The process parameter matching submodule has a built-in structured welding process knowledge database, which stores the optimal combination of process parameters for different base material combinations, different plate thickness ranges, different weld types, different groove sizes and different welding postures (flat welding, vertical welding, horizontal welding, overhead welding). These data usually come from welding procedure qualification tests and expert experience.

[0058] The process parameter matching module uses the geometric parameters output by the bevel feature recognition module as query conditions. First, it filters applicable records in the database based on the workpiece material, and then matches them step by step according to the weld type, plate thickness, and bevel size. For example, for a V-groove butt weld with a plate thickness of 12 mm, a bevel angle of 45 degrees, and a root gap of 2 mm, the process parameter matching module will search the database for the closest process qualification record with the same material, similar plate thickness, and matching bevel form, and extract the recommended welding current, arc voltage, welding speed, oscillation mode, oscillation amplitude, oscillation frequency, and layer number arrangement.

[0059] For the matched process parameters, the process parameter matching module will perform interpolation fine-tuning based on the specific characteristics of the current weld. If the current gap is slightly larger than the standard gap in the database, the process parameter matching module will appropriately increase the wire feed speed and welding current. If the welding position is vertical welding, the process parameter matching module will reduce the welding speed and adjust the oscillation parameters to ensure that the molten pool does not flow downward. Finally, the process parameter matching module will output a complete set of welding process parameter instructions for each segment or layer of the weld.

[0060] Once the process parameters for all welds are determined, the adaptive path planning module enters the final global path optimization and path generation stage. For complex workpieces containing multiple welds, the choice of welding sequence directly affects welding deformation and production efficiency. The adaptive path planning module first establishes a topology graph based on the workpiece's 3D model. Nodes in the graph represent welds or key locations, such as the start and end points of welds, the clamping points of tooling fixtures, transitional locations that the robot end effector must pass through, and points near obstacles that need to be avoided due to structural interference. Edges represent possible paths for the robot end effector to move between two points. The adaptive path planning module comprehensively considers multiple optimization objectives: calculating the shortest transfer path based on the Euclidean distance between welds; eliminating transfer paths that could cause collisions between the robot arm and the workpiece or fixture using collision detection algorithms; evaluating the robot's accessibility at the start and end points of each weld and prioritizing welding sequences with good robot posture and no singularities; and considering the heat effect distribution to avoid excessive heat concentration caused by continuous welding in the same area of ​​the workpiece.

[0061] Specifically, taking into account the heat effect distribution, a heat effect model is established to avoid excessive heat concentration caused by continuous welding in the same area of ​​the workpiece. Specifically: First, the module constructs a heat-affected zone (HAZ) mesh based on the workpiece's 3D model, discretizing the workpiece's surface and internal regions into 3D mesh cells. The side length of each mesh cell is set according to the workpiece's wall thickness and thermal conductivity characteristics. For thin plates (wall thickness less than 5mm), the mesh side length is set to 20mm; for thick plates (wall thickness greater than 20mm), the mesh side length is set to 50mm. Each mesh cell is assigned a cumulative heat input value H, initially set to 0, which represents the relative level of welding heat absorbed in that region.

[0062] When a weld is planned for welding, the module calculates the weld's contribution to the heat input of surrounding mesh cells based on the weld's length, welding speed, heat input power, and spatial position on the workpiece. The influence radius R of the weld is determined based on the material's thermal conductivity: for carbon steel with good thermal conductivity, the influence radius is set to 80 mm; for stainless steel with poor thermal conductivity, the influence radius is set to 120 mm. For each mesh cell within the influence radius, the increase in its cumulative heat input H is calculated according to a distance decay function; mesh cells closer to the weld experience a larger heat input increment, which approaches zero when the distance equals the influence radius.

[0063] The heat concentration threshold is set using "cumulative heat input value H" as the criterion. This is because the heat input varies significantly between different welds; the heat input of a long weld may be several times that of a short weld, and simply counting the number of times cannot accurately reflect the actual heat accumulation. The module's preset heat concentration threshold is... = 1.0. When the H value of a certain grid cell exceeds this threshold, it means that the area has been subjected to excessive heat input. If welding continues in this area, there will be a greater risk of thermal deformation or even burn-through.

[0064] Based on the aforementioned thermal impact model, the adaptive path planning module, when performing global path optimization, not only considers the shortest path, obstacle avoidance, and robot accessibility, but also incorporates "thermal distribution balance" as an independent optimization objective into the algorithm. The module employs a multi-objective optimization algorithm to determine the welding sequence, and the specific process is as follows: The first step is to calculate the initial priority of all weld seams to be welded. The priority is determined by the current cumulative heat input value H of the mesh cells surrounding the weld seam location; the lower the H value, the higher the priority. This strategy ensures that welding starts from the "cold zone," avoiding concentrated welding in the same area from the beginning.

[0065] The second step involves the module immediately updating the H value of all mesh elements within the influence range of each selected weld seam and recalculating the priority of the remaining weld seams. If the workpiece has two sets of weld seams symmetrically arranged on the left and right sides, the module will select to weld one of the weld seams in the left set first, then switch to the right set to weld one, and then return to the left set. This alternating welding process ensures that heat is evenly distributed across the workpiece, avoiding heat concentration on one side.

[0066] Third, when the cumulative heat input value H of a certain area approaches the threshold (reaching 0.8), the module will force the priority of the remaining welds in that area to be reduced to the lowest level, and instead prioritize welding the welds in other areas, even if these welds require the robot to make a long idle transfer, to ensure the balance of heat distribution.

[0067] The fourth step is that when the cumulative heat input value H of a certain area reaches the threshold of 1.0, the module will forcibly insert a cooling waiting time. The system will calculate the required natural cooling time or trigger an external cooling device to force cooling of the area. Welding in that area will continue after the H value drops below 0.6.

[0068] Taking a large box girder structural component as an example, this component contains four symmetrically distributed longitudinal fillet welds and six transverse butt welds. Traditional sequential welding (welding all longitudinal welds first, then the transverse welds) leads to a high concentration of heat in the longitudinal weld areas, causing significant bending deformation of the box girder. The adaptive path planning module of this invention, based on a thermal effect model, plans the following welding sequence: first, weld the first half of the top longitudinal weld; then weld the first half of the bottom longitudinal weld; then weld the first half of the left longitudinal weld; then weld the first half of the right longitudinal weld; then switch to the transverse welds, proceeding sequentially and alternately; finally, return to complete the latter half of each weld. This sequence ensures that the cumulative heat input value H of the mesh elements in any region of the component remains below 0.7 throughout the welding process, without reaching the threshold. Ultimately, the overall deformation of the component after welding is reduced by approximately 60% compared to the traditional sequence, effectively controlling thermal deformation.

[0069] Based on the aforementioned multi-objective constraints, the adaptive path planning module employs a graph search algorithm to find the globally optimal welding sequence. Once the sequence is determined, the module generates the final optimized welding path. This path includes not only the position and attitude sequence of the robot's TCP (tool center point) in three-dimensional space, but also the oscillation trajectory planning of the welding torch as it moves along the weld seam. Simultaneously, the path file embeds welding process parameter instructions generated by the process parameter matching module in the form of data packets. Each path segment is dynamically associated with corresponding parameters such as current, voltage, and wire feed speed.

[0070] The final executable welding path file is transmitted to the real-time feedback control module and the industrial robot arm controller, waiting to be executed. The entire adaptive path planning process is fully automated and requires no human intervention. It can complete complex path planning tasks that would take hours to complete using traditional teaching methods in just a few minutes, and it can intelligently adapt to the individual differences of each workpiece.

[0071] Please see Figure 4 and Figure 5 In this embodiment, the real-time feedback control module will be described in detail: The real-time feedback control module is responsible for dynamic correction and adaptive control during the welding process. When the industrial robot arm begins to move according to the initial optimized welding path, the 3D vision sensor continuously collects real-time point cloud data of the area to be welded in front of the welding torch, as well as image information of the vicinity of the molten pool, at a frequency of tens of frames per second or even higher. The real-time feedback control module receives this real-time data and performs the following operations: The actual position and contour of the current weld are extracted from the real-time point cloud using image processing algorithms and compared with the theoretical weld path generated by the adaptive path planning module. The module calculates the lateral, height, and angular deviations caused by workpiece thermal deformation, clamping errors, or robot motion errors. When the deviation exceeds a preset small threshold, the module immediately generates a trajectory correction command, which is sent to the industrial robot arm's controller via the communication bus. This allows for real-time adjustment of the robot's motion trajectory, ensuring the welding torch is always aligned with the weld center.

[0072] The module also monitors changes in weld gap in real time. During welding, the weld gap dynamically changes due to heat accumulation. The real-time feedback control module dynamically adjusts the output power or wire feed speed of the welding equipment based on the detected real-time gap value. Specifically, when an increase in gap is detected, the module sends instructions to the welding equipment to increase the wire feed speed and welding current to increase the amount of deposited metal and ensure full filling; conversely, when the gap decreases, the module correspondingly reduces the wire feed speed and welding current to prevent burn-through or excessive weld reinforcement, thus achieving adaptive welding under variable gap conditions.

[0073] To achieve higher precision dynamic tracking, the real-time feedback control module of this invention adopts a prediction-correction-based control strategy. This strategy not only corrects the deviation based on the current moment's error, but also combines the deviation trend over a period of time, using Kalman filtering or other prediction algorithms to predict the possible deviation at the next moment and perform feedforward compensation in advance. This prediction-correction mechanism can effectively overcome the response delay of the control system, enabling the robot to track complex curved welds more smoothly and accurately.

[0074] Furthermore, the predictive-correction-based control strategy will be explained in detail: When the industrial robot arm guides the welding torch along the initial optimized welding path, the 3D vision sensor installed at the end of the robot arm continuously collects real-time point cloud data of the area in front of the molten pool at high frequency. After receiving these data streams, the real-time feedback control module immediately starts the preprocessing process.

[0075] First, a fast filtering algorithm removes instantaneous noise caused by arc light and spatter. Then, an edge extraction algorithm is used to identify the actual position and contour of the weld seam from the real-time point cloud. Specifically, the real-time feedback control module analyzes the feature points of curvature abrupt changes in the point cloud, fits the centerline position and left and right edges of the weld seam, and accurately measures the actual weld seam gap at that moment. After this processing, the module can obtain the actual deviation value of the weld seam relative to the welding torch (including lateral deviation and height deviation) and the real-time gap value at the current moment within each control cycle.

[0076] After obtaining the actual deviation at the current moment, the real-time feedback control module enters the prediction calculation stage. Internally, the real-time feedback control module maintains a fixed-length historical deviation buffer, which continuously stores deviation sequence data from the past dozens of control cycles. Based on this historical data, the module uses time series analysis to analyze the trend of deviation changes: it determines whether the deviation is increasing or decreasing by calculating the first difference of the deviation sequence, and analyzes the acceleration of the deviation change by using the second difference.

[0077] Based on the analysis of historical trends, the real-time feedback control module uses a prediction algorithm to estimate the possible deviation value at the next moment. This prediction is not a simple linear extrapolation, but rather a comprehensive consideration of the average rate and acceleration of deviation changes over a recent period, combined with welding process knowledge to impose reasonable constraints on the prediction results. The prediction results represent the possible deviation value of the system at the next moment without external correction.

[0078] Based on the predicted future deviation value, the real-time feedback control module calculates the feedforward compensation amount. Since a certain amount of deviation has been predicted to occur at the next moment, the system can send a correction command to the robot controller in advance, so that the robot can make a slight movement in the opposite direction of the deviation in advance, thereby offsetting part of the deviation before it actually occurs.

[0079] Specifically, the real-time feedback control module maps the predicted deviation value to a correction vector in the robot's motion space. For example, if it predicts that the weld will shift 0.5 mm to the right in the next moment, the module generates a feedforward correction command of 0.5 mm to the left. The amplitude of this command is not simply equal to the predicted deviation, but is dynamically adjusted by the gain: when the predicted deviation changes slowly, the gain coefficient is small, and the correction is gradual; when the predicted deviation increases sharply, the gain coefficient increases accordingly to quickly respond to drastic changes. The calculated feedforward compensation amount is temporarily stored, ready to be fused with subsequent feedback correction amounts.

[0080] While calculating the feedforward compensation, the real-time feedback control module calculates the feedback correction based on the actual deviation value collected in the current control cycle. The feedback correction follows classic control logic: the current actual deviation is compared with a preset allowable error threshold; the larger the deviation, the larger the correction. The feedback control, working in conjunction with the predictive-correction strategy, employs a variable gain mechanism. When the system predicts an increase in deviation, the feedback gain is appropriately increased to enhance the correction force; when the predicted deviation tends to converge, the feedback gain is decreased to avoid oscillations caused by over-correction.

[0081] After the feedback correction calculation is completed, the module weights and fuses the feedforward compensation and feedback correction. The weight allocation depends on the current system state. During the welding start-up phase or in areas of abrupt weld shape changes, the feedback correction has a higher weight to ensure rapid response. During the stable welding phase and when the predicted trend is reliable, the feedforward compensation has a slightly higher weight to achieve smooth tracking. This dynamic weight allocation mechanism allows the control strategy to adapt to the needs of different welding stages.

[0082] After fusing the feedforward compensation and feedback correction, the real-time feedback control module generates the final trajectory correction command. This command is a six-dimensional spatial vector containing fine-tuning amounts for the robot's position and attitude. The real-time feedback control module sends the correction command to the servo driver of the industrial robot arm via a high-speed fieldbus, requiring it to execute the micro-motion within the next control cycle. The typical cycle of this process is within 10 milliseconds, achieving true real-time correction. Simultaneously, based on real-time detected changes in the weld gap, the real-time feedback control module generates welding process parameter adjustment commands in parallel. When it predicts that the gap will continue to increase, the real-time feedback control module sends commands to the welding equipment to gradually increase the wire feed speed and welding power, synchronizing the increase in deposited metal with the gap change. When a sudden decrease in the gap is detected, the module immediately issues a command to reduce the heat input to prevent overfilling of the molten pool. The synchronous issuance of process parameter adjustment commands and trajectory correction commands achieves coordinated control of motion and energy.

[0083] After each control cycle, the actual correction effect is verified using real-time point cloud data from the next cycle. The real-time feedback control module continuously monitors the corrected actual deviation value to evaluate the effectiveness of the prediction-correction strategy. If a systematic error is found between the predicted deviation and the actual deviation, the module automatically adjusts the parameters in the prediction model to achieve online self-optimization of the control strategy. This self-learning capability enables the control strategy to adapt to the dynamic characteristics of different workpieces, welding positions, and welding stages. Over time, the system's tracking accuracy and stability continuously improve.

[0084] Please see Figure 6In this embodiment, throughout the entire welding operation, regardless of whether the robot is moving, the module continuously uses 3D vision sensors to acquire dynamic environmental point cloud data in the welding unit's workspace in real time. Based on this data, the safety monitoring module constructs and updates a dynamic three-dimensional scene model that includes the industrial robot arm, welding torch, welding workpiece, fixture, peripheral equipment, and operator in real time.

[0085] Based on this model, the module runs a space collision detection algorithm to build dynamic virtual safety fences for all moving parts and obstacles. For example, it sets an inviolable safety distance between each link of the robot, the welding torch and the workpiece. The algorithm calculates the shortest distance between the robot's current trajectory and the surrounding static and dynamic obstacles in real time. Once it predicts that any part will enter the preset dangerous approach distance in the next control cycle, or detects a potential collision risk in the simulation prediction mode, the safety monitoring module will immediately send a deceleration command or emergency stop signal to the controller of the industrial robot arm to ensure the safety of human-robot collaborative operation. In the no-load operation mode before welding begins, the module can also simulate the entire welding trajectory, detect potential collision risks in advance and issue an alarm to prevent accidents caused by program errors.

[0086] Please see Figure 7 In this embodiment, a robot intelligent welding method based on 3D vision guidance is described, specifically: (1) The workpiece to be welded is placed at the welding station and roughly positioned using a fixture. The operator starts the scanning program on the vision processing controller. Based on the size and complexity of the workpiece, the controller plans the scanning path of the 3D vision sensor. If the sensor is fixedly installed, the overall point cloud can be obtained with just one shot. If the sensor is installed at the end of the robot arm, the robot arm will carry the sensor and move along a preset trajectory to perform a full-coverage scan of the workpiece from multiple angles. During the scanning process, the 3D vision sensor projects structured light or emits laser lines onto the surface of the workpiece and receives reflected signals to generate high-density three-dimensional point cloud data. This data is transmitted to the vision processing controller in real time, where its three-dimensional reconstruction module performs filtering, registration, and surface reconstruction to finally generate a full-size three-dimensional model that is consistent with the height of the actual object.

[0087] (2) After obtaining the full-size 3D model of the workpiece, the adaptive path planning module starts working. First, the module automatically identifies one or more weld seams to be welded from the model using algorithms such as edge detection and curvature analysis, and clarifies the starting and ending points of each weld seam. Then, the groove feature recognition module performs a detailed analysis of the local area of ​​each weld seam, automatically extracting key geometric features such as groove angle, root gap, and blunt edge height. Then, the process parameter matching module inputs these feature information into the welding process knowledge base for retrieval and matching, finding the most suitable welding current, voltage, speed, and oscillation parameters for the current weld seam. Finally, the adaptive path planning module integrates these process parameters with the robot motion trajectory to generate an initial optimized welding path. In the case of multiple weld seams, the adaptive path planning module also comprehensively considers factors such as the shortest path, obstacle avoidance, robot posture accessibility, and thermal effects, and automatically plans the optimal welding sequence and idle transfer path to ensure that the entire welding process is efficient and stable.

[0088] (3) After the planning is completed, the industrial robot arm begins to execute the optimized welding path, guiding the welding torch to weld the first weld seam. At the same time, the 3D vision sensor starts the real-time acquisition mode to continuously acquire point cloud data of the area to be welded in front of the welding torch. The real-time feedback control module quickly processes this data, extracts the position and gap information of the actual weld seam, and calculates the accurate deviation value by comparing the actual position with the theoretical path point by point. Using the predictive-correction control strategy, the module generates a composite correction command of feedforward and feedback, which is sent to the robot controller in real time to fine-tune the robot's motion trajectory so that the welding torch always accurately tracks the actual weld seam. At the same time, according to the gap changes monitored in real time, the module dynamically adjusts the output of the welding equipment to achieve adaptive filling under variable gap conditions. This process is repeated at a very high frequency to ensure that the welding process can continue with high quality even under strong thermal deformation interference.

[0089] Throughout the welding process, the safety monitoring module remains active. Utilizing real-time environmental point cloud data from 3D vision sensors, it continuously updates a dynamic 3D scene encompassing the robot, welding torch, workpiece, and surrounding environment. An internal collision detection algorithm calculates the distances between each robot component and surrounding obstacles in real time and dynamically constructs virtual safety fences. If any component is detected about to enter a danger zone, or if a trajectory intersection risk is predicted, the safety monitoring module immediately intervenes, sending deceleration or emergency stop commands to the robot controller via hardware I / O or a high-speed bus. This effectively prevents collisions and ensures the safety of equipment and personnel.

[0090] (4) After a weld is completed, or after all welding operations are completed, the 3D vision sensor scans the weld area again to obtain the post-weld point cloud data. The vision processing controller compares the post-weld point cloud with the standard weld model. By analyzing data such as the height and width of the point cloud, it automatically evaluates quality indicators such as weld reinforcement height, weld width, and surface ripple uniformity. The evaluation results can be displayed in real time and a quality inspection report can be generated. If a quality deviation is found to be out of tolerance, the system can immediately alarm and prompt the operator to intervene for inspection or re-welding. At the same time, the data is fed back to the adaptive path planning module for optimization of process parameters for subsequent similar welds, realizing closed-loop quality control.

[0091] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A robotic intelligent welding unit based on 3D vision guidance, characterized in that, include: An industrial robot arm with a welding torch mounted at its end; Welding equipment, connected to a welding torch, used to provide welding power and welding wire; At least one 3D vision sensor is installed at a fixed position in the intelligent welding unit or at the end of an industrial robot arm to scan the workpiece and its surrounding environment before and during welding to acquire three-dimensional point cloud data. A vision processing controller, communicating with a 3D vision sensor and an industrial robot arm, includes: The 3D reconstruction module is used to receive and process 3D point cloud data to reconstruct a full-size 3D model of the welded workpiece. The adaptive path planning module has a built-in weld feature extraction algorithm, which is used to identify and extract the weld area from the full-size 3D model. Based on the geometric features of the weld area and the preset welding process database, it plans an optimized welding path in real time. The optimized welding path not only includes the robot's motion trajectory, but also the welding process parameter instructions that are dynamically associated with the motion trajectory. The real-time feedback control module is used to simultaneously receive real-time point cloud data from a 3D vision sensor during the process of the industrial robot arm guiding the welding torch to perform welding along the optimized welding path. By comparing the real-time point cloud data with the full-size 3D model or the optimized welding path, it calculates the deviation of the weld position and gap caused by workpiece thermal deformation or clamping error, and generates compensation instructions for real-time correction of the motion trajectory of the industrial robot arm and / or the process parameters of the welding equipment. The safety monitoring module is used to construct a dynamic virtual safety fence containing the industrial robot arm, welding torch, welding workpiece and surrounding equipment based on real-time environmental point cloud data acquired by 3D vision sensors throughout the welding process. When it is predicted that any part will enter the preset dangerous approach distance, it immediately sends a deceleration or emergency stop signal to the industrial robot arm.

2. The robotic intelligent welding unit based on 3D vision guidance according to claim 1, characterized in that: The 3D vision sensor is a laser profile scanner or a structured light 3D camera, which is installed at the end of an industrial robot arm and moves with the robot arm. It can scan the welding workpiece from multiple angles to obtain high-precision point cloud data, which is used to eliminate scanning blind spots under a single viewpoint.

3. The robotic intelligent welding unit based on 3D vision guidance according to claim 1, characterized in that, The adaptive path planning module includes: The bevel feature recognition module is used to analyze point cloud data of the weld area and identify the weld type, bevel angle, blunt edge height and root gap. The process parameter matching module, based on the output of the bevel feature recognition module, retrieves and matches the optimal welding current, voltage, welding speed, and oscillation parameters from the preset welding process database, and embeds them into the instruction for optimizing the welding path.

4. The robotic intelligent welding unit based on 3D vision guidance according to claim 1, characterized in that, The real-time feedback control module is specifically used for: extracting the weld point cloud in the area near the current molten pool in real time during the welding process, and comparing it with the theoretical weld position in the optimized welding path; if the deviation in the lateral or height direction exceeds the preset threshold, sending a position correction amount to the industrial robot arm; if the weld gap change is detected, sending an instruction to the welding equipment to adjust the wire feeding speed or welding power.

5. The robotic intelligent welding unit based on 3D vision guidance according to claim 1, characterized in that: The safety monitoring module is also used to simulate the robot's movement trajectory in the no-load operation mode before welding begins, and to predict collision risks based on a dynamic virtual safety fence. If a potential collision risk is detected, an alarm is issued and the startup program is locked.

6. A 3D vision-guided robotic intelligent welding method, applied in any one of claims 1-5, characterized in that, Includes the following steps: S1 controls the 3D vision sensor to scan the workpiece placed on the welding station, obtain its complete three-dimensional point cloud data, and reconstruct a high-precision full-size three-dimensional model through the vision processing controller. S2, the vision processing controller extracts features from the full-size 3D model, identifies one or more weld seams to be welded, and generates an initial optimized welding path based on the geometry of the weld seams and the preset welding process knowledge base. This path data not only includes the robot's motion trajectory points, but also embeds welding process parameter instructions that match the path segment. S3, the industrial robot arm starts to move according to the optimized welding path, guiding the welding torch to weld. During this process, the 3D vision sensor collects point cloud data of the welding area and the front of the molten pool in real time. The vision processing controller compares the real-time point cloud data with the full-size 3D model, calculates the deviation between the actual position of the weld and the planned path caused by factors such as thermal deformation, and immediately generates a trajectory correction command to send to the robot arm. At the same time, it adjusts the output power or wire feeding speed of the welding equipment in real time according to the change of the actual weld gap. S4. During the welding process in step S3, the safety monitoring module uses the environmental data acquired in real time by the 3D vision sensor to construct and update a dynamic virtual safety fence that includes the moving robot, welding torch, workpiece and surrounding equipment. Through the spatial collision detection algorithm, it judges in real time whether there is a collision risk. Once the risk value exceeds the safety threshold, it immediately triggers an emergency stop.

7. The robotic intelligent welding method based on 3D vision guidance according to claim 6, characterized in that, In step S2, the process of generating the initial optimized welding path specifically includes: among the identified multiple weld seams, based on multi-objective constraints, planning a globally optimal welding sequence and robot transfer path to minimize the robot's idle travel time and reduce overall heat input. The multi-objective constraints include at least the shortest path, obstacle avoidance, and robot reachability.

8. The robotic intelligent welding method based on 3D vision guidance according to claim 6, characterized in that, In step S3, the real-time adjustment of the output power or wire feeding speed of the welding equipment specifically includes: when the weld gap is detected to increase, the vision processing controller increases the amount of deposited metal by increasing the welding current or wire feeding speed; when the weld gap is detected to decrease, the welding current or wire feeding speed is reduced accordingly to prevent weld burn-through or excessive weld height, thereby achieving adaptive welding under variable gap conditions.

9. A robotic intelligent welding method based on 3D vision guidance according to claim 6, characterized in that, In step S3, the trajectory correction instruction is generated based on the prediction-correction control strategy. Specifically, the vision processing controller first predicts the deviation amount at the next moment based on the deviation change trend over a recent period and performs feedforward compensation, and then performs feedback correction based on the actual deviation at the current moment, thereby achieving high-precision tracking of complex trajectories.

10. A robotic intelligent welding method based on 3D vision guidance according to claim 6, characterized in that: During the welding process or after a single weld seam is completed, the 3D vision sensor will scan the weld area again to obtain the post-weld formed point cloud data. The vision processing controller compares the post-weld formed point cloud data with the standard weld seam model to evaluate the weld seam's reinforcement height, weld width, and surface forming quality, and generates a quality inspection report for real-time process adjustment or subsequent traceability.

Images