A welding robot homing method and system based on dynamic non-parametric point cloud processing

Abstract

The application discloses a welding robot positioning method and system based on dynamic non-parametric point cloud processing, which comprises the following steps: extracting multi-scale geometric statistical features, performing mean aggregation, and constructing an initial prototype memory bank; extracting geometric statistical features; performing adaptive gated noise reduction and feature purification; performing a welding seam extraction strategy based on an exponential function weighting method; performing unsupervised dynamic incremental updating of the prototype memory bank to realize online iteration; based on the segmented pure welding seam point cloud, the tangent and normal directions of the welding seam are calculated by using the local direction vector in the multi-scale geometric statistics, the posture instruction is generated, and after the hand-eye matrix conversion, the posture instruction is sent to the mechanical arm to drive the robot to complete adaptive welding work. The application realizes the welding robot adaptive positioning with no pre-training, low algorithm power dependence and strong noise resistance, and can realize real-time reasoning on an ordinary industrial computer, and is suitable for the welding scene with strong noise and dynamic drift of working conditions in an industrial field.

Description

Technical Field

[0001] This invention belongs to the field of industrial robots and relates to industrial robot visual perception and automatic control technology, specifically to a welding robot positioning method and system based on dynamic non-parametric point cloud processing. Background Technology

[0002] With the rapid development of intelligent manufacturing and industrial automation, welding robots based on 3D vision are widely used in modern manufacturing. In automated welding processes, accurate segmentation and feature extraction of the weld seam from the original 3D point cloud of the workpiece are core prerequisites for achieving autonomous positioning, online path planning, and stable welding formation of the welding robot. Currently, there are three main technical shortcomings in weld seam extraction and robot positioning technologies for 3D point clouds:

[0003] The first type is the point cloud weld extraction method based on deep learning, with representative solutions such as PointNet++ and DGCNN. Although this type of method can achieve high weld recognition accuracy, it relies on gradient training and parameter learning on large-scale labeled datasets. In complex industrial workshop applications, it suffers from high computational power dependence and high hardware deployment costs. At the same time, the black box effect leads to poor engineering controllability and the need to shut down and retrain when changing production. It cannot adapt to flexible welding production scenarios with multiple varieties and small batches, and its engineering implementation is significantly limited.

[0004] The second type is the existing non-parametric point cloud processing method, represented by the Point-NN framework proposed in recent years. With its non-parametric computing mode, it has gotten rid of its dependence on high-end computing power and training data. However, its core computing method is too simple and has extremely poor noise robustness under the strong spatter, strong reflection and point cloud outlier noise interference in industrial welding scenarios. At the same time, it has not optimized for the problem of extreme imbalance between weld seam and base material point cloud categories, and has extremely low extraction accuracy for micro-sized weld seams with weak geometric features and narrow gap weld seams, which cannot meet the accuracy requirements of industrial welding.

[0005] The third type involves weld seam location methods. Currently, most algorithms only focus on the recognition and segmentation of the weld seam area, failing to fully utilize the local geometric normal features of the point cloud. Additional manual attitude calibration and path planning are required, resulting in insufficient automation and flexible adaptability.

[0006] In summary, to address the core pain points of limited computing power, frequent production changes, strong noise interference, and dynamic drift of working conditions in industrial welding sites, there is an urgent need to develop a 3D point cloud adaptive positioning system and method that requires no pre-training, consumes very little computing power, has strong noise resistance, can dynamically update the prototype online as the workshop environment changes, and ultimately directly generates 6D action postures that the robot can perform. Summary of the Invention

[0007] Purpose of the invention: To address the shortcomings of existing parametric deep learning models, such as high computational consumption and the need for retraining during production changes when deployed in industrial settings, poor noise resistance of non-parametric methods, inability to handle operational drift and class imbalance issues, and the lack of a complete closed-loop welding process in existing solutions, this invention provides a welding robot positioning method and system based on dynamic non-parametric point cloud processing. This method achieves adaptive positioning of welding robots without pre-training, with low computational dependence and strong noise resistance. It can perform real-time inference on ordinary industrial control computers, is plug-and-play for production changes, and is suitable for welding scenarios with high noise and dynamic operational drift in industrial settings. It can be widely applied in automated welding fields such as automobiles and ships.

[0008] Technical Solution: To achieve the above objectives, this invention provides a welding robot positioning method based on dynamic non-parametric point cloud processing, comprising the following steps:

[0009] S1: By collecting unlabeled point cloud samples of the target weld and base material, extracting multi-scale geometric statistical features of the samples, aggregating the mean by category, and constructing an initial prototype memory library, which includes feature prototypes of weld and base material.

[0010] S2: Real-time acquisition of the original 3D point cloud of the workpiece to be welded, construction of a multi-scale local neighborhood for each downsampled point, and extraction of geometric and statistical features;

[0011] S3: Based on the extracted geometric statistical features, calculate the feature variance and distance normalization weight of the local point cloud, and assign fusion weights to outliers with high variance.

[0012] S4: Calculate the similarity between the denoised point cloud features and various prototypes in the prototype memory. Use an exponentially weighted soft prototype voting strategy to calculate the classification probability of each point, exponentially amplify the weight of high similarity features, and extract the local weld point cloud with weak geometric features.

[0013] S5: In real time, high-confidence prediction results with classification probabilities higher than the preset threshold in step S4 are selected, and the features of these high-quality point clouds are extracted and fused into the existing prototype memory online using the moving average algorithm to achieve unsupervised online iteration.

[0014] S6: Based on the segmented pure weld point cloud, the tangent and normal directions of the weld are directly calculated using the local direction vector in multi-scale geometric statistics. This generates 6D attitude commands with spatial three-dimensional coordinates and yaw, pitch, and roll angles. After hand-eye matrix conversion, these commands are sent to the robotic arm to drive the robot to complete adaptive welding operations.

[0015] Furthermore, the method for performing mean aggregation by category in step S1 includes:

[0016] The feature prototypes for weld seam and base material classes are calculated, while the inverse class frequency weighting coefficient is introduced to address the class imbalance problem. The calculation of the feature prototype and the inverse class frequency weighting coefficient is shown below:

[0017]

[0018] In the formula: t is the category index, t=1 represents the weld category, and t=2 represents the base material category; This is the feature prototype of class t; Let be the set of sample points of class t; Let be the number of sample points in class t; Contextual enhancement features for sample point i; This is the inverse class frequency weighting coefficient; the smaller the sample size of a class, the higher its basic weight is assigned. C represents the total number of points in the sample set; C represents the total number of categories.

[0019] Furthermore, the extraction of geometric statistical features in step S2 includes:

[0020] S2-1: Basic position encoding: For the initial input point cloud, the K-neighborhood point set of each center point is obtained through the FPS+KNN algorithm. Multi-band sine and cosine functions are used to encode the absolute and relative positions of the input point cloud to generate basic position features.

[0021] S2-2: Multi-scale geometric encoder encoding: based on dual-scale neighborhood , It extracts neighborhood statistical features, distance encoding features, and unit direction vector features to construct multi-dimensional geometric features. The entire process is a deterministic geometric operation with no learnable parameters.

[0022] S2-3: Feature Fusion: Extend global statistical features to the same dimension as neighborhood features, align neighborhood features at different scales to a unified number of points through linear interpolation, and finally fuse them into multi-scale geometric statistical features with consistent dimensions.

[0023] Furthermore, the neighborhood statistical features in step S2-2 are obtained by calculating the mean, standard deviation, and minimum distance of the neighborhood point cloud. The calculation formula is as follows:

[0024]

[0025] In the formula: This represents the number of neighborhood points at the corresponding scale. This is the robust offset. The output is the neighborhood statistical features;

[0026] The formulas for calculating distance coding features and unit direction vector features are as follows:

[0027]

[0028] In the formula: For the first The unit direction vector of each neighboring point relative to the center point; Distance-encoded features; This represents the number of frequency bands.

[0029] Further, step S3 includes:

[0030] S3-1: Two-branch feature construction: Two branches are constructed: residual features and original features. This captures the geometric differences between the neighborhood and the center point while preserving the original structural information of the neighborhood. The calculation formula is as follows:

[0031]

[0032]

[0033] In the formula: Neighborhood-center residual characteristics; Features of the original neighborhood;

[0034] S3-2: Adaptive Gating Coefficient Calculation: Based on neighborhood distance normalization and local feature variance, a gating coefficient is constructed. Higher weights are assigned to spatially close and feature-stable neighboring points, while lower weights are assigned to distant outliers and noisy points with large feature fluctuations. The calculation formula is as follows:

[0035]

[0036] In the formula: For the first Local variance of features of neighborhood points; Preset weighting coefficients; This is the scaling factor; For the final adaptive gating coefficient;

[0037] S3-3: Adaptive Feature Fusion and Refinement: Adaptive fusion of bi-branch features is performed using gating coefficients to obtain refined features of the center point. The calculation formula is as follows:

[0038]

[0039] In the formula: For the first Adaptive fusion features of neighboring points; To refine the features of the center point obtained from the final aggregation; This is a feature splicing operation.

[0040] Further, step S4 includes:

[0041] S4-1: Cosine Similarity Calculation: After normalizing the test point features and the prototypes in the prototype memory, the cosine similarity matrix is ​​calculated using the following formula:

[0042]

[0043] In the formula: For the first The characteristics of the test point and the first test point Cosine similarity between class prototypes;

[0044] S4-2: Similarity Scaling and Exponential Weighting: By using scaling factors and inverse class frequency weighting factors, the contribution of high-confidence weld features is exponentially amplified. The calculation formula is as follows:

[0045]

[0046] In the formula: This is the similarity scaling factor;

[0047] S4-3: Soft Voting Class Prediction: Normalization is achieved using the Softmax function to obtain the class prediction probability for each point, thus completing the soft voting decision. The calculation formula is as follows:

[0048]

[0049] In the formula: For the first The point belongs to the th point The predicted probability of a class is obtained by taking the maximum probability to determine the segmentation category of each point.

[0050] Further, step S5 includes:

[0051] S5-1: Initial screening of high-confidence samples and first-level point verification: Select points in the weld seam category and base material category whose predicted probability is higher than the preset high-confidence threshold respectively to form the corresponding high-confidence sample set. If the number of points in a single category sample set is lower than the preset minimum sample threshold, then abandon the sample update for this category.

[0052] S5-2: Sample anomaly removal and secondary point count verification: For high-confidence sample sets, outliers are removed based on neighborhood feature variance, and only valid samples with stable geometric features are retained. If the number of sample points does not meet the minimum sample threshold after secondary verification, the update process for this category is terminated.

[0053] S5-3: Moving Average Prototype Update: This method uses a moving average algorithm to achieve smooth iteration of the prototype. The calculation formula is as follows:

[0054]

[0055] In the formula: For the updated number Class feature prototype; The original feature prototype before the update; The moving average coefficient; This represents the number of points in the valid high-confidence sample set for this category.

[0056] S5-4: Update Anomaly Validation and Disk Storage: Calculate the cosine similarity between the prototype before and after the update. If the similarity is lower than the preset minimum similarity threshold, discard the update result. After the validation is passed, overwrite the original prototype library and store it on disk, while retaining historical version backups and update logs.

[0057] Further, step S6 includes:

[0058] S6-1: Key point screening of weld path: Denoise and smooth the pure weld point cloud, fit the weld centerline, sample key points of the path, and form an ordered set of key points;

[0059] S6-2: Calculation of the three-axis vector of the local coordinate system of the weld: For each key point of the path, calculate the X-axis tangent vector and Z-axis normal vector respectively, obtain the Y-axis subnormal vector through cross product, and obtain the orthogonal local coordinate system three-axis after Schmidt orthogonalization correction;

[0060] S6-3: 6D Pose Calculation in Camera Coordinate System: Construct the rotation matrix from the local coordinate system to the camera coordinate system. Following the common fixed coordinate system external rotation and ZYX rotation sequence for welding robots, calculate the 3D coordinates, yaw, pitch, and roll angles in the camera coordinate system. The calculation formulas are as follows:

[0061]

[0062] In the formula: This is the rotation matrix from the local coordinate system to the camera coordinate system; These are the X, Y, and Z axis vectors of the local coordinate system, respectively. For the rotation matrix, the first Line number Column elements;

[0063] S6-4: Hand-Eye Matrix Transformation to Robot Base Coordinate System: Based on the eye-on-hand mounting method, a homogeneous transformation matrix link is used to convert the pose in the camera coordinate system into an executable pose in the robot base coordinate system. The core transformation link is as follows:

[0064]

[0065] In the formula: This is the homogeneous transformation matrix from the local coordinate system of the weld to the robot's base coordinate system; Homogeneous transformation from the tool coordinate system to the base coordinate system for real-time feedback from the robot controller; For the pre-calibrated hand-eye matrix; This is the homogeneous transformation matrix from the local coordinate system of the weld to the camera coordinate system;

[0066] S6-5: Welding process adaptation and path smoothing optimization: The attitude angle is adapted and adjusted according to the bevel type, and the 6D pose of adjacent key points is interpolated by a fifth-order polynomial to finally output a 6D pose sequence that the robot can directly execute.

[0067] The present invention also provides a welding robot positioning system based on dynamic non-parametric point cloud processing, comprising:

[0068] The 3D point cloud data acquisition module is used to scan the workpiece to be welded in the industrial site and acquire the original 3D point cloud data, including the weld area and the background of the base material, in real time.

[0069] The initial prototype building module is used in the system cold start phase. Based on a small number of unlabeled point cloud samples of workpieces to be welded or the geometric priors of the bevel, it extracts the geometric statistical features of the weld and the base material and performs mean aggregation to build the initial prototype memory library of weld and base material classes.

[0070] The edge feature calculation module is deployed in a general industrial control computer and communicates with the 3D point cloud data acquisition module and the initial prototype construction module. It is used to receive real-time acquired 3D point cloud data, perform multi-scale feature extraction at the pure geometric and statistical levels, adaptive noise filtering and weld point cloud segmentation, and has no learning parameters throughout the process.

[0071] The dynamic incremental learning module is bidirectionally connected to the edge feature calculation module. It is used to extract high-confidence point cloud features output by the edge feature calculation module during continuous welding operations, perform unsupervised online updates and iterations on the prototype memory, and feed the updated prototype memory back to the edge feature calculation module in real time, forming a dynamic adaptive closed loop.

[0072] The robot execution module communicates with the edge feature calculation module and is used to calculate and generate 6D pose instructions for the welding robot's positioning and trajectory following based on the extracted clean weld point cloud spatial pose, thereby driving the robot to complete adaptive welding operations.

[0073] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0074] 1. Construct a non-parametric point cloud feature extraction and segmentation architecture with no learnable parameters throughout the entire inference process. By building a prototype memory based on a small number of unlabeled samples, the system can achieve high frame rate real-time inference on ordinary industrial control machines without high-end GPUs, which greatly reduces the hardware threshold for deployment.

[0075] 2. Enable plug-and-play, zero-downtime training for welding production changeover scenarios. Through the unannotated cold-start prototype construction mode, manual annotation and model retraining are not required, which greatly reduces the economic and time costs of flexible manufacturing.

[0076] 3. A multi-scale geometric statistics and adaptive gating noise reduction mechanism is proposed. Based on the local geometric and statistical features of point clouds, outlier noise points caused by splashing and reflection are dynamically suppressed at the physical feature level, thereby improving the robustness of the algorithm in industrial high-noise scenarios.

[0077] 4. An exponentially weighted soft prototype voting strategy is proposed, which combines the inverse class frequency weighting coefficient to solve the problem of extreme imbalance between weld seam and base material point cloud categories, and accurately extracts local weld seam regions with weak geometric features in complex point clouds.

[0078] 5. A dynamic incremental update mechanism is provided. In continuous welding tasks, the feature moving average is performed in real time using the prediction results with high confidence, so as to realize the unsupervised online iteration of the prototype memory. This ensures that the system maintains extremely high recognition accuracy in a long-term maintenance-free state and adapts to dynamic drift of working conditions.

[0079] 6. Based on the geometric and statistical features of point cloud data, the 6D pose of the weld is directly calculated and a welding path executable by the robot is generated. The posture is optimized by combining welding process parameters to achieve a closed loop of the entire process from visual perception to welding execution. Attached Figure Description

[0080] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0081] Figure 2 A schematic diagram of the welding platform design;

[0082] Figure 3 The image shows the visualization of the point cloud scan. (a) and (b) are the original scan data, and (c) and (d) are the scan downsampled and preprocessed data. The point cloud marked in red is the point cloud data of the weld area.

[0083] Figure 4 Schematic diagram of multi-dimensional geometric feature calculation principle

[0084] Figure 5 This is a schematic diagram of 6D pose calculation of the weld and transformation of the robot coordinate system. Detailed Implementation

[0085] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0086] Example 1:

[0087] like Figure 1 As shown, this embodiment provides a welding robot positioning method based on dynamic non-parametric point cloud processing. The robot platform design in this embodiment is as follows: Figure 2 As shown, the method includes the following steps:

[0088] S1: Building the prototype memory during the cold start phase:

[0089] By collecting unlabeled point cloud samples of the target weld and base material, multi-scale geometric statistical features of the samples are extracted, and mean aggregation is performed according to the category to construct an initial prototype memory library, which includes feature prototypes of weld and base material.

[0090] S2: Point cloud geometric statistical feature extraction:

[0091] Real-time acquisition of the original 3D point cloud of the workpiece to be welded; the point cloud data scanning effect and downsampling effect are as follows: Figure 3 As shown; for each downsampled point, a multi-scale local neighborhood is constructed, and geometric statistical features are extracted;

[0092] S3: Based on the extracted geometric statistical features, calculate the feature variance and distance normalization weight of the local point cloud, and assign extremely low fusion weights to outliers with high variance.

[0093] S4: Design a weld seam extraction strategy based on exponential function weighting: calculate the similarity between the denoised point cloud features and various prototypes in the prototype memory, use the exponential weighted soft prototype voting strategy to calculate the classification probability of each point, exponentially amplify the weight of high similarity features, and extract local weld seam point clouds with weak geometric features.

[0094] S5: Unsupervised dynamic incremental updates to the prototype memory:

[0095] In real time, high-confidence prediction results with classification probabilities higher than a preset threshold in step S4 are selected, and the features of these high-quality point clouds are extracted and fused into the existing prototype memory online using a moving average algorithm to achieve unsupervised online iteration.

[0096] S6: Based on the segmented pure weld point cloud, the tangent and normal directions of the weld are directly calculated using the local direction vector in multi-scale geometric statistics. This generates 6D attitude commands with spatial three-dimensional coordinates and yaw, pitch, and roll angles. After hand-eye matrix conversion, these commands are sent to the robotic arm to drive the robot to complete adaptive welding operations.

[0097] The following provides a detailed explanation of steps S1 to S6:

[0098] The specific steps of step S1 are as follows:

[0099] S1-1: Input an unlabeled point cloud sample set of the workpiece to be welded, perform FPS-KNN neighborhood grouping on the sample point cloud, and generate basic location features and multi-scale geometric statistical context enhancement features for each point.

[0100] S1-2: Based on the prior geometry of the groove, the weld area and the base metal area samples are distinguished. For conventional grooves such as V-groove and corner joint groove, the weld area samples are automatically screened by local geometric angle and depth features, and flat areas are automatically divided into base metal area samples.

[0101] S1-3: Perform feature aggregation by category, calculate feature prototypes for weld seam and base material categories, and introduce inverse category frequency weighting coefficients to address the category imbalance problem. The calculation of feature prototypes and inverse category frequency weighting coefficients is shown below:

[0102]

[0103] In the formula: t is the category index, t=1 represents the weld category, and t=2 represents the base material category; This is the feature prototype of class t; Let be the set of sample points of class t; Let be the number of sample points in class t; Contextual enhancement features for sample point i; This is the inverse class frequency weighting coefficient; the smaller the sample size of a class, the higher its basic weight is assigned. C represents the total number of points in the sample set; C represents the total number of categories, which is C=2 in this embodiment.

[0104] S1-4: Store the calculated weld seam and base material feature prototypes as an initial prototype memory library, which can be reused in subsequent reasoning processes without repeated construction.

[0105] like Figure 4 As shown, the extraction of geometric statistical features in step S2 includes:

[0106] S2-1: Basic position encoding: For the initial input point cloud, the K-neighborhood point set of each center point is obtained through the FPS+KNN algorithm. Multi-band sine and cosine functions are used to encode the absolute and relative positions of the input point cloud to generate basic position features.

[0107] In this embodiment, the initial point cloud acquired in real time is input. Where N is the total number of points in the point cloud, p i Let p be the three-dimensional spatial coordinates of the i-th point; for any center point p i The K-neighborhood set of points is obtained through the FPS+KNN algorithm. Corresponding neighborhood feature set ;

[0108] Simultaneously, multi-band sine and cosine functions are used to encode the absolute and relative positions of the input point cloud, generating basic positional features to provide underlying geometric input for subsequent modules. The encoding formula is shown below:

[0109]

[0110] In the formula: As the base value of angular frequency, take The frequency band is 4 in this embodiment;

[0111] S2-2: Multi-scale geometric encoder encoding: based on dual-scale neighborhood , It extracts neighborhood statistical features, distance encoding features, and unit direction vector features to construct multi-dimensional geometric features. The entire process is a deterministic geometric operation with no learnable parameters.

[0112] Neighborhood statistical features are obtained by calculating the mean, standard deviation, and minimum distance of the neighborhood point cloud. The calculation formula is as follows:

[0113]

[0114] In the formula: This represents the number of neighborhood points at the corresponding scale. This is the robustness offset, and the value used in this embodiment is... To avoid numerical anomalies when the standard deviation is 0; The output is the neighborhood statistical features;

[0115] The formulas for calculating distance coding features and unit direction vector features are as follows:

[0116]

[0117] In the formula: For the first The unit direction vector of each neighboring point relative to the center point; Distance-encoded features; The number of frequency bands is 4 in this embodiment.

[0118] S2-3: Feature Fusion: Extend global statistical features to the same dimension as neighborhood features, align neighborhood features at different scales to a unified number of points through linear interpolation, and finally fuse them into multi-scale geometric statistical features with consistent dimensions.

[0119] Step S3 includes:

[0120] S3-1: Two-branch feature construction: Two branches are constructed: residual features and original features. This captures the geometric differences between the neighborhood and the center point while preserving the original structural information of the neighborhood. The calculation formula is as follows:

[0121]

[0122]

[0123] In the formula: Neighborhood-center residual characteristics; Features of the original neighborhood;

[0124] S3-2: Adaptive Gating Coefficient Calculation: Based on neighborhood distance normalization and local feature variance, a gating coefficient is constructed. Higher weights are assigned to spatially close and feature-stable neighboring points, while lower weights are assigned to distant outliers and noisy points with large feature fluctuations. The calculation formula is as follows:

[0125]

[0126] In the formula: For the first Local variance of features of neighborhood points; As preset weighting coefficients, in this embodiment, the values ​​are 0.6 and 0.4 respectively; This is the scaling factor, which is set to 8 in this embodiment; The final adaptive gating coefficients are mapped to the 0-1 interval using the Sigmoid function;

[0127] S3-3: Adaptive Feature Fusion and Refinement: Adaptive fusion of bi-branch features is performed using gating coefficients to obtain refined features of the center point. The calculation formula is as follows:

[0128]

[0129] In the formula: For the first Adaptive fusion features of neighboring points; To refine the features of the center point obtained from the final aggregation; This is a feature splicing operation. Step S3 retains key features of the neighborhood through a combination of max pooling and average pooling. The purified features retain the key structural information of the weld while eliminating isolated noise interference.

[0130] Step S4 includes:

[0131] S4-1: Cosine Similarity Calculation: After normalizing the test point features and the prototypes in the prototype memory, the cosine similarity matrix is ​​calculated using the following formula:

[0132]

[0133] In the formula: For the first The characteristics of the test point and the first test point The cosine similarity of class prototypes, with a value range of [-1, 1]. The higher the value, the higher the confidence that the point belongs to class t.

[0134] S4-2: Similarity Scaling and Exponential Weighting: By using scaling factors and inverse class frequency weighting factors, the contribution of high-confidence weld features is exponentially amplified to solve the class imbalance problem. The calculation formula is as follows:

[0135]

[0136] In the formula: This is the similarity scaling factor, which is set to 10 in this embodiment; These are the inverse class frequency weighting coefficients calculated in step S1;

[0137] S4-3: Soft Voting Class Prediction: Normalization is achieved using the Softmax function to obtain the class prediction probability for each point, thus completing the soft voting decision. The calculation formula is as follows:

[0138]

[0139] In the formula: For the first The point belongs to the th point The predicted probability of the class ranges from [0,1]. The segmentation class of each point is obtained by taking the maximum probability, completing the entire process from feature extraction to class decision, and finally extracting the pure weld point cloud with the predicted class of weld.

[0140] Step S5 includes:

[0141] S5-1: Initial screening of high-confidence samples and first-level point verification: Select points in the weld seam category and base material category whose predicted probability is higher than the preset high-confidence threshold respectively to form the corresponding high-confidence sample set. If the number of points in a single category sample set is lower than the preset minimum sample threshold, then abandon the sample update for this category.

[0142] In this embodiment, for points whose segmentation result is weld seam, the corresponding weld seam category prediction probability P is extracted. weld Filter to satisfy P weld ≥P thre The points constitute a high-confidence weld sample set K. weld For points whose segmentation results belong to the parent material class, extract the corresponding parent material class prediction probability P. base Filter to satisfy P base ≥P thre The points constitute the high-confidence parent material sample set K. base ;

[0143] The sample set after initial screening is checked for the number of points. If the number of points in a single category sample set is lower than the preset minimum sample threshold N, the result is considered valid. min If the sample update for this category is not performed, then the update should be abandoned to avoid interference from small sample noise on the prototype library; where P thre The preset high confidence threshold is set to 0.95 in this embodiment, and N min The value is 100.

[0144] S5-2: Sample anomaly removal and secondary point count verification: For high-confidence sample sets, outliers are removed based on neighborhood feature variance, and only valid samples with stable geometric features are retained. If the number of sample points does not meet the minimum sample threshold after secondary verification, the update process for this category is terminated.

[0145] In this embodiment, for the high-confidence weld sample set, the neighborhood statistical features extracted in step S2 are reused to calculate the neighborhood feature variance of each sample point. Outliers with variances greater than the preset weld variance threshold are removed, and only valid weld samples with stable geometric features are retained. Similarly, for the high-confidence base material sample set, outliers with variances less than the preset base material variance threshold are removed, and misclassified weld edge points are excluded, retaining only valid samples in flat base material areas. The sample set after the second verification is verified again. If the minimum sample size requirement is not met, the update process for this category is terminated.

[0146] S5-3: Moving average prototype update: The moving average algorithm is used to achieve smooth iteration of the prototype, which not only adapts to the feature changes of the current working condition, but also retains the historical feature information of the prototype library, avoids prototype mutation, and ensures the stability of the updated prototype.

[0147] The core calculation formula is:

[0148]

[0149] In the formula: P is the feature prototype of the t-th class (weld, base material) in the prototype memory; α is the moving average coefficient, which is selected as 0.95 in this embodiment. The larger the coefficient, the smoother the prototype update and the stronger the anti-interference ability; K is the effective high-confidence sample set of the t-th class, and N is the number of effective points in the sample set; F i This represents the context enhancement feature corresponding to the i-th valid sample, which is completely consistent with the dimension and type of the feature used in step S1 to build the initial prototype library, ensuring update compatibility;

[0150] S5-4: Update Anomaly Validation and Disk Storage: Calculate the cosine similarity between the prototypes before and after the update. If the similarity is lower than the preset minimum similarity threshold, discard the update result. After the validation passes, overwrite the original prototype library and store it on disk, while retaining historical version backups and update logs.

[0151] In this embodiment, the cosine similarity between the prototypes before and after the update is calculated. If the similarity is lower than the preset minimum similarity threshold (set to 0.9 in this paper), the update is deemed abnormal, the update result is discarded, and the original prototype library is retained. If the verification passes, the updated prototype is used to overwrite the original library and the data is stored on disk. At the same time, the previous version of the prototype library is retained as a backup to support rollback operations under abnormal conditions. During the synchronization process, the number of updated samples, update time, and prototype similarity are recorded to form an update log, ensuring that the entire process is traceable.

[0152] The entire process has no learning parameters, requires no manual annotation intervention, does not interrupt the welding operation process, and does not increase the system deployment and maintenance costs; the dual sample verification mechanism effectively avoids abnormal samples from contaminating the prototype library and ensures the reliability of prototype updates; the moving average update mechanism takes into account both the adaptability of the working conditions and the stability of the prototype, avoids prototype mutations, and adapts to the working condition drift of long-term industrial operations.

[0153] like Figure 5 As shown, step S6 includes:

[0154] S6-1: Input Features and Pre-calibration Parameters: Input the clean weld point cloud output from step S4. ,in The coordinates are in the camera coordinate system (3D spatial coordinates); the set of neighborhood unit direction vectors corresponding to each weld point. Pre-calibrated hand-eye matrix This embodiment adopts an eye-in-hand installation method and uses the homogeneous transformation matrix from the tool coordinate system to the base coordinate system, which is fed back in real time by the robot controller. This step reuses the geometric features extracted by the previous module throughout, eliminating the need for repeated calculations and fully guaranteeing the real-time performance of the system.

[0155] S6-2: Key Point Screening for Weld Path: Denoise and smooth the clean weld point cloud, fit the weld centerline using principal component analysis, and sample key points along the path at equal intervals of 1mm to 5mm to form an ordered set of key points. , The total number of key points;

[0156] S6-3: Calculation of the three-axis vector in the local coordinate system of the weld: for each path key point The three orthogonal normal axis vectors of the local coordinate system are calculated according to the following logic, serving as the core reference for attitude angle calculation:

[0157] The X-axis, or tangent vector, corresponds to the direction of the welding torch's movement.

[0158] Take the key points Three ordered centerline points at the front and three at the back are used to fit the initial tangent vector using the least squares method. ;extract S2 neighborhood unit direction vector set Filter out those that are related to Effective direction vectors with an angle less than 30° are optimized for tangent vectors through mean aggregation, calculated using the following formula:

[0159]

[0160] In the formula: It is a vector normalization function; To optimize the weights, a value of 0.3 is used; This is the set of valid direction vectors after filtering. The number of valid vectors; This is the final normalized X-axis tangent vector;

[0161] The Z-axis, or normal vector, corresponds to the calculation of the welding torch working angle reference:

[0162] by Centered on the target material, the search radius is expanded to a base material region of 10mm-20mm, and the neighborhood unit direction vectors of all points within this region are extracted. Principal component analysis is performed on the expanded direction vector set, and the eigenvector corresponding to the smallest eigenvalue is taken as the initial normal. Based on prior knowledge of the welding scenario, it is ensured that the normal direction points outward from the workpiece, i.e., the angle with the camera optical axis is less than 90°. Finally, Schmidt orthogonalization correction is applied, and the calculation formula is as follows:

[0163]

[0164] In the formula: The Z-axis normal vector after orthogonalization correction;

[0165] The Y-axis, or binormal vector, is calculated by the cross product of the X-axis and Z-axis, ensuring orthogonality of the coordinate systems. The formula is as follows:

[0166]

[0167] The final solution yields the orthogonal three axes of the local coordinate system of the weld: X-axis Y-axis Z-axis The three axes are pairwise orthogonal and the modulus is 1;

[0168] S6-4: 6D pose calculation in camera coordinate system:

[0169] Based on the relationship between the local coordinate system of the weld and the camera coordinate system, the 6D pose of each path key point in the camera coordinate system is calculated. ,in In three-dimensional space coordinates, For yaw angle, For pitch angle, This refers to the roll angle;

[0170] 3D coordinates are directly derived from path key points The coordinates are obtained in the camera coordinate system and no transformation is required.

[0171] Construct the rotation matrix from the local coordinate system to the camera coordinate system The three columns of the rotation matrix correspond to the vectors of the three axes of the local coordinate system in the camera coordinate system, and the calculation formula is as follows:

[0172]

[0173] In the formula: It is a 3×3 orthogonal rotation matrix that satisfies ;

[0174] Euler angle calculation: Following the common fixed coordinate system rotation sequence (ZYX) and yaw-pitch-roll sequence for welding robots, the three attitude angles in the camera coordinate system are calculated using the following formulas:

[0175]

[0176] In the formula: Rotation matrix No. Line number Column elements; The arctangent function is defined in the fourth quadrant, and its range is [range missing]. ; The amplitude limiting function restricts the input value to the interval [-1, 1] to avoid overflow of the arcsin function's domain; finally, the radian value is converted to degrees. ;

[0177] S6-5: Hand-eye matrix transformation to robot base coordinate system:

[0178] Based on the eye-on-hand mounting method, the pose in the camera coordinate system is converted into an executable pose in the robot's base coordinate system through a homogeneous transformation matrix link. The core transformation link is as follows:

[0179]

[0180] In the formula: The homogeneous transformation matrix from the weld local coordinate system to the camera coordinate system is constructed as follows:

[0181]

[0182] in These are the 3D coordinates of the path key points in the camera coordinate system.

[0183] For the pre-calibrated hand-eye matrix, a homogeneous transformation from the camera coordinate system to the tool coordinate system; Homogeneous transformation from the tool coordinate system to the base coordinate system for real-time feedback from the robot controller; The homogeneous transformation matrix from the weld local coordinate system to the robot base coordinate system; from Extract the translation vector Using the three-dimensional coordinates in the base coordinate system, extract the rotation matrix. The three attitude angles in the base coordinate system are obtained by solving according to the ZYX external rotation sequence. Complete the 6D pose calculation in the base coordinate system;

[0184] S6-6: Welding process adaptation and path smoothing optimization:

[0185] Bevel compatibility: V-bevel, which will adjust the roll angle. Adjust to a half-angle bevel, such as a 30° V-bevel. Set to 15°; for fillet welds, set the pitch angle to 15°. Set to 45°;

[0186] Smooth transition: Perform fifth-order polynomial interpolation on the 6D pose of adjacent key points to ensure that the movement of the welding torch from arc initiation to arc termination is free from position and posture changes, and the speed planning is adapted to the welding process requirements.

[0187] S6-7: The final output is a 6D pose sequence that the robot can directly execute:

[0188]

[0189] This sequence can be directly converted into instruction codes that the robot controller can recognize, such as ABB's RAPID and KUKA's KRL. After being issued, the robot can complete adaptive positioning and welding execution, realizing a closed loop of the entire process from visual perception to welding execution.

[0190] This stage achieves a complete closed loop from visual perception to robotic welding execution. This module fully reuses the geometric features extracted in previous steps S2 and S4: neighborhood unit direction vectors and clean weld point clouds, eliminating the need for recalculation and ensuring system real-time performance. Through a standardized coordinate system transformation process, the pose in the camera coordinate system is converted to a robot-executable base coordinate system pose. Combined with welding process parameters, the path and posture are optimized to ensure weld quality, ultimately achieving adaptive positioning and automated welding for the welding robot.

[0191] In summary, the method of this invention achieves adaptive positioning of welding robots with no pre-training, low computational power dependence, and strong noise resistance through multi-scale geometric statistical feature extraction, adaptive gating noise reduction, exponentially weighted soft prototype voting weld segmentation, unsupervised dynamic prototype updating, 6D weld pose calculation and robot path planning. It can realize real-time inference on ordinary industrial control computers, and is plug-and-play for production changes. It is suitable for welding scenarios with strong noise and dynamic drift in industrial environments and can be widely used in automated welding fields such as automobiles and ships.

[0192] Example 2:

[0193] This embodiment provides a welding robot positioning system based on dynamic non-parametric point cloud processing, including:

[0194] The 3D point cloud data acquisition module is used to scan the workpiece to be welded in the industrial site and acquire the original 3D point cloud data, including the weld area and the background of the base material, in real time.

[0195] The initial prototype building module is used in the system cold start phase. Based on a small number of unlabeled point cloud samples of workpieces to be welded or the geometric priors of the bevel, it extracts the geometric statistical features of the weld and the base material and performs mean aggregation to build the initial prototype memory library of weld and base material classes.

[0196] The edge feature calculation module is deployed in a general industrial control computer and communicates with the 3D point cloud data acquisition module and the initial prototype construction module. It is used to receive real-time acquired 3D point cloud data, perform multi-scale feature extraction at the pure geometric and statistical levels, adaptive noise filtering and weld point cloud segmentation, and has no learning parameters throughout the process.

[0197] The dynamic incremental learning module is bidirectionally connected to the edge feature calculation module. It is used to extract high-confidence point cloud features output by the edge feature calculation module during continuous welding operations, perform unsupervised online updates and iterations on the prototype memory, and feed the updated prototype memory back to the edge feature calculation module in real time, forming a dynamic adaptive closed loop.

[0198] The robot execution module communicates with the edge feature calculation module and is used to calculate and generate 6D pose instructions for the welding robot's positioning and trajectory following based on the extracted clean weld point cloud spatial pose, thereby driving the robot to complete adaptive welding operations.

[0199] Example 3:

[0200] To verify the effectiveness of the present invention, comparative experiments and performance data analysis were conducted in this embodiment:

[0201] Data was collected using a RealSense depth camera to construct a self-built dataset. Five typical weld types were selected for data collection: butt joint I-type bevel, butt joint U-type bevel, butt joint V-type bevel, corner joint I-type, and corner joint K-type. The environment simulated actual workshop conditions, with 100 scans for each weld type, totaling 500 samples. Five evaluation metrics were set: mean IoU, weld IoU, precision, recall, and frames per second (FPS) for weld extraction and monitoring experiments. All experiments were repeated five times, and the mean ± standard deviation was used as the final result to reduce the impact of random errors on the experimental conclusions. The experimental results are shown in Table 1.

[0202] Table 1

[0203]

[0204] Weld extraction test results

[0205] Analysis of the experimental results shows that:

[0206] From an overall benchmark perspective, the method of this invention maintains stable and excellent levels in all indicators under five mainstream industrial welding scenarios. The core performance conclusions are as follows:

[0207] Excellent global segmentation: The mIoU is stable at over 85% under all working conditions, reaching a maximum of 89.38%, with an overall average of 87.13%, indicating that the method of the present invention has a stable global segmentation effect on both weld seam and base material targets, without any class bias problem;

[0208] Accurate weld extraction: The Weld IoU model maintains an accuracy of over 79% across all working conditions, reaching a maximum of 84.75%, with an overall average of 81.61%. This accurately achieves effective segmentation of the weld area and solves the problem of extreme imbalance between the weld and base material point cloud categories.

[0209] High adaptability to industrial scenarios: The recall rate is consistently above 97% under all working conditions, with an overall average of 98.35%. The high recall rate of nearly 100% means a low missed detection rate, which can avoid welding defects such as weld deviation and incomplete penetration caused by missed weld detection.

[0210] Excellent noise resistance and purity performance: The accuracy rate remains above 80% under all working conditions, with an overall average of 82.73%, indicating that the proportion of real weld points in the segmentation results is high, and there are few false detections, over-segmentation spatter, and reflective noise points, which verifies that it has an excellent suppression effect on the noise environment of industrial sites.

[0211] High real-time performance: The FPS remained stable at 73 frames per second under all operating conditions, verifying the core advantage of this invention that the non-parametric architecture can achieve real-time inference on ordinary industrial control computers without relying on high-end GPUs.

Claims

1. A welding robot positioning method based on dynamic non-parametric point cloud processing, characterized in that, Includes the following steps: S1: By collecting unlabeled point cloud samples of the target weld and base material, extracting multi-scale geometric statistical features of the samples, aggregating the mean by category, and constructing an initial prototype memory library, which includes feature prototypes of weld and base material. S2: Real-time acquisition of the original 3D point cloud of the workpiece to be welded, construction of a multi-scale local neighborhood for each downsampled point, and extraction of geometric and statistical features; S3: Based on the extracted geometric statistical features, calculate the feature variance and distance normalization weight of the local point cloud, and assign fusion weights to outliers with high variance. S4: Calculate the similarity between the denoised point cloud features and various prototypes in the prototype memory. Use an exponentially weighted soft prototype voting strategy to calculate the classification probability of each point, exponentially amplify the weight of high similarity features, and extract the local weld point cloud with weak geometric features. S5: In real time, high-confidence prediction results with classification probabilities higher than the preset threshold in step S4 are selected, and the features of these high-quality point clouds are extracted and fused into the existing prototype memory online using the moving average algorithm to achieve unsupervised online iteration. S6: Based on the segmented pure weld point cloud, the tangent and normal directions of the weld are directly calculated using the local direction vector in multi-scale geometric statistics. This generates 6D attitude commands with spatial three-dimensional coordinates and yaw, pitch, and roll angles. After hand-eye matrix conversion, these commands are sent to the robotic arm to drive the robot to complete adaptive welding operations.

2. The welding robot positioning method based on dynamic non-parametric point cloud processing according to claim 1, characterized in that, The method for aggregation of mean values ​​by category in step S1 includes: The feature prototypes for weld seam and base material classes are calculated, while the inverse class frequency weighting coefficient is introduced to address the class imbalance problem. The calculation of the feature prototype and the inverse class frequency weighting coefficient is shown below: ； In the formula: t is the category index, t=1 represents the weld category, and t=2 represents the base material category; This is the feature prototype of class t; Let be the set of sample points of class t; Let be the number of sample points in class t; Contextual enhancement features for sample point i; This is the inverse class frequency weighting coefficient; the smaller the sample size of a class, the higher its basic weight is assigned. C represents the total number of points in the sample set; C represents the total number of categories.

3. The welding robot positioning method based on dynamic non-parametric point cloud processing according to claim 2, characterized in that, The extraction of geometric statistical features in step S2 includes: S2-1: Basic position encoding: For the initial input point cloud, the K-neighborhood point set of each center point is obtained through the FPS+KNN algorithm. Multi-band sine and cosine functions are used to encode the absolute and relative positions of the input point cloud to generate basic position features. S2-2: Multi-scale geometric encoder encoding: based on dual-scale neighborhood , Extract neighborhood statistical features, distance coding features, and unit direction vector features to construct multi-dimensional geometric features; S2-3: Feature Fusion: Extend global statistical features to the same dimension as neighborhood features, align neighborhood features at different scales to a unified number of points through linear interpolation, and finally fuse them into multi-scale geometric statistical features with consistent dimensions.

4. The welding robot positioning method based on dynamic non-parametric point cloud processing according to claim 3, characterized in that, In step S2-2, the neighborhood statistical features are obtained by calculating the mean, standard deviation, and minimum distance of the neighborhood point cloud. The calculation formula is as follows: ； In the formula: This represents the number of neighborhood points at the corresponding scale. This is the robust offset. The output is the neighborhood statistical features; The formulas for calculating distance coding features and unit direction vector features are as follows: ； In the formula: For the first The unit direction vector of each neighboring point relative to the center point; Distance-encoded features; This represents the number of frequency bands.

5. The welding robot positioning method based on dynamic non-parametric point cloud processing according to claim 4, characterized in that, Step S3 includes: S3-1: Two-branch feature construction: Two branches are constructed: residual features and original features. This captures the geometric differences between the neighborhood and the center point while preserving the original structural information of the neighborhood. The calculation formula is as follows: ； ； In the formula: Neighborhood-center residual characteristics; Features of the original neighborhood; S3-2: Adaptive Gating Coefficient Calculation: Based on neighborhood distance normalization and local feature variance, a gating coefficient is constructed. Higher weights are assigned to spatially close and feature-stable neighboring points, while lower weights are assigned to distant outliers and noisy points with large feature fluctuations. The calculation formula is as follows: ； In the formula: For the first Local variance of features of a neighborhood point; Preset weighting coefficients; This is the scaling factor; For the final adaptive gating coefficient; S3-3: Adaptive Feature Fusion and Refinement: Adaptive fusion of bi-branch features is performed using gating coefficients to obtain refined features of the center point. The calculation formula is as follows: ； In the formula: For the first Adaptive fusion features of neighboring points; To refine the features of the center point obtained from the final aggregation; This is a feature splicing operation.

6. The welding robot positioning method based on dynamic non-parametric point cloud processing according to claim 5, characterized in that, Step S4 includes: S4-1: Cosine Similarity Calculation: After normalizing the test point features and the prototypes in the prototype memory, the cosine similarity matrix is ​​calculated using the following formula: ； In the formula: For the first The characteristics of the test point and the first test point Cosine similarity between class prototypes; S4-2: Similarity Scaling and Exponential Weighting: By using a scaling factor and an inverse class frequency weighting factor, the contribution of high-confidence weld features is exponentially amplified. The calculation formula is as follows: ； In the formula: This is the similarity scaling factor; S4-3: Soft Voting Class Prediction: Normalization is achieved using the Softmax function to obtain the class prediction probability for each point, thus completing the soft voting decision. The calculation formula is as follows: ； In the formula: For the first The point belongs to the th point The predicted probability of a class is obtained by taking the maximum probability to determine the segmentation category of each point.

7. The welding robot positioning method based on dynamic non-parametric point cloud processing according to claim 6, characterized in that, Step S5 includes: S5-1: Initial screening of high-confidence samples and first-level point verification: Select points in the weld seam category and base material category whose predicted probability is higher than the preset high-confidence threshold respectively to form the corresponding high-confidence sample set. If the number of points in a single category sample set is lower than the preset minimum sample threshold, then abandon the sample update for this category. S5-2: Sample anomaly removal and secondary point count verification: For high-confidence sample sets, outliers are removed based on neighborhood feature variance, and only valid samples with stable geometric features are retained. If the number of sample points does not meet the minimum sample threshold after secondary verification, the update process for this category is terminated. S5-3: Moving Average Prototype Update: This method uses a moving average algorithm to achieve smooth iteration of the prototype. The calculation formula is as follows: ； In the formula: For the updated number Class feature prototype; The original feature prototype before the update; The moving average coefficient; This represents the number of points in the valid high-confidence sample set for this category. S5-4: Update Anomaly Validation and Disk Storage: Calculate the cosine similarity between the prototype before and after the update. If the similarity is lower than the preset minimum similarity threshold, discard the update result. After the validation is passed, overwrite the original prototype library and store it on disk, while retaining historical version backups and update logs.

8. The welding robot positioning method based on dynamic non-parametric point cloud processing according to claim 7, characterized in that, Step S6 includes: S6-1: Key point screening of weld path: Denoise and smooth the pure weld point cloud, fit the weld centerline, sample key points of the path, and form an ordered set of key points; S6-2: Calculation of the three-axis vector of the local coordinate system of the weld: For each key point of the path, calculate the X-axis tangent vector and Z-axis normal vector respectively, obtain the Y-axis subnormal vector through cross product, and obtain the orthogonal local coordinate system three-axis after Schmidt orthogonalization correction; S6-3: 6D Pose Calculation in Camera Coordinate System: Construct the rotation matrix from the local coordinate system to the camera coordinate system. Following the common fixed coordinate system external rotation and ZYX rotation sequence for welding robots, calculate the 3D coordinates, yaw, pitch, and roll angles in the camera coordinate system. The calculation formulas are as follows: ； In the formula: This is the rotation matrix from the local coordinate system to the camera coordinate system; These are the X, Y, and Z axis vectors of the local coordinate system, respectively. For the rotation matrix, the first Line number Column elements; S6-4: Hand-Eye Matrix Transformation to Robot Base Coordinate System: Based on the eye-on-hand mounting method, a homogeneous transformation matrix link is used to convert the pose in the camera coordinate system into an executable pose in the robot base coordinate system. The core transformation link is as follows: ； In the formula: This is the homogeneous transformation matrix from the local coordinate system of the weld to the robot's base coordinate system; Homogeneous transformation from the tool coordinate system to the base coordinate system for real-time feedback from the robot controller; For the pre-calibrated hand-eye matrix; This is the homogeneous transformation matrix from the local coordinate system of the weld to the camera coordinate system; S6-5: Welding process adaptation and path smoothing optimization: The attitude angle is adapted and adjusted according to the bevel type, and the 6D pose of adjacent key points is interpolated by a fifth-order polynomial to finally output a 6D pose sequence that the robot can directly execute.

9. A welding robot positioning system based on dynamic non-parametric point cloud processing, characterized in that, For implementing the method of claim 1, the system comprises: The 3D point cloud data acquisition module is used to scan the workpiece to be welded in the industrial site and acquire the original 3D point cloud data, including the weld area and the background of the base material, in real time. The initial prototype building module is used in the system cold start phase. Based on a small number of unlabeled point cloud samples of workpieces to be welded or the geometric priors of the bevel, it extracts the geometric statistical features of the weld and the base material and performs mean aggregation to build the initial prototype memory library of weld and base material classes. The edge feature calculation module is deployed in a general industrial control computer and communicates with the 3D point cloud data acquisition module and the initial prototype construction module. It is used to receive real-time acquired 3D point cloud data, perform multi-scale feature extraction at the pure geometric and statistical levels, adaptive noise filtering and weld point cloud segmentation, and has no learning parameters throughout the process. The dynamic incremental learning module is bidirectionally connected to the edge feature calculation module. It is used to extract high-confidence point cloud features output by the edge feature calculation module during continuous welding operations, perform unsupervised online updates and iterations on the prototype memory, and feed the updated prototype memory back to the edge feature calculation module in real time, forming a dynamic adaptive closed loop. The robot execution module communicates with the edge feature calculation module and is used to calculate and generate 6D pose instructions for the welding robot's positioning and trajectory following based on the extracted clean weld point cloud spatial pose, thereby driving the robot to complete adaptive welding operations.

Images