Groove welding seam treatment method
By combining deep learning algorithms and point cloud preprocessing with the RANSAC algorithm, the problem of insufficient accuracy in groove welding by traditional methods is solved, and high-precision welding is achieved, especially in centrifuge groove welding, where the accuracy rate reaches 99%.
Patent Information
- Application Number
- CN202511916130.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-18
- Publication Date
- 2026-03-17
AI Technical Summary
When welding grooved parts with complex shapes, the accuracy of traditional geometric feature segmentation methods decreases, making it difficult to adapt to complex and ever-changing application scenarios. Furthermore, traditional algorithms are sensitive to noise and changes in point cloud density.
A point cloud segmentation method based on deep learning algorithms is adopted, which combines point cloud preprocessing and contour extraction algorithms. The point cloud data is segmented by deep learning algorithms, the weld trajectory is fitted by RANSAC algorithm, and GPU hardware acceleration processing is combined.
It achieves high-precision welding in complex scenarios, especially in centrifuge groove welding, achieving an accuracy rate of 99% and meeting the accuracy requirement of ±0.3mm, thus providing high-precision welding quality control.
Smart Images
Figure CN121685509A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of machine recognition of weld seams, specifically a method for processing grooved weld seams. Background Technology
[0002] With the rise of industrial welding robots, using robots to replace manual welding offers unparalleled consistency and reliability in terms of both welding efficiency and weld formation. When robotic arms weld non-standard circular grooved parts, the complex shape of the grooves, the large amount of point cloud data, and the presence of noise necessitate high precision in the welding trajectory.
[0003] Existing technologies for handling such problems largely rely on traditional geometric feature segmentation and trajectory extraction. These methods typically involve the following steps: calculating the average distance and number of neighboring points using a fixed-radius search to remove outliers and noise points for point cloud filtering; next, analyzing the geometric features of the point cloud, such as normal vectors and curvature, and segmenting feature regions like grooves by setting normal vector angles or curvature thresholds; finally, using the RANSAC algorithm or least squares method to geometrically fit the segmented point cloud, extracting key contours and generating smooth welding trajectories to ensure welding accuracy and stability. However, traditional algorithms are sensitive to noise and changes in point cloud density, leading to a significant decrease in segmentation accuracy in this scenario. Furthermore, the generalization ability of traditional algorithms is limited by the designed geometric features, requiring new feature extraction methods for entirely new scenarios, making them unsuitable for complex and ever-changing applications. Summary of the Invention
[0004] To address the above issues, this application proposes a method for processing grooved welds, based on point cloud segmentation using deep learning algorithms. This method demonstrates better performance and stronger generalization ability in complex scenarios. The time consumption issue of the algorithm can also be mitigated by using hardware devices such as GPUs.
[0005] A method for treating grooved welds, characterized by comprising the following steps: S1: Point cloud data acquisition: Point clouds are acquired using a line structured light camera, converted into three-dimensional coordinates of the actual position, and then merged to obtain a complete point cloud. S2: Point cloud data preprocessing: Random downsampling and clustering are performed on the raw point cloud data to remove discrete points. Based on the center point of the point cloud bounding box, the point cloud is rotated to unify the point cloud orientation. S3: Deep point cloud segmentation: The preprocessed point cloud data is segmented using deep learning algorithms into three categories: the left and right sides of the groove and the bottom of the groove; S4: Contour extraction and fitting: The key contour is obtained by using the edge contour extraction algorithm on the left and right point clouds. After the key contour is filtered by moving average, it is segmented. The segmented point clouds are then fitted with circular arcs to obtain the left and right boundaries of the weld. S5: Weld position and size calculation: The left and right boundaries of the weld are uniformly sampled and sorted, and the midpoint of the same index point is used to construct the weld trajectory; the vertical distance from the midpoint to the boundary is used as the width, and after fitting the bottom plane of the groove by the RANSAC algorithm, the distance from the midpoint to the plane is used as the height.
[0006] The deep learning algorithm in step S3 includes: Deep learning algorithms consist of three parts: downsampling, upsampling, and fully connected layers. Downsampling is composed of three sets of ensemble abstraction layers, and upsampling is composed of three sets of feature fusion modules. The ensemble abstraction layers consist of a sampling module, a grouping module, and a feature extraction module.
[0007] The edge contour extraction algorithm in step S4 includes: Iterate through each point and find the point in the entire point cloud whose y coordinate is in the range [yn, y+n]. Within this strip region, the point with the smallest x value is the left contour point, and the point with the largest x value is the right contour point. Complete the projection of all points onto the corresponding side contour to obtain the key contour point set.
[0008] The fitting step for the key contour in step S4 is as follows: After filtering the key contour point cloud by moving average, the points are segmented according to their coordinates. The RANSAC algorithm is used to fit the arcs of each segment. The fitted arcs are then smoothly spliced together to obtain the left and right boundaries of the weld.
[0009] The mathematical notation for the edge contour extraction algorithm is as follows: Let the point cloud be a set P containing n points, each point pi having coordinates (xi, yi, zi). For the threshold; The first step is to define the set of all y-values appearing in the point cloud: The second step is to define the neighborhood set of each y value (i.e., all points whose y coordinates are within the range of ±δ): The third step is to find the points with the smallest or largest x-coordinates within the neighborhood; this set of points is the key contour.
[0010] The acquisition of the weld trajectory in step S5 includes: The left and right boundaries of the weld are sampled and sorted evenly. The left and right points of the same index point are added together to obtain the midpoint between the left and right points. All the midpoints are spliced together to obtain a new arc, which is the weld trajectory.
[0011] Beneficial effects: This invention proposes a method for processing grooved welds, based on point cloud segmentation using deep learning algorithms. This method demonstrates better performance and stronger generalization ability in complex scenarios. The algorithm's time consumption can be mitigated by using GPU hardware. The model tested according to this invention meets accuracy requirements. Particularly in a centrifuge groove welding project, the measurement of point cloud trajectory, width, and depth achieved an accuracy rate of no less than 99% within a ±0.3mm precision requirement. This fully demonstrates the precise and reliable technical advantages of this invention in high-precision welding inspection scenarios, providing strong support for welding quality control. Attached Figure Description
[0012] Figure 1 This is a flowchart of the groove weld treatment method; Figure 2 This is a photograph of the actual weld seam of the grooved component; Figure 3 This is a diagram of the network structure of a deep learning algorithm; Figure 4 This is a diagram showing the result of the groove weld treatment. Detailed Implementation
[0013] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described in this application are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0014] This invention proposes a method for treating grooved welds. The following is an example... Figure 2 Taking a centrifuge grooved workpiece as an example, the implementation steps of this method are explained in detail.
[0015] S1: Point cloud data acquisition: A point cloud is acquired using a line structured light camera. The camera emits a laser to scan the weldment, which will acquire a frame of image. Then, based on a preset threshold (used to distinguish effective feature points from background noise points and invalid reflection points), feature point filtering processing is performed on the image to extract effective feature points that meet the threshold requirements. These points are then converted into the three-dimensional coordinates of their corresponding actual positions and finally merged to obtain a complete point cloud.
[0016] S2: Point cloud data preprocessing: First, the original point cloud data is randomly downsampled to effectively reduce the amount of point cloud data and lower the computational complexity of subsequent processing. Then, a clustering algorithm is used to filter noise in the downsampled point cloud data, removing discretely distributed noise points and optimizing the point cloud data quality. Finally, the bounding box corresponding to the point cloud data is calculated and the center point of the bounding box is determined. The spatial orientation of the point cloud data is adjusted based on the center point of the bounding box to achieve uniform standardization of the spatial orientation of the point cloud.
[0017] S3: Deep point cloud segmentation: The preprocessed point cloud data was segmented using a deep learning algorithm into three categories: the left and right sides of the groove and the bottom of the groove.
[0018] Deep learning algorithms (see network structure diagram) Figure 3 The algorithm comprises three parts: downsampling, upsampling, and fully connected layers. Downsampling consists of three sets of ensemble abstraction layers, and upsampling consists of three sets of feature fusion modules. The ensemble abstraction layers consist of a sampling module, a grouping module, and a feature extraction module. The ensemble abstraction layers utilize the existing PointNet++ algorithm.
[0019] In the upsampling process, the inverse distance weighted interpolation algorithm was used to propagate features from a subset of points to the entire subset of points. The formula for the inverse distance weighted interpolation algorithm is shown below: It calculates the weight, which is inversely proportional to the distance; the closer the distance, the greater the influence. For the eigenvalues of the unknown points, These are the characteristic values of a known point. Indicates taking from the set of known points Interpolation calculations are performed at each point. This indicates the degree to which distance affects the weight.
[0020] The inverse distance weighting method can dynamically adjust weights based on distance, better preserve local features when point cloud density is uneven, and can also fuse features of different scales to improve the model's adaptability to complex scenes.
[0021] The deep learning algorithm processing procedure is as follows: A point cloud data augmentation method is proposed, which involves randomly shifting and rotating the input point cloud data, adding noise, and scaling the data up and down based on the centroid. Hundreds of point cloud datasets were obtained, with 80% used as the training set and 20% as the validation set.
[0022] The enhanced point cloud data is input into the deep learning algorithm (e.g.) Figure 3As shown in the figure, 16 represents the batch size, 2048 represents the number of point clouds, and 6 represents the point cloud coordinates x, y, z and the corresponding normal features. The point cloud data is processed through three ensemble abstraction layers, which means that multiple sampling modules and grouping modules are used to extract features. Each ensemble abstraction layer samples different sampling point data and sampling point radii. As the number of layers increases, the number of samples and the radius also increase.
[0023] After feature extraction, the hierarchical features are propagated using the inverse distance interpolation algorithm to complete upsampling and feature fusion. Finally, through a fully connected layer and a sigmoid activation function, the feature vector is transformed into a 16*2048*num_classes vector, where num_classes is the number of classes in the point cloud segmentation. This completes the point cloud segmentation. In this embodiment, the value of num_classes is 3, representing the left and right sides of the groove and the bottom of the groove, respectively.
[0024] S4: Contour extraction and fitting: An edge contour extraction algorithm is used for the point clouds on the left and right sides. Each point is traversed, and points with y coordinates in the range [yn, y+n] are found in the entire point cloud. Within this strip-shaped region, the point with the smallest x value is the left contour point, and the point with the largest x value is the right contour point. The projection of all points onto the corresponding side contours is completed to obtain the key contours.
[0025] The proposed edge contour extraction algorithm differs from ordinary edge detection algorithms, which can only extract the contours of the entire point cloud and require further segmentation in subsequent processing. This algorithm, combined with prior information about the point cloud shape, can extract only the key contours.
[0026] The mathematical notation for the edge contour extraction algorithm is as follows: Let the point cloud be a set P containing n points, each point pi having coordinates (xi, yi, zi). For the threshold; The first step is to define the set of all y-values appearing in the point cloud: The second step is to define the neighborhood set of each y value (i.e., all points whose y coordinates are within the range of ±δ): The third step is to find the points with the smallest or largest x-coordinates within the neighborhood; this set of points is the key contour.
[0027] For the extracted left and right key contours, a moving average filter is used to eliminate noise interference. Since the key contours are not standard circular arc structures, direct overall fitting is prone to error accumulation. Therefore, the point cloud coordinates are divided into segments, and a segmented fitting strategy is adopted to improve the fitting accuracy. Then, the RANSAC algorithm is used to fit the circular arcs of each segment of the point cloud. This algorithm effectively removes abnormal data in the point cloud and finally accurately obtains the left and right boundaries of the weld.
[0028] S5: Calculation of weld location and dimensions: First, uniformly sample the left and right boundaries of the weld and sort them according to a preset spatial rule. Take the sampling points of the left and right boundaries corresponding to the same index and calculate the midpoint of the coordinates of the two points. Then, connect all the midpoints of the coordinates in sequence according to the sampling order to form an arc. This arc is the trajectory of the target weld (see...). Figure 4 Subsequently, for each coordinate midpoint, the vertical distance from it to the left and right boundary arcs is calculated, and the sum of the two sets of vertical distances is determined as the weld width at the corresponding position. Finally, the RANSAC algorithm is used to perform plane fitting on the groove bottom point cloud data output by the deep learning algorithm to obtain the groove bottom reference plane. By calculating the vertical distance from each coordinate midpoint to this reference plane, the weld height at the corresponding position is determined.
[0029] The output of this example is as follows:
[0030] In a centrifuge groove welding project, this invention outputs point cloud trajectories, widths, and depths with an accuracy of 99% within ±0.3 mm.
[0031] The model's test results met the accuracy requirements. After deployment on industrial PCs and other devices, the model needs optimization operations such as trimming and quantization to meet the real-time requirements of industrial applications. Due to the different hardware configurations of various industrial PCs, this paper uses two deployment methods based on the Tensorrt and OpenVino frameworks.
Claims
1. A method for treating grooved welds, characterized in that, Includes the following steps: S1: Point cloud data acquisition: Point clouds are acquired using a line structured light camera, converted into three-dimensional coordinates of the actual position, and then merged to obtain a complete point cloud. S2: Point cloud data preprocessing: Random downsampling and clustering are performed on the raw point cloud data to remove discrete points. Based on the center point of the point cloud bounding box, the point cloud is rotated to unify the point cloud orientation. S3: Deep point cloud segmentation: The preprocessed point cloud data is segmented using deep learning algorithms into three categories: the left and right sides of the groove and the bottom of the groove; S4: Contour extraction and fitting: The key contour is obtained by using the edge contour extraction algorithm on the left and right point clouds. After the key contour is filtered by moving average, it is segmented. The segmented point clouds are then fitted with circular arcs to obtain the left and right boundaries of the weld. S5: Weld position and size calculation: The left and right boundaries of the weld are uniformly sampled and sorted, and the midpoint of the same index point is used to construct the weld trajectory; the vertical distance from the midpoint to the boundary is used as the width, and after fitting the bottom plane of the groove by the RANSAC algorithm, the distance from the midpoint to the plane is used as the height.
2. The method for treating grooved welds as described in claim 1, characterized in that, The deep learning algorithm in step S3 includes: Deep learning algorithms consist of three parts: downsampling, upsampling, and fully connected layers. Downsampling is composed of three sets of ensemble abstraction layers, and upsampling is composed of three sets of feature fusion modules. The ensemble abstraction layers consist of a sampling module, a grouping module, and a feature extraction module.
3. The method for treating grooved welds as described in claim 1, characterized in that, The edge contour extraction algorithm in step S4 includes: Iterate through each point and find the point in the entire point cloud whose y coordinate is in the range of [yn, y+n]. Within this strip region, the point with the smallest x value is the left contour point, and the point with the largest x value is the right contour point. Complete the projection of all points onto the corresponding side contours to obtain the key contours.
4. The method for treating grooved welds as described in claim 1, characterized in that, The fitting step for the key contour in step S4 is as follows: After filtering the key contour point cloud by moving average, the points are segmented according to their coordinates. The RANSAC algorithm is used to fit the arcs of each segment. The fitted arcs are then smoothly spliced together to obtain the left and right boundaries of the weld.
5. The method for treating grooved welds as described in claim 3, characterized in that, The mathematical notation for the edge contour extraction algorithm is as follows: Let the point cloud be a set P containing n points, each point pi having coordinates (xi, yi, zi). For the threshold; The first step is to define the set of all y values that appear in the point cloud; The second step is to define the neighborhood point set (i.e., all points whose y coordinates are within ±δ) for each y value. The third step is to find the points with the smallest or largest x-coordinates in the neighborhood; this set of points is the key contour.
6. The method for treating grooved welds as described in claim 1, characterized in that, The acquisition of the weld trajectory in step S5 includes: The left and right boundaries of the weld are sampled and sorted evenly. The left and right points of the same index point are added together to obtain the midpoint between the left and right points. All the midpoints are spliced together to obtain a new arc, which is the weld trajectory.