A welding seam point cloud recognition method based on improved gradient boosting tree

Abstract

The present application relates to the field of three-dimensional point cloud processing and machine learning, especially to a welding seam point cloud recognition method based on improved gradient boosting tree, which comprises the following steps: collecting three-dimensional point cloud data of welding seams with different shapes, pre-processing and extracting ROI region, calculating geometric features under a preset neighborhood scale to construct training set and test set, taking LightGBM as a base classifier, optimizing key hyperparameters by using improved glide snake optimization algorithm, obtaining LFMGSO-LightGBM model and verifying it, and realizing accurate recognition and classification of input welding seam point cloud. The welding seam point cloud recognition method based on improved gradient boosting tree provided by the present application improves the population initialization, search step and guide mechanism of the improved glide snake optimization algorithm, improves the initial solution space distribution quality, realizes the dynamic balance of global exploration and local development, and enhances the stability and anti-local optimal ability of the optimization process.

Description

Technical Field

[0001] This invention relates to the fields of 3D point cloud processing and machine learning technology, and in particular to a method for weld seam point cloud recognition based on an improved gradient boosting tree. Background Technology

[0002] With the rapid development of high-end manufacturing fields such as aerospace, automobile manufacturing, and shipbuilding, welding quality inspection and automatic weld identification have become key links in intelligent manufacturing and robotic welding. Due to the irregular shape, significant dimensional fluctuations, and local noise interference of weld surfaces, traditional inspection methods relying on human experience are difficult to meet the industrial demands for automation, precision, and high reliability. Therefore, high-precision automatic identification methods for welds have significant engineering application value.

[0003] Existing weld seam recognition methods mainly include those based on two-dimensional visual images and those based on three-dimensional point clouds. Two-dimensional image-based methods typically employ threshold segmentation, edge detection, or deep learning object detection networks to extract the weld seam region, but these methods are easily affected by changes in illumination, glare, occlusion, and viewing angle, making it difficult to accurately represent the true three-dimensional geometry of the weld seam. Three-dimensional point cloud methods based on lasers or structured light can directly acquire the spatial geometric information of the weld seam surface, making them more promising for applications in complex welding scenarios.

[0004] However, existing methods for 3D point cloud weld identification still have significant shortcomings. On the one hand, traditional point cloud processing methods mostly rely on threshold segmentation, geometric fitting, or manually set rules. For example, while methods based on RANSAC or surface fitting can extract welds in some regular scenarios, they are prone to undersegmentation, oversegmentation, and insufficient robustness when the weld shape is complex, the background noise is strong, or the point cloud distribution is uneven. On the other hand, weld point clouds have obvious local geometric differences. Relying solely on a single height feature or simple curvature feature often fails to fully characterize the subtle structural differences between the weld area and the base material area.

[0005] With the development of machine learning methods, point cloud classification methods based on 3D geometric features and ensemble learning models have gradually attracted attention. LightGBM, in particular, as an efficient gradient boosting tree model, boasts advantages such as fast training speed, good classification performance, and suitability for handling multi-dimensional feature inputs, showing promising application prospects in point cloud classification tasks. However, the classification performance of LightGBM largely depends on hyperparameter settings. Inappropriate configuration of key parameters such as learning rate, tree depth, number of leaf nodes, and number of base learners can easily lead to decreased model accuracy or insufficient generalization ability. Existing commonly used parameter search methods, such as empirical hyperparameter tuning, grid search, or random search, are not only computationally expensive but also prone to getting trapped in local optima in complex parameter spaces, making it difficult to balance search efficiency with the quality of the found data. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] This invention provides a weld seam point cloud recognition method based on an improved gradient boosting tree, which overcomes the problems of insufficient representation of local geometric structure of 3D point clouds, low recognition accuracy, low efficiency of key hyperparameter optimization of LightGBM model and easy getting trapped in local optima in the existing technology.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, this invention provides a weld seam point cloud recognition method based on an improved gradient boosting tree, comprising the following steps:

[0010] Step S1: Collect three-dimensional point cloud data of welds of different shapes using a surface structured light camera, and preprocess the three-dimensional point cloud data to obtain the ROI region.

[0011] Step S2: Calculate the geometric features of the ROI region at a preset neighborhood scale, and construct a training set and a test set based on the geometric features of the ROI region;

[0012] Step S3: Using the LightGBM model as the base classifier, the key hyperparameters of the LightGBM model are optimized using the improved gliding snake optimization algorithm to obtain the optimized LFMGSO-LightGBM model.

[0013] The LFMGSO-LightGBM model was trained using the training set and validated using the test set.

[0014] Step S4: Use the LFMGSO-LightGBM model trained in step S3 to identify and classify the input weld point cloud.

[0015] Preferably, step S3, which involves optimizing the key hyperparameters of the LightGBM model using an improved gliding snake optimization algorithm, includes:

[0016] Step S31: Generate a chaotic sequence using the Logistic chaotic mapping mathematical model to initialize the algorithm's population;

[0017] Step S32: The individual search step size is dynamically adjusted according to the iteration progress, population diversity and stagnation state through a feedback adaptive step size mathematical model.

[0018] Step S33: Introduce a mathematical model of multi-leadership collaboration mechanism, using the top K best individuals to jointly guide the population search direction.

[0019] Preferably, the Logistic chaotic mapping mathematical model in step S31 is:

[0020]

[0021] in, The value at the nth iteration is in the range [0, 1]. For the control parameters of the mapping, In the first The value after the next iteration;

[0022] Based on the chaotic sequence, the position of each individual in the population is initialized using the following formula:

[0023]

[0024] in, For the first The individual in the first The initial position in the 3D search space. and They represent the first The lower and upper bounds of the dimensional search space. To map to the The individual Chaotic variables of dimension.

[0025] Preferably, the mathematical model formula for the feedback adaptive step size in step S32 is:

[0026]

[0027] in, Indicates the first The individual in the first Step size in the next iteration Indicates the current iteration number. Indicates the maximum number of iterations. , as well as These are non-negative weighting coefficients, used to control the influence of the time progress term, population diversity term, and stagnation term on the step size, respectively. The time decay exponent, For the first Population diversity index at the next iteration For the first Stagnation indicators during iteration.

[0028] Preferably, the mathematical model formula for the multi-leadership collaboration mechanism in step S33 is:

[0029]

[0030] in, For the first The individual in the first The position vector at the next iteration. In the first The new position in the next iteration For the first The individual in the first Step size in the next iteration and These are the weighting coefficients for the multi-leadership guidance term and the neighbor collaboration term, respectively. The preset location of the neighboring individuals;

[0031] The For multi-leadership collaboration center, ,in For the first The position vector of each leader The number of elite individuals participating in collaborative guidance; For the first The weight of each leader.

[0032] (III) Beneficial Effects

[0033] This invention provides a weld point cloud recognition method based on an improved gradient boosting tree. By employing an improved gliding snake optimization algorithm, it automatically and efficiently optimizes the key hyperparameters of the LightGBM model, overcoming the shortcomings of traditional empirical parameter tuning, grid, or random search methods, which suffer from high computational costs and susceptibility to local optima. This optimization process finds a better combination of hyperparameters for the classification model, thereby directly improving the accuracy of the final model in the binary classification task of weld and base material point clouds.

[0034] This improved gliding snake optimization algorithm significantly enhances optimization performance by comprehensively improving population initialization, search step size strategy, and guidance mechanism. Specifically, it enhances the algorithm's early global exploration capability by improving the distribution quality of the initial population in the parameter space; it better balances global exploration and local development by enabling the search process to be dynamically adjusted according to the iteration progress and population state; and it reduces the dependence on a single optimal solution by utilizing elite information in the population to collaboratively guide the search direction, thereby improving the stability, efficiency, and resistance to local optimum traps in complex parameter spaces. Attached Figure Description

[0035] Figure 1 The flowchart illustrates the process of a weld point cloud recognition method based on an improved gradient boosting tree according to the present invention.

[0036] Figure 2 The flowchart of the LFMGSO-LightGBM algorithm of the present invention is shown. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0038] In the description of this invention, it is necessary to understand that the orientations or positional relationships indicated by terms such as "upper," "lower," "left," "right," "inner," "outer," "top," and "bottom" are based on the orientations or positional relationships shown in the accompanying drawings. They are intended only to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the components referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0039] like Figure 1-2 As shown, this invention provides a weld point cloud recognition method based on an improved gradient boosting tree, comprising the following steps:

[0040] Step S1: Collect three-dimensional point cloud data of welds of different shapes using a surface structured light camera, and preprocess the three-dimensional point cloud data to obtain the ROI region.

[0041] Step S1 involves preprocessing the 3D point cloud data, including:

[0042] Step S11: Remove the background of the parent material from the 3D point cloud data to obtain the initial ROI region;

[0043] It is important to note that in actual work, by using an AUBO-E5 robot equipped with an S028120 surface structured light camera, three different types of real weld 3D point cloud data of straight welds, arc welds, and S-shaped welds are collected respectively. By setting a spatial clipping threshold, based on the prior position information of the weld area, a large area of ​​the base material background point cloud is quickly removed, and the weld and its adjacent areas are retained to obtain the initial ROI area. This clipping process can effectively reduce the amount of data for subsequent processing and reduce the interference of background noise on geometric feature extraction.

[0044] Step S12: The initial ROI region is denoised and outlier removed using a centroid-based voxel filtering method to obtain the preprocessed ROI region.

[0045] This paper employs a centroid-based voxel filtering method for noise reduction. By dividing the initial ROI region into several three-dimensional voxel grids, the centroid coordinates of all points within each voxel are calculated, and the centroid is used to replace all points within the voxel. This achieves data downsampling and noise suppression while maintaining the overall geometric structure of the point cloud. It can more accurately preserve the local geometric features of the weld area, while effectively eliminating outliers caused by scanning environment or sensor errors, thus improving the stability and accuracy of subsequent geometric feature calculations.

[0046] The geometric features of the ROI region include: linearity, flatness, sphericity, total variance, anisotropy, sum of eigenvalues, eigenentiation, umbrella curvature, and mean absolute deviation of the Z coordinate.

[0047] For each point in the preprocessed ROI region, a neighborhood point set is constructed with that point as the center and a set of support radii. Let this neighborhood contain n points, and its coordinate set be... Calculate the center point of the neighborhood. And construct a decentralized point set matrix. Then calculate the covariance matrix. :

[0048]

[0049] Eigenvalue decomposition of the covariance matrix yields three eigenvalues. , , And sorted in descending order ( The following nine geometric features are calculated: linearity, flatness, sphericity, total variance, anisotropy, sum of eigenvalues, eigenentiation, umbrella curvature, and mean absolute deviation of Z coordinate. These features characterize the differences in three-dimensional geometric structure between the weld area and the base material area from the perspectives of linear, planar, and volumetric distribution characteristics and curvature changes of the local point cloud, providing discriminative input for subsequent classification models.

[0050] Step S2: Calculate the geometric features of the ROI region at a preset neighborhood scale, and construct training and testing sets based on the geometric features of the ROI region.

[0051] The formula for calculating the geometric features is shown in Table 1 below:

[0052] Table 1. Formulas for calculating geometric features

[0053]

[0054] In Table 1 For the sum of eigenvalues, As the center point of this neighborhood, For the i-th neighboring point, The normal vector corresponding to the current point. This represents the total number of points within the neighborhood. Let Z be the mean of the neighboring points. Let Z be the Z coordinate of the i-th neighboring point.

[0055] In real-world welding scenarios, weld seam morphologies are diverse, including straight weld seams, curved weld seams, and S-shaped weld seams. To enhance the model's generalization ability and applicability in practical industrial settings, geometric feature samples from these three types of weld seams are merged to construct a unified training set. This allows the model to learn more universal geometric feature representations across various weld seam morphologies.

[0056] Step S3: Using the LightGBM model as the base classifier, the improved gliding snake optimization algorithm is used to optimize the key hyperparameters of the LightGBM model to obtain the optimized LFMGSO-LightGBM model.

[0057] The LFMGSO-LightGBM model was trained using this training set and validated using this test set.

[0058] Step S3 employs an improved glider optimization algorithm to optimize the key hyperparameters of the LightGBM model, including:

[0059] Step S31: Generate a chaotic sequence using the Logistic chaotic mapping mathematical model to initialize the algorithm's population;

[0060] Step S32: The individual search step size is dynamically adjusted according to the iteration progress, population diversity and stagnation state through a feedback adaptive step size mathematical model.

[0061] Step S33: Introduce a mathematical model of multi-leadership collaboration mechanism, using the top K best individuals to jointly guide the population search direction.

[0062] It should be noted that the hyperparameter search space needs to be initialized before step S31. The hyperparameters include: n_estimators, learning rate, max_depth, and num_leaves.

[0063] Specifically, based on the performance impact of the LightGBM model in actual point cloud classification tasks, four key hyperparameters were selected as optimization targets: the number of base learners n_estimators, the learning rate, the maximum tree depth max_depth, and the number of leaf nodes num_leaves.

[0064] The search ranges for each hyperparameter are set as follows: n_estimators ranges from (50 to 300), learning rate ranges from (0.01 to 0.2), max_depth ranges from (1 to 12), and num_leaves ranges from (20 to 100). This range covers a reasonable parameter configuration range in engineering applications and provides a clear search boundary for subsequent optimization algorithms.

[0065] Step S31: Generate a chaotic sequence using the Logistic chaotic mapping mathematical model to initialize the algorithm's population;

[0066] To improve the uniformity of the initial population distribution in the parameter space and avoid insufficient early search ability due to overly concentrated initial positions, a Logistic chaotic mapping is used to initialize the population. The mathematical model of this Logistic chaotic mapping is as follows:

[0067]

[0068] in, The value at the nth iteration is in the range [0, 1]. For the control parameters of the mapping, In the first The value after the next iteration;

[0069] Based on the chaotic sequence, the position of each individual in the population is initialized using the following formula:

[0070]

[0071] in, For the first The individual in the first The initial position in the 3D search space. and They represent the first The lower and upper bounds of the dimensional search space. To map to the The individual Chaotic variables of dimension.

[0072] Step S32: The individual search step size is dynamically adjusted according to the iteration progress, population diversity and stagnation state through a feedback adaptive step size mathematical model.

[0073] To overcome the shortcomings of the original GSO algorithm, where the step size only depends on the linear decay of the iteration count and cannot reflect the actual search state of the population, a feedback adaptive step size mechanism is introduced. This mechanism comprehensively considers the iteration progress, population diversity, and stagnation state, and dynamically adjusts the search step size of each individual. The mathematical model of this feedback adaptive step size is as follows:

[0074]

[0075] in, Indicates the first The individual in the first Step size in the next iteration Indicates the current iteration number. Indicates the maximum number of iterations. , as well as These are non-negative weighting coefficients, used to control the influence of the time progress term, population diversity term, and stagnation term on the step size, respectively. The time decay exponent, For the first Population diversity index at the next iteration For the first Stagnation indicators during iteration;

[0076]

[0077]

[0078] in, For population size, For the first The current position vector of each individual For population mass center, For the first The distance from an individual to the center of the population It is the maximum value among all individuals' distances to the population center. This is a very small positive number set to avoid a denominator of zero. The larger the value, the more dispersed the population is, and the search is still highly exploratory; conversely, the smaller the value, the more the population tends to converge.

[0079]

[0080] in, For the first The current globally optimal fitness value at the next iteration. The length of the stall detection window. The threshold for determining stagnation is when the most recent The change in optimal fitness is less than the threshold. When the algorithm is considered to have entered a stagnant state, it is considered to have entered a stagnant state. 1. Increase the step size to enhance the ability to escape local optima.

[0081] Step S33: Introduce a mathematical model for a multi-leadership collaboration mechanism, utilizing the previous... The optimal individuals collectively guide the population search direction;

[0082] The mathematical model for this multi-leadership collaboration mechanism is as follows:

[0083]

[0084] in, For the first The individual in the first The position vector at the next iteration. In the first The new position in the next iteration For the first The individual in the first Step size in the next iteration and The weighting coefficients for the multi-leadership guidance term and the neighbor collaboration term, respectively. , , Let the direction vector of the current individual be the direction of the multi-leadership collaboration center. The direction vector from the current individual to its neighboring individuals. To determine the location of neighboring individuals according to the preset chained neighborhood rules, In order to be with the first The neighbor index function corresponding to each individual;

[0085] Should For multi-leadership collaboration center, ,in For the first The position vector of each leader The number of elite individuals participating in collaborative guidance; For the first The weight of each leader, and satisfying Through a multi-leader collaboration mechanism, the population can comprehensively utilize information from multiple superior individuals during the update process, thereby improving the stability and robustness of the search direction.

[0086] After updating the positions of all individuals according to the above model, a greedy selection strategy is used to compare the fitness values ​​of individuals before and after the update, retaining individuals with better fitness for the next generation of the population. To avoid individuals exceeding the search boundary after the update, boundary truncation or random repositioning strategies are used to correct out-of-bounds individuals. Then, the current globally optimal individual is updated, and it is determined whether the maximum number of iterations has been reached; if not, the iterative search continues until the optimal parameter combination is output.

[0087] Step S4: Use the LFMGSO-LightGBM model trained in step S3 to identify and classify the input weld point cloud.

[0088] It is understood that the various embodiments mentioned above in this invention can be combined with each other to form combined embodiments without violating the principle and logic. Due to space limitations, this invention will not elaborate further.

[0089] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0090] This invention provides a weld point cloud recognition method based on an improved gradient boosting tree. By employing an improved gliding snake optimization algorithm, it automatically and efficiently optimizes the key hyperparameters of the LightGBM model, overcoming the shortcomings of traditional empirical parameter tuning, grid, or random search methods, which suffer from high computational costs and susceptibility to local optima. This optimization process finds a better combination of hyperparameters for the classification model, thereby directly improving the accuracy of the final model in the binary classification task of weld and base material point clouds.

[0091] This improved gliding snake optimization algorithm significantly enhances optimization performance by comprehensively improving population initialization, search step size strategy, and guidance mechanism. Specifically, it enhances the algorithm's early global exploration capability by improving the distribution quality of the initial population in the parameter space; it better balances global exploration and local development by enabling the search process to be dynamically adjusted according to the iteration progress and population state; and it reduces the dependence on a single optimal solution by utilizing elite information in the population to collaboratively guide the search direction, thereby improving the stability, efficiency, and resistance to local optimum traps in complex parameter spaces.

[0092] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for weld point cloud recognition based on an improved gradient boosting tree, characterized in that, The steps include the following: Step S1: Collect three-dimensional point cloud data of welds of different shapes using a surface structured light camera, and preprocess the three-dimensional point cloud data to obtain the ROI region. Step S2: Calculate the geometric features of the ROI region at a preset neighborhood scale, and construct a training set and a test set based on the geometric features of the ROI region; Step S3: Using the LightGBM model as the base classifier, the key hyperparameters of the LightGBM model are optimized using the improved gliding snake optimization algorithm to obtain the optimized LFMGSO-LightGBM model. The LFMGSO-LightGBM model was trained using the training set and validated using the test set. Step S3 involves optimizing the key hyperparameters of the LightGBM model using an improved glider optimization algorithm, including: Step S31: Generate a chaotic sequence using the Logistic chaotic mapping mathematical model to initialize the algorithm's population; Step S32: The individual search step size is dynamically adjusted according to the iteration progress, population diversity and stagnation state through a feedback adaptive step size mathematical model. Step S33: Introduce a mathematical model of multi-leadership collaboration mechanism, and use the top K best individuals to jointly guide the population search direction; The key hyperparameters include: n_estimators, learning rate, max_depth, and num_leaves; Step S4: Use the LFMGSO-LightGBM model trained in step S3 to identify and classify the input weld point cloud.

2. The weld point cloud recognition method based on improved gradient boosting tree according to claim 1, characterized in that, The Logistic chaotic mapping mathematical model in step S31 is: in, The value at the nth iteration is in the range [0, 1]. For the control parameters of the mapping, In the first The value after the next iteration; Based on the chaotic sequence, the position of each individual in the population is initialized using the following formula: in, For the first The individual in the first The initial position in the 3D search space. and They represent the first The lower and upper bounds of the dimensional search space. To map to the The individual Chaotic variables of dimension.

3. The weld point cloud recognition method based on improved gradient boosting tree according to claim 1, characterized in that, The mathematical model formula for the feedback adaptive step size in step S32 is as follows: in, Indicates the first The individual in the first Step size in the next iteration Indicates the current iteration number. Indicates the maximum number of iterations. , as well as These are non-negative weighting coefficients, used to control the influence of the time progress term, population diversity term, and stagnation term on the step size, respectively. The time decay exponent, For the first Population diversity index at the next iteration For the first Stagnation indicators during iteration.

4. The weld point cloud recognition method based on improved gradient boosting tree according to claim 1, characterized in that, The mathematical model formula for the multi-leadership collaboration mechanism in step S33 is as follows: in, For the first The individual in the first The position vector at the next iteration. In the first The new position in the next iteration For the first The individual in the first Step size in the next iteration and These are the weighting coefficients for the multi-leadership guidance term and the neighbor collaboration term, respectively. The preset location of the neighboring individuals; The For multi-leadership collaboration center, ,in For the first The position vector of each leader The number of elite individuals participating in collaborative guidance; For the first The weight of each leader.

Images