CN-122023406-A - Weld joint point cloud identification method based on improved gradient lifting tree

Abstract

The invention relates to the technical field of three-dimensional point cloud processing and machine learning, in particular to a weld point cloud identification method based on an improved gradient lifting tree. The method comprises the steps of collecting three-dimensional point cloud data of welding seams of different shapes, extracting an ROI (region of interest) area through preprocessing, and calculating geometric features under a preset neighborhood scale to construct a training set and a testing set. And using LightGBM as a base classifier, optimizing key super parameters by adopting an improved glide snake optimization algorithm, obtaining a LFMGSO-LightGBM model, verifying, and realizing accurate identification and classification of the input weld point cloud. According to the weld point cloud identification method based on the improved gradient lifting tree, the initial solution space distribution quality is improved by improving the population initialization, the search step length and the guiding mechanism of the glide snake optimization algorithm, the dynamic balance of global exploration and local development is realized, and the stability and the local optimization resistance of the optimizing process are enhanced.

Inventors

• YANG HONGTAO
• BI ZIQIANG
• LI XIULAN
• FANG ZHIQIANG
• LU SHAN
• BAI HAOTIAN
• CHENG YUNLONG
• WANG JUNYANG
• YI QIUSHI

Assignees

• 长春工业大学

Dates

Publication Date

20260512

Application Date

20260413

Claims (5)

1. 1. The weld joint point cloud identification method based on the improved gradient lifting tree is characterized by comprising the following steps of: step S1, respectively acquiring three-dimensional point cloud data of welding seams with different shapes through a surface structure light camera, and preprocessing the three-dimensional point cloud data to obtain an ROI (region of interest); s2, calculating geometrical characteristics of the ROI region under a preset neighborhood scale, and respectively constructing a training set and a testing set based on the geometrical characteristics of the ROI region; Step S3, optimizing key super parameters of the LightGBM model by using a LightGBM model as a base classifier and adopting an improved glide snake optimization algorithm to obtain an optimized LFMGSO-LightGBM model; Training the LFMGSO-LightGBM model by using the training set, and verifying by using the testing set; And S4, identifying and classifying the input weld point cloud by using the LFMGSO-LightGBM model trained in the step S3.
2. 2. The weld point cloud identification method based on the improved gradient lifting tree according to claim 1, wherein the optimizing key super parameters of the LightGBM model by adopting the improved glide snake optimization algorithm in the step S3 includes: s31, generating a chaotic sequence by using a Logistic chaotic mapping mathematical model, and initializing the population of an algorithm; Step S32, dynamically adjusting individual searching step length according to iteration progress, population diversity and stagnation state through a feedback type self-adaptive step length mathematical model; and S33, introducing a multi-leader collaborative mechanism mathematical model, and jointly guiding the population searching direction by using the first K optimal individuals.
3. 3. The weld point cloud identification method based on the improved gradient lifting tree according to claim 2, wherein the Logistic chaotic mapping mathematical model in step S31 is as follows: Wherein, the For values at the nth iteration, the range is within 0,1, For the control parameters of the mapping, To at the first Numerical values after the iteration; based on the chaotic sequence, the positions of the individuals in the population are initialized by the following formula: Wherein, the Is the first Individual at the first The initial position in the dimensional search space, And Respectively represent the first The lower and upper bounds of the dimensional search space, To map to the first Individual first Chaos variables of dimensions.
4. 4. The weld point cloud identifying method based on the improved gradient lifting tree according to claim 2, wherein the feedback type adaptive step length mathematical model formula in step S32 is: Wherein, the Represent the first Individual at the first Step size at the time of the iteration, Representing the number of current iterations and, The number of iterations of the maximum is indicated, 、 And Is a non-negative weight coefficient and is respectively used for controlling the influence degree of a time progress item, a population diversity item and a stagnation item on the step length, In order to be an index of the time decay, Is the first The population diversity index at the time of the iteration, Is the first And (5) a stagnation index during iteration.
5. 5. The weld point cloud identification method based on the improved gradient lifting tree according to claim 2, wherein the multi-leader collaborative mechanism mathematical model formula in step S33 is: Wherein, the Is the first Individual at the first The position vector at the time of the iteration, To at the first The new position at the time of the iteration, Is the first Individual at the first Step size at the time of the iteration, And The weighting coefficients of the multi-leader and neighbor co-terms respectively, The method comprises the steps of setting preset neighbor individual positions; the said In order to be a multi-leaded collaboration center, Wherein Is the first The position vector of the individual leader is used, The number of elite individuals participating in collaborative guidance; is the first The weight of the individual leader.

Description

Weld joint point cloud identification method based on improved gradient lifting tree Technical Field The invention relates to the technical field of three-dimensional point cloud processing and machine learning, in particular to a weld point cloud identification method based on an improved gradient lifting tree. Background With the rapid development of high-end manufacturing fields such as aerospace, automobile manufacturing, ship processing and the like, welding quality detection and automatic weld joint identification have become key links in intelligent manufacturing and robot welding. Because the surface of the welding line has the characteristics of irregular shape, obvious size fluctuation, local noise interference and the like, the traditional detection mode relying on manual experience is difficult to meet the industrial requirements of automation, refinement and high reliability, and therefore, the high-precision automatic identification method for the welding line has important engineering application value. The existing weld joint identification method mainly comprises a method based on a two-dimensional visual image and a method based on a three-dimensional point cloud. The method based on the two-dimensional image generally adopts a threshold segmentation, edge detection or deep learning target detection network to extract the weld joint region, but is easily influenced by illumination change, reflection interference, shielding and visual angle change, so that the real three-dimensional geometric form of the weld joint is difficult to accurately represent. The three-dimensional point cloud method based on laser or structured light can directly acquire the space geometric information of the surface of the welding seam, and has more application potential in complex welding scenes. However, the existing three-dimensional point cloud weld joint identification method still has obvious defects. On the one hand, most of the traditional point cloud processing methods rely on threshold segmentation, geometric fitting or manual rule setting, for example, although the method based on RANSAC or surface fitting can realize weld extraction under partial rule scenes, the problems of under segmentation, over segmentation and insufficient robustness are easy to occur when the weld is complex in shape, strong in background noise or uneven in point cloud distribution. On the other hand, the weld point cloud has obvious local geometric differences, and if only single height features or simple curvature features are relied on, the microstructure differences between the weld region and the base material region are often difficult to fully describe. With the development of machine learning methods, a point cloud classification method based on three-dimensional geometric features and an integrated learning model is attracting attention. Particularly LightGBM is used as a high-efficiency gradient lifting tree model, has the advantages of high training speed, good classification performance, suitability for processing multidimensional feature input and the like, and has a good application prospect in point cloud classification tasks. However, the classification performance of LightGBM depends on the super-parameter setting to a great extent, and if the configuration of key parameters such as learning rate, tree depth, leaf node number, base learner number and the like is unreasonable, the model accuracy is easily reduced or the generalization capability is insufficient. The existing common parameter searching method, such as empirical parameter tuning, grid searching or random searching, has high calculation cost, is easy to sink into local optimum in a complex parameter space, and is difficult to consider searching efficiency and optimizing quality. Disclosure of Invention First, the technical problem to be solved The invention provides a weld point cloud identification method based on an improved gradient lifting tree, which aims to solve the problems of insufficient representation of a three-dimensional point cloud local geometric structure, low identification precision, low key super-parameter optimization efficiency of LightGBM models and easiness in sinking into local optimization in the prior art. (II) technical scheme In order to achieve the above purpose, the invention provides a weld point cloud identification method based on an improved gradient lifting tree, which comprises the following steps: step S1, respectively acquiring three-dimensional point cloud data of welding seams with different shapes through a surface structure light camera, and preprocessing the three-dimensional point cloud data to obtain an ROI (region of interest); s2, calculating geometrical characteristics of the ROI region under a preset neighborhood scale, and respectively constructing a training set and a testing set based on the geometrical characteristics of the ROI region; Step S3, optimizing key super parameters of the LightGBM mo