A method and device for on-line monitoring of deformation amount of plasma laser strengthening of a thin-walled part

Abstract

A thin-walled part plasma laser strengthening deformation online monitoring method and device are disclosed. In the method, a thin-walled part plasma laser strengthening measurement system is constructed; during the intermittent period of laser impact processing, full-field three-dimensional point cloud data of the workpiece surface and three-dimensional profile data of the partition edge lap area are collected in real time; the cumulative deformation is calculated in real time based on the full-field three-dimensional point cloud data, and the geometric quality is quantitatively evaluated based on the three-dimensional profile data; the future deformation trend of the workpiece is predicted through a deep learning network based on an encoder-decoder architecture, the input is deformation field data arranged in time sequence, and the output is deformation prediction results; according to the deformation prediction results and the partition edge lap quality evaluation results, a hierarchical correction strategy is generated, the subsequent partition planning of the galvanometer scanning path and the robot is dynamically adjusted, and closed-loop control is realized.

Description

Technical Field

[0001] This invention relates to the field of laser surface strengthening technology, and in particular to a method and apparatus for online monitoring of deformation in thin-walled parts subjected to plasma laser strengthening. Background Technology

[0002] In the field of high-end equipment manufacturing, laser shock peening is a key surface modification process for improving the fatigue strength and stress corrosion resistance of thin-walled components. The distribution characteristics and uniformity of the residual stress field on the surface of the strengthened material directly determine the fatigue resistance and service life of the product. However, the inherent high-energy pulse shock characteristics of this process can lead to two types of coupled serious quality problems when using robot-galvanometer collaborative zonal strengthening: First, the non-uniform residual stress introduced by laser shock into the component can cause cumulative macroscopic deformation of the workpiece. This deformation evolves dynamically throughout the processing, with different regions influencing each other, ultimately leading to workpiece instability and severely affecting its assembly accuracy and aerodynamic shape. Second, at the overlapping edges of the zonal strengthening, due to the combined effects of robot positioning errors, path planning deviations, and the aforementioned overall deformation, geometric defects such as overlapping gaps, boundary misalignments, and steps are prone to occur. These defects not only damage the integrity of the structure but also form stress concentration points, becoming sources of fatigue crack initiation and directly weakening the gain effect of the strengthening process.

[0003] For full-field deformation monitoring, the current industrial practice generally relies on offline sampling inspection using coordinate measuring machines (CMMs) after machining. This method cannot capture the dynamic evolution process, and by the time a problem is detected, the workpiece has already undergone irreversible plastic deformation, resulting in high correction costs. Although some studies have attempted to use laser displacement sensors and other methods to measure the displacement of several discrete points online, they cannot obtain high-density full-field three-dimensional topography, making it difficult to establish an accurate deformation field model, let alone achieve pre-emptive predictive control.

[0004] In terms of monitoring the quality of overlap at the edges of workpieces, existing technologies lack effective on-machine 3D measurement methods. Traditional methods mostly rely on manual visual inspection after processing or comparison with contact templates, which is inefficient and cannot accurately quantify 3D geometric deviations. Some machine vision solutions have achieved preliminary results for static workpieces in laboratory environments, but their adaptability, stability, and quantification accuracy still face challenges when faced with surface morphology changes caused by deformation in actual processing, 3D geometric deviations, and the complex measurement environment of industrial sites, making it difficult to meet the reliability requirements of online monitoring.

[0005] Furthermore, in terms of control strategies, existing technologies generally suffer from a disconnect between "monitoring and control" and a mismatch between "goal and execution." Most systems employ fixed processing paths and parameters, lacking the ability to dynamically adjust based on 3D measurement results during intervals. For robot-galvanometer composite systems, an effective mechanism has not yet been established that can simultaneously respond to full-field deformation prediction and assess the overlapping quality of partition edges, and perform hierarchical collaborative control of galvanometer micro-scanning and robot macro-motion.

[0006] The information disclosed in the background section is only intended to enhance the understanding of the background of the present invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To address the shortcomings or defects of existing technologies, a method and apparatus for online monitoring of deformation in thin-walled parts subjected to plasma laser strengthening are provided. This method can rapidly acquire the overall state during processing intervals, and uses intelligent algorithms to predict deformation trends and quantify three-dimensional overlap quality. Based on this, hierarchical closed-loop control is implemented, thereby improving the process quality and consistency of laser shock strengthening of thin-walled parts at the system level.

[0008] The objective of this invention is achieved through the following technical solutions.

[0009] A method for online monitoring of deformation in thin-walled parts subjected to plasma laser strengthening includes,

[0010] A plasma laser-enhanced measurement system for thin-walled components was constructed. A surface structured light sensor was used to perform full-field three-dimensional surface scanning, and a line structured light sensor was used to perform three-dimensional line scanning of the overlapping areas at the partition edges. The surface structured light sensor and the line structured light sensor were jointly calibrated to unify their measurement data in the same coordinate system.

[0011] During the intervals of laser shock processing, real-time acquisition of full-field three-dimensional point cloud data and three-dimensional contour data of the overlapping areas of partition edges on the workpiece surface is performed.

[0012] The cumulative deformation is calculated in real time based on the full-field three-dimensional point cloud data, and the geometric quality is quantitatively evaluated based on the three-dimensional contour data.

[0013] The future deformation trend of the workpiece is predicted by a deep learning network based on an encoder-decoder architecture. The input is deformation field data arranged in time sequence, and the output is deformation prediction results.

[0014] Based on the deformation prediction results and the evaluation results of the overlap quality of the partition edges, a graded correction strategy is generated to dynamically adjust the galvanometer scanning path and the robot's subsequent partition planning, thereby achieving closed-loop control.

[0015] In the method described, a surface structured light camera is used as a full-field monitoring sensor; a line structured light profile measuring instrument based on the laser triangulation principle and equipped with a precision linear guide is used as a sensor for monitoring the overlapping quality of partition edges; the surface structured light camera and the line structured light profile measuring instrument are jointly calibrated to unify the coordinate system.

[0016] In the method described, during the interval of laser shock processing, the surface structured light camera is triggered to acquire the full-field point cloud, and the line structured light profile measuring instrument is controlled to move and scan along the edge of the partition to acquire its three-dimensional profile point cloud, so that the surface structured light camera, the line structured light profile measuring instrument and the laser strengthening equipment are triggered in sequence.

[0017] In the method described, the currently acquired full-field three-dimensional point cloud data is registered with the workpiece's reference three-dimensional model, and the spatial deviation of each point is calculated to generate a deformation distribution map; multiple three-dimensional geometric feature parameters are extracted from the three-dimensional contour data for geometric quality assessment, including step height, overlap gap, and contour straightness.

[0018] The method described above employs statistical filtering and radius filtering algorithms to denoise the original point cloud, voxel grid method to downsample the point cloud, covariance analysis method to estimate the point cloud normal, and Euclidean clustering to separate the ultrathin component from the fixture. Initial alignment is achieved through the SAC-IA algorithm based on FPFH features, and the improved ICP algorithm is used to register the current frame point cloud with the reference model. The registered point cloud data is then used to calculate the spatial offset between the corresponding points of the current point cloud and the reference model through KD-tree nearest neighbor search. A full-field deformation distribution map is generated based on hue mapping technology.

[0019] In the method described, if the deformation prediction result exceeds the threshold, the laser process parameters and / or galvanometer scanning path of the subsequent partitions are adjusted to compensate for the stress in the predicted deformation area; if the evaluation shows that the overlap quality does not meet the requirements, the robot path planning of the subsequent partitions is adjusted to correct the spatial position and orientation of the overlap area.

[0020] The method employs hierarchical dynamic adjustment of laser power, scanning speed, and overlap spacing. Galvanometer-level adjustment is used to suppress predicted deformation, while robot-level adjustment is used to correct overlap deviations. The galvanometer scanning path and robot partitioning path are optimized in real time to compensate for morphology and overlap deviations caused by deformation. Multi-parameter collaborative closed-loop control is achieved based on a feedforward-feedback composite control algorithm.

[0021] In the method described, the overlap spacing is detected and sampled at 0.5mm intervals along the partition edge direction. The minimum three-dimensional Euclidean distance between adjacent partition point clouds is calculated. The overlap rate in three-dimensional space is calculated based on the ratio of the overlapping area volume of the point clouds on both sides of the partition edge to the theoretical overlap volume.

[0022] An apparatus for implementing the online monitoring method for deformation of thin-walled parts by plasma laser strengthening includes,

[0023] The hardware acquisition module is used to acquire full-field three-dimensional point cloud data and three-dimensional contour data of the surface to be tested of the thin-walled part during processing intervals.

[0024] The data processing module is used to process the collected data in real time, including full-field deformation calculation and quantitative evaluation of the overlapping quality of partition edges based on three-dimensional topography.

[0025] The predictive analytics module is used to predict the overall deformation trend based on historical data.

[0026] The control execution module is used to generate a graded correction strategy based on the prediction and evaluation results, and to adjust the processing parameters of the galvanometer scanning system and the robot motion system respectively.

[0027] In the device, the control execution module constructs a hierarchical optimization objective function using an improved NSGA-II multi-objective genetic algorithm and achieves adaptive closed-loop optimization.

[0028] Compared with existing technologies, the beneficial effects of this invention are as follows: By integrating two three-dimensional measurement technologies, this invention achieves a more accurate and comprehensive perception of the overall deformation and the overlapping quality of the partition edges; by predicting deformation trends, it provides a forward-looking basis for control; by quantitatively evaluating the overlapping quality based on three-dimensional morphology, it achieves accurate diagnosis of the overlapping state of the partition edges; and finally, through hierarchical collaborative control, it simultaneously optimizes deformation suppression and edge overlapping quality, achieving a technological leap from offline, sampling, two-dimensional detection to online, full-domain, three-dimensional monitoring.

[0029] The description provided is merely an overview of the technical solution of this invention. In order to make the technical means of this invention clearer and more understandable, so that those skilled in the art can implement it according to the contents of the specification, and to make the described and other objects, features and advantages of this invention more obvious and understandable, specific embodiments of this invention are described below. Attached Figure Description

[0030] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0031] In the attached diagram:

[0032] Figure 1 This is a schematic diagram of the overall system structure of the present invention;

[0033] Figure 2 This is a flowchart of the data processing and closed-loop control logic of an embodiment of the method of the present invention;

[0034] Figure 3 This is a flowchart illustrating the deformation calculation algorithm based on point cloud processing in an embodiment of the method of the present invention.

[0035] Figure 4 This is a schematic diagram of intelligent quality assessment of the overlapping area of ​​partition edges based on the GBDT machine learning model in an embodiment of the method of the present invention;

[0036] Figure 5 This is a schematic diagram of the temporal deformation prediction network architecture based on the CNN-LSTM-Attention deep learning model in the embodiment of the method of the present invention;

[0037] Figure 6 This is a schematic diagram of the deformation monitoring and lap joint quality monitoring process for thin-walled components in Embodiment 6 of the present invention.

[0038] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0039] Specific embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0040] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0041] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments, and the accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0042] To better understand, such as Figures 1 to 6 As shown, the method includes,

[0043] A plasma laser strengthening measurement system for thin-walled parts was constructed. A surface structured light sensor was used for full-field 3D surface scanning, and a line structured light sensor was used for 3D line scanning of the overlapping areas at the edges of partitioned parts. Both the surface and line structured light sensors were jointly calibrated to unify their measurement data within the same coordinate system. Thin-walled parts refer to components with a thickness not exceeding 1 mm. In the field of machining, thin-walled parts are defined as components whose wall thickness to their outline dimension ratio is less than 1:20, or whose absolute wall thickness is less than 2.5 mm. Further, thin-walled parts refer to four typical ultra-thin components: film disks (thickness 0.35-1 mm), ducts (thickness 0.8 mm), ultra-thin blade edging (thickness 0.5-1 mm), and lattice skin structures (thickness 1 mm). The overlapping area at the edges of partitioned parts refers to the boundary overlap between two adjacent laser-impacted areas during robot-galvanometer collaborative partitioned strengthening. Partitioning is necessary because the laser spot size and galvanometer scanning range are limited, making it impossible to cover the entire surface of large-area or complex curved workpieces in one pass. The workpiece surface needs to be divided into multiple continuous blocks, and the robotic arm sequentially performs laser impact on each block. Overlapping refers to the requirement that, to ensure the integrity and uniformity of the strengthening process and to prevent unstrengthened "missed" gaps, a certain overlap must be set between the scanning paths of two adjacent blocks. This overlapping part is called the "block edge overlap area."

[0044] During the intervals of laser shock processing, real-time acquisition of full-field three-dimensional point cloud data and three-dimensional contour data of the overlapping areas of partition edges on the workpiece surface is performed.

[0045] The cumulative deformation is calculated in real time based on the full-field 3D point cloud data, and the geometric quality is quantitatively evaluated based on the 3D contour data. The real-time calculation of cumulative deformation first involves filtering, downsampling, and registration preprocessing the full-field point cloud to ensure precise alignment with the reference model. Then, based on the registration results, spatial indexing technology is used to calculate the 3D spatial offset between the current point cloud and the corresponding points in the reference model, generating a full-field deformation distribution map. Quantitative indicators such as maximum deformation, average deformation, and deformation gradient are statistically analyzed to characterize the degree of cumulative deformation during processing. The quantitative evaluation of geometric quality based on 3D contour data first involves filtering and completing the edge overlap contour point cloud obtained by line structured light scanning. Then, 3D geometric feature parameters such as step height, overlap gap, contour straightness, and effective overlap rate are extracted. Finally, the feature parameters are input into a pre-trained machine learning classification model, which outputs the overlap quality level and generates a structured evaluation report containing defect type, location coordinates, and parameter deviation values, achieving an objective quantitative diagnosis of overlap quality.

[0046] A deep learning network based on an encoder-decoder architecture is used to predict the future deformation trend of a workpiece. The input is deformation field data arranged in chronological order, and the output is the deformation prediction result. The deep learning network based on the encoder-decoder architecture is a CNN-LSTM-Attention hybrid model used to predict the future deformation trend of the workpiece. The network input is a 128×128×3-dimensional feature map sequence of the deformation field arranged in processing time sequence, and the corresponding global statistical features. The encoder part first extracts the spatial features of each time step through three convolutional layers with 3×3 kernels and 32 / 64 / 128 channels, respectively, and concatenates them with the global features processed by fully connected layers. Then, it is input into a two-layer stacked bidirectional LSTM to capture long-term temporal dependencies. Finally, a multi-head self-attention layer dynamically focuses on key historical states and outputs a context feature vector. The decoder part maps the context features to the predicted deformation features of the future multiple steps through fully connected layers, and then reconstructs the complete future deformation field sequence. The network training uses a time-weighted Huber loss as the objective function, giving higher weight to recent predictions to improve the predictive practicality. The optimization process uses the AdamW optimizer combined with cosine annealing learning rate scheduling to smoothly orient the optimal solution.

[0047] Based on the deformation prediction results and the geometric quality assessment results of the overlap at the partition edges, a hierarchical correction strategy is generated to dynamically adjust the galvanometer scanning path and the robot's subsequent partition planning, achieving closed-loop control. The hierarchical correction strategy is a macro-micro collaborative control strategy generated through a multi-objective optimization algorithm. Its core is to perform two-level dynamic adjustments to the galvanometer and robot based on the deformation prediction and overlap quality assessment results. The galvanometer-level adjustment executes feedforward control, real-time correcting the laser power, scanning speed, and scanning point position of the subsequent scanning area based on the predicted deformation to compensate for local stress. The robot-level adjustment executes feedback control, replanning the processing path, overlap rate, and robot pose of the subsequent partitions based on the measured overlap gap, step height, and straightness deviation to correct macro-geometric defects. Closed-loop control is achieved through a feedforward-feedback composite architecture. The system synchronously processes sensor data, performs optimization, and issues commands at fixed intervals, evaluating the control effect after processing each partition and adaptively updating model parameters to form a continuously optimized process closed loop.

[0048] In a preferred embodiment of the method, a surface structured light camera is used as a full-field monitoring sensor; a line structured light profile measuring instrument based on the laser triangulation principle and equipped with a precision linear guide is used as a sensor for monitoring the overlapping quality of partition edges; the surface structured light camera and the line structured light profile measuring instrument are jointly calibrated to unify the coordinate system.

[0049] In a preferred embodiment of the method, during the interval of laser shock processing, a surface structured light camera is triggered to acquire a full-field point cloud, and a line structured light profile measuring instrument is controlled to move and scan along the edge of the partition to acquire its three-dimensional profile point cloud, so that the surface structured light camera, the line structured light profile measuring instrument and the laser strengthening equipment are triggered in sequence.

[0050] In a preferred embodiment of the method, the currently acquired full-field three-dimensional point cloud data is registered with the workpiece's reference three-dimensional model, and the spatial deviation of each point is calculated to generate a deformation distribution map; multiple three-dimensional geometric feature parameters are extracted from the three-dimensional contour data for geometric quality assessment, including step height, overlap gap, and contour straightness.

[0051] In a preferred embodiment of the method, statistical filtering and radius filtering algorithms are used to denoise the original point cloud; voxel grid method is used to downsample the point cloud; covariance analysis method is used to estimate the point cloud normal; Euclidean clustering is used to separate the ultrathin component from the fixture; initial alignment is completed by SAC-IA algorithm based on FPFH features; improved ICP algorithm is used to register the current frame point cloud with the reference model; the registered point cloud data is used to calculate the spatial offset between the corresponding points of the current point cloud and the reference model through KD-tree nearest neighbor search; and a full-field deformation distribution map is generated based on hue mapping technology. First, statistical filtering is used to remove outliers that deviate from the point cloud mean by more than 2.5 standard deviations. Then, radius filtering with a radius of 0.3 mm is used to remove noise points in areas with density below a threshold. Next, a 0.15 mm grid size is used, and voxel downsampling is employed to preserve features while reducing data volume. Euclidean clustering is also used for precise segmentation of the thin-walled component, background, and fixture. The registration stage employs a coarse-to-fine registration strategy. The initial alignment stage uses the SAC-IA algorithm based on FPFH features. First, a multi-scale FPFH algorithm is used to calculate point cloud features and generate FPFH feature descriptors. Then, a KD-tree is used to accelerate feature matching, establishing a feature correspondence between the source and target point clouds. Finally, the SAC-IA algorithm is used to output the initial rotation matrix and the summation matrix. The transformation matrix is ​​used in the precise registration stage. Based on the initial alignment, the GICP algorithm is employed. This algorithm is based on a probability model where the point positions follow a Gaussian distribution. It solves for the transformation matrix that maximizes the matching degree between the two point cloud distributions through optimization methods in Lie algebra space. Finally, deformation quantization and visualization are performed. The registered point cloud data is used to construct a spatial index through KD-tree nearest neighbor search. The nearest neighbor of each point in the currently scanned point cloud in the reference model is searched point by point, and the three-dimensional Euclidean distance between the two is calculated as the offset of the point. The scalar field is linearly mapped to the color space through hue mapping technology. The ideal deformation range is set to correspond to green, positive convex deformation is mapped to the red system, and negative concave deformation is mapped to the blue system, thereby generating a color cloud map that intuitively reflects the deformation distribution and pattern of the entire field.

[0052] In a preferred embodiment of the method, if the deformation prediction result exceeds the threshold, the laser process parameters and / or galvanometer scanning path of the subsequent partitions are adjusted to compensate for the stress in the predicted deformation area; if the evaluation shows that the overlap quality does not meet the requirements, the robot path planning of the subsequent partitions is adjusted to correct the spatial position and orientation of the overlap area.

[0053] In a preferred embodiment of the method, the laser power, scanning speed, and overlap spacing are dynamically adjusted in stages. The galvanometer-level adjustment is used to suppress predicted deformation, and the robot-level adjustment is used to correct overlap deviations. The galvanometer scanning path and the robot partitioning path are optimized in real time to compensate for the morphology and overlap deviations caused by deformation. Multi-parameter collaborative closed-loop control is achieved based on a feedforward-feedback composite control algorithm.

[0054] In a preferred embodiment of the method, the overlap spacing is detected and sampled at 0.5mm intervals along the partition edge direction. The minimum three-dimensional Euclidean distance between adjacent partition point clouds is calculated. The overlap rate in three-dimensional space is calculated based on the ratio of the overlapping area volume of the point clouds on both sides of the partition edge to the theoretical overlap volume.

[0055] An apparatus for implementing an online monitoring method for deformation of thin-walled parts through plasma laser strengthening includes,

[0056] The hardware acquisition module is used to acquire full-field three-dimensional point cloud data and three-dimensional contour data of the surface to be tested of the thin-walled part during processing intervals.

[0057] The data processing module is used to process the collected data in real time, including full-field deformation calculation and quantitative evaluation of the overlapping quality of partition edges based on three-dimensional topography.

[0058] The predictive analytics module is used to predict the overall deformation trend based on historical data.

[0059] The control execution module is used to generate a graded correction strategy based on the prediction and evaluation results, and to adjust the processing parameters of the galvanometer scanning system and the robot motion system respectively.

[0060] In a preferred embodiment of the device, the control execution module constructs a hierarchical optimization objective function using an improved NSGA-II multi-objective genetic algorithm and achieves adaptive closed-loop optimization.

[0061] In one embodiment, deformation calculation and analysis includes: calculating the spatial offset between the registered point cloud data and the corresponding points of the reference model using KD-tree nearest neighbor search; generating a full-field deformation distribution map based on hue mapping technology to intuitively reflect the deformation area and deformation mode. The quantitative evaluation of the overlap quality of partition edges includes: extracting three-dimensional geometric features such as step height, overlap gap, contour straightness, and roughness from the three-dimensional point cloud of the partition edges; constructing a comprehensive evaluation system for edge overlap quality based on three-dimensional features. A multi-objective optimization correction strategy is generated based on the deformation prediction results and the partition edge overlap quality evaluation results; process parameters such as laser power, scanning speed, and overlap spacing are dynamically adjusted in stages, where galvanometer-level adjustment is used to suppress predicted deformation, and robot-level adjustment is used to correct overlap deviations; the galvanometer scanning path and robot partition path are optimized in real time to compensate for the morphology and overlap deviations caused by deformation; and multi-parameter collaborative closed-loop control is achieved based on a feedforward-feedback composite control algorithm. Example 1

[0062] according to Figure 1 As shown, the full-field deformation monitoring section uses a dual-lens grating structured light camera, specifically the LMIGOcator 3820M intelligent 3D camera, with the following specifications: resolution: 2448×2048; scan rate: 10 Hz; Z-axis repeatability: 2μm. The auxiliary subsystem uses DLP blue light projection to project coded stripe structured light, enabling full-field high-definition imaging from a fixed position. The monitoring section for the overlapping areas at the partition edges uses a Keyence LJ-X8000 series high-precision line structured light profile measuring instrument. The specific model is LJ-X8080, with the following specifications: accuracy: maximum repeatability up to 0.1 μm; scan speed: 64,000 profiles / second; Z-axis measurement range: ±8 mm. It is equipped with a precision linear guide rail and positioned above the overlapping areas at the partition edges.

[0063] The specific implementation plan for the hardware system is as follows:

[0064] S1. First, calibrate the system. This includes verifying the factory self-calibration accuracy of the area structured light camera and the line structured light profilometer, as well as the joint calibration between the two sensors, to facilitate unifying the point clouds collected by the two devices into the same coordinate system.

[0065] Factory calibration accuracy verification of the surface structured light camera: This system uses a highly integrated commercial surface structured light camera. The sensor has undergone rigorous internal parameter calibration at the factory, including camera intrinsic parameters, distortion coefficients, and the binocular stereo vision system. During system deployment, internal parameter calibration is not required, but its factory calibration accuracy must be verified to meet technical specifications by measuring standard gauge blocks, etc.

[0066] Factory self-calibration accuracy verification of the line structured light profiler: A high-precision line structured light profiler pre-calibrated at the factory was used for monitoring the overlap quality of zoned edges. This sensor has already undergone the manufacturer's calibration process at the factory, establishing a precise "pixel coordinates-3D profile" mapping relationship, meaning it incorporates camera intrinsic parameters, distortion coefficients, and the light plane equation. During system integration, there is no need to repeat its internal calibration, but accuracy verification is still required.

[0067] Joint calibration between the two sensors: Although both sensors have been calibrated independently, they each establish their own independent measurement coordinate systems. To achieve data unification, a joint system calibration is necessary to determine the pose transformation matrix from the line structured light sensor coordinate system to the surface structured light camera coordinate system. The specific steps are as follows:

[0068] Step 1: Place a standard sphere or cube calibration object with rich 3D features within the common field of view of both sensors simultaneously; Step 2: Simultaneously trigger both sensors to acquire the 3D point cloud of the calibration object. The calibration point cloud acquired by the surface structured light camera is... The calibration point cloud collected by the line structured light sensor is Step 3: Use point cloud registration algorithms such as ICP to accurately calculate the data from... arrive The rotation matrix R and translation vector T are used to obtain the transformation relationship. ={R, T}; Step 4: Subsequently, any point cloud of the overlapping region at the partition edge measured by line structured light. All can be accessed through By switching to the global coordinate system, seamless integration with the full-field deformation data and precise coordinate feedback for the robot can be achieved.

[0069] in, These are the coordinates of a point in the global coordinate system, and also the coordinates of a point in the coordinate system of the surface structured light camera; R represents the point cloud coordinates of the overlapping area at the partition edge measured by the line structured light; R and T are the rotation matrix and translation vector between the line structured light sensor and the area structured light camera.

[0070] S2. Perform full-field point cloud acquisition. The DLP projection system projects an 8-bit phase-shifted grating fringe sequence in HDR mode, achieving high-precision phase resolution through an 8-step phase-shifting method. The exposure time is adaptively adjusted according to the ambient light intensity, and automatic gain control ensures optimal fringe contrast under various operating conditions. Phase information is extracted from the fringe pattern using image processing algorithms, and the principal phase value is calculated using the 8-step phase-shifting method. The calculation formula is:

[0071]

[0072] in, Image pixel coordinates, The fringe image acquired at the nth phase shift is in The light intensity values ​​at the location, n = 1, 2, …, 8.

[0073] The phase is expanded using the multi-frequency heterodyne method to resolve phase ambiguities. Based on calibration parameters, the three-dimensional coordinates are calculated using stereo matching and triangulation principles. The calculation formula is as follows:

[0074]

[0075] B: Baseline distance; f: Focal length; d: Parallax; , This is the phase-to-parallax conversion coefficient.

[0076] Finally, the full-field 3D point cloud obtained from the scan is unified to the global coordinate system through coordinate transformation, and a full-field deformed point cloud model is generated.

[0077] S3. Perform contour scanning of the overlapping area at the edge of the overlapping section. A line structured light profile measuring instrument emits a red laser line with a wavelength of 650nm, projecting it onto the overlapping area at the edge of the section on the workpiece surface. A precision linear guide drives the sensor to scan at a uniform speed perpendicular to the overlapping boundary, with the scanning speed set to 50mm / s. During the scanning process, the sensor continuously acquires images of the deformed laser line at a sampling rate of 64,000 profiles / second. The laser power and camera exposure time are automatically optimized based on the reflectivity of the workpiece surface, ensuring clear laser stripe images with a signal-to-noise ratio better than 45dB can be obtained under different materials and surface conditions. The center of the light stripe is accurately extracted from the acquired laser line image using a real-time image processing algorithm, and sub-pixel-level positioning accuracy is achieved using Gaussian fitting. Based on the factory-calibrated internal parameters, the two-dimensional pixel coordinates are converted into three-dimensional contour coordinates using the laser triangulation principle. The calculation formula is:

[0078]

[0079] : Pixel coordinates of the center of the light stripe; K: Camera intrinsic parameter matrix; : Three-dimensional coordinates in the camera coordinate system; : Coefficients of the light plane equation.

[0080] S4. Data Synchronization and Transmission. Multi-device hard triggering is achieved via EtherCAT bus. After the laser stops, structured light acquisition is intermittently triggered after the thin-walled component stabilizes. The line structured light profilometer is triggered when the overlapping area at the partition edge cools to a visible state. Real-time scanning is performed as the overlapping area at the partition edge increases. All timestamps have an accuracy ≤100%. .

[0081] Example 2: Implementation of Deformation Calculation Algorithm Based on Point Cloud Processing

[0082] according to Figure 3 The implementation scheme for deformation calculation based on point cloud processing is as follows:

[0083] S1. Point Cloud Preprocessing. First, statistical filtering is used to remove outliers that deviate from the point cloud mean by more than 2.5 standard deviations. Then, radius filtering with a radius of 0.3 mm is used to remove noise points in areas with a density below the threshold. Next, a 0.15 mm grid size is used to reduce the amount of data while preserving features through voxel downsampling. Finally, Euclidean clustering is used to accurately segment the thin-walled parts, background, and fixtures.

[0084] S2. Perform point cloud registration. The point cloud registration implementation plan is divided into two stages: initial alignment and precise registration.

[0085] In the initial alignment stage, the SAC-IA algorithm based on FPFH features is adopted. First, the multi-scale FPFH algorithm is used to calculate the point cloud features and generate FPFH feature descriptors. Then, KD-tree is used to accelerate feature matching and establish feature correspondence between the source point cloud and the target point cloud. Finally, the SAC-IA algorithm is used to output the initial rotation matrix and translation matrix.

[0086] The precise registration stage employs the GICP algorithm. First, a probabilistic model is established, assuming that the location of each point in the source point cloud P and the target point cloud Q follows a Gaussian distribution. and The covariance matrix C is calculated through local neighborhood principal component analysis, and its specific expression is as follows:

[0087]

[0088] in, It is the j-th neighbor of the query point. It is the centroid of these k neighboring points.

[0089] Based on this probability model, an objective function is constructed to minimize the negative log-likelihood between the transformed source point cloud distribution and the target point cloud distribution. The objective function is defined as follows:

[0090]

[0091] Where T is the transformation matrix to be determined.

[0092] To improve computational efficiency, a systematic distribution aggregation strategy is adopted, dividing the point cloud into a 3mm×3mm×3mm voxel grid. The point distribution within each voxel V is then aggregated to form an aggregated Gaussian distribution. , where the mean covariance .

[0093] In the optimization stage, the Gauss-Newton method is used to solve the problem in the Lie algebra space se(3). This is achieved through Lie algebras... Parameterized transformation matrix, i.e. Constructing the residual function And calculate the corresponding Jacobian matrix. By solving the incremental equations Complete parameter update, including the Hessian matrix. gradient vector .

[0094] Finally, convergence criteria are set, including the relative transformation norm. Rate of change of objective function The maximum number of iterations was set to 30 to ensure that the algorithm meets real-time requirements while maintaining accuracy.

[0095] S3. Deformation Calculation and Visualization. Based on the precisely registered point cloud data, a KD-tree spatial index structure is established to search for the nearest neighbor correspondence between the currently scanned point cloud and the benchmark CAD model point by point. Normal distance is used as the deformation evaluation index to calculate the deformation at each measurement point. To the reference surface point The spatial offset, where Determined through nearest neighbor search:

[0096]

[0097] The formula for calculating deformation is:

[0098]

[0099] in, , , Each is the current point Its corresponding point in the baseline model The difference between coordinate components in the X, Y, and Z directions.

[0100] The normal deformation component is then passed through The result is obtained, where N is the unit normal vector of the reference surface at the corresponding point. To eliminate the influence of measurement noise, a Gaussian filtering algorithm is used to smooth the deformation field, with the filter kernel size set to 3×3 and the standard deviation... This ensures that random noise is effectively suppressed while preserving the true deformation characteristics.

[0101] The visualization engine is deeply customized based on the open-source point cloud library Open3D and the CloudCompare plugin, constructing a thickness-hue mapping model. This model linearly maps the deformation scalar field to the hue channel of the HSV color space, setting ideal deformation intervals to correspond to the green family (H=120°), negative deformation (concavity) to the blue family (H=240°), and positive deformation (convexity) to the red family (H=0°), forming a complete blue-green-red gradient color spectrum. The system uses OpenGL shader technology to dynamically update the RGB attributes of the point cloud and optimizes the rendering process through GPU parallel computing, achieving high performance even at point cloud scales. It can still maintain a high refresh rate, ensuring smooth real-time visualization.

[0102] Example 3: Algorithm Implementation of Partition Edge Overlap Quality Assessment Scheme

[0103] S1. Preprocessing of 3D Point Cloud in the Overlapping Region of Partition Edges. The original 3D point cloud of the overlapping region of partition edges, obtained by the line structured light profilometer, is preprocessed. First, a statistical filtering algorithm is used to remove discrete noise points that deviate from the point cloud mean by more than 2.5 standard deviations. Then, a radius filtering algorithm with a radius of 0.1 mm is used to remove sparse noise points, preserving continuous and complete boundary contours. For any minor data gaps that may exist during the scanning process, a K-nearest neighbor-based linear interpolation algorithm is used to complete the data, ensuring contour continuity.

[0104] S2. Precise extraction of 3D geometric features of the overlapping region at the partition edge. Based on the preprocessed high-density 3D point cloud, the geometric feature parameters of the overlapping boundary are precisely quantified:

[0105] Step height measurement: Establish an analysis area with a width of 2mm on both sides of the overlapping edge. Fit the best fitting plane of the two areas respectively using the RANSAC algorithm, and calculate the average distance between the two planes in the normal direction as the step height value.

[0106] Overlap gap detection: Sampling is performed at 0.5mm intervals along the edge of the partition. The overlap gap is accurately quantified by calculating the minimum three-dimensional Euclidean distance between the point clouds of adjacent partitions.

[0107] 3D edge straightness assessment: Project the 3D edge point cloud onto the best-fit plane, obtain the principal direction of the boundary through principal component analysis, and calculate the root mean square value of the distance from each edge point to the best-fit line as the straightness index.

[0108] Effective overlap rate calculation: The effective overlap rate in three-dimensional space is calculated based on the ratio of the overlapping area volume of the point clouds on both sides of the partition edge to the theoretical overlap volume.

[0109] S3. Intelligent Assessment Model for Overlap Quality. A three-dimensional geometric feature-based assessment model for overlap quality is constructed, using the previously extracted step height, overlap gap, edge straightness, and effective overlap rate as input features. A gradient boosting decision tree (GBDT) algorithm is used to construct a classifier, classifying overlap quality into four levels: "Excellent," "Good," "Medium," and "Poor." The training dataset contains 2000 labeled samples, covering typical overlap states under different materials and process parameters. Model parameters are optimized through five-fold cross-validation to ensure an assessment accuracy greater than 95%. The processing flow is as follows: Figure 4 As shown.

[0110] S4. Online Assessment and Decision Support for Overlap Quality. The system inputs real-time extracted 3D geometric features into a trained assessment model to obtain the quality level and deviation of various parameters in the current partition edge overlap area. When a step height greater than 10μm, overlap gap greater than 15μm, edge straightness deviation greater than 0.1mm, or a quality level of "poor" is detected, the system automatically triggers a quality alarm. The assessment results are bound to spatial coordinates to generate a structured report containing defect type, location coordinates, and deviation amount. This report is mapped to the global coordinate system through coordinate transformation, providing accurate decision-making basis for subsequent partition path planning and overlap parameter adjustment. The system also establishes an overlap quality evolution database, supporting quality trend analysis and process parameter optimization. Example 4: Implementation Plan for the Full-Field Deformation Trend Prediction Model

[0111] according to Figure 5 The temporal deformation prediction network architecture is used to predict the overall deformation trend. The process is as follows:

[0112] S1. Deformation Field Feature Sequence Construction. Based on the full-field 3D point cloud data acquired by the surface structured light camera during processing intervals, a deformation field feature sequence is constructed. First, the full-field point cloud at each time step is converted into a 128×128×3 deformation field feature map, with the three channels representing the deformation in the X, Y, and Z directions, respectively. Simultaneously, a global statistical feature vector containing key indicators such as maximum deformation, average deformation, deformation gradient, and energy distribution of the principal components of the deformation field is extracted. The deformation field feature maps and corresponding global feature vectors from 20 consecutive time steps are arranged chronologically to form a multimodal input sequence with a temporal resolution of 100 ms / frame, providing the model with deformation evolution information that combines local details and global statistics.

[0113] S2. Spatiotemporal Fusion Prediction Network Architecture. The network adopts a CNN-LSTM-Attention hybrid architecture to achieve accurate prediction of deformation trends. The encoder part consists of a spatial feature extraction module and a temporal modeling module. The spatial feature extraction adopts a two-branch CNN structure. The main branch processes the deformation field feature map through a three-layer convolutional network with a kernel size of 3×3 and the number of channels of 32, 64 and 128 respectively. It extracts spatial features with ReLU activation and batch normalization. The auxiliary branch processes the global statistical feature vector through a fully connected layer. The temporal modeling module uses a two-layer stacked bidirectional LSTM with a hidden state dimension of 512 to capture the long-term dependencies of deformation evolution. The attention mechanism adopts an 8-head self-attention to dynamically focus on key deformation states, and finally outputs a 1024-dimensional comprehensive feature vector as the basis for decoding.

[0114] S3. Unified Sequence Prediction Decoder. The decoder uses a single prediction head structure to directly output the complete future sequence, simplifying model design and improving computational efficiency. A fully connected layer directly maps the comprehensive feature vector output by the encoder to the deformation feature prediction values ​​for the next five time steps, maintaining the continuity of the output dimension.

[0115] S4. Model Training and Loss Function. The model is trained as a multi-step temporal regression task, and the loss function is designed as a time-weighted uniform form. The loss function is composed of the weighted sum of the Huber losses at each time step:

[0116]

[0117] For time step weighting coefficients, Let be the Huber loss at the t-th prediction time step.

[0118] The weight coefficients were set to decrease over time as 0.25, 0.22, 0.20, 0.18, and 0.15 to enhance recent prediction accuracy. The AdamW optimizer was used with an initial learning rate of 0.001, and a cosine annealing strategy was employed to dynamically adjust the learning rate, along with gradient clipping to prevent gradient explosion. The model was trained for 250 epochs with a batch size of 12 to ensure stable convergence.

[0119] S5. Prediction Result Generation and Visualization. The model output is a sequence of predicted deformation features for the next 5 time steps. The system uses a shared deconvolutional network to batch reconstruct the predicted feature values ​​into a 128×128×3 deformation field. The visualization interface overlays the predicted deformation field with real-time monitoring data in the same 3D view, using color mapping to represent the magnitude of the deformation, and displaying the deformation evolution process through semi-transparent rendering and contour line flow animation. The user interface provides multi-time-step prediction confidence assessment and deformation curve analysis of key areas, supporting real-time verification and historical backtracking of prediction results. Example 5: Implementation plan for a hierarchical collaborative closed-loop control system

[0120] according to Figure 1 and Figure 2 The implementation process of hierarchical collaborative closed-loop control is as follows:

[0121] S1. Generation of Multi-Objective Optimization Correction Strategy. Based on the overall deformation trend and the evaluation results of the overlapping quality of partition edges output by the prediction model, a hierarchical optimization objective function is constructed. Deformation control objective. The focus is on suppressing predictive deformations and aligning them with quality objectives. Correction of boundary geometric defects, processing efficiency target To ensure process continuity, the multi-objective optimization problem is transformed into: [The following is a separate, unrelated section:] Using a dynamic weighted linear weighting method, the multi-objective optimization problem is transformed into:

[0122]

[0123] The weighting coefficients are adaptively adjusted according to the processing stage: set in the normal area. , , Adjusted in the predicted deformation-sensitive area , , In areas with high risk of overlap quality, it is set as follows: , , The optimization algorithm uses an improved NSGA-II multi-objective genetic algorithm, with a population size of 50 and a maximum number of iterations of 80.

[0124] S2. Galvanometer-level deformation suppression control. Based on the deformation prediction results, a feedforward control strategy is implemented at the galvanometer level: when the predicted deformation... In the current and subsequent galvanometer scanning paths, a sparse scanning mode is used in the predicted convex region, reducing laser power by 8-12% and increasing scanning speed by 10-15%; when the predicted deformation... In the predicted concave area, an enhanced scanning mode is adopted, which increases the laser power by 5-8%, reduces the scanning speed by 8-12%, and increases the local impact density. Based on the predicted deformation field, the scanning path of the galvanometer is corrected in real time through a thin plate spline interpolation algorithm. The compensation coefficient k is set according to the material properties: 0.6-0.8 for titanium alloy and 0.4-0.6 for aluminum alloy.

[0125] S3. Robot-level overlap quality correction. Based on the overlap quality assessment results, feedback correction is implemented at the robot level: when an overlap gap is detected... When this happens, in subsequent partition path planning, the overlap compensation is increased, improving the overlap rate by 3-5%; when step height is detected... At that time, adjust the Z-axis positioning coordinates of the robot's subsequent zones to compensate for height deviations; when the boundary straightness deviations... At that time, the robot path is replanned.

[0126] S4. Feedforward-Feedback Composite Control Architecture. A feedforward-feedback composite control system is established. The feedforward channel adjusts the galvanometer parameters in advance based on deformation prediction results, while the feedback channel corrects the robot path based on the overlap quality assessment results.

[0127] S5. System Integration and Real-Time Performance Assurance. The control algorithm is deployed on an embedded controller, enabling real-time communication with the laser, galvanometer system, and robot via the EtherCAT bus. A control command priority mechanism is established: galvanometer deformation suppression commands have the highest priority and the highest real-time requirements; robot path correction has a medium priority; and process parameter optimization has a normal priority. After completing each zone of processing, the system automatically performs a control effect evaluation, dynamically adjusting the control parameters based on the deviation between the actual deformation measurement value and the predicted value, achieving adaptive closed-loop optimization.

[0128] Example 6: Deformation monitoring and overlap quality monitoring, as follows Figure 6 As shown:

[0129] S1. Full-field point cloud acquisition and reference model establishment for thin-walled parts. A thin-walled part was selected as the sample and fixed on the processing table. The workpiece surface was divided into multiple adjacent processing zones according to the actual laser strengthening path. Before the strengthening process began, a full-field 3D scan of the unstrengthened thin-walled part was performed using a surface structured light camera to obtain the reference point cloud of the undeformed thin-walled part, which was used as the reference model for subsequent deformation monitoring. During the processing, during the interval after each processing zone was completed, the surface structured light camera was triggered to perform a full-field scan of the current workpiece surface to obtain the deformation point cloud data at the corresponding time.

[0130] S2. Preprocessing of the full-field point cloud for thin-walled parts. The original full-field 3D point cloud acquired by structured light scanning is preprocessed. First, a statistical filtering algorithm is used to remove outlier noise points to eliminate isolated and abnormal points generated during the scanning process. Then, radius filtering is used to further remove sparse noise points, preserving the continuous and complete workpiece surface contour. Next, voxel downsampling is used to reduce the point cloud density, reducing the computational load for subsequent registration. Finally, Euclidean clustering is used to separate the thin-walled part point cloud from the background, fixtures, and other non-target point clouds, obtaining the target workpiece point cloud for deformation analysis.

[0131] S3. Coarse Registration and Initial Pose Adjustment of Point Clouds for Thin-Walled Parts. The preprocessed deformed point cloud and the reference point cloud at the current moment are imported into the registration module. Coarse registration is performed first to complete the initial pose adjustment between the two. The coarse registration stage adopts an initial registration method based on feature descriptors. By extracting local geometric features of the point cloud and establishing initial correspondences, the current deformed point cloud and the reference point cloud achieve preliminary alignment in overall pose, thus providing initial transformation parameters for subsequent precise registration.

[0132] S4. Precise Registration and Spatial Correspondence Establishment of Point Clouds for Thin-Walled Parts. Based on the coarse registration results, precise registration is performed between the current deformed point cloud and the reference point cloud. An iterative nearest-neighbor registration algorithm or a generalized iterative nearest-neighbor registration algorithm is used to further optimize the rigid transformation relationship between the two sets of point clouds, achieving higher spatial overlap in corresponding regions. After registration, a nearest-neighbor search method is used to establish the spatial correspondence between the current point cloud and the reference point cloud, providing a point-to-point matching basis for subsequent deformation calculations.

[0133] S5. Calculation and Distribution Characterization of Full-Field Deformation of Thin-Walled Parts. Based on the correspondence of the point cloud after accurate registration, the spatial offset of each measurement point in the current deformed point cloud relative to the corresponding point in the reference point cloud is calculated. The spatial offset can be characterized by three-dimensional Euclidean distance and / or normal direction distance. Based on the local offset of all points, a full-field deformation distribution map of the thin-walled part surface is generated, and the characterization results such as maximum deformation, average deformation, and deformation concentration area are output, thereby realizing online monitoring of full-field deformation during the strengthening process of thin-walled parts.

[0134] S6. Acquisition of 3D Contour Data for the Overlapping Area of ​​the Partition Edge. After each processed partition is completed, a line structured light profile measuring instrument is used to locally scan the overlapping area of ​​the current partition edge to obtain a high-density 3D contour point cloud of the area. The line structured light profile measuring instrument is based on the principle of laser triangulation. By projecting laser lines and collecting reflected contour information, it reconstructs the local 3D topographic data of the overlapping area of ​​the partition edge for subsequent overlap quality analysis.

[0135] S7. Preprocessing of 3D Point Cloud in the Overlapping Region of Partition Edges. The original 3D point cloud of the overlapping region of partition edges, obtained by the line structured light profilometer, is preprocessed. First, a statistical filtering algorithm is used to remove discrete noise points; then, radius filtering is used to remove sparse noise points, preserving continuous and complete boundary contours. For any minor local data loss that may exist during the scanning process, an interpolation completion method is used to reconstruct the missing areas, improving contour continuity and feature extraction stability.

[0136] S8. Precise Extraction of 3D Geometric Features of the Overlapping Region at the Partition Edge. Based on the preprocessed high-density 3D point cloud of the overlapping region at the partition edge, feature parameters reflecting the geometric quality of the overlapping region are extracted: Step Height Measurement: Analysis regions are established on both sides of the overlapping edge, and the point clouds of the two regions are fitted to a plane. The distance between the two fitted planes in the normal direction is calculated as the step height value; Overlap Gap Measurement: Sampling is performed at set intervals along the overlapping edge direction, and the minimum 3D Euclidean distance between adjacent partition contour point clouds is calculated as the overlap gap value; Contour Straightness Measurement: The edge contour point cloud is projected onto the fitting plane, and the distance deviation from each edge point to the best fitted line is calculated to characterize the straightness of the boundary contour; Effective Overlap Rate Calculation: Based on the correspondence between the actual overlapping part of the overlapping region and the theoretical overlapping region, the effective overlap ratio is calculated to characterize the coverage effectiveness of the overlapping region.

[0137] S9. Comparative Analysis of Surface Morphology Before and After Strengthening of the Overlap Area. A comparative analysis is performed on the surface roughness and morphology results of the overlapping area at the partition edge before and after strengthening to assist in assessing the changes in the surface condition of the local area. The results of this comparative analysis can serve as a supplementary basis for assessing the local geometric quality of the overlapping area, reflecting the changes in the surface morphology of the local area before and after strengthening treatment.

[0138] S10. Monitoring Results Output and Process Analysis. Based on the overall deformation distribution results and the geometric characteristic parameters of the overlapping areas at the edges of the partitions, the deformation monitoring results and overlapping quality monitoring results of the thin-walled parts during laser strengthening are output. For areas with concentrated deformation or abnormal overlapping quality, corresponding process analysis information can be further generated, providing a measurement basis for subsequent processing path verification, local supplementary measurement, or process parameter inspection.

[0139] S11. Optional Result Analysis and Extended Applications. In one optional implementation, statistical analysis, time-series comparison, or trend judgment can be performed on the continuously collected deformation monitoring results and overlap quality monitoring results to assist in evaluating the closed-loop control of subsequent processing stages. The result analysis is an extended application based on the online monitoring results and is not a necessary limiting condition for achieving full-field deformation monitoring and overlap quality monitoring in this embodiment.

[0140] Furthermore, the surface structured light system of this invention acquires the full-domain 3D morphology of the workpiece at a high frame rate, capturing the macroscopic deformation evolution trend; the line structured light system focuses on the critical quality-sensitive area of ​​the overlapping edges of the partitions, extracting microscopic geometric defects such as steps, gaps, and straightness with sub-micron Z-axis accuracy. Both systems are unified to the same global coordinate system through joint calibration, avoiding the inherent contradiction between resolution and field of view of a single sensor, and ensuring strict spatial and temporal alignment between the macroscopic deformation field and local overlapping features, providing a reliable data foundation for subsequent accurate evaluation and correction. By combining FPFH+SAC-IA initial registration with GICP fine registration, the non-rigid offset problem of point cloud caused by dynamic deformation of the workpiece is effectively solved, significantly improving registration robustness and accuracy. Based on this, point-by-point deformation is calculated using normal distance and KD-tree nearest neighbor search, and combined with hue mapping to achieve full-field deformation visualization, allowing operators to intuitively identify high-risk areas. Simultaneously, for the overlapping edges, the system automatically extracts three-dimensional geometric indicators such as step height, overlapping gap, and effective overlapping rate, and inputs them into a GBDT classification model trained with a large number of samples to achieve objective, quantitative, and graded evaluation of overlapping quality, completely replacing subjective visual inspection and significantly improving detection consistency and reliability. By constructing a CNN-LSTM-Attention hybrid network, the model can learn the nonlinear dynamic characteristics of material response from historical deformation field sequences, accurately predicting the full-field deformation evolution trend of multiple processing steps in the future. This prediction result serves as a feedforward signal, guiding the galvanometer to adjust the energy density and scanning density in areas where severe deformation has not yet occurred, thereby actively suppressing warping caused by residual stress accumulation, fundamentally reducing the occurrence of irreversible plastic deformation, and significantly improving process robustness. The system generates a multi-objective optimization strategy based on deformation prediction and overlapping evaluation: at the micro level, the galvanometer corrects the scanning path and adjusts the laser parameters in real time according to the predicted deformation, achieving adaptive control of the local energy field; at the macro level, the robot dynamically adjusts the pose, overlap rate, and boundary orientation of subsequent zones according to the measured overlapping deviation, ensuring overlapping continuity and geometric consistency. The two-level control operates in concert through a unified time-triggered mechanism and priority scheduling strategy, which not only ensures the real-time response of the high-frequency galvanometer but also takes into account the feasibility of robot path updates, forming an efficient, stable, and adaptive closed-loop control system.

[0141] Furthermore, this invention establishes a hardware foundation for the collaborative perception of macroscopic deformation and microscopic defects by integrating surface structured light and line structured light 3D measurement technologies. The surface structured light system acquires the full-domain 3D morphology of the workpiece at a high frame rate, dynamically capturing the evolution trend of macroscopic deformation caused by accumulated residual stress; the line structured light system focuses on the process quality-sensitive area of ​​the overlapping edge, accurately extracting key microscopic geometric defects such as step height, overlapping gap, and contour straightness with sub-micron Z-axis repeatability. Both systems are unified to the same global coordinate system through the joint calibration method described in Example 1, thereby achieving strict alignment between the macroscopic deformation field and local overlapping features in space and time, providing a reliable multi-scale data foundation for subsequent accurate evaluation and correction.

[0142] At the data processing level, this invention employs a hierarchical algorithm strategy to ensure the accuracy and efficiency of the analysis. For the full-field point cloud, a strategy combining FPFH+SAC-IA initial registration and GICP fine registration, as described in Example 2, is adopted. This effectively overcomes the non-rigid offset of the point cloud caused by the dynamic deformation of the workpiece during processing, significantly improving the robustness and accuracy of the registration. Based on the results of the precise registration, the point-by-point deformation is quantified through normal distance calculation and KD-tree nearest neighbor search. A visually intuitive full-field deformation distribution map is generated using hue mapping technology to assist in quickly locating high-risk deformation areas. For the overlapping edges, as described in Example 3, the system automatically extracts three-dimensional geometric features such as step height, overlapping gap, and effective overlapping rate, and inputs them into a GBDT classification model trained with a large number of samples. This enables an objective, quantitative, and graded evaluation of the overlapping quality from "excellent" to "poor," completely replacing the traditional inefficient and subjective visual inspection.

[0143] At the prediction and control level, this invention constructs an intelligent closed-loop system of "feedforward prediction - feedback evaluation - hierarchical control". As described in Example 4, by constructing a CNN-LSTM-Attention hybrid neural network, the model can learn the nonlinear mapping relationship between material response and process parameters from historical deformation field sequences, thereby accurately predicting the deformation evolution trend of multiple future processing steps. This prediction result, as a feedforward signal, can guide the galvanometer to pre-adjust the laser energy density and scanning path density in areas where severe deformation has not yet occurred, actively suppressing the accumulation of residual stress and reducing irreversible plastic deformation from the source. As described in Example 5, the system integrates deformation prediction and overlap quality evaluation results, and generates a hierarchical correction strategy through multi-objective optimization. At the microscopic execution level, the galvanometer corrects the scanning path and laser parameters in real time and at high frequency based on the predicted deformation, realizing adaptive control of the local energy field; at the macroscopic planning level, the robot dynamically adjusts the spatial pose, motion path, and overlap rate of subsequent zones based on the measured overlap deviation, ensuring the geometric continuity and consistency of the overlap. The two-level control operates collaboratively through a unified time-triggered and priority scheduling mechanism. While ensuring the real-time performance of galvanometer-level control, it also ensures the feasibility of robot-level path updates. Ultimately, this forms an efficient, stable, and self-learning adaptive closed-loop control system, achieving a fundamental technological leap from offline, passive, and spot-check to online, proactive, and full-check.

[0144] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0145] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for online monitoring of deformation in thin-walled parts subjected to plasma laser strengthening, characterized in that, It includes, A plasma laser-enhanced measurement system for thin-walled components was constructed. A surface structured light sensor was used to perform full-field three-dimensional surface scanning, and a line structured light sensor was used to perform three-dimensional line scanning of the overlapping areas at the partition edges. The surface structured light sensor and the line structured light sensor were jointly calibrated to unify their measurement data in the same coordinate system. During the intervals of laser shock processing, real-time acquisition of full-field three-dimensional point cloud data and three-dimensional contour data of the overlapping areas of partition edges on the workpiece surface is performed. The cumulative deformation is calculated in real time based on the full-field three-dimensional point cloud data, and the geometric quality is quantitatively evaluated based on the three-dimensional contour data. The future deformation trend of the workpiece is predicted by a deep learning network based on an encoder-decoder architecture. The input is deformation field data arranged in time sequence, and the output is deformation prediction results. Based on the deformation prediction results and the evaluation results of the overlap quality of the partition edges, a graded correction strategy is generated to dynamically adjust the galvanometer scanning path and the robot's subsequent partition planning, thereby achieving closed-loop control.

2. The method as described in claim 1, characterized in that, Preferably, a surface structured light camera is used as the full-field monitoring sensor; a line structured light profile measuring instrument based on the laser triangulation principle and equipped with a precision linear guide is used as the sensor for monitoring the overlapping quality of the partition edges; the surface structured light camera and the line structured light profile measuring instrument are jointly calibrated to unify the coordinate system.

3. The method as described in claim 2, characterized in that, During the interval of laser shock processing, the surface structured light camera is triggered to acquire the full-field point cloud, and the line structured light profile measuring instrument is controlled to move and scan along the edge of the partition to acquire its three-dimensional profile point cloud, so that the surface structured light camera, the line structured light profile measuring instrument and the laser strengthening equipment are triggered in sequence.

4. The method as described in claim 1, characterized in that, The currently acquired full-field 3D point cloud data is registered with the workpiece's reference 3D model, and the spatial deviation of each point is calculated to generate a deformation distribution map. Multiple three-dimensional geometric feature parameters are extracted from the three-dimensional contour data for geometric quality assessment. The three-dimensional geometric feature parameters include step height, overlap gap, and contour straightness.

5. The method as described in claim 1, characterized in that, The original point cloud was denoised using statistical filtering and radius filtering algorithms; the point cloud was downsampled using voxel grid method; the point cloud normal was estimated using covariance analysis method; and the ultrathin component was separated from the fixture using Euclidean clustering. Initial alignment was completed using the SAC-IA algorithm based on FPFH features; the current frame point cloud was registered with the reference model using an improved ICP algorithm; and the spatial offset between the current point cloud and the corresponding points of the reference model was calculated using KD-tree nearest neighbor search after registration. A full-field deformation distribution map was generated based on hue mapping technology.

6. The method as described in claim 1, characterized in that, If the deformation prediction result exceeds the threshold, the laser process parameters and / or galvanometer scanning path of the subsequent partitions are adjusted to compensate for the stress in the predicted deformation area; if the evaluation shows that the overlap quality does not meet the requirements, the robot path planning of the subsequent partitions is adjusted to correct the spatial position and orientation of the overlap area.

7. The method as described in claim 1, characterized in that, A hierarchical dynamic adjustment of laser power, scanning speed, and overlap spacing is adopted. Among them, galvanometer-level adjustment is used to suppress predicted deformation, and robot-level adjustment is used to correct overlap deviation. The galvanometer scanning path and robot partition path are optimized in real time to compensate for the morphology and overlap deviation caused by deformation. Multi-parameter collaborative closed-loop control is realized based on a feedforward-feedback composite control algorithm.

8. The method as described in claim 1, characterized in that, The overlap spacing is detected and sampled at 0.5mm intervals along the partition edge direction. The minimum three-dimensional Euclidean distance between adjacent partition point clouds is calculated. The overlap rate in three-dimensional space is calculated based on the ratio of the overlapping area volume of the point clouds on both sides of the partition edge to the theoretical overlap volume.

9. An apparatus for implementing the online monitoring method for deformation of thin-walled parts strengthened by plasma laser as described in any one of claims 1-8, characterized in that, It includes, The hardware acquisition module is used to acquire full-field three-dimensional point cloud data and three-dimensional contour data of the surface to be tested of the thin-walled part during processing intervals. The data processing module is used to process the collected data in real time, including full-field deformation calculation and quantitative evaluation of the overlapping quality of partition edges based on three-dimensional topography. The predictive analytics module is used to predict the overall deformation trend based on historical data. The control execution module is used to generate a graded correction strategy based on the prediction and evaluation results, and to adjust the processing parameters of the galvanometer scanning system and the robot motion system respectively.

10. The apparatus as claimed in claim 9, characterized in that, The control execution module constructs a hierarchical optimization objective function using the improved NSGA-II multi-objective genetic algorithm and achieves adaptive closed-loop optimization.

Images