Robot laser vision three-dimensional scanning measurement system for large structural parts

Abstract

The application relates to the technical field of large structure measurement, and discloses a robot laser vision three-dimensional scanning measurement system for large structural parts. The system comprises a robot motion platform, a laser vision sensing module, an ambient light sensing module, a point cloud processing module and a motion control module, wherein the laser vision sensing module is arranged at the tail end of the robot motion platform. The ambient light sensing module collects ambient light intensity distribution data in real time, the motion control module generates multi-view dynamic scanning path instructions in combination with the data and three-dimensional model data of the large structural part to be measured, the robot motion platform executes the instructions to drive the laser vision sensing module to move, the module projects a laser stripe to the surface of the structural part to be measured and synchronously collects images, and the point cloud processing module generates three-dimensional point cloud data after adaptive filtering processing of the images, so that the efficient and accurate measurement requirement of the large structural part in a complex industrial environment can be met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large structure measurement, in particular to a robot laser vision three-dimensional scanning measurement system for large structure. BACKGROUND

[0002] In the fields of aerospace, shipbuilding, heavy machinery, etc., the dimensional accuracy and morphological integrity of large structure directly relate to the overall performance and safe operation of the product. Precise measurement of large structures is an important requirement in the production and quality detection process. At present, the measurement methods for large structures mainly include contact measurement and non-contact measurement. Contact measurement is represented by three-coordinate measuring machines. Although this type of equipment can achieve high-precision measurement, it has obvious limitations in practical application. The measurement range is limited by the structure of the equipment. For large structures with a length of tens of meters, the equipment needs to be moved multiple times and the measurement data needs to be spliced. This not only increases the operation complexity and time consumption, but also easily affects the final measurement accuracy due to splicing errors.

[0003] In non-contact measurement technology, laser scanning measurement gradually becomes an important choice for large structure measurement due to its advantages of fast measurement speed, high precision, non-contact and non-damage. However, the existing laser scanning measurement system still has many shortcomings when facing complex industrial environments. Environmental light changes have a significant impact on the quality of laser stripe image acquisition. In industrial sites, fluctuations in natural light intensity, switching or position changes of lighting equipment in the workshop, etc. will all lead to a decrease in the contrast between laser stripes and the background, and even the laser stripes may be submerged in stray light, thereby affecting the accuracy of subsequent image processing and point cloud generation. Most current systems lack real-time perception and dynamic adaptation capabilities for environmental light, and can only work under preset fixed lighting conditions. Once the environmental light exceeds the preset range, the measurement accuracy will be greatly reduced, making it difficult to meet the stable measurement requirements in complex industrial environments.

[0004] The scan path planning of the existing system is mostly based on the fixed three-dimensional model of the structure to be measured, and a preset fixed view angle scanning mode is adopted. When facing large structure with complex surface morphology and many shielding areas, this mode is prone to have scanning blind area, resulting in that some areas cannot be effectively scanned, and manual adjustment of scanning view angle and path is required, which not only increases the operation difficulty, but also may introduce additional error due to the uncertainty of manual intervention. Meanwhile, the fixed scan path cannot be dynamically adjusted according to the environmental changes in the actual measurement process (such as the decline of the measurement quality of some areas caused by the change of the ambient light intensity), and it is difficult to ensure the consistency and reliability of the data acquisition of each area in the whole measurement process. In addition, in the point cloud processing link, the filtering algorithm of the existing system mostly has fixed parameter setting, and cannot be adaptively adjusted according to the quality of the laser stripe image collected under different ambient light conditions. For the image seriously disturbed by stray light, the filtering effect is not good, and a large number of noise points are easily retained, which affects the precision of the three-dimensional point cloud data, and further adversely affects the subsequent form analysis and size detection of the structure. SUMMARY

[0005] The present application aims to provide a robot laser vision three-dimensional scanning measurement system for large structures to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides a robot laser vision three-dimensional scanning measurement system for large structures, which comprises:

[0007] a robot motion platform, a laser vision sensing module installed at the end of the robot motion platform, an ambient light perception module, a point cloud processing module and a motion control module;

[0008] The ambient light perception module collects real-time ambient light intensity distribution data; the motion control module generates multi-view dynamic scanning path instructions according to the ambient light intensity distribution data and the three-dimensional model data of the large structure to be measured; the robot motion platform executes the multi-view dynamic scanning path instructions to drive the laser vision sensing module to move; the laser vision sensing module projects laser stripes onto the surface of the large structure to be measured and synchronously collects laser stripe images; and the point cloud processing module performs adaptive filtering processing on the laser stripe images and generates three-dimensional point cloud data.

[0009] Preferably, the laser vision sensing module comprises a laser power adjustment unit; the laser power adjustment unit receives the real-time ambient light intensity distribution data sent by the ambient light perception module, and dynamically adjusts the laser emission power of the laser vision sensing module according to a preset laser intensity threshold value; the preset laser intensity threshold value is greater than the maximum value of the current ambient light intensity distribution data and less than half of the maximum emission power of the laser vision sensing module.

[0010] Preferably, the motion control module comprises a curved surface curvature analysis unit and a path planning unit; the curved surface curvature analysis unit calculates the curved surface curvature distribution based on the three-dimensional model data of the large structure to be measured; the path planning unit generates the multi-view dynamic scanning path instruction according to the curved surface curvature distribution, the multi-view dynamic scanning path instruction comprising a scanning speed and a scanning view angle interval; the scanning speed is inversely proportional to the curved surface curvature distribution, and the scanning view angle interval is proportional to the curved surface curvature distribution.

[0011] Preferably, the point cloud processing module comprises a feature point extraction unit and a registration optimization unit; the feature point extraction unit performs multi-scale feature point detection on the generated three-dimensional point cloud data and generates a feature point descriptor; the registration optimization unit constructs a point cloud registration cost function based on the feature point descriptor, and solves an optimal transformation matrix of the point cloud registration cost function by using a multi-objective optimization algorithm; the point cloud registration cost function comprises a feature point matching degree constraint and a point cloud overlap constraint.

[0012] Preferably, the adaptive filtering process comprises: the point cloud processing module constructs a spatial filter according to the ambient light intensity distribution data, suppresses background noise of the laser stripe image by using the spatial filter, and extracts a laser stripe center line through a gray-scale morphological operation.

[0013] Preferably, the system further comprises a multi-station synchronous module; the multi-station synchronous module receives real-time pose data sent by the robot motion platform, and triggers a scanning action of the laser vision sensing module according to the real-time pose data; a triggering time of the scanning action is synchronized with a timestamp of the robot motion platform reaching a preset scanning station in the multi-view dynamic scanning path instruction.

[0014] Preferably, the registration optimization unit stores the optimal transformation matrix and sends the optimal transformation matrix to the motion control module; the motion control module corrects the scanning station coordinates in the subsequent multi-view dynamic scanning path instruction according to the optimal transformation matrix.

[0015] Preferably, the system comprises a closed loop verification module; the closed loop verification module receives the three-dimensional point cloud data generated by the point cloud processing module, calculates a point cloud overlap area error between adjacent scanning stations, and sends the point cloud overlap area error to the registration optimization unit; the registration optimization unit updates a weight coefficient of the point cloud registration cost function according to the point cloud overlap area error.

[0016] Preferably, the multi-objective optimization algorithm adopts a particle swarm optimization algorithm, a particle position vector of the particle swarm optimization algorithm contains rotation and translation parameters, and a fitness function of the particle swarm optimization algorithm is the point cloud registration cost function.

[0017] Preferably, the motion control module comprises a path re-planning unit; when the point cloud overlap area error exceeds a preset error threshold, the path re-planning unit generates a multi-view dynamic scanning path instruction covering the point cloud overlap area based on the updated point cloud registration cost function.

[0018] Compared with the prior art, the present application has the following beneficial effects:

[0019] The ambient light sensing module is arranged to realize real-time acquisition of light intensity distribution data in the measurement environment, breaking the dependence of the existing system on fixed lighting conditions. The module can dynamically capture the changes of ambient light and transmit the relevant data to the motion control module in real time, providing a basis for subsequent dynamic adjustment of the scanning path and optimization of the working parameters of the laser vision sensing module, so that the system can always effectively control the quality of laser stripe image acquisition under complex environmental light conditions such as natural light fluctuations and changes in workshop lighting, avoid problems such as image contrast reduction or stripe flooding caused by environmental light interference, and ensure the stability and reliability of the image acquisition link.

[0020] The motion control module generates a multi-view dynamic scanning path instruction according to the ambient light intensity distribution data and the three-dimensional model data of the large structure to be measured, which has stronger flexibility and adaptability compared with the existing fixed path scanning method. The module can automatically plan a scanning path that can avoid high glare interference areas and focus on covering low interference areas based on the three-dimensional model data in combination with the ambient light distribution, dynamically adjust the scanning angle for the occluded areas and complex shapes on the surface of the structure, ensure that there is no blind area in the scanning process, and does not need manual repeated adjustment of the path, greatly reducing the operation process and time cost. In addition, when the ambient light intensity changes suddenly, the motion control module can quickly respond and adjust the scanning path in real time to ensure that each measurement area can still obtain high-quality image data during the ambient light fluctuation process, further improving the measurement stability of the system in complex industrial environments.

[0021] The robot motion platform executes a multi-view dynamic scanning path instruction to drive the laser vision sensing module to move accurately, so as to ensure that the laser vision sensing module can scan the large structural part according to the planned optimal path and view angle. The platform has high-precision motion control capability, can accurately adjust the position and attitude of the laser vision sensing module, can accurately project the laser stripe to the target area on the surface of the structure to be measured, can ensure the stability of the module position during image acquisition, can avoid measurement errors caused by motion deviation, and can provide reliable protection for high-quality laser stripe image acquisition.

[0022] The laser vision sensing module synchronously completes laser stripe projection and image acquisition, and its working process is closely linked with the ambient light sensing module and the motion control module. Under the guidance of the dynamic path planned by the motion control module, the module can project clear laser stripes to the surface of the structure at the best view angle, and synchronously acquire laser stripe images, so as to ensure that the acquired images can truly reflect the morphological characteristics of the surface of the structure, and provide high-quality original data for subsequent point cloud processing.

[0023] The point cloud processing module performs adaptive filtering processing on the laser stripe images. Different from the existing fixed parameter filtering algorithm, the module can automatically adjust the filtering parameters and algorithm according to the ambient light intensity distribution data acquired by the ambient light sensing module and the actual quality (such as contrast, degree of stray light interference, etc.) of the laser stripe images, so as to realize targeted filtering processing of images of different qualities. For images with less stray light interference and high quality, the filtering process can retain the detail information of the laser stripe while removing a small amount of noise; for images with serious stray light interference and low quality, the filtering process can effectively suppress stray light noise, enhance the contrast between the laser stripe and the background, accurately extract the laser stripe features, and then generate three-dimensional point cloud data with high precision and less noise, so as to truly restore the surface morphology and size information of the large structure to be measured, and meet the needs of various application scenarios such as subsequent structural part quality detection and morphological analysis. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 A timing diagram of the robot laser vision three-dimensional scanning measurement system for large structural parts described in the present application;

[0025] Figure 2 A flowchart for feature extraction and registration optimization of the point cloud processing module;

[0026] Figure 3 A flowchart for updating the point cloud registration cost function weight of the closed-loop verification module. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0028] Please refer to Figure 1 The present application provides a robot laser vision three-dimensional scanning measurement system for large structural parts, which comprises a robot motion platform, a laser vision sensing module, an ambient light perception module, a point cloud processing module and a motion control module.

[0029] The robot motion platform adopts a six-axis industrial robot, the end of which is provided with the laser vision sensing module through a flange. The ambient light perception module adopts a high dynamic range light intensity sensor, which is arranged at multiple positions in a measurement site to collect ambient light intensity distribution data. The motion control module is realized based on an industrial computer, runs a path planning algorithm and generates control instructions. The point cloud processing module is provided with a GPU computing unit for real-time processing of image and point cloud data. In implementation, the ambient light perception module continuously collects ambient light intensity distribution data and transmits the data to the motion control module. The motion control module generates multi-view dynamic scanning path instructions according to the ambient light intensity distribution data and a pre-input three-dimensional CAD model of the large structural part to be measured. The instructions are sent to a robot controller through Ethernet to drive the robot motion platform to move the laser vision sensing module along the planned path. The laser vision sensing module projects a laser stripe to the workpiece surface and synchronously collects laser stripe images during the movement. The image data is transmitted to the point cloud processing module through a gigabit network. The point cloud processing module processes the laser stripe images by using an adaptive filtering algorithm, extracts a laser stripe center line and calculates three-dimensional point cloud data, and finally completes full-size three-dimensional reconstruction of the large structural part.

[0030] In embodiment 1, the laser power adjustment unit of the laser vision sensing module continuously receives real-time ambient light intensity distribution data from the ambient light perception module through a special data interface. The data is transmitted in the form of a two-dimensional array, each array element corresponding to an illumination intensity value of a sampling point. The sampling points are distributed according to a spatial grid to cover the entire measurement area. A data processing chip inside the laser power adjustment unit iterates through the entire array and finds the maximum value in the array. This maximum value represents the strongest background light interference level in the current environment. The chip then calls a pre-set laser intensity threshold value stored in a non-volatile memory. The threshold value is set during system initialization according to the calibration parameters of the laser. Its value must be strictly greater than the maximum intensity of the ambient light and less than half of the maximum output power of the laser. This setting ensures that the laser stripe has sufficient contrast under any ambient light conditions, and avoids image overexposure or laser life loss due to excessive power.

[0031] The data processing chip generates a voltage signal proportional to the required power through a digital-to-analog conversion circuit, and this analog voltage signal is sent to the driving circuit of the laser, which accurately adjusts the injection current of the laser diode according to the voltage value, thereby controlling the size of the laser output power. The whole adjustment process is dynamic and continuous. The ambient light sensing module updates the light intensity data at a frequency of ten times per second, and the laser power adjustment unit adjusts the output in real time, ensuring that the laser stripe is always clear and visible in the case of rapid changes in ambient light. For example, when a moving light source passes through the scene, the ambient light intensity will suddenly increase, and the adjustment unit can increase the laser power within milliseconds to maintain stable image quality. The curvature analysis unit of the motion control module reads the three-dimensional CAD model file of the large structure to be measured from the system database. This file usually contains complete surface geometry information. The analysis unit uses a surface fitting algorithm based on the moving least squares method to discretize the model surface. The algorithm selects a neighborhood point set around each calculation point, obtains the principal curvature parameters of the point through local surface fitting, and then calculates the Gaussian curvature reflecting the bending degree of the surface. All the curvature values of the calculation points are organized into a curvature distribution map corresponding to the CAD model grid. The map uses color gradients to intuitively identify different curvature regions. High curvature areas such as edges and corners are usually marked with warm tones.

[0032] The path planning unit receives the curvature distribution data and generates multi-view dynamic scanning path instructions accordingly. The path planning algorithm first performs regional segmentation on the curvature distribution, dividing continuous surfaces with similar curvature values into the same scanning region. For each scanning region, the algorithm automatically calculates two key parameters: scanning speed and scanning view angle interval. In high curvature regions, due to the complex surface geometry, the scanning speed needs to be reduced and the view angle interval needs to be decreased to obtain denser measurement data. The scanning speed is inversely proportional to the curvature value, and for every unit increase in curvature, the scanning speed decreases by a certain percentage. The scanning view angle interval is proportional to the curvature value, and the greater the curvature, the smaller the angle interval between adjacent scanning stations. The generated path instructions include a series of spatial coordinates and attitude parameters of scanning station points, each station point is accompanied by corresponding scanning speed and view angle interval values. These parameters are encoded into a motion instruction format that can be recognized by the robot controller, and sent to the robot motion platform through industrial Ethernet. When the robot executes the scanning task, it will automatically adjust the motion speed and the angle change between adjacent station points according to the curvature region to which the current position belongs. This dynamic adjustment mechanism enables the measurement resources to focus on areas with complex geometric features, ensuring measurement accuracy while significantly improving overall scanning efficiency. The laser power adjustment unit and the path planning unit have a data interaction channel. When the path planning unit instructs the robot to enter a high curvature region for fine scanning, it will simultaneously notify the laser power adjustment unit to pay special attention to the ambient light conditions in that region. Because high curvature regions often have complex reflection and shadow effects, more accurate laser power control is needed. The coordinated work of the two units ensures optimal scanning data quality under different curvature conditions.

[0033] The operation of the entire system is based on a strict time synchronization mechanism. The sampling period of the ambient light perception module, the laser power adjustment period, and the robot motion control period maintain an integer multiple relationship. This design avoids control errors caused by timing mismatches between subsystems. All modules align data through precise timestamps to ensure that each scanning action is performed under preset ambient light conditions and motion parameters.

[0034] Embodiment 2: see Figure 2, the feature point extraction unit of the point cloud processing module performs multi-scale feature point detection on the generated three-dimensional point cloud data. This process first establishes a multi-level resolution data structure, creates multiple point cloud levels from coarse to fine through voxel grid downsampling method, calculates the normal vector variation rate of each point in each level of point cloud, and obtains the normal vector through local plane fitting. The neighborhood radius changes with the point cloud level. A larger neighborhood radius is used in the coarse resolution layer to capture macro features, and a smaller neighborhood radius is used in the fine resolution layer to capture microscopic details. When the angle between the normal vector of a point and the average normal vector of its neighborhood points exceeds a certain threshold, the point is marked as a feature point. This multi-scale detection mechanism can comprehensively identify geometric features of different sizes. Based on the extracted feature point set, the feature point descriptor generation algorithm starts to run. This algorithm uses the fast point feature histogram method to calculate the relative position and normal vector angle relationship between all point pairs in the neighborhood of each feature point, and puts these relationship parameters into a statistical histogram to form a descriptor vector with rotation and translation invariance. The descriptor vector contains multiple dimensions of statistical values, which fully represent the local geometric distribution characteristics around the feature point. The descriptors of all feature points form a descriptor database, which is used for feature matching in the subsequent point cloud registration process. The point cloud registration cost function constructed by the registration optimization unit includes two core constraint terms. The feature point matching degree constraint term is realized by calculating the Euclidean distance between the feature point descriptors in the two pieces of point cloud. This distance reflects the similarity of the corresponding feature points. The point cloud overlap constraint term is realized by calculating the projection coincidence rate of the point cloud. The point cloud is projected into the perspective space of the other piece of point cloud, and the proportion of the projection points falling within the neighborhood range of the target point cloud is calculated. This proportion reflects the spatial coverage consistency of the two pieces of point cloud. The cost function weights and combines these two constraint terms to form a comprehensive evaluation index.

[0035] A multi-objective optimization algorithm is applied to solve the optimal transformation matrix of the point cloud registration cost function. This algorithm searches for the optimal solution in a six-dimensional space, and the solution space includes three rotation degrees of freedom and three translation degrees of freedom. The optimization process has two objectives: minimizing the feature point matching distance and maximizing the point cloud overlap rate. Through iterative calculation, the pose transformation parameters that minimize the weighted cost function value are found. The optimization result is output as a four-order homogeneous transformation matrix, which accurately describes the spatial transformation relationship between the two pieces of point cloud. The adaptive filtering process constructs a spatial filter based on the ambient light intensity distribution data. The light intensity data collected by the ambient light perception module is mapped to the image coordinate system to generate a background light intensity distribution map consistent with the size of the laser stripe image. This distribution map marks the ambient light interference intensity at each pixel position of the image. Based on this, the adaptive median filter has a dynamically adjusted window size, which is proportional to the light intensity variance of the local area. In uniform light areas, small windows are used for fast filtering, and in light abrupt areas, large windows are used to suppress strong noise.

[0036] The filtered laser stripe image enters the gray-scale morphological processing stage, first uses the open operation to eliminate scattered point noise, the size of the structure element is dynamically determined according to the image signal-to-noise ratio, a smaller structure element is used for connection repair in the laser stripe fracture area, and a larger structure element is used to enhance the profile in the stripe edge fuzzy area. The binary image after morphological processing enters the skeleton extraction link, and the thinning algorithm is used for iteration to remove the stripe edge pixels until the single-pixel width center line is obtained. The center line coordinates are recorded and used for subsequent three-dimensional point cloud reconstruction calculation. A data feedback channel is established between the feature point extraction unit and the registration optimization unit. When the registration optimization unit detects that the matching rate of the feature points in a certain area is low, it will trigger the feature point extraction unit to re-extract the multi-scale features in that area. The re-extraction process uses adjusted neighborhood parameters, and the neighborhood radius is reduced in the feature missing area to increase the feature point density. This cooperative mechanism effectively solves the problem of uneven distribution of feature points on complex curved surfaces. The adaptive filtering processing module and the laser power adjustment unit share the ambient light data. When the ambient light sensing module detects local strong light interference, this information is sent to the laser power adjustment unit and the point cloud processing module at the same time. The laser power adjustment unit increases the laser power in the corresponding area, and the point cloud processing module increases the filter window size accordingly. This cross-module linkage control significantly enhances the robustness of the system in harsh lighting conditions.

[0037] The entire point cloud processing flow adopts a pipeline architecture, and the image filtering, center line extraction, three-dimensional reconstruction, feature extraction and registration optimization links operate in parallel. Each link processes the data output by the previous link and passes the results to the next link. The pipeline design fully utilizes the parallel computing capability of the GPU, enabling the system to process high-frame-rate scanning image data streams in real time. In the point cloud registration process, a multi-resolution registration strategy is adopted. First, initial registration is performed on the coarse resolution point cloud layer, and then gradually refined to the fine layer for accurate registration. This hierarchical strategy significantly improves the registration efficiency and global convergence. At the same time, the registration optimization unit records historical registration error data and automatically starts the re-registration process when it detects abnormal error change trends, ensuring the accuracy of the final point cloud model.

[0038] Taking the scanning measurement of a wind turbine blade as an example, the point cloud processing module starts the feature point extraction process. The blade is 60 meters long, covered with a special coating and has local damage areas. The feature point extraction unit first performs multi-level grid division on the three-dimensional point cloud. The initial layer uses 50mm voxel size downsampling to capture the overall aerodynamic profile of the blade. The second layer switches to 20mm voxel size to focus on the leading edge reinforcement zone geometry. The final layer uses 5mm voxel size to finely identify the blade root bolt hole edge. When calculating the point cloud normal vector at each level, the 200mm neighborhood radius is used to analyze the overall bending trend in the tip area, while the 50mm radius is used to capture the flange mounting hole details in the root area. When the angle between the normal vector of a point and the average value of the neighborhood exceeds 15 degrees, the feature label is triggered. In the serrated structure at the trailing edge of the blade, the feature points are regularly distributed in bands; in the lightning damage area, irregular point clusters are formed. In the feature point descriptor generation stage, for the leading edge erosion-resistant coating area, the extended FPFH algorithm is used to increase the curvature change rate statistical dimension; for the coating shedding area, the surface roughness histogram component is additionally calculated. The generated descriptor database contains 12-dimensional feature vectors, which fully characterize the various geometric abnormalities on the blade surface. When the registration optimization unit processes the point clouds of adjacent stations, the double-constraint cost function constructed in the tip area performs significantly: the feature point matching degree constraint successfully associates the lightning protection features captured by the two stations, and the Euclidean distance of their descriptors is less than 0.1; the point cloud overlap constraint effectively handles the blade waving deformation problem, and through the projection coincidence rate calculation, 0.8mm of elastic deformation is compensated. In the blade root flange area, the bolt hole feature matching distance reaches 0.3, at which time the overlap constraint weight automatically increases to 0.7, and the feature missing problem is compensated through spatial coverage consistency. The adaptive filtering process faces the challenge of high reflectivity on the blade surface. When the environmental light perception module detects changes in coating reflectivity, the spatial filter window size is expanded to 9x9 pixels in the strong reflection area and maintained at 3x3 pixels in the matte area. In a certain scan, the cloud layer suddenly changes, and the light intensity increases by 300 lux within 3 seconds. The filter increases the window size in real time to suppress glare noise, and the morphological processing adjusts the structure element size simultaneously: a 5-pixel disc structure element is used for opening operation in the glare area, and a 3-pixel rectangular structure element is used in the shadow area. The laser stripe centerline extraction encounters a special condition - water erosion pits on the blade leading edge cause the stripe to break, and the morphological processing starts multi-level repair: first use a 3-pixel diameter disc for closing operation to connect the broken points, then use the skeleton extraction algorithm for 5 iterations to obtain a continuous centerline. In the pit depth mutation area, the algorithm automatically marks the abnormal point cloud data, triggering the local rescan mechanism. During the registration optimization process, it is found that the feature point density in the tip area is insufficient, and the feedback mechanism triggers multi-scale feature re-extraction: a 10mm voxel level is added at 40m span position, the neighborhood radius is reduced to 30mm, and 127 new aerodynamic vortex generator feature points are added.The optimization algorithm adopts a hierarchical registration strategy when solving the splicing from the blade root to the blade tip: first, the initial alignment is completed in the coarse resolution point cloud, then the accuracy is gradually improved, and finally the complete point cloud reconstruction of the 56-meter-long blade is realized at an accuracy of 0.05 millimeters. The point cloud processing pipeline maintains real-time performance under GPU acceleration, and when the scanning speed reaches 150 millimeters per second, the single-frame image processing time is controlled within 33 milliseconds. The system specially handles the problem of uneven point cloud density caused by the twist angle change of the blade, automatically increases the feature point detection sensitivity threshold in the high twist angle area, and avoids false matching caused by sparse point cloud. The entire scanning process generates about 12 million point cloud data, and the registration optimization unit performs 37 times of multi-objective optimization calculation. In the lightning conductor area of the blade, the feature point matching degree constraint weight is maintained above 0.6; while in the large curvature web area, the overlap constraint weight is increased to 0.8 to cope with the feature symmetry problem. The final point cloud model completely presents the depth gradient of the leading edge erosion pit and the micro-geometric characteristics of the trailing edge serration.

[0039] In embodiment 3, the multi-station synchronous module receives the real-time pose data stream sent by the robot motion platform through the high-speed industrial Ethernet receiver, the data stream is updated at a frequency of kilohertz, and contains a position vector and a rotation matrix of a coordinate system of a flange at the end of the robot relative to a base coordinate system. The position vector records three-dimensional space coordinates, and the rotation matrix stores attitude information in the form of a rotation vector. The module internally maintains a preset scanning station queue, each station point in the queue contains space coordinates, attitude angles, and time stamp information accurate to milliseconds. When the time stamp difference between the real-time pose data and the time stamp of a certain station point in the queue is less than the synchronization threshold, the synchronous module immediately sends a high-level pulse signal to the trigger port of the laser vision sensing module.

[0040] After receiving the pulse signal, the hardware trigger circuit of the laser vision sensing module synchronously starts the exposure period of the laser projector and the industrial camera. The laser emits the modulated laser stripe within a microsecond-level delay, and the global shutter of the camera is opened at the accurately calculated time. This hardware-level synchronization mechanism ensures that each scanning action occurs at the moment when the robot end reaches the preset space position. The generation process of the scanning station queue is verified through kinematics simulation to ensure that the robot is in a stable state when it reaches each station point, avoiding image blur caused by mechanical vibration. After completing the point cloud registration calculation, the registration optimization unit writes the obtained optimal transformation matrix into the shared memory area. The matrix is expressed in homogeneous coordinates, containing 3x3 rotation matrix components and 3x1 translation vector components. At the same time, the matrix data is sent to the path planning unit of the motion control module through the cross-process communication mechanism. The path planning unit analyzes the matrix parameters and calculates the coordinate deviation. The deviation calculation adopts a spatial transformation model:

[0041] ;

[0042] wherein: preset station coordinates, actual point cloud center coordinates, and respectively represent the rotation transformation and translation transformation, and the calculation result represents the position compensation vector of the current station.

[0043] The motion control module dynamically corrects the subsequent scanning station coordinates based on the compensation vector. The correction process uses a recursive compensation algorithm. First, an error propagation model under the global coordinate system is established. The compensation vector of the current station is decomposed into a translation component and a rotation component. The translation component is directly superimposed on the spatial coordinates of the subsequent station point, and the rotation component is distributed to adjacent stations through an attitude interpolation algorithm. For example, when a 0.5-degree deviation around the Z-axis is detected, the yaw angles of the subsequent ten station points are increased by 0.05 degrees of compensation in turn. This progressive compensation strategy avoids mechanical impact caused by sudden attitude changes. A bidirectional data channel is established between the multi-station synchronization module and the motion control module. When the path planning unit updates the scanning station coordinates, it immediately sends a station queue update instruction to the synchronization module. The synchronization module receives the new queue data and overwrites the original queue. The update process uses a transaction processing mechanism to ensure data integrity. The scanning trigger signal is paused at the moment of queue switching, and the synchronization trigger function is restored after the new queue is fully loaded. The entire update process is completed when the robot moves to the transition path between two stations, without affecting continuous scanning operations. The trigger accuracy of the laser vision sensing module is guaranteed through a time calibration mechanism. A special timing calibration program is executed during system initialization to measure the delay time between the trigger signal and the stable laser output, as well as the time deviation between the camera exposure signal and the mechanical shutter opening. These delay parameters are written into the compensation register of the synchronization module to add corresponding time compensation during actual triggering, ensuring that the laser stripe projection and image acquisition strictly match the robot pose during dynamic scanning. The optimal transformation matrix stored in the registration optimization unit has a version identifier and a timestamp. The motion control module checks the time sequence when receiving the matrix data and only enables compensation when the matrix timestamp is later than the current path instruction generation timestamp, avoiding compensation errors caused by using outdated transformation parameters. The matrix data is stored in a ring buffer structure in the shared memory area. The latest ten sets of transformation matrices are retained for compensation trend analysis.

[0044] When the robot motion platform receives the corrected path instructions, its trajectory planner recalculates the joint space path. The trajectory planning uses a seven-degree polynomial interpolation algorithm to ensure motion smoothness. In the case of large compensation amounts, a transition path point is automatically inserted to avoid overshoot or vibration of the mechanical arm during high-speed motion. All coordinate correction operations are completed in the Cartesian space and converted to joint angle instructions through the robot inverse kinematics. The pose data, transformation matrix, and compensation parameters during the entire scanning process are recorded in the operation log. The log file contains detailed time series and spatial coordinate information for offline analysis of scanning accuracy and system stability. The log record is stored in binary format and can be reconstructed to show the spatial relationship of each scanning station by using a special analysis tool. The multi-station synchronization mechanism and the coordinate compensation mechanism together form a closed-loop control system that eliminates the effects of robot positioning errors and environmental disturbances in real time. The spatial positioning accuracy of the scanning station is always maintained within the sensor measurement accuracy range. This dynamic correction capability is particularly suitable for large structural thermal deformation or foundation settlement and other slowly changing working conditions.

[0045] Example 4: refer to Figure 3 The closed-loop verification module continuously receives the three-dimensional point cloud data stream generated by the point cloud processing module. The data stream is organized in the order of the scanning stations, and each station's point cloud data packet is labeled with spatial coordinates and a timestamp. A point cloud buffer area is established within the module to store the complete point cloud data of the last five scanning stations. When new point cloud data arrives, the verification algorithm automatically calculates the overlap area between the current point cloud and the previous one. The overlap area is determined by the spatial grid mapping method, which divides the point cloud space into cubic grid units and calculates the proportion of points in the same grid unit between the two point clouds. The point cloud overlap area error calculation uses an improved iterative closest point algorithm. This algorithm first randomly samples corresponding point pairs in the overlap area, calculates the Euclidean distance between these points, removes outliers with a distance exceeding three standard deviations, and then calculates the average distance value of the remaining inliers as the error indicator. The error value is recorded in millimeters, and the inlier proportion is calculated as the overlap quality coefficient. All error data and corresponding station numbers are sent to the registration optimization unit. The registration optimization unit maintains a dynamic weight coefficient table that records the weight distribution of the feature point matching degree constraint and the point cloud overlap constraint in the point cloud registration cost function. The initial weight value is set to 0.5 for each. When receiving the point cloud overlap area error data from the closed-loop verification module, the weight adjustment algorithm automatically updates the weight coefficients based on the error value. The update strategy is based on the error threshold mechanism, with two critical values of 0.05 mm and 0.1 mm. When the error is below the lower limit, the overlap constraint is emphasized, and when the error exceeds the upper limit, the matching degree constraint is emphasized.

[0046] The particle swarm optimization algorithm in point cloud registration adopts an adaptive parameter adjustment mechanism, the population size is dynamically set according to the point cloud density, usually one percent of the number of point cloud points, the number of iterations is proportional to the complexity of the point cloud, for the point cloud of a structure with complex geometric features, the number of iterations is appropriately increased to ensure convergence, the particle position vector is defined as a six-dimensional solution space, the first three dimensions represent the rotation Euler angle, and the last three dimensions represent the translation vector, the initial value of the velocity vector is automatically set according to the size of the point cloud. The fitness function, i.e. the point cloud registration cost function, calculates a large amount of computing resources, for this purpose, a parallel computing architecture is adopted, and the fitness calculation task of each particle in the population is distributed to multiple computing cores at the same time, and a space partition tree is used to accelerate the nearest neighbor search during the calculation process, greatly reducing the time overhead of corresponding point matching, and the historical trajectory of the global optimal solution is recorded after each iteration, and the iteration is automatically terminated when it is detected that the optimal solution has not improved for ten consecutive generations. Referring to Table 1, the adjustment of the weight coefficient under different point cloud overlap error conditions is shown.

[0047] Table 1: Point cloud registration weight coefficient adjustment table

[0048] Error range (mm) Matching constraint weight Overlap constraint weight Applicable scenario description <0.05 0.4 0.6 High-precision matching phase 0.05-0.1 0.5 0.5 Standard registration process >0.1 0.7 0.3 Initial coarse registration phase

[0049] The weight coefficient adjustment process has a smooth transition feature, when the error value fluctuates around the critical value, the weight value is calculated by linear interpolation, avoiding the instability of registration caused by sudden change of the coefficient, and the historical weight change trend is recorded, when it is detected that the weight value is frequently and greatly adjusted in a short time, the system self-checking program is triggered to investigate the point cloud quality or environmental interference factors. The inertia weight parameter of the particle swarm optimization algorithm is adjusted in conjunction with the weight coefficient, when the matching degree constraint weight increases, the inertia weight is increased to enhance the global search ability, when the overlap degree constraint weight increases, the inertia weight is reduced to improve the local precision, this linkage mechanism enables the algorithm to dynamically balance the relationship between exploration and utilization according to the actual registration requirements. In addition to calculating the point cloud overlap error, the closed-loop verification module also performs point cloud quality evaluation, the evaluation indicators include point cloud density uniformity, noise point proportion, feature integrity, etc., these quality indicators are packaged together with the error data to provide a more comprehensive decision basis for weight adjustment, when the point cloud quality is detected to be severely degraded, a rescan request is automatically sent to the motion control module. A feedback channel is established between the registration optimization unit and the point cloud processing module, when the registration result after weight adjustment still cannot meet the precision requirements, the feature point extraction unit is triggered to recalculate the feature point descriptor, different feature detection parameters are used, and special attention is paid to the areas with larger previous registration errors, and the newly generated descriptor database is used for a new round of registration calculation.

[0050] The whole closed-loop verification process is synchronized with the point cloud collection, and the registration parameters are optimized in real time during the scanning measurement process to ensure that the final generated overall point cloud model has consistency and high precision. Especially for long-period scanning tasks of large structural parts, this real-time optimization mechanism effectively compensates for mechanical deformation and measurement errors caused by environmental temperature changes. After the particle swarm optimization algorithm outputs the optimal transformation matrix, the registration optimization unit performs result verification, calculates the coincidence degree index and feature matching error of the registered point cloud, and automatically starts re-registration when the verification index exceeds the acceptance standard. The re-registration uses adjusted weight coefficients and optimization parameters to ensure the reliability of the output results, and all registration results are labeled with quality evaluation.

[0051] In embodiment 5, the path re-planning unit of the motion control module continuously monitors the point cloud overlap area error data stream sent by the closed-loop verification module. The data stream is updated at a fixed frequency, and each data packet contains error values, corresponding station number and timestamp information. When the point cloud overlap area error between some adjacent stations is monitored to exceed the preset threshold of 0.2 mm for three consecutive times, the path re-planning unit immediately starts the path re-planning process. The process first locks the spatial area with error exceeding the standard, and determines the three-dimensional spatial range where the error occurs by analyzing the station coordinates. Taking the scanning of the leading edge of an aircraft wing as an example, when the closed-loop verification module detects that the point cloud overlap area error between stations S15 and S16 reaches 0.25 mm, the path re-planning unit retrieves the original scanning path parameters of the two stations from the database. The original path sets the inter-station distance in this area to 120 mm, the scanning speed to 100 mm per second, and the angle interval to 8 degrees. At the same time, the curvature distribution data of the area is obtained, and the leading edge curvature value is displayed as 0.08 mm. The path re-planning unit accesses the updated point cloud registration cost function of the registration optimization unit, extracts the latest adjusted weight combination in the function, and the current weight shows that the feature point matching degree constraint weight is 0.65 and the point cloud overlap degree constraint weight is 0.35. This weight distribution indicates that there is a significant feature matching problem in this area. The unit calls the three-dimensional CAD model interface to load the original design surface data of the wing leading edge area, and identifies that the maximum deviation occurs at the wing transition at 25% chord length by comparing the deviation distribution diagram of the actual point cloud and the design surface.

[0052] Based on the deviation analysis results, the re-planning algorithm recalculates the scanning path parameters for the covered problem area, the new path reduces the station spacing to 60 mm, doubles the scanning station density, reduces the viewing angle interval from 8 degrees to 4 degrees, and reduces the scanning speed to 70 mm per second. Additionally, an extra ring-shaped scanning path is added for the airfoil transition, which includes eight equally distributed station points, forming a 360-degree coverage around the transition. The newly generated dynamic scanning path instructions are sent to the multi-station synchronization module to replace the original path segment after passing the safety check. The synchronization module loads the new path queue immediately after the robot completes the current station S16 scanning. The new queue starts from station S16.1 and ends with the ring-shaped scanning of stations S16.8, finally connecting to the original station S17. The queue switching process is completed during the robot's movement, ensuring motion continuity through real-time pose interpolation. When the robot motion platform executes the new path, adaptive speed control is used at the airfoil transition. The end effector automatically reduces the movement speed when approaching high-curvature areas and restores the standard speed in flat areas. The laser vision sensing module synchronously adjusts the sampling frequency, increasing the image acquisition frame rate from 30 Hz to 60 Hz in high-density station areas to ensure sufficient data points are obtained.

[0053] The path re-planning unit continuously monitors the closed-loop check data after sending the new instruction. When scanning to the new station S16.5, the check module feedbacks that the point cloud overlap area error between this station and the previous station S16.4 has been reduced to 0.12 mm, reaching the system accuracy requirement. The unit records the parameter adjustment records of this re-planning, including the inter-station distance reduction ratio, speed adjustment amplitude, and other key data, forming a re-planning case and storing it in the knowledge base. The re-planning knowledge base uses a tree structure to organize case data, storing it by structure type and surface characteristics. When scanning a new type of structure, the system automatically matches historical cases with similar features, preloading optimized parameters as initial settings. For example, when scanning the tenon of an engine blade, the system automatically calls the path parameter template in the turbine disc mortise scanning case. The event log generated during the entire re-planning process is recorded in detail. The log contains original error data, CAD model analysis screenshots, new path parameter tables, and execution process status codes. The log file is associated with the point cloud data through a timestamp, allowing the user to trace back and view the scanning path version corresponding to each point cloud during post-processing. The path re-planning mechanism has an abnormal interruption processing capability. When the robot encounters unexpected obstacles during the execution of the new path, it triggers an emergency stop in real-time and saves the current state. After the obstacles are removed, it automatically continues scanning from the breakpoint. The scanning data seamlessly connects with the point cloud obtained before the interruption. This robust design ensures the continuity of long-time scanning of large-scale structures. For the segmented scanning scenario of super-large structures, the re-planning unit supports cross-region collaborative optimization. When multiple regions simultaneously report error exceeding, the unit analyzes the spatial relationship of each region, generates a global optimization path scheme, and avoids path conflicts caused by local optimization. For example, in the segmented scanning of a ship body, the scanning path sequences of the deck and the side regions are simultaneously optimized.

[0054] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A robot laser vision three-dimensional scanning measurement system for large structural members, characterized by, The application relates to a robot motion platform, a laser vision sensing module installed at the end of the robot motion platform, an ambient light sensing module, a point cloud processing module and a motion control module. The ambient light sensing module adopts a high-dynamic-range light intensity sensor and is arranged at multiple positions in a measurement field to collect ambient light intensity distribution data; the motion control module generates multi-view dynamic scanning path instructions according to the ambient light intensity distribution data and three-dimensional model data of a large structure to be measured; the robot motion platform executes the multi-view dynamic scanning path instructions to drive the laser vision sensing module to move; the laser vision sensing module projects a laser stripe on the surface of the large structure to be measured and synchronously collects a laser stripe image; and the point cloud processing module performs adaptive filtering processing on the laser stripe image and generates three-dimensional point cloud data. The motion control module comprises a curved surface curvature analysis unit and a path planning unit; the curved surface curvature analysis unit calculates a curved surface curvature distribution based on the three-dimensional model data of the large structure to be measured; the path planning unit generates the multi-view dynamic scanning path instructions according to the curved surface curvature distribution, wherein the multi-view dynamic scanning path instructions contain a scanning speed and a scanning view angle interval; the scanning speed is inversely proportional to the curved surface curvature distribution, and the scanning view angle interval is proportional to the curved surface curvature distribution. The point cloud processing module comprises a feature point extraction unit and a registration optimization unit; the feature point extraction unit performs multi-scale feature point detection on the generated three-dimensional point cloud data and generates a feature point descriptor; the registration optimization unit constructs a point cloud registration cost function based on the feature point descriptor and solves an optimal transformation matrix of the point cloud registration cost function by using a multi-objective optimization algorithm; the point cloud registration cost function contains a feature point matching degree constraint and a point cloud overlap constraint. The laser vision sensing module comprises a laser power adjustment unit; the laser power adjustment unit receives real-time ambient light intensity distribution data sent by the ambient light sensing module and dynamically adjusts the laser emission power of the laser vision sensing module according to a preset laser intensity threshold value; the preset laser intensity threshold value is greater than the maximum value of the current ambient light intensity distribution data and less than half of the maximum emission power of the laser vision sensing module; there is a data interaction channel between the laser power adjustment unit and the path planning unit; when the path planning unit instructs the robot to enter a high-curvature area for fine scanning, the laser power adjustment unit is synchronously notified to focus on the ambient light condition of the area. The adaptive filtering processing comprises the following steps: the point cloud processing module constructs a spatial filter according to the ambient light intensity distribution data, uses the spatial filter to suppress background noise of the laser stripe image, extracts a laser stripe center line through a gray-scale morphological operation, and shares ambient light data with the laser power adjustment unit; when the ambient light sensing module detects local strong light interference, the information is simultaneously sent to the laser power adjustment unit and the point cloud processing module. ​ 2. The robot laser vision 3D scanning measurement system for large structural parts according to claim 1, characterized in that, The system further comprises a multi-station synchronization module; the multi-station synchronization module receives real-time position data sent by the robot motion platform, and triggers a scanning action of the laser vision sensing module according to the real-time position data; a triggering time of the scanning action is synchronized with a timestamp of a time when the robot motion platform reaches a preset scanning station in the multi-view dynamic scanning path instruction.

3. The robot laser vision three-dimensional scanning measurement system for large structural parts according to claim 2, wherein, The registration optimization unit stores the optimal transformation matrix and sends the optimal transformation matrix to the motion control module; the motion control module corrects scanning station coordinates in subsequent multi-view dynamic scanning path instructions according to the optimal transformation matrix.

4. The robot laser vision three-dimensional scanning measurement system for large structural parts according to claim 3, wherein, The system comprises a closed-loop verification module; the closed-loop verification module receives three-dimensional point cloud data generated by the point cloud processing module, calculates a point cloud overlap area error between adjacent scanning stations, and sends the point cloud overlap area error to the registration optimization unit; the registration optimization unit updates a weight coefficient of the point cloud registration cost function according to the point cloud overlap area error.

5. The robot laser vision three-dimensional scanning measurement system for large structural parts according to claim 4, wherein, The multi-objective optimization algorithm adopts a particle swarm optimization algorithm, a particle position vector of the particle swarm optimization algorithm contains rotation and translation parameters, and a fitness function of the particle swarm optimization algorithm is the point cloud registration cost function.

6. The robot laser vision three-dimensional scanning measurement system for large structural parts according to claim 5, wherein, The motion control module comprises a path re-planning unit; when the point cloud overlap area error exceeds a preset error threshold, the path re-planning unit regenerates multi-view dynamic scanning path instructions covering the point cloud overlap area based on the updated point cloud registration cost function.

Images