Intelligent positioning and assembly system based on ship block construction

Abstract

The present application belongs to the technical field of ships or other water vessels, and specifically relates to an intelligent positioning and assembling system based on ship segment construction. A solid-state laser radar and a linear array infrared camera are fixed in the butt joint area of the outer plate of the segment to be assembled to obtain local point cloud data, and the theoretical position of the rib plate and the theoretical trend of the weld are extracted from a digital template as prior topological constraint features. For the point cloud missing area caused by physical shielding of adjacent segments, a non-uniform rational B-spline surface fitting algorithm based on a moving least square method is used to reconstruct the virtual point cloud with the prior topological constraint features as the boundary conditions. The complete local point cloud containing the virtual point cloud is iteratively closest point registered with the target point cloud to calculate the pose deviation, and the deviation is input into a three-degree-of-freedom servo driving mechanism to perform pose adjustment. The present scheme overcomes the defect of missing butt joint feature information caused by structural shielding in a narrow space, and realizes the calculation of segment pose deviation and adjustment of the actuator in a restricted space.

Description

Technical Field

[0001] This invention belongs to the technical field of ships or other watercraft, and specifically relates to an intelligent positioning and assembly system based on the segmented construction of ships. Background Technology

[0002] In the process of ship segment construction, the docking and positioning of segments to be assembled and those already assembled typically employs laser scanning and visual measurement technologies. A conventional approach involves fixing a solid-state lidar and a linear infrared camera in the joint area of ​​the outer plate of the segment to be assembled. These sensors collect local point cloud data between the segments. Subsequently, the collected local point cloud data is matched with the target point cloud in a pre-stored digital template to calculate the rotation and translation transformation matrix between the current assembly pose and the target pose. Finally, the transformation matrix is ​​decomposed into translational and deflection deviations, which are then input into a three-degree-of-freedom servo drive mechanism to perform pose approximation adjustments.

[0003] When performing segmented docking in narrow, confined spaces such as the bottom of a dry dock or the bow and stern, the structures of adjacent segments can physically obstruct each other. Based on the aforementioned conventional technical solutions, the sensor's line of sight is blocked, making it impossible to acquire complete point cloud data of the docking surface. The local point cloud lacking spatial feature information cannot be registered with the target point cloud in the digital sample image, causing conventional technical solutions to fail in such obstructed scenarios. This necessitates manual intervention for auxiliary measurements or multiple blind, trial-and-error adjustments. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent positioning and assembly system based on ship section construction, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The intelligent positioning and assembly system based on ship segment construction includes a solid-state lidar, a linear infrared camera, an edge computing processor, and a three-degree-of-freedom servo drive mechanism. The solid-state lidar and the linear infrared camera are fixed to the outer plate butt joint area of ​​the segment to be assembled. The edge computing processor is communicatively connected to the solid-state lidar, the linear infrared camera, and the three-degree-of-freedom servo drive mechanism. The edge computing processor acquires local three-dimensional point cloud data between the segment to be assembled and the assembled segment, and extracts the theoretical position of the rib plate and the theoretical orientation of the weld in the butt joint area from a pre-stored digital sample drawing as a priori topology. For the missing point cloud regions caused by physical occlusion in the local 3D point cloud data, the prior topological constraint features are used as boundary conditions. A non-uniform rational B-spline surface fitting algorithm based on moving least squares is used to reconstruct the virtual point cloud of the missing point cloud regions. The complete local point cloud containing the virtual point cloud is iteratively registered with the target point cloud in the digital sample image to calculate the rotation and translation transformation matrix between the current assembly pose and the target pose. The rotation and translation transformation matrix is ​​decomposed into translation deviation and deflection deviation and input to the three-degree-of-freedom servo drive mechanism to perform pose iterative approximation adjustment.

[0006] Preferably, when acquiring local 3D point cloud data, the edge computing processor controls the timing synchronization of the laser emitted by the solid-state lidar and the exposure of the linear infrared camera by configuring a hardware synchronization trigger signal. The edge computing processor receives the original distance data output by the solid-state lidar and the grayscale image data output by the linear infrared camera. Using a pre-calibrated joint extrinsic parameter matrix of the lidar and camera, the two-dimensional edge feature points extracted from the grayscale image data are projected onto the three-dimensional spatial coordinate system corresponding to the original distance data to generate a 3D sparse point cloud that incorporates texture edge information. The 3D sparse point cloud and the 3D dense point cloud generated from the original distance data are then spatially fused and aligned to generate the local 3D point cloud data.

[0007] Preferably, in the process of extracting prior topological constraint features, the edge computing processor parses the structural data of the digital sample image, extracts the spatial linear equation parameters of the neutral axis of the rib profile in the butt joint area, extracts the spatial cubic spline curve parameters of the weld centerline on the outer plate surface, discretizes the spatial linear equation parameters and the spatial cubic spline curve parameters into a spatial discrete point set, rasterizes the spatial discrete point set according to the coordinate axis direction of the shipbuilding coordinate system, generates a two-dimensional topological mask matrix containing the rib distribution position code and the weld direction gradient code, and maps the two-dimensional topological mask matrix into a prior topological constraint feature point set in three-dimensional space.

[0008] Preferably, during the reconstruction of the virtual point cloud, the edge computing processor extracts the actual boundary contour point set of the missing region of the point cloud in the local 3D point cloud data, extracts the prior topological constraint feature point set adjacent to the missing region of the point cloud as theoretical boundary conditions, merges and sorts the actual boundary contour point set and the theoretical boundary condition point set to generate a joint boundary point set, calculates the node vector of the non-uniform rational B-spline surface based on the spatial distribution density of the joint boundary point set, uses the 3D coordinates of the joint boundary point set as shape points, obtains the control point grid of the non-uniform rational B-spline surface by solving a system of linear equations, and generates a virtual point cloud covering the missing region of the point cloud based on the control point grid.

[0009] Preferably, during iterative nearest-neighbor registration, the edge computing processor assigns initial weight coefficients to each data point in the complete local point cloud, assigns a first fixed weight coefficient to the real measurement points in the local 3D point cloud data, and assigns a second fixed weight coefficient to the reconstructed data points in the virtual point cloud. In each iteration of searching for the nearest neighbor, the current iteration weight coefficient of the reconstructed data points is dynamically adjusted according to the rate of curvature change of the corresponding matching points in the target point cloud. A registration error cost function is constructed based on the point-to-point distance with the current iteration weight coefficient, and the rotation and translation transformation matrix that makes the registration error cost function converge is solved by singular value decomposition.

[0010] Preferably, during pose iteration approximation adjustment, the edge computing processor decouples the rotation and translation transformation matrix into translational deviation components along the X, Y, and Z axes of the ship construction coordinate system and deflection deviation components around the X, Y, and Z axes. The three-degree-of-freedom servo drive mechanism includes an X-axis lifting hydraulic cylinder, a Y-axis translation hydraulic cylinder, and a Z-axis lateral thrust hydraulic cylinder. The edge computing processor maps the translational deviation components to the target stroke displacement of the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis lateral thrust hydraulic cylinder, and maps the deflection deviation components to the relative speed difference between the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis lateral thrust hydraulic cylinder.

[0011] Preferably, during the timing synchronization process, the edge computing processor has a built-in field-programmable gate array (FPGA). The FPGA outputs a square wave pulse signal with a fixed frequency as the hardware synchronization trigger signal. The rising edge of the square wave pulse signal triggers the start of the laser emitter modulation signal generator inside the solid-state lidar. After passing through a preset delay counter, the square wave pulse signal triggers the global shutter reset signal of the linear array infrared camera. The count value of the preset delay counter is set based on the difference between the round-trip propagation time of the solid-state lidar laser beam and the photoelectric conversion response time of the linear array infrared camera.

[0012] Preferably, in the process of calculating the node vector based on the spatial distribution density of the joint boundary point set, the edge computing processor calculates the chord length parameters between adjacent data points in the joint boundary point set, accumulates the chord length parameters to obtain the total chord length, normalizes each chord length parameter with the total chord length to obtain a parameter sequence, sets the span of the node interval according to the distribution interval of the parameter sequence, inserts a second-order re-node in the parameter interval corresponding to the missing area of ​​the point cloud, and uses the parameter sequence after inserting the second-order re-node as the node vector of the non-uniform rational B-spline surface.

[0013] Preferably, during the process of dynamically adjusting the current iteration weight coefficient of the reconstructed data point, the edge computing processor extracts the corresponding matching point in the target point cloud that matches the reconstructed data point, calculates the normal vector dispersion of the corresponding matching point within a preset neighborhood range, inputs the normal vector dispersion into a pre-constructed inverse proportional mapping function, calculates the weight decay coefficient of the reconstructed data point, and uses the product of the second fixed weight coefficient and the weight decay coefficient as the current iteration weight coefficient of the reconstructed data point in the current iteration.

[0014] Preferably, during the mapping to relative velocity differences, the edge computing processor constructs a Jacobian matrix containing the kinematic parameters of the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis side-push hydraulic cylinder. The deflection deviation component is input as an input vector into the pseudo-inverse matrix of the Jacobian matrix to calculate the instantaneous velocity vectors of the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis side-push hydraulic cylinder. Based on the instantaneous velocity vectors and the target stroke displacement, a cubic polynomial interpolation algorithm is used to generate the servo motor control pulse sequence of the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis side-push hydraulic cylinder.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This scheme extracts the theoretical positions of ribs and the theoretical orientation of welds from digital sample images to form prior topological constraint features. In areas where point clouds are missing due to physical occlusion of adjacent segments, the prior topological constraint features are used as theoretical boundary conditions, and a non-uniform rational B-spline surface fitting algorithm based on moving least squares is employed to reconstruct the virtual point cloud. The complete local point cloud containing the virtual point cloud is iteratively registered with the target point cloud at the nearest point, and the pose deviation is calculated. This overcomes the defect of incomplete spatial feature information of the docking surface caused by structural occlusion in narrow and confined spaces, and realizes the calculation of segmented pose deviation and the pose adjustment of the actuator in confined spaces.

[0016] 2. By controlling the hardware synchronization triggering timing of the solid-state lidar and the linear array infrared camera, the consistency of spatial alignment between the 3D sparse point cloud and the 3D dense point cloud with fused texture edge information is ensured. During the registration iteration process, the current iteration weight coefficient of the reconstructed data points is dynamically adjusted according to the curvature change rate of the corresponding matching points in the target point cloud. The weight attenuation coefficient is calculated by combining the normal vector dispersion of the corresponding matching points within the preset neighborhood range. This reduces the interference of the reconstructed virtual point cloud region on the overall registration error cost function, ensures the reliability of the pose deviation calculation results, and realizes the accurate mapping of translational deviation and deflection deviation along each coordinate axis of the ship construction coordinate system to the target stroke displacement and relative speed difference of each hydraulic cylinder. Attached Figure Description

[0017] Figure 1 This is an overall execution flowchart of an intelligent positioning and assembly system based on ship section construction provided in an embodiment of the present invention; Figure 2 This is a flowchart of a local three-dimensional point cloud data acquisition process provided in an embodiment of the present invention; Figure 3 A flowchart for prior topological constraint feature extraction provided in an embodiment of the present invention; Figure 4 A flowchart of virtual point cloud reconstruction provided in an embodiment of the present invention; Figure 5 A flowchart of iterative nearest point registration with dynamic weight adjustment provided in an embodiment of the present invention; Figure 6 The flowchart of pose iterative approximation adjustment control provided in the embodiment of the present invention. Detailed Implementation

[0018] refer to Figure 1 The technical solution disclosed in this embodiment provides a detailed description of the intelligent positioning and assembly scenario during the segmented construction of ships. Those skilled in the art can implement the corresponding technical solution based on the following content, solving the technical problems of incomplete point cloud data and pose registration failure caused by structural occlusion in narrow, confined spaces. In this embodiment, all coordinate systems adopt a unified ship construction coordinate system. The X-axis of the coordinate system represents the ship's length, the Y-axis represents the ship's width, and the Z-axis represents the ship's depth. The origin of the coordinate system is located at the intersection of the ship's stern perpendicular and the baseline.

[0019] In one embodiment, the intelligent positioning and assembly system based on ship segment construction includes a solid-state lidar, a linear infrared camera, an edge computing processor, and a three-degree-of-freedom servo drive mechanism. The solid-state lidar and linear infrared camera are fixed to the outer plate joint area of ​​the segment to be assembled. Specifically, the segment to be assembled can be any one of a double-bottom segment, a side segment, or a deck segment. The outer plate joint area is the edge of the outer plate at the joint end face between the segment to be assembled and the already assembled segment. Multiple sets of solid-state lidar and linear infrared cameras are arranged at intervals along the joint. The optical axis angle between each set of solid-state lidar and linear infrared camera is fixed, and their field of view completely covers the structural areas such as ribs, outer plates, and stiffeners within a predetermined width on both sides of the joint. The edge computing processor communicates with the solid-state lidar, linear infrared camera, and three-degree-of-freedom servo drive mechanism. Specifically, the edge computing processor establishes a full-duplex communication link with each hardware device through an industrial Ethernet bus. The communication cycle is consistent with the line scan cycle of the solid-state lidar and the line exposure cycle of the linear infrared camera, ensuring real-time transmission of control commands and collected data.

[0020] The edge computing processor acquires local 3D point cloud data between the segment to be assembled and the already assembled segment. Specifically, the segment to be assembled is moved by hoisting equipment to the docking station above the already assembled segment within the dock. The distance between the docking end faces of the segment to be assembled and the already assembled segment is within the effective ranging range of the solid-state lidar. The edge computing processor sends a sampling start command to the solid-state lidar and the linear infrared camera. The solid-state lidar emits a line laser along the docking seam, scans the structural surface of the docking end face, and outputs raw distance data. The linear infrared camera simultaneously acquires infrared grayscale images of the docking end face and outputs grayscale image data. The edge computing processor preprocesses and spatially fuses the acquired raw distance data and grayscale image data to generate local 3D point cloud data with both high-density spatial coverage and high-precision edge features.

[0021] The edge computing processor extracts the theoretical positions of the ribs and the theoretical orientation of the welds in the butt joint area from a pre-stored digital sample image as prior topological constraint features. The pre-stored digital sample image is a digital model file exported from the ship's 3D design model, conforming to shipbuilding standards, and contains structural geometric data, coordinate system definitions, and assembly datum information for all ship sections. The edge computing processor parses the structural data of the digital sample image and extracts the model fragment corresponding to the butt joint area. This model fragment contains the theoretical geometric data of all ribs, outer plates, and welds at the butt joint positions of the sections to be assembled and those already assembled. From this, the spatial position of the central axis of the rib profile is extracted as the theoretical position of the rib, and the spatial orientation of the centerline of the outer plate butt weld is extracted as the theoretical orientation of the weld. The above geometric data is transformed into a feature point set containing spatial positions and topological relationships, which serves as prior topological constraint features for subsequent reconstruction constraints of missing point cloud areas.

[0022] The edge computing processor addresses the issue of missing point cloud regions caused by physical occlusion in local 3D point cloud data. Using prior topological constraints as boundary conditions, it reconstructs virtual point clouds for these missing regions using a non-uniform rational B-spline surface fitting algorithm based on moving least squares. Specifically, the edge computing processor preprocesses the acquired local 3D point cloud data, including statistical filtering to remove outliers and voxel downsampling to reduce point cloud density. Then, it segments the effective measurement region and missing region in the point cloud data using a region growing algorithm. Spatial regions with consecutive data points without any data points are marked as missing point cloud regions. These missing regions are caused by adjacent structural components blocking the emitted beam of the solid-state lidar and the imaging line of the linear infrared camera, preventing the acquisition of 3D measurement data for that region. The edge computing processor extracts the actual boundary contour point set of the missing point cloud region and simultaneously extracts the feature point set adjacent to the missing region from the prior topological constraints as theoretical boundary conditions. The actual boundary contour point set and the theoretical boundary condition point set are then fused and sorted to generate a joint boundary point set. Using the joint boundary point set as the model points, the moving least squares method is used to smooth the model points and eliminate boundary fluctuations caused by measurement noise. Then, the control point mesh of the surface is solved by the non-uniform rational B-spline surface fitting algorithm. Uniform sampling is performed on the fitted surface to generate a virtual point cloud covering the missing area of ​​the point cloud. The density of the virtual point cloud is consistent with the density of the real measured points in the local three-dimensional point cloud data.

[0023] The edge computing processor performs iterative nearest-neighbor registration between the complete local point cloud containing the virtual point cloud and the target point cloud in the digital sample image to calculate the rotation and translation transformation matrix between the current assembly pose and the target pose. The target point cloud in the digital sample image is a theoretical 3D point cloud extracted from a model fragment of the docking seam area in the digital sample image, and the coordinate system of this point cloud is the global coordinate system for shipbuilding. The edge computing processor first transforms the complete local point cloud to the global coordinate system for shipbuilding, and then uses an iterative nearest-neighbor algorithm for registration. In each iteration, for each data point in the complete local point cloud, a corresponding nearest neighbor matching point is searched in the target point cloud to construct matching point pairs. Based on the spatial distance of the matching point pairs, a registration error cost function is constructed, and the rotation and translation transformation matrix that minimizes the cost function is solved. The iteration process is repeated until the change in the cost function is less than a preset convergence threshold, or the number of iterations reaches a preset maximum number of iterations. The final rotation and translation transformation matrix is ​​then output. This matrix is ​​a 4×4 homogeneous transformation matrix that describes the spatial transformation relationship between the current actual pose of the segment to be assembled and the target pose defined in the digital sample image.

[0024] The edge computing processor decomposes the rotation-translation transformation matrix into translational and deflection deviations and inputs these deviations into a three-degree-of-freedom servo drive mechanism to perform iterative pose approximation adjustment. Specifically, the rotation-translation transformation matrix contains a 3×3 rotation matrix and a 3×1 translation vector. The edge computing processor decouples the homogeneous transformation matrix, decomposing it into translational deviation components along the three orthogonal axes of the shipbuilding coordinate system, and deflection deviation components around these axes. The three-degree-of-freedom servo drive mechanism includes an X-axis lifting hydraulic cylinder, a Y-axis translation hydraulic cylinder, and a Z-axis side thrust hydraulic cylinder, corresponding to displacement adjustments along the three axes of the shipbuilding coordinate system. The edge computing processor maps the translational deviation components to the target stroke displacement of each hydraulic cylinder and the deflection deviation components to the relative velocity difference between the cylinders, generating servo control commands and sending them to the three-degree-of-freedom servo drive mechanism. The three-degree-of-freedom servo drive mechanism drives each hydraulic cylinder to perform corresponding extension and retraction actions according to the control commands, causing the assembly segment to adjust its pose, thus approximating the target pose. After a pose adjustment is completed, the complete process of point cloud data acquisition, prior feature extraction, point cloud reconstruction, registration calculation, and deviation decomposition is repeated to perform the next pose iteration adjustment until both translation deviation and deflection deviation are less than the assembly tolerance threshold required by shipbuilding specifications.

[0025] Table 1. Definition of Assembly Attitude Parameters in Ship Construction Coordinate System

[0026] Table 1 defines the coordinate system references and parameter meanings used in all pose calculation processes in this embodiment, ensuring the consistency of the coordinate system during pose transformation calculations and avoiding calculation deviations caused by inconsistent coordinate system definitions. This embodiment fully realizes intelligent positioning and assembly in the ship segment construction process. By reconstructing the virtual point cloud of the occluded area through prior topological constraint features, it solves the problem of incomplete point cloud data caused by structural occlusion in narrow and confined spaces, and realizes accurate calculation and iterative adjustment of the pose of the segments to be assembled.

[0027] refer to Figure 2In another embodiment, when acquiring local 3D point cloud data, the edge computing processor controls the timing synchronization of the solid-state lidar's emitted line laser and the linear infrared camera's exposure by configuring a hardware synchronization trigger signal. The edge computing processor incorporates a field-programmable gate array (FPGA). The hardware logic unit of the FPGA outputs a fixed-frequency square wave pulse signal as the hardware synchronization trigger signal. The frequency of the square wave pulse signal is consistent with the line scan frequency of the solid-state lidar and the line exposure frequency of the linear infrared camera. The rising edge of the square wave pulse signal is split into two paths. The first path is directly input to the trigger signal interface of the solid-state lidar, triggering the modulation signal generator of the laser emitter inside the solid-state lidar to start, causing the laser emitter to emit a line laser beam according to a preset modulation frequency. The line laser beam is projected onto the structural surface of the docking end face, and after diffuse reflection, returns to the receiving unit of the solid-state lidar. The receiving unit calculates the original distance data corresponding to each laser emission direction using the time-of-flight method. The second square wave pulse signal is input to the preset delay counter inside the field programmable gate array. After the preset delay count, it is output to the trigger signal interface of the linear infrared camera, triggering the global shutter reset signal of the linear infrared camera, so that all pixel units of the linear infrared camera start exposure simultaneously, acquire infrared grayscale images of the docking end face structure surface, and output grayscale image data.

[0028] The preset delay counter value is set based on the difference between the round-trip propagation time of the solid-state lidar laser beam and the photoelectric conversion response time of the linear infrared camera. The corresponding delay time calculation formula is as follows:

[0029] in, The preset delay time corresponds to the time length corresponding to the count value of the delay counter; This represents the round-trip propagation time of the solid-state lidar laser beam at the maximum ranging distance. This refers to the response time of the photoelectric conversion unit in the linear infrared camera. The fixed offset time for hardware circuit transmission is obtained from the transmission delay test of the hardware link. The delay time set by the above formula can ensure that the exposure time of the linear infrared camera and the time when the laser beam of the solid-state lidar arrives at the surface being measured are completely synchronized, eliminating point cloud motion distortion and spatial misalignment caused by asynchronous sampling timing.

[0030] The edge computing processor receives raw distance data from a solid-state lidar and grayscale image data from a linear infrared camera. Using a pre-calibrated joint extrinsic parameter matrix of the lidar and camera, it projects the 2D edge feature points extracted from the grayscale image data onto the 3D spatial coordinate system corresponding to the raw distance data, generating a 3D sparse point cloud that incorporates texture edge information. The pre-calibrated joint extrinsic parameter matrix of the lidar and camera is obtained by acquiring point cloud data from the lidar and image data from the linear infrared camera under multiple preset calibration poses using a calibration board, and solved through a hand-eye calibration algorithm. It describes the spatial transformation relationship between the 3D coordinate system of the lidar and the 2D pixel coordinate system of the linear infrared camera, including a 3×3 rotation matrix. Translation vector of 3×1 This forms a 4×4 homogeneous joint extrinsic parameter matrix. .

[0031] The edge computing processor extracts edge features from grayscale image data using the Canny edge detection algorithm. It extracts strong textured edge features such as rib edges, weld edges, and profile edges to obtain the pixel coordinates of two-dimensional edge feature points. The two-dimensional pixel coordinates are converted to three-dimensional coordinates in the camera coordinate system using the camera intrinsic parameter matrix, and then converted to three-dimensional coordinates in the radar coordinate system using the joint extrinsic parameter matrix. The corresponding projection calculation formula is as follows:

[0032] in, These are three-dimensional coordinates in the radar coordinate system. This is the joint extrinsic parameter matrix for radar and camera. This is the intrinsic parameter matrix of the linear infrared camera; These are the pixel coordinates of the two-dimensional edge feature points; The depth value corresponding to the feature point in the camera coordinate system is obtained by interpolation of the original distance data output by the solid-state LiDAR under the same line of sight.

[0033] All two-dimensional edge feature points are converted into a three-dimensional coordinate point set in the radar coordinate system using the above projection formula. This point set is a three-dimensional sparse point cloud that incorporates texture edge information. This point cloud contains the spatial location information of strong texture edges in the grayscale image, which makes up for the lack of measurement of low reflectivity edge structures in the original range data of solid-state lidar.

[0034] The edge computing processor spatially fuses and aligns a 3D sparse point cloud with a 3D dense point cloud generated from the original distance data to produce local 3D point cloud data. The original distance data, after coordinate transformation, generates a 3D dense point cloud in a radar coordinate system. This point cloud has high spatial density but insufficient accuracy in describing edge structures. The 3D sparse point cloud, on the other hand, has high edge localization accuracy but lower point density. The edge computing processor employs a feature-matching-based point cloud registration algorithm to spatially fuse and align the 3D sparse and dense point clouds. Specifically, it extracts edge feature points from the 3D dense point cloud and matches them with feature points from the 3D sparse point cloud to construct matching point pairs. It then solves for the fine-tuning transformation matrix between the two point clouds, transforming the 3D sparse point cloud to the same coordinate system as the 3D dense point cloud. Finally, it merges the two point clouds, removing duplicate spatial points to obtain local 3D point cloud data that simultaneously possesses high-density spatial coverage and high-precision edge features.

[0035] Table 2. Comparison of Feature Matching Errors in Point Cloud Fusion Process

[0036] Table 2 records the feature matching errors during the fusion and alignment process of the 3D sparse point cloud and the 3D dense point cloud in this embodiment. This is used to verify the spatial alignment accuracy of the fused point cloud and ensure the geometric accuracy of the local 3D point cloud data. This embodiment achieves sampling timing synchronization between the solid-state lidar and the linear infrared camera through a hardware synchronization trigger signal, eliminating motion distortion and spatial misalignment caused by asynchronous sampling timing. Joint calibration of the radar and camera achieves the fusion of texture edge information and distance information, improving the edge structure description accuracy of the local 3D point cloud data and providing a high-quality data source for subsequent point cloud reconstruction and registration calculations.

[0037] refer to Figure 3 and Figure 4 In another embodiment, during the extraction of prior topological constraint features, the edge computing processor parses the structural data of the digital sample image, extracts the spatial linear equation parameters of the vertical axis of the rib profiles within the butt joint region, and extracts the spatial cubic spline curve parameters of the weld centerline on the outer plate surface. The structural data of the digital sample image is three-dimensional model data conforming to the STEP standard. The edge computing processor extracts the geometric model of all rib profiles within the butt joint region through a geometric analysis engine. The rib profiles are rolled profiles with symmetrical cross-sections, where the vertical axis is a spatial straight line formed by extending the symmetrical centerline of the profile cross-section along the length of the profile. For each rib, the three-dimensional coordinates of the two endpoints of its vertical axis are extracted to generate a spatial linear equation, expressed as:

[0038] in, The three-dimensional coordinates of the starting point of the vertical axis in the rib plate; is the direction vector component of the neutral axis in the rib; Let be the three-dimensional coordinates of any point on the neutral axis.

[0039] The edge computing processor extracts the geometric data of the weld centerline on the outer plate surface. The weld centerline is the theoretical centerline of the butt joint of the outer plate, located on the intersection line of the outer plate surfaces, and is a smooth spatial curve described using a cubic spline curve. The coordinates of the shape points, node vectors, and basis function coefficients of the cubic spline curve are extracted to obtain the parametric equation of the spatial cubic spline curve, expressed as:

[0040] in, Let be the three-dimensional coordinate vector of any point on the center line of the weld; The control point vector of the cubic spline curve; The basis functions are cubic B-spline functions. represents the normalization parameter of the curve.

[0041] The edge computing processor discretizes the parameters of the spatial linear equation and the spatial cubic spline curve into a set of discrete points. This set is then rasterized according to the coordinate axes of the shipbuilding coordinate system, generating a two-dimensional topological mask matrix containing codes for the rib distribution location and the weld orientation gradient. This two-dimensional topological mask matrix is ​​then mapped to a set of prior topological constraint feature points in three-dimensional space. Specifically, the spatial linear equation of the rib and the spatial cubic spline curve of the weld are discretized with equal parameters. Following a preset discretization step size, a series of three-dimensional coordinate points are sampled on the line and curve, respectively. All sampled points are merged to obtain a set of discrete points. This set of discrete points is projected onto the XOY horizontal reference plane of the shipbuilding coordinate system, resulting in a two-dimensional projected point set. The XOY plane is then uniformly rasterized according to a preset grid size, with each grid cell corresponding to an element of a two-dimensional matrix. For each grid cell, if it contains a projection point of a rib, the rib distribution position encoding of that matrix element is set to 1; otherwise, it is set to 0. If it contains a projection point of a weld, the angle between the tangent direction of the weld curve at that point and the X-axis is calculated. The normalized value of the angle is used as the weld orientation gradient encoding and assigned to the corresponding matrix element, ultimately generating a two-dimensional topological mask matrix. Each non-zero element in the two-dimensional topological mask matrix is ​​mapped back to its corresponding theoretical height position in three-dimensional space according to the inverse transformation of projection, resulting in a set of prior topological constraint feature points in three-dimensional space. This set of points contains the theoretical spatial positions and topological relationships of the ribs and welds within the joint area, serving as boundary constraints for subsequent point cloud reconstruction.

[0042] In the process of reconstructing the virtual point cloud, the edge computing processor extracts the actual boundary contour point set of the missing region in the local 3D point cloud data, and extracts the prior topological constraint feature point set adjacent to the missing region as theoretical boundary conditions. The actual boundary contour point set and the theoretical boundary condition point set are then fused and sorted to generate a joint boundary point set. Specifically, the edge computing processor uses the alphashapes algorithm to extract the actual boundary contour point set of the missing region. This point set represents the intersection of the effectively measured point cloud and the missing region, containing the actual geometric boundary of the missing region. In the prior topological constraint feature point set, feature points with a spatial distance less than a preset threshold from the actual boundary contour point set are extracted and used as the theoretical boundary condition point set adjacent to the missing region. This point set contains the theoretical structural boundary corresponding to the missing region. The actual boundary contour point set and the theoretical boundary condition point set are merged and sorted according to the curvature change of their spatial positions, so that the fused point set is arranged sequentially along the direction of the boundary contour, eliminating spatial intersections and overlaps of the point sets, and generating a joint boundary point set. This point set simultaneously contains the boundary information of actual measurements and the boundary constraints of theoretical design, providing accurate boundary conditions for surface fitting.

[0043] The edge computing processor calculates the node vectors of the non-uniform rational B-spline surface based on the spatial distribution density of the joint boundary point set. Using the 3D coordinates of the joint boundary point set as shape points, it obtains the control point mesh of the non-uniform rational B-spline surface by solving a system of linear equations. Based on this control point mesh, it generates a virtual point cloud covering the missing areas of the point cloud. Specifically, the joint boundary point set is parameterized using a chord length parameterization method. The chord length parameters between adjacent data points in the joint boundary point set are calculated, and these parameters are accumulated to obtain the total chord length. The total chord length is then used to normalize each chord length parameter to obtain a parameter sequence. The corresponding chord length parameterization calculation formula is as follows:

[0044] in, The normalization parameter corresponding to the kth type value point; Let k be the three-dimensional coordinate vector of the k-th point in the set of joint boundary points; The Euclidean distance between two adjacent data points is the chord length. The total chord length of the joint boundary point set is the sum of the chord lengths of all adjacent points; This represents the number of points in the joint boundary point set.

[0045] Based on the distribution range of the parameter sequence, the span of the node interval is set. Secondary repetition nodes are inserted within the parameter interval corresponding to the missing regions of the point cloud. The parameter sequence after inserting these secondary repetition nodes is used as the node vector of the non-uniform rational B-spline surface. For a p-th degree non-uniform rational B-spline surface, the node vector... Among them, the parameter range corresponding to the missing area of ​​the point cloud Insert a quadratic repeating node into the interval so that the node repeatability is p, thus ensuring the interpolation properties of the surface at the boundary.

[0046] The parametric equation of a non-uniform rational B-spline surface is expressed as follows:

[0047] in, Let be the three-dimensional coordinate vector of any point on the non-uniform rational B-spline surface; The control point vectors of the control point grid; The weight factor corresponding to the control point; Let p be the B-spline basis function in the u direction; Let q be the B-spline basis function in the v direction; These are two normalized parameters for the surface.

[0048] Using the 3D coordinates of the joint boundary point set as the model points, and substituting them into the parametric equations of the non-uniform rational B-spline surface, a system of linear equations is constructed. This system is then solved using the LU decomposition method to obtain the vectors of all control points in the control point mesh and their corresponding weight factors. The solved control point mesh is then smoothed using the moving least squares method to eliminate the influence of measurement noise on the control point positions. The corresponding fitting formula is as follows:

[0049] in, The smoothed control point vector; For compactly supported kernel functions; These are the parametric coordinates of the surface; These are the parameter coordinates corresponding to the kth type value point; Let k be the three-dimensional coordinate vector of the k-th shape value point; This represents the total number of type value points.

[0050] On the smoothed non-uniform rational B-spline surface obtained by the solution, uniform grid sampling is performed in the u and v parameter directions according to the preset sampling step size to obtain a series of three-dimensional coordinate points. These points constitute a virtual point cloud covering the missing area of ​​the point cloud. The density of the virtual point cloud is consistent with the density of the real measured point cloud in the local three-dimensional point cloud data.

[0051] Table 3. Parameters corresponding to the fitting points and control points of the non-uniform rational B-spline surface.

[0052] Table 3 records the correspondence and parameters between shape points and control points during the non-uniform rational B-spline surface fitting process in this embodiment. This is used to verify the interpolation accuracy of the surface fitting and ensure the geometric consistency between the reconstructed virtual point cloud and the theoretical design model. This embodiment extracts the theoretical geometric parameters of the ribs and welds by analyzing the geometric data of the digital sample image and generates a set of prior topological constraint feature points. This provides accurate theoretical boundary conditions for the reconstruction of the missing point cloud regions. The node vectors of the non-uniform rational B-spline surface are generated using the chord length parameterization method, and the surface fitting and virtual point cloud reconstruction of the missing regions are achieved by combining the moving least squares method. This ensures the geometric continuity and accuracy of the reconstructed point cloud and provides complete point cloud data for subsequent point cloud registration.

[0053] refer to Figure 5 and Figure 6 In another embodiment, during iterative nearest-point registration, the edge computing processor assigns initial weight coefficients to each data point in the complete local point cloud. The actual measured points in the local 3D point cloud data are assigned a first fixed weight coefficient, while the reconstructed data points in the virtual point cloud are assigned a second fixed weight coefficient. The complete local point cloud consists of two parts: actual measured points and reconstructed virtual points. The actual measured points are point cloud data actually acquired by solid-state LiDAR and linear infrared cameras, possessing high measurement reliability. Therefore, the value of the first fixed weight coefficient is greater than the second fixed weight coefficient, making the influence of actual measured points on the registration result greater than that of virtual points during the registration process.

[0054] In each iteration of the nearest neighbor search, the edge computing processor dynamically adjusts the current iteration weight coefficients of the reconstructed data points based on the rate of curvature change of the corresponding matching point in the target point cloud. It constructs a registration error cost function based on the point-to-point distance with the current iteration weight coefficients and solves for the rotation and translation transformation matrix that makes the registration error cost function converge using singular value decomposition. Specifically, in each iteration, for each data point in the complete local point cloud, the point with the closest Euclidean distance in the target point cloud is searched as the corresponding matching point, and a matching point pair is constructed. For the matching point pair corresponding to a real measurement point, its weight coefficient remains unchanged at the first fixed weight coefficient. For the matching point pair corresponding to a virtual point, all points within a preset neighborhood of the corresponding matching point in the target point cloud are extracted, the rate of curvature change at that matching point is calculated, and the current iteration weight coefficient of the virtual point is dynamically adjusted based on the rate of curvature change.

[0055] During the dynamic adjustment of the current iteration weight coefficients of the reconstructed data points, the edge computing processor extracts the corresponding matching points in the target point cloud that match the reconstructed data points, calculates the normal vector dispersion of the corresponding matching points within a preset neighborhood, inputs the normal vector dispersion into a pre-constructed inverse proportional mapping function, calculates the weight decay coefficient of the reconstructed data points, and uses the product of the second fixed weight coefficient and the weight decay coefficient as the current iteration weight coefficient of the reconstructed data points in the current iteration. The formula for calculating the normal vector dispersion is:

[0056] in, The normal vector dispersion of the corresponding matching point; The number of points within the preset neighborhood; Let be the unit normal vector of the k-th point in the neighborhood; It is the average of the unit normal vectors of all points in the neighborhood.

[0057] The formula for calculating the weight decay coefficient is:

[0058] in, This is the weight decay coefficient; This is a preset proportionality coefficient; The discreteness of the normal vector.

[0059] The formula for calculating the current iteration weight coefficient of the reconstructed data points is:

[0060] in, The current iteration weight coefficient of the virtual point in the t-th iteration; This is the second fixed weighting coefficient; This is the weight decay coefficient.

[0061] The greater the dispersion of the normal vector, the greater the change in surface curvature at the matching point, indicating a region with strong geometric features. This results in lower matching reliability of the virtual point, a smaller weight decay coefficient, and a lower current iteration weight coefficient for the virtual point, thus reducing the interference of the virtual point on the registration result in the region with strong features. Conversely, the smaller the dispersion of the normal vector, the smoother the surface at the matching point, resulting in higher matching reliability of the virtual point. The weight decay coefficient is closer to 1, and the weight of the virtual point is closer to the initial second fixed weight coefficient.

[0062] The edge computing processor constructs a registration error cost function based on the point-to-point distance with the current iteration weight coefficients, expressed as:

[0063] in, The registration error cost function; It is a 3×3 rotation matrix; It is a 3×1 translation vector; This represents the total number of matching point pairs; Let be the weight coefficient of the i-th matching point pair, for the matching point pair of the actual measurement points. For matching point pairs of virtual points, ; The three-dimensional coordinate vector of the i-th point in the complete local point cloud; In the target point cloud The three-dimensional coordinate vector of the corresponding matching point.

[0064] Solving the cost function using singular value decomposition yields the desired result. Minimal rotation matrix With translation vector Specifically, the weighted centroids of the matching point pairs are first calculated, then a covariance matrix is ​​constructed. Singular value decomposition is performed on the covariance matrix to obtain the rotation matrix. The translation vector is then calculated using the centroids, ultimately resulting in a 4×4 homogeneous rotation-translation transformation matrix. This iterative process is repeated until the change in the cost function is less than a preset convergence threshold, at which point the final rotation-translation transformation matrix is ​​output.

[0065] During pose iteration and approximation adjustments, the edge computing processor decouples the rotation and translation transformation matrix into translational deviation components along the X, Y, and Z axes of the shipbuilding coordinate system and deflection deviation components around the X, Y, and Z axes. The three-degree-of-freedom servo drive mechanism includes an X-axis lifting hydraulic cylinder, a Y-axis translation hydraulic cylinder, and a Z-axis lateral thrust hydraulic cylinder. The edge computing processor maps the translational deviation components to the target stroke displacement of the X-axis lifting hydraulic cylinder, Y-axis translation hydraulic cylinder, and Z-axis lateral thrust hydraulic cylinder, and maps the deflection deviation components to the relative speed difference between the X-axis lifting hydraulic cylinder, Y-axis translation hydraulic cylinder, and Z-axis lateral thrust hydraulic cylinder. Specifically, the X-axis lifting hydraulic cylinder is used to adjust the lifting displacement of the segment to be assembled along the Z-axis, the Y-axis translation hydraulic cylinder is used to adjust the longitudinal displacement of the segment to be assembled along the X-axis, and the Z-axis lateral thrust hydraulic cylinder is used to adjust the lateral displacement of the segment to be assembled along the Y-axis. The 3×3 rotation matrix in the rotation-translation transformation matrix is ​​decomposed into deflection angles about the X, Y, and Z axes through the Rodrigues transform, i.e., deflection deviation components, denoted as […]. The translation vector is decomposed into translational deviation components along the X, Y, and Z axes, denoted as follows: The translational deviation components are mapped to the target stroke displacement of each hydraulic cylinder. Specifically, The target stroke displacement corresponding to the X-axis lifting hydraulic cylinder. The target stroke displacement corresponding to the Y-axis translational hydraulic cylinder The target stroke displacement of the Z-axis side thrust hydraulic cylinder.

[0066] In the process of mapping to relative velocity differences, the edge computing processor constructs a Jacobian matrix containing the kinematic parameters of the X-axis lifting hydraulic cylinder, the Y-axis translating hydraulic cylinder, and the Z-axis lateral thrust hydraulic cylinder. The deflection deviation component is input as an input vector into the pseudo-inverse matrix of the Jacobian matrix to calculate the instantaneous velocity vectors of the X-axis lifting hydraulic cylinder, the Y-axis translating hydraulic cylinder, and the Z-axis lateral thrust hydraulic cylinder. Based on the instantaneous velocity vectors and the target stroke displacement, a cubic polynomial interpolation algorithm is used to generate the servo motor control pulse sequence for the X-axis lifting hydraulic cylinder, the Y-axis translating hydraulic cylinder, and the Z-axis lateral thrust hydraulic cylinder. The formula for calculating the instantaneous velocity vector is:

[0067] in, Let be the instantaneous velocity vectors of the three hydraulic cylinders. Let X be the instantaneous velocity of the lifting hydraulic cylinder in the X direction. Let be the instantaneous velocity of the Y-axis translational hydraulic cylinder. The instantaneous velocity of the Z-axis lateral thrust hydraulic cylinder; Jacobian matrix The Moore-Penrose pseudo-inverse matrix; It is the rate of change vector of the deflection deviation component, which is calculated from the deflection deviation component and the preset adjustment period.

[0068] The expression for the cubic polynomial interpolation algorithm is:

[0069] in, Let be the stroke displacement of the hydraulic cylinder at time t; This is the preset adjustment cycle; The coefficients of the cubic polynomial are determined by the boundary conditions: ,in This represents the current stroke of the hydraulic cylinder. This represents the target stroke of the hydraulic cylinder.

[0070] Based on the curve of hydraulic cylinder stroke versus time obtained by cubic polynomial interpolation, the displacement increment of hydraulic cylinder in each control cycle is calculated. The displacement increment is converted into the number of control pulses of servo motor, generating a servo motor control pulse sequence, which is sent to the servo driver of the three-degree-of-freedom servo drive mechanism. The servo driver drives the servo motor to run according to the control pulse sequence, which drives the hydraulic cylinder to perform the corresponding extension and retraction action, thereby realizing the iterative approximation adjustment of the pose of the segment to be assembled.

[0071] Table 4. Correspondence between weight coefficients and registration errors during iterative registration.

[0072] Table 4 records the correspondence between the weight coefficients and registration errors in each iteration of the weighted iterative nearest-point registration process in this embodiment. This is used to verify the convergence and accuracy of the registration algorithm, ensuring the reliability of the pose transformation matrix solution. This embodiment reduces the interference of the reconstructed virtual point cloud on the registration results by assigning different weight coefficients to the real measurement points and virtual points, and dynamically adjusting the iteration weights of the virtual points according to the curvature change rate of the target point cloud. This improves the convergence speed and solution accuracy of the registration algorithm. Through the Jacobian matrix and cubic polynomial interpolation algorithm, accurate mapping of translational and deflection deviations to hydraulic cylinder control parameters is achieved, ensuring the stability and accuracy of the pose adjustment of the assembled segments.

Claims

1. An intelligent positioning and assembly system based on ship section construction, characterized in that, The system includes a solid-state lidar, a linear infrared camera, an edge computing processor, and a three-degree-of-freedom servo drive mechanism. The solid-state lidar and the linear infrared camera are fixed to the butt joint area of ​​the outer plate of the segment to be assembled. The edge computing processor is communicatively connected to the solid-state lidar, the linear infrared camera, and the three-degree-of-freedom servo drive mechanism. The edge computing processor acquires local three-dimensional point cloud data between the segment to be assembled and the assembled segment, and extracts the theoretical position of the rib plate and the theoretical direction of the weld in the butt joint area from a pre-stored digital sample image as prior topological constraint features. For the missing point cloud regions caused by physical occlusion in the 3D point cloud data, a virtual point cloud is reconstructed using the prior topological constraint features as boundary conditions and a non-uniform rational B-spline surface fitting algorithm based on moving least squares. The complete local point cloud containing the virtual point cloud is iteratively registered with the target point cloud in the digital sample to calculate the rotation and translation transformation matrix between the current assembly pose and the target pose. The rotation and translation transformation matrix is ​​decomposed into translation deviation and deflection deviation and input to the three-degree-of-freedom servo drive mechanism to perform pose iterative approximation adjustment. During pose iteration and approximation adjustment, the edge computing processor decouples the rotation and translation transformation matrix into translational deviation components along the X, Y, and Z axes of the ship construction coordinate system and deflection deviation components around the X, Y, and Z axes. The three-degree-of-freedom servo drive mechanism includes an X-axis lifting hydraulic cylinder, a Y-axis translation hydraulic cylinder, and a Z-axis side thrust hydraulic cylinder. The edge computing processor maps the translational deviation components to the target stroke displacement of the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis side thrust hydraulic cylinder, and maps the deflection deviation components to the relative speed difference between the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis side thrust hydraulic cylinder.

2. The intelligent positioning and assembly system based on ship section construction according to claim 1, characterized in that, When acquiring local 3D point cloud data, the edge computing processor controls the timing synchronization of the laser emitted by the solid-state lidar and the exposure of the linear infrared camera by configuring a hardware synchronization trigger signal. The edge computing processor receives the raw distance data output by the solid-state lidar and the grayscale image data output by the linear infrared camera. Using a pre-calibrated joint extrinsic parameter matrix of the radar and camera, the two-dimensional edge feature points extracted from the grayscale image data are projected onto the three-dimensional spatial coordinate system corresponding to the raw distance data to generate a 3D sparse point cloud that incorporates texture edge information. The 3D sparse point cloud is then spatially fused and aligned with the 3D dense point cloud generated from the raw distance data to generate the local 3D point cloud data.

3. The intelligent positioning and assembly system based on ship section construction according to claim 1, characterized in that, In the process of extracting prior topological constraint features, the edge computing processor parses the structural data of the digital sample image, extracts the spatial linear equation parameters of the neutral axis of the rib profile in the butt joint area, extracts the spatial cubic spline curve parameters of the weld centerline on the outer plate surface, discretizes the spatial linear equation parameters and the spatial cubic spline curve parameters into a spatial discrete point set, and rasterizes the spatial discrete point set according to the coordinate axis direction of the shipbuilding coordinate system to generate a two-dimensional topological mask matrix containing the rib distribution position code and the weld direction gradient code. The two-dimensional topological mask matrix is ​​then mapped to a prior topological constraint feature point set in three-dimensional space.

4. The intelligent positioning and assembly system based on ship section construction according to claim 1, characterized in that, In the process of reconstructing the virtual point cloud, the edge computing processor extracts the actual boundary contour point set of the missing region of the point cloud in the local three-dimensional point cloud data, extracts the prior topological constraint feature point set adjacent to the missing region of the point cloud as theoretical boundary conditions, merges and sorts the actual boundary contour point set and the theoretical boundary condition point set to generate a joint boundary point set, calculates the node vector of the non-uniform rational B-spline surface based on the spatial distribution density of the joint boundary point set, uses the three-dimensional coordinates of the joint boundary point set as shape points, obtains the control point grid of the non-uniform rational B-spline surface by solving a system of linear equations, and generates a virtual point cloud covering the missing region of the point cloud based on the control point grid.

5. The intelligent positioning and assembly system based on ship section construction according to claim 1, characterized in that, During iterative nearest-neighbor registration, the edge computing processor assigns initial weight coefficients to each data point in the complete local point cloud, assigns a first fixed weight coefficient to the real measurement points in the local 3D point cloud data, and assigns a second fixed weight coefficient to the reconstructed data points in the virtual point cloud. In each iteration of searching for the nearest neighbor, the current iterative weight coefficient of the reconstructed data points is dynamically adjusted according to the rate of curvature change of the corresponding matching points in the target point cloud. A registration error cost function is constructed based on the point-to-point distance with the current iterative weight coefficient, and the rotation and translation transformation matrix that makes the registration error cost function converge is solved by singular value decomposition.

6. The intelligent positioning and assembly system based on ship section construction according to claim 2, characterized in that, During the timing synchronization process, the edge computing processor has a built-in field-programmable gate array (FPGA). The FPGA outputs a square wave pulse signal with a fixed frequency as the hardware synchronization trigger signal. The rising edge of the square wave pulse signal triggers the start of the modulation signal generator of the laser emitter inside the solid-state lidar. After passing through a preset delay counter, the square wave pulse signal triggers the global shutter reset signal of the linear infrared camera. The count value of the preset delay counter is set according to the difference between the round-trip propagation time of the solid-state lidar laser beam and the photoelectric conversion response time of the linear infrared camera.

7. The intelligent positioning and assembly system based on ship section construction according to claim 4, characterized in that, In the process of calculating the node vector based on the spatial distribution density of the joint boundary point set, the edge computing processor calculates the chord length parameter between adjacent data points in the joint boundary point set, accumulates the chord length parameters to obtain the total chord length, normalizes each chord length parameter with the total chord length to obtain a parameter sequence, sets the span of the node interval according to the distribution interval of the parameter sequence, inserts a second-order repetition node in the parameter interval corresponding to the missing area of ​​the point cloud, and uses the parameter sequence after inserting the second-order repetition node as the node vector of the non-uniform rational B-spline surface.

8. The intelligent positioning and assembly system based on ship section construction according to claim 5, characterized in that, During the process of dynamically adjusting the current iteration weight coefficient of the reconstructed data point, the edge computing processor extracts the corresponding matching point in the target point cloud that matches the reconstructed data point, calculates the normal vector dispersion of the corresponding matching point within a preset neighborhood range, inputs the normal vector dispersion into a pre-constructed inverse proportional mapping function, calculates the weight decay coefficient of the reconstructed data point, and uses the product of the second fixed weight coefficient and the weight decay coefficient as the current iteration weight coefficient of the reconstructed data point in the current iteration.

9. The intelligent positioning and assembly system based on ship section construction according to claim 1, characterized in that, During the mapping to relative speed difference, the edge computing processor constructs a Jacobian matrix containing the kinematic parameters of the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis side-push hydraulic cylinder. The deflection deviation component is input as an input vector into the pseudo-inverse matrix of the Jacobian matrix to calculate the instantaneous speed vectors of the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis side-push hydraulic cylinder. Based on the instantaneous speed vectors and the target stroke displacement, a cubic polynomial interpolation algorithm is used to generate the servo motor control pulse sequence of the X-axis lifting hydraulic cylinder, the Y-axis translation hydraulic cylinder, and the Z-axis side-push hydraulic cylinder.

Images