Digital Twin-Adaptive Assembly Correction Method and System for Prefabricated Segments of Composite Structure

Abstract

This invention relates to the field of digital twin control technology, and discloses a digital twin-adaptive assembly correction method and system for prefabricated segments of composite structures. The method includes: reading BIM geometric model data, constructing an assembly reference coordinate system and digital twin geometry for the prefabricated segments; acquiring composite data in real time and extracting the feature point cloud of the end tooth groove boundary; matching the measured point cloud with the digital twin geometry through dynamic point cloud registration technology, calculating the six-degree-of-freedom pose error vector, and generating a predicted total pose error vector by combining it with a pre-constructed pose drift prediction model; driving a mechanical fine-tuning jack to perform pose adjustment based on the error vector until the pose converges and the tooth groove precision meshing and mechanical locking are completed; this invention achieves high-precision dynamic correction of the assembly process of multiple composite structural segments without changing the original mechanical connection structure through the synergistic effect of dynamic mapping of the digital twin model and adaptive correction algorithm.

Description

Technical Field

[0001] This invention relates to the field of digital twin control technology, and more specifically, to a digital twin-adaptive assembly correction method and system for prefabricated segments of combined structures. Background Technology

[0002] In the field of precise assembly of prefabricated segments of composite structures, the integration of digital twin technology and real-time pose correction algorithms provides a new path to solve the problems of insufficient precision and low efficiency in traditional assembly. Through the dynamic mapping between virtual models and physical entities, it can realize the digital control of the assembly process.

[0003] In the prior art, Chinese patent application CN119397844A discloses a deep learning-based method for geometric digital twin modeling and pose calculation of assemblies. Through point cloud measurement and neural network training, it establishes a mapping model between geometric distribution error and assembly pose, focusing on error modeling of precision mechanical products in static scenarios, and providing a data processing foundation for high-precision assembly. Chinese patent application CN118673619A proposes an online prediction method and system for the assembly accuracy of complex products based on digital twins. It constructs a multi-dimensional error digital twin assembly model by integrating point cloud geometric models and deformation mechanism models, and realizes assembly accuracy prediction based on the Jacobi spinor error propagation model, focusing on the analysis of error propagation laws in the assembly process of complex products.

[0004] However, the aforementioned existing technologies still have specific limitations: the deep learning modeling in the Chinese patent application with publication number CN119397844A relies on static point cloud data and batch training samples, and does not involve the prediction and closed-loop correction of real-time pose drift during dynamic assembly, making it difficult to cope with the six-degree-of-freedom dynamic changes during the hoisting of prefabricated segments of the composite structure; although the accuracy prediction model in the Chinese patent application with publication number CN118673619A integrates multi-dimensional errors, it does not construct a real-time point cloud registration closed loop between the digital twin geometry and the physical segment, and cannot directly convert the prediction error into instructions to drive the mechanical fine-tuning device. As a result, the cumulative error during the assembly of multiple segments in series cannot be proactively compensated through real-time data processing. In particular, there is a break in the link between tooth groove boundary feature extraction and jack drive instruction generation, making it difficult to meet the dynamic control requirements for millimeter-level assembly accuracy of prefabricated segments of the composite structure. Summary of the Invention

[0005] This invention is applicable to the assembly construction of precast segments in bridges, buildings, and municipal engineering, including steel-concrete composite beams, steel-wood composite components, and precast concrete composite structures. It is particularly suitable for high-precision docking requirements in multi-segment series assembly, enabling real-time correction and locking of segment posture during hoisting. To overcome the aforementioned shortcomings of existing technologies, this invention provides a digital twin-adaptive assembly correction method and system for precast composite structure segments. By constructing a closed-loop control system using a digital twin model and real-time sensing by composite sensors, it achieves high-precision dynamic correction of precast composite structure segments, significantly improving assembly accuracy and efficiency. Its integration of dynamic point cloud registration and posture drift prediction technologies effectively eliminates accumulated errors across multiple segments, ensuring precise docking at connection points and providing a reliable solution for the intelligent prefabricated construction of large composite structures.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A digital twin-adaptive assembly correction method for prefabricated segments of composite structures includes:

[0008] Read the BIM geometric model data and establish a reference coordinate system for prefabricated segment assembly and a digital twin geometry of the prefabricated segment;

[0009] A composite sensor array is deployed at the four corners of the prefabricated segment to collect composite data of the prefabricated segment in real time and extract the feature point cloud of the tooth groove boundary at the end of the prefabricated segment;

[0010] The point cloud of the tooth groove boundary feature of the prefabricated segment is dynamically registered with the digital twin geometry of the prefabricated segment, the six-degree-of-freedom pose error vector is calculated, and the pose drift prediction is performed based on the pre-built pose drift prediction model to obtain the total predicted pose error vector.

[0011] The jack's driving command is generated based on the predicted total pose error vector, and pose adjustment is performed.

[0012] Determine whether the position adjustment has met the position convergence condition. If the position convergence condition is met, then perform precise gear meshing and mechanical locking.

[0013] Furthermore, the composite sensor combination includes a lidar sensor and an IMU inertial measurement unit; the composite data of the prefabricated segment includes three-dimensional point cloud data of the surface of the prefabricated segment and six-degree-of-freedom pose data of the prefabricated segment; the six-degree-of-freedom pose data includes three-axis acceleration data and three-axis angular velocity data;

[0014] The method for extracting the feature point cloud of the tooth groove boundary at the end of the prefabricated segment includes: performing statistical filtering to remove noise from the three-dimensional point cloud data and converting it to the prefabricated segment assembly reference coordinate system to form integrated point cloud data; and using a supervoxel-principal curvature segmentation algorithm to automatically extract the feature point cloud of the tooth groove boundary at the end of the prefabricated segment from the integrated point cloud data.

[0015] Furthermore, the method for automatically extracting the feature point cloud of the prefabricated segment end tooth groove boundary from the integrated point cloud data using the supervoxel-principal curvature segmentation algorithm includes:

[0016] An octree decomposition algorithm is used to divide the space of the integrated point cloud data into a multi-scale voxel grid to generate an initial voxel set; local feature clustering is performed on the point cloud within the initial voxel set to form multiple super voxel units.

[0017] Calculate the covariance matrix of all points within each supervoxel unit, obtain the curvature values ​​of the three principal directions through eigenvalue decomposition, and define the principal curvature κ based on the curvature values ​​of the three principal directions;

[0018] Set a curvature threshold and select supervoxel units with principal curvature κ greater than the curvature threshold as high curvature supervoxel units;

[0019] Adjacent high-curvature supervoxel units are aggregated to form a continuous tooth groove boundary feature region. All point cloud data within the tooth groove boundary feature region are extracted as the tooth groove boundary feature point cloud of the prefabricated segment end.

[0020] Furthermore, the curvature values ​​of the three principal directions are respectively the curvature value λ1 of the first principal direction, the curvature value λ2 of the second principal direction, and the curvature value λ3 of the third principal direction; the principal curvature κ is the ratio of the curvature value of the first principal direction to the sum of the curvature values ​​of the three principal directions, where λ1≥λ2≥λ3.

[0021] Furthermore, the method for dynamically registering the point cloud of the end tooth groove boundary features of the prefabricated segment with the digital twin geometry of the prefabricated segment includes:

[0022] The point cloud of the tooth groove boundary features at the end of the prefabricated segment is hierarchically layered using a spatial octree to construct a multi-scale point cloud pyramid structure; the multi-scale point cloud pyramid structure includes coarse-scale, medium-scale, and fine-scale levels.

[0023] At the coarse-scale level, the FPFH feature descriptor is used to perform initial registration between the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry of the prefabricated segment, obtaining the coarse registration transformation matrix T. coarse ;

[0024] Based on the coarse registration transformation matrix T coarseAt the mesoscale level, the ICP iterative nearest point algorithm is executed to optimize the correspondence between the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry of the prefabricated segment.

[0025] During the ICP iteration process, the six-degree-of-freedom pose data provided by the IMU inertial measurement unit is integrated, and the state is predicted and updated through the EKF extended Kalman filter.

[0026] At a fine-scale level, a point-to-surface distance metric is used instead of a point-to-point distance metric to output the final fine registration transformation matrix T. fine .

[0027] Furthermore, the method for calculating the six-degree-of-freedom pose error vector includes:

[0028] The fine registration transformation matrix T fine With the design target pose matrix T of the prefabricated segment target Perform matrix operations to calculate the pose deviation matrix ΔT;

[0029] Singular value decomposition is performed on the pose deviation matrix ΔT to extract a 3×3 rotation matrix R and a 3×1 translation vector t;

[0030] Based on the rotation matrix R, the Rodriguez formula is used to calculate the rotation axis vector n and the rotation angle θ.

[0031] The rotation axis vector n and rotation angle θ are decomposed into the X, Y, and Z axes of the prefabricated segment assembly reference coordinate system to obtain the three rotational components of the pose error.

[0032] The three translation components of the pose error are directly extracted from the translation vector t;

[0033] The three rotational components of the combined pose error and the three translational components of the combined pose error form a complete six-degree-of-freedom pose error vector ε.

[0034] Furthermore, the method for performing pose drift prediction includes:

[0035] Simultaneously collect environmental factor data and fuse it with the six-degree-of-freedom pose error vector to form a comprehensive state vector;

[0036] The integrated state vector is input into the pre-constructed pose drift prediction model to obtain the predicted pose error vector.

[0037] Furthermore, the method for obtaining the predicted total pose error vector includes: adding the predicted pose error vector to the six-degree-of-freedom pose error vector of the current position to obtain the predicted total pose error vector.

[0038] Furthermore, the method for determining whether the pose adjustment has met the pose convergence condition includes:

[0039] Obtain the real-time pose error vector, and determine whether the three translational components of the real-time pose error vector converge, and whether the three rotational components of the real-time pose error vector converge. If both the translational convergence condition and the rotational convergence condition are met, then the pose is determined to be converged.

[0040] A digital twin-adaptive assembly correction system for prefabricated segments of composite structures, used to implement the aforementioned digital twin-adaptive assembly correction method for prefabricated segments of composite structures, the system comprising:

[0041] Twin model construction module; used to read BIM geometric model data and establish a digital twin geometry of the prefabricated segment assembly reference coordinate system and the prefabricated segment;

[0042] Feature point cloud extraction module: used to deploy composite sensor combinations at the four corners of the prefabricated segment, collect composite data of the prefabricated segment in real time, and extract feature point clouds of the tooth groove boundary at the end of the prefabricated segment;

[0043] Pose prediction module: used to dynamically register the feature point cloud of the tooth groove boundary of the prefabricated segment with the digital twin geometry of the prefabricated segment, calculate the six-degree-of-freedom pose error vector, and perform pose drift prediction based on the pre-built pose drift prediction model to obtain the total predicted pose error vector.

[0044] Pose adjustment drive module: Generates drive instructions for the jack based on the predicted total pose error vector and performs pose adjustment;

[0045] Position convergence judgment module: used to determine whether the position adjustment has reached the position convergence condition. If the position convergence condition is reached, then the gear precision meshing and mechanical locking are executed.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] This invention combines composite sensor arrays with dynamic point cloud registration technology to capture the three-dimensional spatial pose deviation of the tooth groove boundary at the end of prefabricated segments in real time. Combined with a pose drift prediction model, it achieves proactive compensation for accumulated errors, significantly improving the accuracy and stability of multi-segment serial assembly. The driving commands of the mechanical fine-tuning jacks are generated based on a six-degree-of-freedom pose error vector, realizing multi-dimensional coordinated adjustment of the segment's spatial attitude and avoiding local stress concentration and channel eccentricity problems caused by traditional manual hammering or single-point pushing. The software-hardware coupled closed-loop control system integrates the single-segment "scan-algorithm-fine-tuning-locking" process into an automated operation unit, significantly shortening assembly time and reducing the risk of manual intervention. Simultaneously, the pose convergence judgment mechanism ensures the effectiveness of each adjustment step, ultimately achieving precise docking of the tooth groove meshing surface and reliable anchoring of the mechanical locking device. This provides an intelligent solution that combines efficiency and precision for the assembly of composite structures in bridges, buildings, and other fields. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of the digital twin-adaptive assembly correction method for prefabricated segments of the composite structure in this invention;

[0050] Figure 2 This is a schematic diagram illustrating the deployment principle of the prefabricated segment four-corner composite sensor of the present invention;

[0051] Figure 3 This is a flowchart of the method for dynamically registering the point cloud of the end tooth groove boundary feature of the prefabricated segment with the digital twin geometry of the prefabricated segment according to the present invention.

[0052] Figure 4 This is a flowchart illustrating the principle of determining whether pose adjustment has reached pose convergence in this invention.

[0053] Figure 5 This is a flowchart illustrating the principle of determining whether the precision meshing accuracy has passed verification in this invention.

[0054] Figure 6 This is a flowchart illustrating the principle of determining whether the coaxiality accuracy of the channel has passed verification in this invention.

[0055] Figure 7 This is a functional block diagram of the digital twin-adaptive assembly correction system for prefabricated composite structures in this invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example 1:

[0058] Please see Figure 1 As shown, this embodiment provides a digital twin-adaptive assembly correction method for prefabricated segments of a composite structure, including:

[0059] Step S10: Read the BIM geometric model data and establish the prefabricated segment assembly reference coordinate system and the digital twin geometry of the prefabricated segment;

[0060] Further, step S10 includes:

[0061] Step S11: Read the BIM geometric model data containing the three-dimensional geometric information, material properties and assembly constraints of the prefabricated segments and establish the prefabricated segment assembly reference coordinate system;

[0062] Step S12: Based on the prefabricated segment assembly reference coordinate system, construct a three-dimensional virtual space environment, load the BIM geometric model data into the virtual space environment, and form a digital twin geometry of the prefabricated segment.

[0063] Specifically, the three-dimensional geometric information of prefabricated segments includes contour dimensions, groove shape, and duct location; material properties include density and elastic modulus; and assembly constraints include the connection sequence and relative positional restrictions of each segment. Composite structures encompass prefabricated component assemblies using steel-concrete, steel-wood, or other composite methods in bridges, buildings, and other fields. A right-handed coordinate system is established with the design origin of the composite structure as the virtual coordinate origin. The X-axis points in the longitudinal direction of the composite structure (e.g., the extension direction of a bridge), the Y-axis points in the transverse direction (e.g., the width direction of a bridge), and the Z-axis points vertically upwards, perpendicular to the ground. This coordinate system is chosen because in engineering practice, the assembly of composite structures is often carried out longitudinally, with the transverse and vertical directions being the primary adjustment directions. Using this as a benchmark provides a unified spatial reference for the assembly process of all prefabricated segments, ensuring that the relative positional relationships between different segments can be accurately expressed and avoiding the accumulation of assembly errors caused by inconsistent coordinate systems. After establishing a reference coordinate system for prefabricated segment assembly, a three-dimensional virtual space environment is constructed based on this coordinate system. The aforementioned BIM geometric model data is then loaded into this constructed three-dimensional virtual space environment to form a digital twin geometry of the prefabricated segment. The construction of the three-dimensional virtual space environment needs to simulate real construction environment parameters, including basic physical parameters such as gravity field and temperature field, so that the virtual environment has physical consistency with the physical construction environment. Loading the BIM geometric model data is not a simple import; it requires format conversion, such as converting the commonly used .ifc format to an internal format suitable for the virtual space environment. It also requires accuracy verification to ensure that the deviation between the geometric dimensions of the model and the design values ​​is within the allowable range, as well as topological relationship reconstruction to clarify the connection relationships of each component in the model, such as the spatial positional relationship between the tooth groove and the hole, so as to ensure that the digital twin geometry can accurately map the physical properties and geometric features of the prefabricated segment.

[0064] In existing technologies, traditional static BIM can only perform pre-assembly verification and cannot reflect the assembly process in real time. Furthermore, different segments may use their own independent coordinate systems, requiring multiple coordinate transformations during assembly. Each transformation introduces errors, and as the number of segments increases, these errors accumulate, affecting the precise alignment of connections. Step S10 establishes a unified prefabricated segment assembly reference coordinate system, ensuring that all prefabricated segments use the same coordinate system as a reference during assembly. This eliminates the source of errors caused by coordinate transformations, making the positional information of different segments directly comparable. This provides a consistent benchmark for subsequent pose error calculations and solves the problem of error accumulation in multi-segment serial assembly. Simultaneously, the construction of the digital twin geometry transforms the static BIM model into a dynamic model that can be synchronously mapped with the physical entity in virtual space. This provides a precise virtual reference for subsequent real-time point cloud data acquisition, giving subsequent steps such as point cloud registration a clear comparison standard. A unified reference coordinate system is not only applicable to a single type of composite structure, but also universally applicable in the assembly of composite structures in multiple fields such as bridges, buildings, and municipal works. This is because the definition of its coordinate axes is consistent with the main extension directions, lateral and vertical directions, of the structure in each field, eliminating the need to redefine the coordinate system for different fields and improving the method's versatility. The simulation of physical parameters during the construction of digital twin geometry ensures that the virtual model is not only geometrically consistent with the physical entity, but also approximately reflects the physical characteristics of the physical entity under different environments. For example, minute deformations caused by temperature changes can be reflected in the virtual model through corresponding adjustments to physical parameters, providing a foundation for subsequently considering the impact of environmental factors on the structure's orientation. Without step S10, subsequent point cloud acquisition and registration would lack a unified coordinate reference. Point cloud data from different segments would reside in their respective coordinate systems, making direct comparison impossible to calculate pose errors. This would make the point cloud registration process extremely complex and inaccurate. Furthermore, without a digital twin geometry as a reference, the real-time acquired point cloud data would lack a comparison object, making it impossible to determine the deviation between the physical segments and the design state. The entire adaptive assembly correction process would lose its foundation, making it impossible to achieve the goal of dynamically capturing pose errors and driving the fine-tuning device for correction.

[0065] Step S20: Deploy composite sensor combinations at the four corners of the prefabricated segment to collect composite data of the prefabricated segment in real time and extract the feature point cloud of the tooth groove boundary at the end of the prefabricated segment.

[0066] Further, step S20 includes:

[0067] Step S21: Deploy a composite sensor combination containing a lidar sensor and an IMU inertial measurement unit at the four corners of the prefabricated segment to collect composite data of the prefabricated segment in real time. The composite data of the prefabricated segment includes three-dimensional point cloud data of the surface of the prefabricated segment and six-degree-of-freedom pose data of the prefabricated segment; the six-degree-of-freedom pose data includes three-axis acceleration data and three-axis angular velocity data.

[0068] Step S22: Perform statistical filtering to remove noise from the three-dimensional point cloud data and convert it to the prefabricated segment assembly reference coordinate system to form integrated point cloud data;

[0069] Step S23: The supervoxel-principal curvature segmentation algorithm is used to automatically extract the feature point cloud of the prefabricated segment end tooth groove boundary from the integrated point cloud data.

[0070] Further, step S23 includes:

[0071] Step S231: The octree decomposition algorithm is used to divide the space of the integrated point cloud data into a multi-scale voxel grid to generate an initial voxel set.

[0072] Step S232: Perform local feature clustering on the point cloud within the initial voxel set to form multiple super voxel units;

[0073] Step S233: Calculate the covariance matrix of all points within each supervoxel unit, and obtain the curvature values ​​λ1, λ2, and λ3 in three principal directions through eigenvalue decomposition, where λ1 ≥ λ2 ≥ λ3. Define the principal curvature κ = λ1 / (λ1 + λ2 + λ3); where λ1 is the curvature value in the first principal direction, λ2 is the curvature value in the second principal direction, and λ3 is the curvature value in the third principal direction.

[0074] Step S234, set the curvature threshold κ thre Supervoxel units with principal curvature κ greater than curvature threshold are selected as high curvature supervoxel units, and high curvature supervoxel units correspond to the geometric abrupt change region of the tooth groove boundary.

[0075] Step S235: Perform spatial connectivity analysis on the selected high curvature supervoxel units, use a region growing algorithm to aggregate adjacent high curvature supervoxel units to form a continuous tooth groove boundary feature region, and extract all point cloud data within the tooth groove boundary feature region as the tooth groove boundary feature point cloud of the prefabricated segment end.

[0076] Specifically, such as Figure 2As shown, composite sensor combinations, including lidar sensors and IMU inertial measurement units, are deployed at the four corner points of the prefabricated segment. There are four composite sensor combinations in total: composite sensor combination 1, composite sensor combination 2, composite sensor combination 3, and composite sensor combination 4. The lidar sensors are used to acquire the three-dimensional spatial coordinate information of the prefabricated segment surface, and the IMU inertial measurement units are used to sense the motion state of the prefabricated segment. The two work together to achieve omnidirectional monitoring of the six-degree-of-freedom pose of the prefabricated segment. The configuration of the four corner points can completely capture the translation and rotation states of the segment because the four vertices of the quadrilateral structure can uniquely determine the position and orientation of a plane through spatial geometric relationships. Multi-point measurement can reduce the influence of random errors in single-point measurement through data redundancy, thereby improving the overall measurement accuracy. When the tower crane begins hoisting the prefabricated segments, four sets of deployed composite sensing units are activated to simultaneously acquire three-dimensional point cloud data of the prefabricated segment surface. Each lidar sensor generates a point cloud dataset containing tens of thousands of three-dimensional coordinate points in a single scan. The simultaneous operation of the four sets of sensors can obtain dense point cloud coverage of the prefabricated segment surface, with a point cloud density sufficient to reflect the surface geometric details at the millimeter level. At the same time, the IMU inertial measurement unit outputs three-axis acceleration data and three-axis angular velocity data of the prefabricated segment in real time. The three-axis acceleration data includes linear acceleration along the X-axis, linear acceleration along the Y-axis, and linear acceleration along the Z-axis, reflecting the translational acceleration of the segment in the three coordinate axes, respectively. The three-axis angular velocity data includes angular velocity around the X-axis, angular velocity around the Y-axis, and angular velocity around the Z-axis, reflecting the rotational velocity of the segment around the three coordinate axes, respectively. These data together constitute the six-degree-of-freedom pose data of the prefabricated segment, which can completely describe the instantaneous motion state of the segment.

[0077] After acquiring 3D point cloud data, statistical filtering is performed to remove noise, and the data is then transformed to the prefabricated segment assembly reference coordinate system to form integrated point cloud data. The specific process of statistical filtering for noise removal is as follows: for each point in the point cloud data, the average distance to neighboring points within a set radius is calculated. Points whose distance deviation from the average exceeds a set distance threshold are identified as noise points and removed. The distance threshold is determined based on the inherent noise level of the lidar sensor. For example, noise distribution characteristics are obtained through multiple static scanning experiments, and the threshold is set to three times the noise standard deviation to ensure that more than 99% of valid points are retained. Coordinate transformation is performed using the extrinsic parameter matrix obtained during sensor calibration. This matrix transforms the point cloud data from the lidar sensor's own coordinate system to the prefabricated segment assembly reference coordinate system. The extrinsic parameter matrix is ​​solved by collecting data in a calibration field with known precise coordinates. The calibration field coordinates must be strictly aligned with the prefabricated segment assembly reference coordinate system. The least squares method is used to solve the transformation relationship between the sensor coordinate system and the reference coordinate system, ensuring that the transformed point cloud data is within the same coordinate frame as the digital twin geometry.

[0078] A super-voxel-principal curvature segmentation algorithm is used to automatically extract the feature point cloud of the prefabricated segment end tooth groove boundary from the integrated point cloud data. The algorithm first uses an octree decomposition algorithm to divide the space of the integrated point cloud data into a multi-scale voxel grid, generating an initial voxel set. The voxel size is selected based on the geometric abrupt changes in the tooth groove boundary; excessively small voxels increase computational cost, while excessively large voxels may lose detailed features. Geometric abrupt changes in the tooth groove boundary are typically evident at the millimeter scale; for example, a voxel size of 5 mm can be set. Local feature clustering is then performed on the point cloud within the initial voxel set to form multiple super-voxel units. Each super-voxel unit contains a subset of the point cloud from a local region and its centroid coordinates. The clustering process is based on the spatial distance between voxels and the angle between the point cloud normal vectors. Voxels that are close together and have the same normal vector direction are clustered into a single super-voxel unit to preserve local geometric continuity. The covariance matrix of all points within each supervoxel is calculated. This covariance matrix is ​​obtained through statistical calculation of the coordinate values ​​of all points within the cell, reflecting the spatial distribution characteristics of the point cloud. Curvature values ​​in three principal directions are obtained through eigenvalue decomposition, where the curvature value of the first principal direction is greater than or equal to the second, and the second is greater than or equal to the third. The principal curvature is defined as the ratio of the curvature value of the first principal direction to the sum of the curvature values ​​of the three principal directions. A curvature threshold is set, determined through statistical analysis of the principal curvature of a large number of toothed boundary samples; for example, it is set to 0.7. Toothed boundary regions have larger principal curvature values ​​due to drastic geometric changes, while non-boundary regions have smaller principal curvature values. Supervoxel cells with principal curvature greater than the curvature threshold are thus selected as high-curvature supervoxel cells. Spatial connectivity analysis was performed on the selected high-curvature supervoxel units. A region growing algorithm was used to aggregate adjacent high-curvature supervoxel units to form a continuous tooth groove boundary feature region. The adjacency determination criterion was that the Euclidean distance between the supervoxel mass centers was less than the set adjacency spatial threshold. The adjacency spatial threshold was determined based on the average size of the supervoxel units. After aggregation, all point cloud data in the tooth groove boundary feature region were extracted as the tooth groove boundary feature point cloud of the prefabricated segment end.

[0079] In existing technologies, traditional in-situ calibration relies on manual tapping or inefficient jack fine-tuning, lacking real-time pose monitoring methods, resulting in low calibration accuracy and time consumption. Simple laser ranging can only acquire positional information of a limited number of points, failing to comprehensively reflect the overall pose and surface features of the segment. Step S20, through the deployment of a composite sensor combination, utilizes the dense point cloud data provided by the lidar to fully reflect the three-dimensional geometric information of the segment surface, and the six-DOF pose data provided by the IMU to capture the segment's motion state in real time. The combination of these two technologies enables dynamic and comprehensive monitoring of the segment's pose, solving the problems of incomplete and delayed information acquisition in traditional methods. Statistical filtering and coordinate transformation ensure the accuracy and uniformity of the point cloud data, laying the foundation for subsequent registration with the digital twin geometry. The supervoxel-principal curvature segmentation algorithm, through multi-scale voxel partitioning and principal curvature analysis, can accurately extract tooth boundary features from complex surface point clouds, avoiding interference from non-critical area point clouds and improving the targeting and efficiency of feature extraction. The four-corner deployment method of composite sensor combination is not only applicable to prefabricated segments with regular shapes, but also effective for irregularly shaped composite structures. This is because the four vertices of the quadrilateral can be adapted to segments with different cross-sectional shapes through spatial geometry. Only the installation angle of the sensor needs to be adjusted according to the actual contour of the segment, without the need to redesign the sensor layout, thus enhancing the adaptability of the method.

[0080] The prefabricated segment assembly reference coordinate system established in steps S20 and S10 is coordinated. The point cloud data after coordinate transformation can be directly compared with the digital twin geometry within the same spatial framework, ensuring the accuracy of subsequent registration processes. The tooth groove boundary feature point cloud obtained by feature extraction corresponds precisely with the tooth groove model in the digital twin geometry, enabling pose error calculation to focus on key connection parts and improving the relevance of error assessment. Without step S20, subsequent pose error calculation would lack reliable raw data support, making it impossible to accurately obtain the deviation between the actual pose of the prefabricated segment and the design target pose, resulting in a lack of basis for adaptive correction. At the same time, the lack of tooth groove boundary features would cause the registration process to be based only on the overall contour, failing to reflect the subtle deviations of key connection parts, thus affecting the accuracy and reliability of mechanical locking. Step S20, by providing real-time and accurate feature data, enables the digital twin model to dynamically map the state of the physical segment, providing a key feedback link for closed-loop control and driving the entire assembly process from static planning to dynamic adjustment.

[0081] Step S30: Dynamically register the feature point cloud of the tooth groove boundary at the end of the prefabricated segment with the digital twin geometry of the prefabricated segment, calculate the six-degree-of-freedom pose error vector, and perform pose drift prediction based on the pre-built pose drift prediction model to obtain the total predicted pose error vector.

[0082] Further, step S30 includes:

[0083] Step S31: The multi-scale ICP-EKF fusion algorithm is used to dynamically register the feature point cloud of the tooth groove boundary at the end of the prefabricated segment with the digital twin geometry of the prefabricated segment.

[0084] Please see Figure 3 As shown, step S31 further includes:

[0085] Step S311: The point cloud of the tooth groove boundary feature is spatially octree-layered to construct a multi-scale point cloud pyramid structure; the multi-scale point cloud pyramid structure includes coarse-scale level, medium-scale level and fine-scale level.

[0086] Step S312: At the coarse-scale level, the FPFH feature descriptor is used to perform initial registration between the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry of the prefabricated segment, obtaining the coarse registration transformation matrix T. coarse ;

[0087] Step S313, based on the coarse registration transformation matrix T coarse At the mesoscale level, the ICP iterative nearest point algorithm is executed to optimize the correspondence between the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry of the prefabricated segment.

[0088] Step S314: During the ICP iteration process, the six-degree-of-freedom pose data provided by the IMU inertial measurement unit is fused, and the state is predicted and updated through the EKF extended Kalman filter;

[0089] Step S315: At the fine-scale level, the point-to-surface distance metric is used instead of the point-to-point distance to refine the registration result, and the final fine registration transformation matrix T is output. fine .

[0090] Specifically, the octree hierarchical structure recursively divides the 3D space of the feature point cloud of the prefabricated segment end tooth groove boundary into eight sub-cubes. Each cube has a side length half that of the previous one, forming coarse-scale, meso-scale, and fine-scale layers. The coarse-scale layer preserves the overall outline of the point cloud, the meso-scale layer contains moderate detail, and the fine-scale layer preserves the subtle features of the tooth groove boundary. This structure reduces the computational cost of registration while maintaining accuracy. At the coarse-scale layer, the FPFH feature descriptor is used for initial registration of the feature point cloud and the digital twin geometry. The FPFH feature descriptor is generated by calculating the angle between the normal vectors of a point and its neighbors and the distance distribution. It is invariant under rotation and translation transformations. Corresponding point pairs are found through feature matching, and the coarse registration transformation matrix is ​​calculated. Based on the coarse registration transformation matrix, the ICP iterative nearest point algorithm is executed at the mesoscale level. This algorithm iteratively finds the nearest point of each point in the point cloud on the digital twin geometry, calculates the transformation matrix that minimizes the sum of squared distances between corresponding points, and optimizes the correspondence between the point cloud and the model. The number of iterations is determined based on the coarse registration error, typically 10–20 times to balance efficiency and accuracy. During the ICP iteration, six-DOF pose data provided by the IMU (Inertial Measurement Unit) are fused, and state prediction and updates are performed using an EKF (Extended Kalman Filter). The EKF state vector contains the position and attitude of segments. The position of a segment is the coordinates on the X, Y, and Z axes, and the attitude of a segment is the rotation angle around the X, Y, and Z axes. The prediction step utilizes the IMU's motion model, i.e., estimates the state and covariance matrix of the next time step based on the integral of acceleration and angular velocity. The update step uses the pose correction calculated by ICP as the observation value, and fuses the predicted value and the observation value through Kalman gain to correct the state estimate. This fusion mechanism compensates for the lag of the ICP algorithm in dynamic scenarios. For example, when a segment sways rapidly due to wind load, the real-time data from the IMU can predict pose changes in advance, enabling faster convergence of the ICP iteration and avoiding matching errors caused by point cloud motion blur. At the fine-scale level, a point-to-surface distance metric is used instead of a point-to-point distance metric. The point-to-surface distance is calculated by measuring the perpendicular distance from the point to the corresponding triangular facet, which is more resistant to noise and local shape deviations than the point-to-point distance metric. This refines the registration results and outputs the final fine registration transformation matrix, which describes the optimal spatial transformation relationship from the feature point cloud to the digital twin geometry.

[0091] Multi-scale registration accurately focuses on key connection points, avoiding interference from non-feature area point clouds, and ensuring that registration errors mainly originate from the tooth groove mating surface. Traditional ICP algorithms are prone to convergence to local optima when processing dynamic point clouds due to excessive initial pose deviations. Relying solely on geometric registration cannot handle real-time motion during hoisting, resulting in low registration accuracy and time consumption. The multi-scale strategy, through hierarchical registration from coarse to fine, first rapidly narrows the pose deviation range and then gradually refines the details, solving the problems of low efficiency and easy divergence of traditional algorithms when searching a large area. The fusion of EKF and IMU data organically combines geometric and kinematic information, enabling the registration process to dynamically adapt to the segment's motion state and remain stable under disturbances such as wind loads and cable vibrations, improving accuracy compared to pure geometric registration. The octree hierarchical multi-scale structure is robust to local defects at the tooth groove boundary. For example, when there are small gaps caused by construction in the tooth groove, the registration at the coarse-scale level is not affected by local defects, and the fine-scale level can still maintain overall accuracy through interpolation of surrounding intact features. In contrast, traditional single-scale registration will produce significant errors due to local defects. The state prediction mechanism of EKF can compensate for the inertial motion of segments in advance. For example, at the moment of starting and stopping the tower crane, the small sway of the segment caused by inertia can be captured by the IMU and corrected in advance by EKF, so that the registration result is ahead of the actual pose change, providing time redundancy for subsequent error prediction and correction.

[0092] Step S32: Based on the registration results, calculate the six-degree-of-freedom pose error vector ε between the current pose of the prefabricated segment and the design target pose through transformation matrix decomposition;

[0093] Further, step S32 includes:

[0094] Step S321, the fine registration transformation matrix T fine With the design target pose matrix T of the prefabricated segment target Perform matrix operations to calculate the pose deviation matrix. ;

[0095] Step S322: Perform singular value decomposition on the pose deviation matrix ΔT to extract a 3×3 rotation matrix R and a 3×1 translation vector t;

[0096] Step S323: Based on the rotation matrix R, calculate the rotation axis vector n and rotation angle θ using the Rodriguez formula;

[0097] Step S324: Decompose the rotation axis vector n and rotation angle θ to the X, Y, and Z axes of the prefabricated segment assembly reference coordinate system to obtain the three rotational components of the pose error.

[0098] Step S325: Directly extract the three translation components of the pose error from the translation vector t;

[0099] Step S326: Combine the three rotational components of the pose error with the three translational components of the pose error to form a complete six-degree-of-freedom pose error vector ε.

[0100] Step S33: Simultaneously collect environmental factor data and fuse it with the six-degree-of-freedom pose error vector to form a comprehensive state vector;

[0101] Step S34: Input the integrated state vector into the pre-constructed pose drift prediction model to obtain the predicted pose error vector. Add the predicted pose error vector to the six-degree-of-freedom pose error vector of the current position to obtain the total predicted pose error vector.

[0102] Specifically, the target pose matrix is ​​the theoretical position matrix of the digital twin geometry in the reference coordinate system, and the pose deviation matrix is ​​the product of the fine registration transformation matrix and the inverse of the target pose matrix, reflecting the spatial transformation relationship between the actual pose and the theoretical pose. Singular value decomposition (SVD) is performed on the pose deviation matrix to extract a 3×3 rotation matrix and a 3×1 translation vector. SVD is achieved by decomposing the matrix into the product of orthogonal, diagonal, and orthogonal matrices. The rotation matrix describes the rotation deviation of a segment, and the translation vector describes the translation deviation of a segment. Based on the rotation matrix, the Rodrigues formula is used to calculate the rotation axis vector and rotation angle. The Rodrigues formula is achieved by converting the rotation matrix into a rotation vector, where the direction of the rotation vector is the rotation axis and the magnitude is the rotation angle. The rotation axis vector and rotation angle are decomposed onto the X, Y, and Z axes of the prefabricated segment assembly reference coordinate system to obtain three rotational components of the pose error. The decomposition process calculates the angle between the rotation axis and each coordinate axis, and proportionally distributes the rotation angle to the corresponding axis. The three translational components of the pose error are directly extracted from the translation vector, which are the projections of the translation vector onto the X, Y, and Z axes. The three rotational components and three translational components of the pose error are combined to form a complete six-degree-of-freedom pose error vector.

[0103] Environmental factor data is simultaneously collected and fused with the six-degree-of-freedom pose error vector to form a comprehensive state vector. Environmental factor data includes parameters such as real-time wind speed, sling tension, and tower crane rotation angle. Real-time wind speed data is obtained from an anemometer installed on the top of the tower crane, reflecting the impact of wind load on segment stability. Sling tension data is measured by tension sensors on the slings, reflecting the stress state of the hoisting system. Tower crane rotation angle data is obtained from the angle encoder of the tower crane control system, reflecting changes in the hoisting position. After standardization, this data is concatenated dimensionally with the six-degree-of-freedom pose error vector to form a comprehensive state vector containing pose error and environmental information. Standardization is achieved by converting the data into a normal distribution with a mean of 0 and a standard deviation of 1, ensuring effective fusion of data from different magnitudes.

[0104] The pose drift prediction model employs a lightweight GNN-Transformer neural network. The GNN captures the spatial relationships between segments through node embedding; node features include the segment's geometric parameters and historical pose data, while edge features represent the relative positional relationships between segments. The Transformer network uses a self-attention mechanism to handle temporal variation patterns, learning the evolution of pose error over time. The GNN-Transformer neural network is trained using training data from historical assembly processes. This training data includes comprehensive state vectors under different environmental conditions and corresponding actual pose drift values. Gradient descent is used to optimize network parameters, and the loss function is the mean squared error between the predicted and actual drift values. Given the current comprehensive state vector, the neural network outputs a predicted pose error vector within a preset timeframe, such as 3 seconds into the future. This vector contains three translational and three rotational components. Each component of the total predicted pose error vector is the sum of the current pose error component and its corresponding predicted drift component, reflecting the total pose deviation after considering future drift.

[0105] In existing technologies, traditional static BIM cannot reflect the assembly process in real time, and simple laser ranging or manual correction is difficult to handle pose changes in dynamic environments, and errors are prone to accumulate in multi-segment assembly. Step S30 achieves dynamic point cloud registration through a multi-scale ICP-EKF fusion algorithm. The coarse-to-fine registration strategy improves computational efficiency while ensuring accuracy, resolving the contradiction between real-time performance and accuracy in dynamic scenarios. EKF fusion with IMU data reduces the lag in point cloud registration during rapid movement, making pose calculation more consistent with the actual motion state. The decomposition of the pose error vector is achieved through rigorous matrix operations and geometric transformations, ensuring that the physical meaning of the error components is clear and providing accurate quantitative basis for subsequent correction. The fusion of environmental data considers the influence of external factors on segment pose, avoiding prediction bias caused by relying solely on geometric information, making pose drift prediction more consistent with the actual construction environment. The lightweight GNN-Transformer neural network, combined with the learning of spatial correlation and temporal changes, can better capture the error evolution law in complex environments compared to traditional temporal models, improving prediction accuracy. The multi-scale registration strategy is not only applicable to regular features such as tooth grooves, but also maintains high registration accuracy for segments with surface defects. This is because the coarse-scale layer filters out the influence of local defects, while the fine-scale layer focuses on key feature regions, giving the algorithm a certain degree of robustness. After environmental data fusion, the pose drift prediction model can adapt to different weather conditions. By learning error patterns under different wind speeds from historical data, it can still provide reliable predictions when wind speeds change suddenly, reducing correction lag caused by sudden environmental changes. The lightweight design of GNN-Transformer allows the model to run on edge computing devices without relying on cloud computing power, reducing the impact of network instability at construction sites and improving the system's independence and response speed.

[0106] Dynamic registration is achieved based on a unified coordinate system, ensuring the consistency of the reference for pose error calculation. Accurate extraction of feature point clouds provides high-quality input for registration, reducing interference from redundant data and making the registration results more reliable. Without step S30, subsequent jack drive commands would lack accurate error basis, and the correction process would be blind, leading to problems such as insufficient meshing of connection parts and duct eccentricity. Simultaneously, the inability to predict pose drift would cause the correction action to lag behind actual error changes, forming a "adjustment-drift-readjustment" cycle and reducing construction efficiency. Step S30, by establishing a quantification and prediction mechanism for pose error, dynamically links the digital twin model with the physical segments, providing a decision-making core for adaptive correction.

[0107] Step S40: Generate the driving command for the jack based on the predicted total pose error vector, and perform pose adjustment;

[0108] Further, step S40 includes:

[0109] Step S41: Initialize a miniature hydraulic jack device with three degrees of freedom adjustment capability at the four corners of the bottom of the prefabricated segment; each miniature hydraulic jack device has a built-in magnetostrictive displacement sensor for real-time monitoring of the extension and retraction displacement of the jack piston rod and a pressure sensor for monitoring the change of oil pressure in the hydraulic cylinder.

[0110] Step S42: Based on the predicted total pose error vector, the inverse kinematics algorithm is used to convert it into the driving commands for the four initialized micro hydraulic jacks.

[0111] Step S43: Send the drive command to the hydraulic control valve array to perform pose adjustment;

[0112] Step S44: During the pose adjustment process, obtain the real-time pose error vector ε. real The real-time pose error vector ε real The adjustment effect deviation is calculated by comparing it with the target pose adjustment amount. If the adjustment effect deviation exceeds the preset adjustment threshold, the conversion of the drive command is re-executed to generate the corrected drive commands for the four micro hydraulic jacks and send them out for execution.

[0113] Specifically, the miniature hydraulic jacks incorporate magnetostrictive displacement and pressure sensors. The magnetostrictive displacement sensor monitors the piston rod's extension and retraction displacement in real time by measuring the propagation time difference of the magnetostrictive wave. The pressure sensor senses changes in oil pressure within the hydraulic cylinder and converts them into electrical signals, reflecting the supporting force output by the jacks. The four jacks, together with the main sling system, constitute a virtual Stewart platform mechanism. The Stewart platform is a six-degree-of-freedom parallel mechanism, achieving precise control of arbitrary spatial orientation by adjusting the length of each branch. The four-corner layout allows each jack to independently undertake horizontal pushing and pulling along the X and Y axes and vertical lifting and lowering along the Z axis. Their installation position coordinates are obtained using a total station before the prefabricated segments are hoisted and recorded in the digital twin system as the basic parameters of the kinematic model.

[0114] Based on the predicted total pose error vector, the inverse kinematics algorithm is used to convert it into drive commands for four miniature hydraulic jacks. The specific process is as follows: A kinematic model of the Stewart platform is constructed, establishing the transformation relationship between the platform coordinate system and the prefabricated segment assembly reference coordinate system. The platform coordinate system has the segment centroid as its origin, and all axes are parallel to the reference coordinate system. The predicted total pose error vector is decomposed into the desired displacement and desired rotation of the platform centroid. The desired displacement consists of three translational components, and the desired rotation consists of three rotational components. Based on the Stewart platform inverse kinematics equations, the change in length of each jack's support chain is calculated according to the jack installation position coordinates and desired pose parameters. The inverse kinematics equations are established through geometric derivation, mapping the spatial pose change to the linear displacement of each support chain. Combined with the prefabricated segment mass distribution... The support force required by each jack is calculated through static analysis of the distribution of the load and center of gravity. The mass distribution is obtained through calculation of material properties and geometric dimensions in the BIM model. The center of gravity is directly extracted by 3D modeling software. The distribution of support force follows the principle of torque balance to ensure that the segment does not overturn during adjustment. Considering the dynamic response characteristics of the hydraulic system, the flow-displacement response curves under different oil temperatures and pressures are measured experimentally. The change in branch length and support force are converted into target displacement commands and target pressure commands for the hydraulic cylinder. The target displacement command corresponds to the piston rod extension and retraction, and the target pressure command corresponds to the hydraulic pump output pressure. A complete drive command set including a timing control strategy is generated. The timing control avoids additional torque caused by synchronous adjustment by setting the sequence and rate of each jack's action, ensuring balanced force on the segment.

[0115] The drive command is sent to the hydraulic control valve array to perform posture adjustment. The hydraulic control valve array includes a proportional directional valve, a flow control valve, and a pressure control valve. The proportional directional valve adjusts the hydraulic oil flow direction according to the target displacement command to control the extension and retraction direction of the jack. The flow control valve controls the piston rod movement speed by adjusting the oil flow rate per unit time, making the adjustment process smooth. The pressure control valve limits the maximum oil pressure according to the target pressure command to prevent overload damage to the structure. During execution, the magnetostrictive displacement sensor collects the actual displacement of the piston rod in real time, and the pressure sensor collects the actual pressure. The data is transmitted to the control unit through the industrial bus to form a closed-loop feedback. During pose adjustment, a real-time pose error vector is acquired and compared with the target pose adjustment amount to calculate the adjustment effect deviation. The real-time pose error vector is obtained by repeatedly executing the point cloud acquisition in step S20 and the registration calculation in step S30. The target pose adjustment amount is the theoretical adjustment value corresponding to the predicted total pose error vector, and the adjustment effect deviation is the difference between the two. If the adjustment effect deviation exceeds the preset adjustment threshold, the conversion of the driving command is re-executed to generate and issue a corrected driving command. The adjustment threshold is determined through multiple debugging sessions, such as testing the adjustment accuracy under different error scenarios. Finally, it is set to a translation deviation of no more than 0.1 mm and a rotation deviation of no more than 0.005 degrees. The adjustment threshold ensures both adjustment accuracy and avoids system oscillation caused by frequent corrections.

[0116] In existing technologies, traditional jack fine-tuning relies on manual operation, resulting in low adjustment accuracy and an inability to handle dynamic errors. In multi-segment assembly, uneven force distribution can easily lead to additional deformation. Step S40 achieves precise six-degree-of-freedom control of the pose through the coordination of the three-degree-of-freedom adjustment of the micro hydraulic jack and the Stewart platform mechanism, solving the problem that traditional single-degree-of-freedom adjustment cannot compensate for complex spatial pose errors. The inverse kinematics algorithm transforms abstract pose errors into specific jack action commands, enabling the virtual adjustment of the digital twin model to be accurately mapped to the physical segments. The real-time feedback and dynamic correction mechanism, through sensor data closed-loop, compensates for the influence of nonlinear factors such as hydraulic system hysteresis and component elastic deformation, ensuring that the adjustment effect is consistent with the theoretical target. The retractable miniature hydraulic jacks can be completely retracted after adjustment. The groove depth at their installation location is greater than the jack's maximum extension length, allowing them to retract flush with the bottom surface of the segment. This not only avoids interference with the permanent stress on the structure but also allows for continuous assembly of subsequent segments without additional disassembly. The coordinated monitoring of magnetostrictive displacement and pressure sensors can identify the contact status between the segment and the installed structure in real time during adjustment. When the pressure sensor detects an abnormal pressure surge, it automatically pauses adjustment and issues an alarm to prevent damage from toothed collisions—a passive protection feature not found in traditional adjustment methods. The parallel mechanism of the Stewart platform provides redundancy in the adjustment forces of the four jacks. When a single jack experiences a minor malfunction, the system can compensate by adjusting the outputs of the other three jacks, maintaining adjustment accuracy and improving the fault tolerance of the construction process. Without step S40, the pose error analysis of the digital twin model cannot be translated into actual adjustment actions, the entire closed-loop control chain breaks, segment assembly still relies on manual operation, and millimeter-level precision cannot be achieved; the lack of a dynamic correction mechanism will cause the deviation between the initial adjustment command and the actual error to increase over time, eventually leading to the recurrence of problems such as insufficient meshing of connection parts and eccentricity of channels. Step S40 establishes a real-time mapping and data interaction mechanism between the digital twin model and the physical construction process by converting the error information in the virtual space into adjustment actions in the physical space, realizing precise correlation and collaborative control between the virtual and physical spaces.

[0117] Step S50: Determine whether the position adjustment has reached position convergence. If position convergence has been achieved, then perform precise gear meshing and mechanical locking.

[0118] Further, step S50 includes:

[0119] For step S51, please refer to... Figure 4As shown, the method for determining whether the pose adjustment has reached pose convergence is as follows: determine whether the three translation components of the real-time pose error vector have converged in translation, and whether the three rotation components of the real-time pose error vector have converged in rotation. If both the translation convergence condition and the rotation convergence condition are met at the same time, it is determined that the pose has converged; otherwise, it is determined that the pose has not converged.

[0120] Step S52: After the pose convergence is achieved, the precision meshing accuracy and channel coaxiality are verified by using the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry of the prefabricated segment.

[0121] The method for verifying whether the precision meshing accuracy and the coaxiality of the channel meet the set accuracy standards includes: determining whether the precision meshing accuracy has been verified and whether the coaxiality accuracy of the channel has been verified. If both the precision meshing accuracy and the coaxiality accuracy of the channel have been verified, then it is determined that the set accuracy standards have been met.

[0122] Please see Figure 5 As shown, the method for determining whether the precision meshing accuracy has passed verification includes: extracting the three-dimensional contour line of the tooth groove contact surface based on the feature point cloud of the tooth groove boundary of the prefabricated segment end after pose convergence; calculating the spatial distance between the extracted three-dimensional contour line and the corresponding tooth groove contour of the adjacent installed segment to generate a contact surface gap distribution map; performing statistical analysis on the contact surface gap distribution map, and calculating the gap standard deviation as the gap uniformity deviation σ. gap The maximum value point in the gap distribution map is identified as the maximum local gap, and the precision meshing accuracy is verified. If the gap uniformity deviation is less than the uniformity limit and the maximum local gap is less than the maximum gap limit, the precision meshing accuracy verification is passed.

[0123] Please see Figure 6 As shown, the method for determining whether the coaxiality accuracy of the duct has been verified includes: extracting the centerline data of the prestressed ducts of the current segment and adjacent segments from the digital twin geometry of the precast segment, and calculating the coaxiality error of the duct; verifying the coaxiality accuracy of the duct based on the coaxiality error; if the coaxiality error of the duct is less than the set coaxiality standard, then the coaxiality accuracy verification of the duct has been passed.

[0124] Step S53: If the set accuracy standard is reached, a locking preparation command is issued, the micro hydraulic jack device is retrieved, and after retrieval, the prefabricated segment is permanently mechanically locked and a locking completion signal is fed back.

[0125] Specifically, the translation convergence condition refers to the absolute values ​​of the three translation components being less than or equal to the set translation error, and the rotation convergence condition refers to the absolute values ​​of the three rotation components being less than or equal to the set rotation error. The translation and rotation errors are determined based on the assembly accuracy requirements of the combined structure in engineering practice. For example, the translation error can refer to the minimum clearance requirement for tooth-groove fits in mechanical connection structures to ensure that the teeth do not interfere during docking. The rotation error can be derived based on the allowable deviation of the coaxiality of the channel to avoid channel misalignment due to rotational deviation. The judgment process also needs to meet the time stability requirement, i.e., the convergence condition must be met for five consecutive data acquisition cycles to prevent misjudgment caused by instantaneous errors reaching the standard but not stabilizing. The data acquisition cycle is determined according to the sensor's sampling frequency to ensure that the dynamic fluctuations of the segment can be captured. If the pose error does not meet the convergence condition, i.e., the pose has not converged, the dynamic correction process continues; if it has been met, the mechanical locking preparation stage begins.

[0126] After pose convergence, the precision meshing accuracy and channel coaxiality are verified using the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry. The specific process is as follows: Based on the feature point cloud of the tooth groove boundary after pose convergence, the three-dimensional contour line of the tooth groove contact surface is extracted. The extraction method is achieved by three-dimensional curve fitting of the feature point cloud, and the least squares method is used to fit the mathematical equation of the contour line to ensure that the contour line accurately reflects the actual shape of the tooth groove. The spatial distance between the extracted three-dimensional contour line and the corresponding tooth groove contour of the adjacent installed segment is calculated to generate a contact surface gap distribution map. The spatial distance is calculated as the Euclidean distance between corresponding point pairs on the contour line, and the distribution map uses color gradients to represent the gap. The size directly reflects the uniformity of contact. Statistical analysis is performed on the gap distribution map of the contact surface, and the standard deviation of the gap is calculated as the gap uniformity deviation. The standard deviation is obtained by taking the square root of the variance of all gap values, reflecting the degree of dispersion of the gap. The maximum value point in the gap distribution map is identified as the maximum local gap for precision meshing accuracy verification. If the gap uniformity deviation is less than the uniformity limit and the maximum local gap is less than the maximum gap limit, the precision meshing accuracy verification is passed. The uniformity limit is set at 0.2 mm, based on the fact that excessive gap non-uniformity will lead to stress concentration under force. The maximum gap limit is set at 0.3 mm, referring to the effective range of the mechanical locking device, to ensure that the gap can be eliminated after locking. The centerline data of the prestressed ducts for the current segment and adjacent segments are extracted from the digital twin geometry. The duct centerlines are generated using duct parameters in the BIM model and are spatial straight line equations. The coaxiality error of the ducts is calculated by uniformly selecting several points along the length of the two ducts, calculating the distance between the two axes at corresponding points, and taking the maximum value as the coaxiality error. If the coaxiality error is less than the set coaxiality standard, such as 1 mm, the coaxiality accuracy verification is passed. This standard is determined based on the diameter of the prestressed steel strand and the allowable deviation, ensuring that the steel strand does not rub against the duct wall when inserted. If both the precision meshing accuracy and the coaxiality accuracy of the ducts are verified as passed, the set accuracy standard is determined to have been met.

[0127] If the set accuracy standard is reached, a locking preparation command is issued, and the miniature hydraulic jack device is retrieved. During the retrieval process, the piston rod is controlled by the hydraulic system to retract until the jack is completely retracted into the reserved groove at the bottom of the precast segment. The groove depth is greater than the maximum extension length of the jack to ensure that it does not protrude from the segment surface after retrieval. After retrieval, the precast segment is permanently mechanically locked and a locking completion signal is fed back. The mechanical locking adopts a locking device that matches the original mechanical connection structure, such as a prestressed toothed groove mechanical locking device. The locking bolts are tightened by a hydraulic wrench until the design preload is reached. The preload value is determined according to the structural stress calculation to ensure the connection strength. After permanent mechanical locking, the final assembly pose matrix containing segment position and attitude information is recorded. This matrix is ​​generated using the coordinates and attitude parameters of the current segment in the reference coordinate system. The precast segment assembly reference coordinate system and digital twin geometry are updated, incorporating the actual pose of the locked segment into the reference as a reference for subsequent segment assembly. The cumulative geometric error of the assembled segment sequence is calculated by superimposing the pose error vectors of each segment and monitoring the cumulative coaxial error of the prestressed ducts to ensure that the overall error is within the allowable range. The accuracy of the pose drift prediction of the deep learning network is monitored. By comparing the deviation between the predicted and actual values, the parameters of the GNN-Transformer neural network model are optimized using an incremental learning algorithm. Incremental learning fine-tunes the network weights by introducing new assembly data to avoid forgetting historical data. After completing data recording and model updates, the updated digital twin model is used as the reference to start the next precast segment assembly task.

[0128] In existing technologies, traditional assembly methods lack clear convergence criteria, often resulting in insufficient accuracy due to premature locking or wasted time due to excessive adjustment. Furthermore, the lack of monitoring for error accumulation after mechanical locking affects subsequent segment assembly. Step S50, through rigorous convergence judgment combined with set translational and rotational errors and time stability requirements, ensures that segments are locked only after reaching stable standards, avoiding accuracy rebound caused by instantaneous errors and solving the problem of inaccurate locking timing in traditional methods. Verification of precision meshing and channel coaxiality specifically checks key accuracy indicators of the connection parts, ensuring full meshing of the teeth and continuity of the channels, providing a reliable foundation for subsequent construction. Model updates and error monitoring after mechanical locking enable the digital twin system to reflect the actual state of the assembled structure in real time, providing a dynamic benchmark for multi-segment series assembly and reducing error accumulation. The time stability requirement not only ensures the assembly accuracy of the current segment but also indirectly reflects the degree of influence of environmental disturbances. If stability is maintained for multiple consecutive cycles, it indicates that the influence of environmental factors such as wind load and cable tension has weakened. At this point, locking can reduce the impact of external disturbances on the connection. This implicit judgment of environmental adaptability is something that traditional static judgment methods do not possess. The design of the retractable jack ensures that the permanent stress of the structure is not affected, and its reserved grooves can serve as observation points for subsequent inspections after locking, facilitating non-destructive testing of the connection and improving the maintainability of the structure. The application of incremental learning algorithms enables the pose drift prediction model to be continuously optimized as construction progresses, maintaining high-precision predictions under different working conditions. Especially in the later stages of multi-segment assembly, the model's ability to predict cumulative errors is significantly improved, further reducing the number of adjustments.

[0129] Convergence judgment provides a clear termination condition for dynamic adjustment, avoiding over- or under-adjustment. Dynamic correction, through continuous approximation, ensures that the error reaches the convergence condition. These two processes form a closed loop of "adjustment-judgment-locking," making precision control more reliable. Precision meshing verification, based on the final state of the error vector and combined with detailed features of point cloud data, reflects the actual connection quality better than simply relying on the error vector, improving the comprehensiveness of precision verification. Without step S50, the segment locking timing cannot be scientifically determined, potentially leading to substandard connection accuracy or low construction efficiency. Furthermore, the lack of a model update mechanism would cause subsequent segment assembly to use the initial benchmark, accumulating errors and ultimately affecting the overall structural performance. Step S50, by completing the transition from dynamic adjustment to static locking, achieves real-time synchronization between the digital twin model and the physical structure, providing a crucial link for the closed-loop control of the entire assembly process, ensuring the consistency and reliability of multi-segment assembly accuracy.

[0130] Example 2:

[0131] This embodiment, based on Embodiment 1, provides a digital twin-adaptive assembly correction system for prefabricated segments of a combined structure, such as... Figure 7 As shown, it includes:

[0132] Twin model construction module; used to read BIM geometric model data and establish a digital twin geometry of the prefabricated segment assembly reference coordinate system and the prefabricated segment;

[0133] Feature point cloud extraction module: used to deploy composite sensor combinations at the four corners of the prefabricated segment, collect composite data of the prefabricated segment in real time, and extract feature point clouds of the tooth groove boundary at the end of the prefabricated segment;

[0134] Pose prediction module: used to dynamically register the feature point cloud of the tooth groove boundary of the prefabricated segment with the digital twin geometry of the prefabricated segment, calculate the six-degree-of-freedom pose error vector, and perform pose drift prediction based on the pre-built pose drift prediction model to obtain the total predicted pose error vector.

[0135] Pose adjustment drive module: Generates drive instructions for the jack based on the predicted total pose error vector and performs pose adjustment;

[0136] Position convergence judgment module: used to determine whether the position adjustment has reached position convergence. If position convergence is achieved, then perform precise gear meshing and mechanical locking.

[0137] In the feature point cloud extraction module, the method of deploying a composite sensor combination at the four corners of the prefabricated segment, collecting composite data of the prefabricated segment in real time, and extracting the feature point cloud of the tooth groove boundary at the end of the prefabricated segment includes:

[0138] Step S21: Deploy a composite sensor combination containing a lidar sensor and an IMU inertial measurement unit at the four corners of the prefabricated segment to collect composite data of the prefabricated segment in real time. The composite data of the prefabricated segment includes three-dimensional point cloud data of the surface of the prefabricated segment and six-degree-of-freedom pose data of the prefabricated segment; the six-degree-of-freedom pose data includes three-axis acceleration data and three-axis angular velocity data.

[0139] Step S22: Perform statistical filtering to remove noise from the three-dimensional point cloud data and convert it to the prefabricated segment assembly reference coordinate system to form integrated point cloud data;

[0140] Step S23: The supervoxel-principal curvature segmentation algorithm is used to automatically extract the feature point cloud of the prefabricated segment end tooth groove boundary from the integrated point cloud data.

[0141] Further, step S23 includes:

[0142] Step S231: The octree decomposition algorithm is used to divide the space of the integrated point cloud data into a multi-scale voxel grid to generate an initial voxel set.

[0143] Step S232: Perform local feature clustering on the point cloud within the initial voxel set to form multiple super voxel units;

[0144] Step S233: Calculate the covariance matrix of all points within each supervoxel unit, and obtain the curvature values ​​λ1, λ2, and λ3 in three principal directions through eigenvalue decomposition, where λ1 ≥ λ2 ≥ λ3. Define the principal curvature κ = λ1 / (λ1 + λ2 + λ3); where λ1 is the curvature value in the first principal direction, λ2 is the curvature value in the second principal direction, and λ3 is the curvature value in the third principal direction.

[0145] Step S234, set the curvature threshold κ thre Supervoxel units with principal curvature κ greater than curvature threshold are selected as high curvature supervoxel units, and high curvature supervoxel units correspond to the geometric abrupt change region of the tooth groove boundary.

[0146] Step S235: Perform spatial connectivity analysis on the selected high curvature supervoxel units, use a region growing algorithm to aggregate adjacent high curvature supervoxel units to form a continuous tooth groove boundary feature region, and extract all point cloud data within the tooth groove boundary feature region as the tooth groove boundary feature point cloud of the prefabricated segment end.

[0147] In the pose prediction module, the method for dynamically registering the point cloud of the prefabricated segment end tooth groove boundary features with the digital twin geometry of the prefabricated segment includes:

[0148] Step S311: The point cloud of the tooth groove boundary feature is spatially octree-layered to construct a multi-scale point cloud pyramid structure; the multi-scale point cloud pyramid structure includes coarse-scale level, medium-scale level and fine-scale level.

[0149] Step S312: At the coarse-scale level, the FPFH feature descriptor is used to perform initial registration between the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry of the prefabricated segment, obtaining the coarse registration transformation matrix T. coarse ;

[0150] Step S313, based on the coarse registration transformation matrix T coarse At the mesoscale level, the ICP iterative nearest point algorithm is executed to optimize the correspondence between the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry of the prefabricated segment.

[0151] Step S314: During the ICP iteration process, the six-degree-of-freedom pose data provided by the IMU inertial measurement unit is fused, and the state is predicted and updated through the EKF extended Kalman filter;

[0152] Step S315: At the fine-scale level, the point-to-surface distance metric is used instead of the point-to-point distance to refine the registration result, and the final fine registration transformation matrix T is output.fine .

[0153] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0154] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0155] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A digital twin-adaptive assembly correction method for prefabricated segments of a composite structure, characterized in that, The method includes: Read the BIM geometric model data and establish a reference coordinate system for prefabricated segment assembly and a digital twin geometry of the prefabricated segment; A composite sensor array is deployed at the four corners of the prefabricated segment to collect composite data of the prefabricated segment in real time and extract the feature point cloud of the tooth groove boundary at the end of the prefabricated segment; The point cloud of the tooth groove boundary feature of the prefabricated segment is dynamically registered with the digital twin geometry of the prefabricated segment, the six-degree-of-freedom pose error vector is calculated, and the pose drift prediction is performed based on the pre-built pose drift prediction model to obtain the total predicted pose error vector. The jack's driving command is generated based on the predicted total pose error vector, and pose adjustment is performed. Determine whether the position adjustment has met the position convergence condition. If the position convergence condition is met, then perform precise gear meshing and mechanical locking.

2. The digital twin-adaptive assembly correction method for prefabricated segments of a composite structure according to claim 1, characterized in that, The composite sensor assembly includes a lidar sensor and an IMU inertial measurement unit; the composite data of the prefabricated segment includes three-dimensional point cloud data of the surface of the prefabricated segment and six-degree-of-freedom pose data of the prefabricated segment; the six-degree-of-freedom pose data includes three-axis acceleration data and three-axis angular velocity data; The method for extracting the feature point cloud of the tooth groove boundary at the end of the prefabricated segment includes: performing statistical filtering to remove noise from the three-dimensional point cloud data and converting it to the prefabricated segment assembly reference coordinate system to form integrated point cloud data; The supervoxel-principal curvature segmentation algorithm is used to automatically extract the feature point cloud of the end tooth groove boundary of the prefabricated segment from the integrated point cloud data.

3. The digital twin-adaptive assembly correction method for prefabricated segments of a composite structure according to claim 2, characterized in that, The method for automatically extracting the feature point cloud of the end tooth groove boundary of the prefabricated segment from the integrated point cloud data using the supervoxel-principal curvature segmentation algorithm includes: An octree decomposition algorithm is used to divide the space of the integrated point cloud data into a multi-scale voxel grid to generate an initial voxel set; local feature clustering is performed on the point cloud within the initial voxel set to form multiple super voxel units. Calculate the covariance matrix of all points within each supervoxel unit, obtain the curvature values ​​of the three principal directions through eigenvalue decomposition, and define the principal curvature κ based on the curvature values ​​of the three principal directions; Set a curvature threshold and select supervoxel units with principal curvature κ greater than the curvature threshold as high curvature supervoxel units; Adjacent high-curvature supervoxel units are aggregated to form a continuous tooth groove boundary feature region. All point cloud data within the tooth groove boundary feature region are extracted as the tooth groove boundary feature point cloud of the prefabricated segment end.

4. The digital twin-adaptive assembly correction method for prefabricated segments of a composite structure according to claim 3, characterized in that, The curvature values ​​of the three principal directions are λ1, λ2, and λ3, respectively, representing the curvature values ​​of the first principal direction, the second principal direction, and the third principal direction. The principal curvature κ is the ratio of the curvature value of the first principal direction to the sum of the curvature values ​​of the three principal directions, where λ1 ≥ λ2 ≥ λ3.

5. The digital twin-adaptive assembly correction method for prefabricated segments of a composite structure according to claim 4, characterized in that, The method for dynamically registering the point cloud of the end tooth groove boundary features of the prefabricated segment with the digital twin geometry of the prefabricated segment includes: The point cloud of the tooth groove boundary features at the end of the prefabricated segment is hierarchically layered using a spatial octree to construct a multi-scale point cloud pyramid structure; the multi-scale point cloud pyramid structure includes coarse-scale, medium-scale, and fine-scale levels. At the coarse-scale level, the FPFH feature descriptor is used to perform initial registration between the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry of the prefabricated segment, obtaining the coarse registration transformation matrix T. coarse ; Based on the coarse registration transformation matrix T coarse At the mesoscale level, the ICP iterative nearest point algorithm is executed to optimize the correspondence between the feature point cloud of the tooth groove boundary at the end of the prefabricated segment and the digital twin geometry of the prefabricated segment. During the ICP iteration process, the six-degree-of-freedom pose data provided by the IMU inertial measurement unit is integrated, and the state is predicted and updated through the EKF extended Kalman filter. At a fine-scale level, a point-to-surface distance metric is used instead of a point-to-point distance metric to output the final fine registration transformation matrix T. fine .

6. The digital twin-adaptive assembly correction method for prefabricated segments of a composite structure according to claim 5, characterized in that, The method for calculating the six-degree-of-freedom pose error vector includes: The fine registration transformation matrix T fine With the design target pose matrix T of the prefabricated segment target Perform matrix operations to calculate the pose deviation matrix ΔT; Singular value decomposition is performed on the pose deviation matrix ΔT to extract a 3×3 rotation matrix R and a 3×1 translation vector t; Based on the rotation matrix R, the Rodriguez formula is used to calculate the rotation axis vector n and the rotation angle θ. The rotation axis vector n and rotation angle θ are decomposed into the X, Y, and Z axes of the prefabricated segment assembly reference coordinate system to obtain the three rotational components of the pose error. The three translation components of the pose error are directly extracted from the translation vector t; The three rotational components of the combined pose error and the three translational components of the combined pose error form a complete six-degree-of-freedom pose error vector ε.

7. The digital twin-adaptive assembly correction method for prefabricated segments of a composite structure according to claim 6, characterized in that, The method for predicting pose drift includes: Simultaneously collect environmental factor data and fuse it with the six-degree-of-freedom pose error vector to form a comprehensive state vector; The comprehensive state vector is input into the pre-constructed pose drift prediction model to obtain the predicted pose error vector.

8. The digital twin-adaptive assembly correction method for prefabricated segments of a composite structure according to claim 7, characterized in that, The method for obtaining the predicted total pose error vector includes: adding the predicted pose error vector to the six-degree-of-freedom pose error vector of the current position to obtain the predicted total pose error vector.

9. The digital twin-adaptive assembly correction method for prefabricated segments of a composite structure according to claim 8, characterized in that, The method for determining whether pose adjustment has met the pose convergence condition includes: Obtain the real-time pose error vector, and determine whether the three translational components of the real-time pose error vector converge, and whether the three rotational components of the real-time pose error vector converge. If both the translational convergence condition and the rotational convergence condition are met, then the pose is determined to be converged.

10. A digital twin-adaptive assembly correction system for prefabricated composite structures, used to implement the digital twin-adaptive assembly correction method for prefabricated composite structures as described in any one of claims 1-9, characterized in that, The system includes: Twin model construction module; used to read BIM geometric model data and establish a digital twin geometry of the prefabricated segment assembly reference coordinate system and the prefabricated segment; Feature point cloud extraction module: used to deploy composite sensor combinations at the four corners of the prefabricated segment, collect composite data of the prefabricated segment in real time, and extract feature point clouds of the tooth groove boundary at the end of the prefabricated segment; Pose prediction module: used to dynamically register the feature point cloud of the tooth groove boundary of the prefabricated segment with the digital twin geometry of the prefabricated segment, calculate the six-degree-of-freedom pose error vector, and perform pose drift prediction based on the pre-built pose drift prediction model to obtain the total predicted pose error vector. Pose adjustment drive module: Generates drive instructions for the jack based on the predicted total pose error vector and performs pose adjustment; Position convergence judgment module: used to determine whether the position adjustment has reached the position convergence condition. If the position convergence condition is reached, then perform precise gear meshing and mechanical locking.

Images