Steel structure teachless welding system based on BIM online analysis and real-time visual registration

Abstract

This invention relates to the field of steel structure welding technology and discloses a teach-free welding system for steel structures based on BIM online analysis and real-time visual registration. The system includes analyzing the Building Information Model (BIM) to extract physical manufacturing parameters, constructing a theoretically weighted undirected graph with mechanical semantic edge weights and calculating the theoretical Laplace matrix, obtaining the workpiece's 3D point cloud to extract the topological skeleton, constructing the actual undirected graph and its corresponding Laplace matrix, performing feature comparison in spectral space to establish a mapping and complete coarse registration, constructing a non-rigid deformation energy functional in manifold space, and using the mechanical semantic edge weights as constraints to generate a continuous deformation field. This invention breaks the assumption of absolute rigidity, endowing the underlying algorithm with physical perception capabilities. It can not only flexibly absorb physical deformations caused by gravity and assembly tolerances but also dynamically offset welding thermal stress, overcoming the misalignment problem caused by traditional geometric registration and improving processing accuracy.

Description

Technical Field

[0001] This invention relates to the field of steel structure welding technology, specifically to a steel structure welding system without teaching based on BIM online analysis and real-time visual registration. Background Technology

[0002] As the framework of modern buildings, bridges, and large industrial facilities, the manufacturing quality of steel structures directly affects the safety and lifespan of the entire project. Welding is the most crucial connection process in steel structure processing and manufacturing. However, traditional construction operations have long relied heavily on manual labor, which is not only labor-intensive and involves harsh working environments, but also subject to significant subjective factors such as the welder's skill level and fatigue level, making it difficult to guarantee the stability and consistency of processing. With the increasing shortage of skilled welders and the urgent need for efficient and high-quality manufacturing in modern industry, the use of industrial robots to replace manual labor for automated welding of steel structures has become an inevitable trend in industrial transformation and upgrading.

[0003] To overcome the drawbacks of traditional point-to-point robot teaching programming, such as long processing time and inability to adapt to multi-variety, small-batch, non-standard production, a teaching-free welding technology based on digital models and 3D vision has emerged. This technology typically directly analyzes the BIM or CAD theoretical model of the steel structure to extract the ideal weld trajectory. At the same time, it uses 3D vision sensors installed at the workstation or robot end effector to scan the workpiece on site and obtain the 3D point cloud data of the actual surface. Through underlying algorithms, the theoretical 3D model and the actual visual point cloud are spatially registered. The system can automatically calculate the actual pose of the workpiece on the worktable and then convert the theoretical trajectory into an actual motion trajectory that the robot can execute, thus initially realizing flexible automatic processing without human intervention.

[0004] However, existing teach-free welding systems are generally based on the ideal assumption that the workpiece is an absolutely rigid body. In actual manufacturing, large non-standard steel structures are prone to significant non-rigid bending and twisting due to their own weight and accumulated tolerances during assembly. Existing technologies mostly use purely geometric rigid coordinate registration algorithms, which are not only highly dependent on the initial placement of the workpiece but also easily affected by noise such as on-site reflections and welding slag. Furthermore, this type of purely mathematical matching completely separates the physical manufacturing attributes such as plate thickness and section moment of inertia contained in the BIM model, making it impossible to perceive the local deformation resistance of the component. Moreover, existing solutions are all static registrations, failing to take into account the dynamic thermal stress deformation trend during the welding process. As a result, the generated rigid trajectory frequently suffers from off-center welding, missed welding, and severe trajectory misalignment when facing workpieces that actually undergo flexible deformation and thermal warping, making it difficult to meet the high-precision teach-free welding requirements of heavy steel structures. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a steel structure teachless welding system based on BIM online analysis and real-time visual registration. This solves the problems of existing teachless welding systems, which are based on the assumption of an absolute rigid body and pure geometric rigid registration, and cannot perceive the mechanical properties of components. They are also unable to adapt to the non-rigid physical deformation of large non-standard steel structures, resulting in serious misalignment of the actual welding trajectory, off-center welding, and missed welding.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a steel structure teachless welding system based on BIM online analysis and real-time visual registration, comprising the following steps:

[0007] S1: Online parsing of steel structure BIM model, extraction of physical manufacturing parameters from steel structure BIM model, construction of theoretical weighted undirected graph containing theoretical nodes and mechanical semantic edge weights, wherein the mechanical semantic edge weights are calculated based on the physical manufacturing parameters, and the corresponding theoretical Laplace matrix is ​​calculated according to the theoretical weighted undirected graph;

[0008] S2: Acquire the three-dimensional point cloud data of the actual steel structure workpiece in real time through a vision sensor, extract the global topological skeleton of the three-dimensional point cloud data, construct an actual undirected graph containing actual nodes based on the global topological skeleton, and calculate the actual Laplacian matrix corresponding to the actual undirected graph.

[0009] S3: Perform generalized eigenvalue decomposition on the theoretical Laplacian matrix and the actual Laplacian matrix respectively, map the eigenvalue decomposition results to the Laplacian spectrum space for feature comparison, establish a global topological mapping relationship between the theoretical node and the actual node, and obtain the coarse registration result between the steel structure BIM model and the actual steel structure workpiece.

[0010] S4: Based on the global topological mapping relationship and the coarse registration result, a non-rigid deformation energy functional is constructed in the Lie group manifold space to map the theoretical nodes to the actual nodes. The mechanical semantic edge weights are added as constraints of the adaptive manifold regularization mechanism to the non-rigid deformation energy functional to participate in the optimization solution, thereby generating a continuous deformation field that fits the shape of the actual steel structure workpiece and completing the real-time visual registration of the actual steel structure workpiece.

[0011] S5: Extract the preset welding process parameters from the steel structure BIM model, calculate the expected heat input of the area to be welded based on the preset welding process parameters, convert the expected heat input into a thermal deformation perturbation term, introduce the thermal deformation perturbation term into the continuous deformation field for reverse pre-compensation, and generate a robot teachless welding trajectory based on the continuous deformation field after reverse pre-compensation.

[0012] Preferably, the physical manufacturing parameters include the local plate thickness parameters, section moment of inertia, and material yield strength of the corresponding component in the steel structure BIM model;

[0013] The construction of the theoretically weighted undirected graph, which includes theoretical nodes and mechanical semantic edge weights, includes:

[0014] The assembly nodes or feature endpoints in the steel structure BIM model are extracted as theoretical nodes, and the potential welds connecting the theoretical nodes are extracted as edges of an undirected graph.

[0015] The stiffness evaluation function, composed of the local plate thickness parameter, the moment of inertia of the section, and the yield strength of the material, is used to calculate the local deformation resistance between adjacent theoretical nodes, and the local deformation resistance is used as the mechanical semantic edge weight of the corresponding edge.

[0016] Preferably, the step of calculating the corresponding theoretical Laplacian matrix based on the theoretically weighted undirected graph includes:

[0017] The weighted adjacency matrix is ​​constructed using the mechanical semantic edge weights as matrix elements;

[0018] Calculate the diagonal matrix, whose diagonal elements are the sum of the weights of the connected mechanical semantic edges;

[0019] The difference between the identity matrix and the normalized adjacency matrix is ​​calculated using the identity matrix, the inverse square root matrix of the angle matrix, and the weighted adjacency matrix to obtain the theoretical Laplace matrix containing the prior physical stiffness.

[0020] Preferably, the step of extracting the global topological skeleton of the 3D point cloud data, constructing an actual undirected graph containing actual nodes based on the global topological skeleton, and calculating the actual Laplacian matrix corresponding to the actual undirected graph includes:

[0021] A graph segmentation algorithm is used to extract the global topological skeleton representing the workpiece topology from the 3D point cloud data containing environmental noise;

[0022] Extract the skeleton endpoints or intersections of the global topology skeleton as the actual nodes, and establish the actual undirected graph;

[0023] The geometric adjacency matrix and actual degree matrix of the actual undirected graph are constructed using a spatial distance function, and the normalized actual Laplacian matrix is ​​calculated based on the geometric adjacency matrix and the actual degree matrix.

[0024] Preferably, the step of mapping the eigenvalue decomposition results to the Laplacian spectral space for feature comparison and establishing a global topological mapping relationship between the theoretical nodes and the actual nodes includes:

[0025] Obtain the first eigenvalue sequence and the first eigenvector signature obtained from the theoretical Laplace matrix decomposition, and the second eigenvalue sequence and the second eigenvector signature obtained from the actual Laplace matrix decomposition;

[0026] In the dimensionality-reduced high-dimensional spectral space, the first eigenvector signature is compared with the second eigenvector signature to construct the spectral domain allocation cost matrix.

[0027] An allocation algorithm is used to solve for the optimal permutation matrix of the spectral domain allocation cost matrix. The optimal permutation matrix is ​​used to realize the one-to-one mapping between the theoretical nodes and the actual nodes in the topological space, so as to get rid of the dependence on the initial three-dimensional pose of the actual steel structure workpiece.

[0028] Preferably, constructing the non-rigid deformable energy functional mapping the theoretical nodes to the actual nodes in the Lie group manifold space includes:

[0029] A control lattice network is constructed along the three-dimensional theoretical weld seam of the steel structure BIM model, and each lattice node in the control lattice network is set to correspond to a Lie algebra pose increment.

[0030] Construct the non-rigid deformation energy functional that includes a data fitting term and a manifold regularization term;

[0031] The data fitting term is used to characterize the error between the theoretical three-dimensional coordinate points and the actual three-dimensional coordinate points after mapping to the Lie group through the exponential mapping operator, and the manifold regularization term is used to control the smoothness of the deformation of adjacent lattice nodes through adaptive constraint parameters.

[0032] Preferably, the step of incorporating the mechanical semantic edge weights as constraints of the adaptive manifold regularization mechanism into the non-rigid deformation energy functional to participate in the optimization solution, thereby generating a continuous deformation field that fits the actual steel structure workpiece morphology, includes:

[0033] The mechanical semantic edge weights are directly mapped to adaptive constraint parameters for controlling the deformation smoothness of adjacent lattice nodes by a monotonically increasing mapping function. Strong penalty constraints are generated for high stiffness regions and weak penalty constraints are generated for low stiffness regions.

[0034] The optimal solution of the non-rigid deformation energy functional is obtained by iteratively solving the Lie algebra tangent space using a nonlinear least squares optimization algorithm, and the continuous deformation field corresponding to the optimal solution is output.

[0035] Preferably, the preset welding process parameters include welding voltage, welding current, and welding speed;

[0036] The step of calculating the expected heat input to the area to be welded based on the preset welding process parameters, and converting the expected heat input into a thermal deformation perturbation term, includes:

[0037] The expected heat input is calculated by dividing the product of the arc thermal efficiency, the welding voltage and the welding current by the welding speed.

[0038] Based on the coupling relationship between the local inverse stiffness of the component and the linear energy per unit length, the thermal deformation perturbation term is calculated and generated in the Lie algebra increment field.

[0039] Preferably, the step of introducing the thermal deformation perturbation term into the continuous deformation field for reverse pre-compensation, and generating a robot teachless welding trajectory based on the continuous deformation field after reverse pre-compensation, includes:

[0040] The thermal deformation perturbation term is used as a compensation value and is inversely added to the Lie algebra pose increment after optimization.

[0041] Using the Lie group exponential mapping rule, the Lie algebraic pose increment after inverse compensation is combined and transformed with the ideal theoretical pose extracted from the steel structure BIM model to calculate the final continuous pose of the center point of the welding robot end tool.

[0042] The six-degree-of-freedom spatial trajectory sequence containing the final continuous pose is used as the robot's teach-free welding trajectory.

[0043] Preferred, a steel structure teachless welding system based on BIM online analysis and real-time visual registration includes:

[0044] The online parsing and mechanical graph construction module is used to parse the steel structure BIM model online, extract physical manufacturing parameters, construct a theoretical weighted undirected graph containing theoretical nodes and mechanical semantic edge weights, and calculate the corresponding theoretical Laplace matrix.

[0045] The visual skeleton extraction and actual graph construction module is used to acquire 3D point cloud data of actual steel structure workpieces through visual sensors, extract the global topological skeleton, construct an actual undirected graph containing actual nodes, and calculate the corresponding actual Laplacian matrix.

[0046] The spectral space matching and coarse registration module is used to perform generalized eigenvalue decomposition on the two Laplacian matrices respectively, perform feature comparison in the Laplacian spectral space, establish a global topological mapping relationship, and obtain the coarse registration result.

[0047] The physical constraint manifold optimization module is used to construct a non-rigid deformation energy functional in the Lie group manifold space, and use the mechanical semantic edge weights as constraints of the adaptive manifold regularization mechanism to participate in the optimization solution, generate a continuous deformation field, and complete real-time visual registration.

[0048] The thermal compensation and trajectory generation module is used to calculate the expected heat input based on the extracted preset welding process parameters, convert the expected heat input into a thermal deformation perturbation term for reverse pre-compensation in the continuous deformation field, and generate a robot teachless welding trajectory.

[0049] This invention provides a steel structure welding system without teaching, based on BIM online analysis and real-time visual registration. It has the following beneficial effects:

[0050] 1. This invention constructs a non-rigid deformation energy functional in the Lie group manifold space and uses the mechanical semantic edge weights extracted from the physical manufacturing parameters of the BIM model as constraints in the optimization solution of the adaptive manifold regularization mechanism. This breaks the conventional thinking of treating large steel structures as absolutely rigid bodies in the prior art and endows the algorithm with the underlying physical perception capability. This allows the generated continuous deformation field to be adaptively stretched or kept rigid according to the actual local deformation resistance of the component. The theoretical welding trajectory is flexibly and accurately fitted to the actual workpiece that undergoes physical deformation such as bending and twisting, thereby solving the problem of welding misalignment and trajectory misalignment caused by self-weight and assembly tolerance in large non-standard steel structures.

[0051] 2. This invention extracts the Laplacian matrix from the theoretical graphical model containing physical stiffness and the actual point cloud topological skeleton, respectively, and establishes a global mapping relationship by performing generalized eigenvalue decomposition and eigenvector signature comparison in the dimensionality-reduced Laplacian spectrum space. This avoids the shortcomings of traditional ICP point cloud registration algorithms based on the three-dimensional Cartesian coordinate system, which have large computational load and are prone to getting trapped in local optima. Since the physical deformation of the same component has high isomorphic invariance in the topological spectrum space, the cross-domain coarse registration mechanism gets rid of the dependence on the initial placement pose of the workpiece on the workbench. At the same time, it shows strong robustness to high-frequency environmental noise such as reflection, welding slag, and rust on the surface of the workpiece, and improves the stability of teaching-free registration in complex workshop environments.

[0052] 3. This invention extracts preset welding process parameters from the BIM model to calculate the expected heat input, and combines them with the local stiffness relationship of the component to convert them into thermal deformation perturbation terms. Then, it performs reverse dynamic pre-compensation in the generated continuous deformation field, realizing closed-loop control of component processing errors. This not only corrects the static geometric deviations and mechanical deformations that already exist before processing, but also offsets the dynamic thermal deformation trend that will be generated due to the local drastic temperature rise during the actual welding operation. This makes the final output robot teaching-free welding trajectory have a high degree of foresight and self-correction ability, ensuring stable forming quality throughout the welding process.

[0053] 4. This invention analyzes the mechanical semantics in the BIM model online and deeply couples them into the entire mathematical derivation process of visual skeleton extraction, spectral space coarse registration, and manifold optimization fine registration. This forms a hardware and software collaborative architecture driven by physical logic. By integrating the pre-design attributes with the post-registration algorithm, it eliminates the simple serial data splicing between processing modules in the past. It eliminates the need for manual intervention, teaching, or repeated debugging and calibration parameters, and opens up a fully automatic control link from the theoretical digital design end to the complex physical execution end. This improves the intelligence level and production execution continuity of non-standard customized steel component manufacturing. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the overall steps of the present invention;

[0055] Figure 2 This is a flowchart of the BIM online analysis and theoretical diagram construction process of the present invention;

[0056] Figure 3 This is a flowchart of the visual point cloud extraction and actual image construction process of the present invention;

[0057] Figure 4 This is a flowchart of the global coarse registration process in spectral space according to the present invention;

[0058] Figure 5 This is a flowchart of the physical constraint manifold optimization fine registration process of the present invention;

[0059] Figure 6 This is a flowchart of the thermal compensation and collision-free trajectory generation process of the present invention. Detailed Implementation

[0060] The technical solutions in 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.

[0061] See attached document Figure 1 - Appendix Figure 6 This invention provides a steel structure teach-free welding system based on BIM online analysis and real-time visual registration, comprising the following steps:

[0062] S1: Online parsing of steel structure BIM model, extraction of physical manufacturing parameters from steel structure BIM model, construction of theoretical weighted undirected graph containing theoretical nodes and mechanical semantic edge weights, wherein the mechanical semantic edge weights are calculated based on the physical manufacturing parameters, and the corresponding theoretical Laplace matrix is ​​calculated according to the theoretical weighted undirected graph;

[0063] The physical manufacturing parameters include the local plate thickness parameters, cross-sectional moment of inertia, and material yield strength of the corresponding components in the steel structure BIM model;

[0064] The construction of the theoretically weighted undirected graph, which includes theoretical nodes and mechanical semantic edge weights, includes:

[0065] The assembly nodes or feature endpoints in the steel structure BIM model are extracted as theoretical nodes, and the potential welds connecting the theoretical nodes are extracted as edges of an undirected graph.

[0066] The stiffness evaluation function, which is composed of the local plate thickness parameter, the moment of inertia of the section and the yield strength of the material, is used to calculate the local deformation resistance between adjacent theoretical nodes, and the local deformation resistance is used as the mechanical semantic edge weight of the corresponding edge.

[0067] The calculation of the corresponding theoretical Laplacian matrix based on the theoretically weighted undirected graph includes:

[0068] The weighted adjacency matrix is ​​constructed using the mechanical semantic edge weights as matrix elements;

[0069] Calculate the diagonal matrix, whose diagonal elements are the sum of the weights of the connected mechanical semantic edges;

[0070] The difference between the identity matrix and the normalized adjacency matrix is ​​calculated using the identity matrix, the inverse square root matrix of the angle matrix, and the weighted adjacency matrix to obtain the theoretical Laplace matrix containing the prior physical stiffness.

[0071] Specifically, the system parses the steel structure BIM model online through API interfaces or IFC standard format files, and extracts assembly nodes or feature endpoints from the BIM model as a theoretical node set. The component entities or potential welds connecting these theoretical nodes are extracted as edge sets. This allows the BIM model to be abstracted into a graph structure. At the same time, the system deeply extracts the physical manufacturing parameters from the BIM attribute information, specifically including the local plate thickness parameters of the corresponding components. Moment of inertia of cross section and material yield strength And construct a stiffness evaluation function. The above physical parameters are transformed into adjacent theoretical nodes. and The local resistance to deformation between them, i.e., the mechanical semantic edge weights :

[0072] ;

[0073] In this embodiment, the stiffness evaluation function A linear weighted summation form can be used, that is: ,in , , The positive real-valued weighting coefficients, preset based on the mechanical properties of the component, can be determined through finite element calibration or empirical adjustment. It is understood that this function can also take other monotonic mapping forms such as product or exponential.

[0074] This weight reflects the physical rigidity properties at the component connection. The system constructs a weighted adjacency matrix using the mechanical semantic edge weights as matrix elements. (Elements corresponding to non-adjacent nodes are 0), further calculate the angle matrix. Its diagonal elements are the sum of the weights of the connected edges, that is:

[0075] ;

[0076] Finally, the identity matrix is ​​used. After normalization, the theoretical Laplace matrix containing the prior physical stiffness is calculated. :

[0077] ;

[0078] S2: Acquire the three-dimensional point cloud data of the actual steel structure workpiece in real time through a vision sensor, extract the global topological skeleton of the three-dimensional point cloud data, construct an actual undirected graph containing actual nodes based on the global topological skeleton, and calculate the actual Laplacian matrix corresponding to the actual undirected graph.

[0079] The process of extracting the global topological skeleton of the 3D point cloud data, constructing an actual undirected graph containing actual nodes based on the global topological skeleton, and calculating the actual Laplacian matrix corresponding to the actual undirected graph includes:

[0080] A graph segmentation algorithm is used to extract the global topological skeleton representing the workpiece topology from the 3D point cloud data containing environmental noise;

[0081] Extract the skeleton endpoints or intersections of the global topology skeleton as the actual nodes, and establish the actual undirected graph;

[0082] The geometric adjacency matrix and actual degree matrix of the actual undirected graph are constructed using a spatial distance function, and the normalized actual Laplacian matrix is ​​calculated based on the geometric adjacency matrix and the actual degree matrix.

[0083] Specifically, a 3D vision sensor (such as a 3D structured light camera) installed at the end of the welding robot or above the workstation is first used to scan the steel structure workpiece on the actual workbench to obtain actual 3D point cloud data with its own weight deformation and environmental noise (such as rust, welding slag, and reflection). To overcome the interference of local noise on feature point extraction, the system introduces a graph segmentation algorithm (such as the Reeb Graph algorithm) to extract a global topological skeleton representing the macroscopic topological structure of the workpiece from the point cloud, and extract the endpoints or structural intersections of the skeleton as actual nodes. The skeleton lines serve as edges. Construct an actual undirected graph The Gaussian spatial distance function is used to calculate the connection relationships between actual nodes, and a geometric adjacency matrix is ​​constructed. With the actual degree matrix And calculate the actual Laplace matrix according to the same normalization formula. :

[0084] ;

[0085] S3: Perform generalized eigenvalue decomposition on the theoretical Laplacian matrix and the actual Laplacian matrix respectively, map the eigenvalue decomposition results to the Laplacian spectrum space for feature comparison, establish a global topological mapping relationship between the theoretical node and the actual node, and obtain the coarse registration result between the steel structure BIM model and the actual steel structure workpiece.

[0086] The step of mapping the eigenvalue decomposition results to the Laplacian spectral space for feature comparison and establishing a global topological mapping relationship between the theoretical nodes and the actual nodes includes:

[0087] Obtain the first eigenvalue sequence and the first eigenvector signature obtained from the theoretical Laplace matrix decomposition, and the second eigenvalue sequence and the second eigenvector signature obtained from the actual Laplace matrix decomposition;

[0088] In the dimensionality-reduced high-dimensional spectral space, the first eigenvector signature is compared with the second eigenvector signature to construct the spectral domain allocation cost matrix.

[0089] An allocation algorithm is used to solve for the optimal permutation matrix of the spectral domain allocation cost matrix. The optimal permutation matrix is ​​used to realize the one-to-one mapping between the theoretical nodes and the actual nodes in the topological space, so as to get rid of the dependence on the initial three-dimensional pose of the actual steel structure workpiece.

[0090] Specifically, we first perform generalized eigenvalue decomposition on the two Laplacian matrices obtained earlier:

[0091] ;

[0092] ;

[0093] Obtain the theoretical first eigenvalue sequence Signature with the first feature vector and the actual second eigenvalue sequence With the second feature vector signature Since the steel structure components are highly isomorphic in terms of topological connection, although they undergo bending deformation in physical space, the eigenvector signatures of these two states in the Laplace spectral space have rotation, translation and non-rigid deformation invariance.

[0094] In the dimensionality-reduced high-dimensional spectral space, the system calculates the difference between the first eigenvector signature and the second eigenvector signature, constructs the spectral domain allocation cost matrix, and then calls the Hungarian Algorithm to solve for the optimal permutation matrix. Through this matrix, the system accurately achieves a one-to-one correspondence between BIM theoretical nodes and actual point cloud skeleton nodes without relying on any 3D initial pose, thus completing coarse registration.

[0095] S4: Based on the global topological mapping relationship and the coarse registration result, a non-rigid deformation energy functional is constructed in the Lie group manifold space to map the theoretical nodes to the actual nodes. The mechanical semantic edge weights are added as constraints of the adaptive manifold regularization mechanism to the non-rigid deformation energy functional to participate in the optimization solution, thereby generating a continuous deformation field that fits the shape of the actual steel structure workpiece and completing the real-time visual registration of the actual steel structure workpiece.

[0096] The construction of the non-rigid deformable energy functional mapping the theoretical nodes to the actual nodes in the Lie group manifold space includes:

[0097] A control lattice network is constructed along the three-dimensional theoretical weld seam of the steel structure BIM model, and each lattice node in the control lattice network is set to correspond to a Lie algebra pose increment.

[0098] Construct the non-rigid deformation energy functional that includes a data fitting term and a manifold regularization term;

[0099] The data fitting term is used to characterize the error between the theoretical three-dimensional coordinate points and the actual three-dimensional coordinate points after mapping to the Lie group through the exponential mapping operator, and the manifold regularization term is used to control the smoothness of the deformation of adjacent lattice nodes through adaptive constraint parameters.

[0100] The step of incorporating the mechanical semantic edge weights as constraints of the adaptive manifold regularization mechanism into the non-rigid deformation energy functional for optimization, thereby generating a continuous deformation field that conforms to the actual morphology of the steel structure workpiece, includes:

[0101] The mechanical semantic edge weights are directly mapped to adaptive constraint parameters for controlling the deformation smoothness of adjacent lattice nodes by a monotonically increasing mapping function. Strong penalty constraints are generated for high stiffness regions and weak penalty constraints are generated for low stiffness regions.

[0102] The optimal solution of the non-rigid deformation energy functional is obtained by iteratively solving the Lie algebra tangent space using a nonlinear least squares optimization algorithm, and the continuous deformation field corresponding to the optimal solution is output.

[0103] Specifically, the system constructs a control lattice network along the BIM 3D theoretical weld seam and sets each lattice node. The pose change is determined by a Lie algebra pose increment. This indicates the construction of a continuous non-rigid deformation energy functional. It consists of a data fitting term and a manifold regularization term:

[0104] ;

[0105] in, Let be the set of all Lie algebra increments. For Lie algebra to Lie groups The exponential mapping matrix, and These are the corresponding theoretical and actual 3D coordinate points for matching.

[0106] In this embodiment, the adaptive constraint parameters of the manifold regularization term It is not a global constant, but rather a mapping function that is monotonically increasing. The mechanical semantic edge weights extracted in step S1 Force binding, i.e. This enables the algorithm to possess physical awareness, allowing it to detect when the lattice is located in a high-stiffness region as determined by BIM analysis. big, This generates strong penalty constraints, forcing adjacent nodes in the region to maintain consistent pose increments, exhibiting rigid body motion. This is especially relevant when the lattice is in a low-stiffness or assembly seam region. Weak penalty constraints are generated, allowing the deformation field to undergo stretching and flexible bending under the traction of the visual data fitting term. Finally, the Levenberg-Marquardt nonlinear least squares algorithm is used to iteratively solve the problem in the manifold tangent space, generating a smooth deformation field that perfectly fits the shape of the actual workpiece.

[0107] S5: Extract the preset welding process parameters from the steel structure BIM model, calculate the expected heat input of the area to be welded based on the preset welding process parameters, convert the expected heat input into a thermal deformation perturbation term, introduce the thermal deformation perturbation term into the continuous deformation field for reverse pre-compensation, and generate a robot teaching-free welding trajectory based on the continuous deformation field after reverse pre-compensation.

[0108] The preset welding process parameters include welding voltage, welding current, and welding speed;

[0109] The step of calculating the expected heat input to the area to be welded based on the preset welding process parameters, and converting the expected heat input into a thermal deformation perturbation term, includes:

[0110] The expected heat input is calculated by dividing the product of the arc thermal efficiency, the welding voltage and the welding current by the welding speed.

[0111] Based on the coupling relationship between the local stiffness reciprocal of the component and the linear energy per unit length, the thermal deformation perturbation term is calculated and generated in the Lie algebra increment field;

[0112] The step of introducing the thermal deformation perturbation term into the continuous deformation field for reverse pre-compensation, and generating a robot-less welding trajectory based on the reverse pre-compensated continuous deformation field, includes:

[0113] The thermal deformation perturbation term is used as a compensation value and is inversely added to the Lie algebra pose increment after optimization.

[0114] Using the Lie group exponential mapping rule, the Lie algebraic pose increment after inverse compensation is combined and transformed with the ideal theoretical pose extracted from the steel structure BIM model to calculate the final continuous pose of the center point of the welding robot end tool.

[0115] The six-degree-of-freedom spatial trajectory sequence containing the final continuous pose is used as the robot's teach-free welding trajectory.

[0116] Specifically, the system extracts preset welding process parameters from the BIM model, including arc thermal efficiency. Welding voltage Welding current and welding speed The expected heat input (linear energy per unit length) of the area to be welded is calculated using the following formula:

[0117] ;

[0118] Since the smaller the local stiffness, the more significant the deformation caused by the same heat input, the system constructs a thermal deformation distribution model, taking the expected heat input... Reciprocal of local stiffness The coupled calculations are performed, transforming it into a thermal deformation vibration term in the Lie algebra increment field. ;

[0119] In this embodiment, the coupling calculation can adopt a linear coupling model, that is:

[0120] ;

[0121] in This is the thermo-mechanical coupling constant obtained through material thermophysical parameters and experimental calibration. This model reflects the basic trend that the smaller the local stiffness and the larger the heat input, the more significant the thermal deformation perturbation, but it is not the only way to achieve this.

[0122] When generating the final trajectory, the thermal deformation vibration term is added as a reverse compensation value to the optimized Lie algebraic pose increment, and the final continuous pose of the welding robot end-tool center point (TCP) is calculated using the following formula. :

[0123] ;

[0124] in, The system extracts the intangible ideal theoretical pose from the BIM model, calculates the pose of all discrete points on the weld in sequence, generates a six-degree-of-freedom trajectory sequence containing spatial XYZ coordinates and three-dimensional attitude angles, and sends it to the robot's underlying controller. This trajectory not only adapts to the physical deviation of the actual workpiece, but also preemptively offsets the thermal stress deformation that will occur during welding, thus achieving high-quality, teach-free autonomous welding operations.

[0125] As a specific implementation method, a steel structure teachless welding system based on BIM online analysis and real-time visual registration includes:

[0126] The online parsing and mechanical graph construction module is used to parse the steel structure BIM model online, extract physical manufacturing parameters, construct a theoretical weighted undirected graph containing theoretical nodes and mechanical semantic edge weights, and calculate the corresponding theoretical Laplace matrix.

[0127] The visual skeleton extraction and actual graph construction module is used to acquire 3D point cloud data of actual steel structure workpieces through visual sensors, extract the global topological skeleton, construct an actual undirected graph containing actual nodes, and calculate the corresponding actual Laplacian matrix.

[0128] The spectral space matching and coarse registration module is used to perform generalized eigenvalue decomposition on the two Laplacian matrices respectively, perform feature comparison in the Laplacian spectral space, establish a global topological mapping relationship, and obtain the coarse registration result.

[0129] The physical constraint manifold optimization module is used to construct a non-rigid deformation energy functional in the Lie group manifold space, and use the mechanical semantic edge weights as constraints of the adaptive manifold regularization mechanism to participate in the optimization solution, generate a continuous deformation field, and complete real-time visual registration.

[0130] The thermal compensation and trajectory generation module is used to calculate the expected heat input based on the extracted preset welding process parameters, convert the expected heat input into a thermal deformation perturbation term for reverse pre-compensation in the continuous deformation field, and generate a robot teachless welding trajectory.

[0131] Specifically, the system first reads the BIM model of the building steel structure through the online parsing and mechanical graph construction module, extracts the physical manufacturing parameters to construct a theoretically weighted undirected graph containing mechanical semantic edge weights, and calculates the theoretical Laplacian matrix with prior information on physical stiffness. At the processing site, the visual skeleton extraction and actual graph construction module scans the workpiece on the workbench using a 3D vision sensor, acquires 3D point cloud data with physical deformation, and extracts its global topological skeleton, thereby generating the corresponding actual undirected graph and actual Laplacian matrix. Subsequently, the spectral space matching and coarse registration module maps the above theoretical and actual Laplacian matrices to a high-dimensional Laplacian spectral space, performs eigenvalue decomposition and vector signature comparison, thereby establishing a global topological mapping relationship between theoretical nodes and actual nodes. A coarse registration is achieved, unaffected by the initial 3D pose of the workpiece. Based on this and the mapping relationship, the physical constraint manifold optimization module constructs a non-rigid deformation energy functional in the Lie group manifold space and introduces the mechanical semantic edge weights calculated by the aforementioned module as constraints for the adaptive manifold regularization mechanism. Through nonlinear optimization, a continuous deformation field that can flexibly fit the actual physical deformation of the workpiece is output, achieving real-time visual fine registration. Finally, the thermal compensation and trajectory generation module extracts the preset welding process parameters from the BIM model to calculate the expected heat input, converts it into a thermal deformation perturbation term, and performs reverse dynamic thermal stress pre-compensation in the generated continuous deformation field. Combining geometric deformation and thermal deformation trends, a high-precision robot-teachable welding trajectory is directly generated and sent to the underlying controller to execute the welding operation.

[0132] This invention, through the aforementioned scheme, breaks through the technical limitations of traditional teach-free welding systems that treat steel components as absolutely rigid bodies and rely solely on geometric features for rigid registration. It transforms the physical properties inherent in the BIM model into mechanical semantic constraints in the graphical model, and achieves cross-domain global topology matching without initial pose dependence by combining Laplace spectral space mapping. Furthermore, in the continuous manifold space of the Lie group, it uses physical prior knowledge to guide the optimization solution of the non-rigid deformation field, enabling the theoretical trajectory to adaptively and flexibly fit the actual workpiece undergoing physical deformation. With the help of a heat input pre-compensation mechanism, it achieves closed-loop absorption of the entire process of structural self-weight bending, assembly tolerance, and dynamic thermal deformation during processing. It opens up the control link from BIM theoretical data to high-precision automatic processing of complex heavy non-standard steel structures, providing a new technical path for solving the non-rigid adaptation problem in the intelligent manufacturing of building steel structures.

[0133] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A steel structure welding method without teaching based on BIM online analysis and real-time visual registration, characterized in that, Includes the following steps: S1: Online parsing of steel structure BIM model, extraction of physical manufacturing parameters from steel structure BIM model, construction of theoretical weighted undirected graph containing theoretical nodes and mechanical semantic edge weights, wherein the mechanical semantic edge weights are calculated based on the physical manufacturing parameters, and the corresponding theoretical Laplace matrix is ​​calculated according to the theoretical weighted undirected graph; S2: Acquire the three-dimensional point cloud data of the actual steel structure workpiece in real time through a vision sensor, extract the global topological skeleton of the three-dimensional point cloud data, construct an actual undirected graph containing actual nodes based on the global topological skeleton, and calculate the actual Laplacian matrix corresponding to the actual undirected graph. S3: Perform generalized eigenvalue decomposition on the theoretical Laplacian matrix and the actual Laplacian matrix respectively, map the eigenvalue decomposition results to the Laplacian spectrum space for feature comparison, establish a global topological mapping relationship between the theoretical node and the actual node, and obtain the coarse registration result between the steel structure BIM model and the actual steel structure workpiece. S4: Based on the global topological mapping relationship and the coarse registration result, a non-rigid deformation energy functional is constructed in the Lie group manifold space to map the theoretical nodes to the actual nodes. The mechanical semantic edge weights are added as constraints of the adaptive manifold regularization mechanism to the non-rigid deformation energy functional to participate in the optimization solution, thereby generating a continuous deformation field that fits the shape of the actual steel structure workpiece and completing the real-time visual registration of the actual steel structure workpiece. S5: Extract the preset welding process parameters from the steel structure BIM model, calculate the expected heat input of the area to be welded based on the preset welding process parameters, convert the expected heat input into a thermal deformation perturbation term, introduce the thermal deformation perturbation term into the continuous deformation field for reverse pre-compensation, and generate a robot teachless welding trajectory based on the continuous deformation field after reverse pre-compensation.

2. The steel structure welding method without teaching based on BIM online analysis and real-time visual registration according to claim 1, characterized in that, The physical manufacturing parameters include the local plate thickness parameters, cross-sectional moment of inertia, and material yield strength of the corresponding components in the steel structure BIM model; The construction of the theoretically weighted undirected graph, which includes theoretical nodes and mechanical semantic edge weights, includes: The assembly nodes or feature endpoints in the steel structure BIM model are extracted as theoretical nodes, and the potential welds connecting the theoretical nodes are extracted as edges of an undirected graph. The stiffness evaluation function, composed of the local plate thickness parameter, the moment of inertia of the section, and the yield strength of the material, is used to calculate the local deformation resistance between adjacent theoretical nodes, and the local deformation resistance is used as the mechanical semantic edge weight of the corresponding edge.

3. The steel structure welding method without teaching based on BIM online analysis and real-time visual registration according to claim 1, characterized in that, The calculation of the corresponding theoretical Laplacian matrix based on the theoretically weighted undirected graph includes: The weighted adjacency matrix is ​​constructed using the mechanical semantic edge weights as matrix elements; Calculate the diagonal matrix, whose diagonal elements are the sum of the weights of the connected mechanical semantic edges; The difference between the identity matrix and the normalized adjacency matrix is ​​calculated using the identity matrix, the inverse square root matrix of the angle matrix, and the weighted adjacency matrix to obtain the theoretical Laplace matrix containing the prior physical stiffness.

4. The steel structure welding method without teaching based on BIM online analysis and real-time visual registration according to claim 1, characterized in that, The process of extracting the global topological skeleton of the 3D point cloud data, constructing an actual undirected graph containing actual nodes based on the global topological skeleton, and calculating the actual Laplacian matrix corresponding to the actual undirected graph includes: A graph segmentation algorithm is used to extract the global topological skeleton representing the workpiece topology from the 3D point cloud data containing environmental noise; Extract the skeleton endpoints or intersections of the global topology skeleton as the actual nodes, and establish the actual undirected graph; The geometric adjacency matrix and actual degree matrix of the actual undirected graph are constructed using a spatial distance function, and the normalized actual Laplacian matrix is ​​calculated based on the geometric adjacency matrix and the actual degree matrix.

5. The steel structure welding method without teaching based on BIM online analysis and real-time visual registration according to claim 1, characterized in that, The step of mapping the eigenvalue decomposition results to the Laplacian spectral space for feature comparison and establishing a global topological mapping relationship between the theoretical nodes and the actual nodes includes: Obtain the first eigenvalue sequence and the first eigenvector signature obtained from the theoretical Laplace matrix decomposition, and the second eigenvalue sequence and the second eigenvector signature obtained from the actual Laplace matrix decomposition; In the dimensionality-reduced high-dimensional spectral space, the first eigenvector signature is compared with the second eigenvector signature to construct the spectral domain allocation cost matrix. An allocation algorithm is used to solve for the optimal permutation matrix of the spectral domain allocation cost matrix. The optimal permutation matrix is ​​used to realize the one-to-one mapping between the theoretical nodes and the actual nodes in the topological space, so as to get rid of the dependence on the initial three-dimensional pose of the actual steel structure workpiece.

6. The steel structure welding method without teaching based on BIM online analysis and real-time visual registration according to claim 1, characterized in that, The construction of the non-rigid deformable energy functional mapping the theoretical nodes to the actual nodes in the Lie group manifold space includes: A control lattice network is constructed along the three-dimensional theoretical weld seam of the steel structure BIM model, and each lattice node in the control lattice network is set to correspond to a Lie algebra pose increment. Construct the non-rigid deformation energy functional that includes a data fitting term and a manifold regularization term; The data fitting term is used to characterize the error between the theoretical three-dimensional coordinate points and the actual three-dimensional coordinate points after mapping to the Lie group through the exponential mapping operator, and the manifold regularization term is used to control the smoothness of the deformation of adjacent lattice nodes through adaptive constraint parameters.

7. The steel structure welding method without teaching based on BIM online analysis and real-time visual registration according to claim 1, characterized in that, The step of incorporating the mechanical semantic edge weights as constraints of the adaptive manifold regularization mechanism into the non-rigid deformation energy functional for optimization, thereby generating a continuous deformation field that conforms to the actual morphology of the steel structure workpiece, includes: The mechanical semantic edge weights are directly mapped to adaptive constraint parameters for controlling the deformation smoothness of adjacent lattice nodes by a monotonically increasing mapping function. Strong penalty constraints are generated for high stiffness regions and weak penalty constraints are generated for low stiffness regions. The optimal solution of the non-rigid deformation energy functional is obtained by iteratively solving the Lie algebra tangent space using a nonlinear least squares optimization algorithm, and the continuous deformation field corresponding to the optimal solution is output.

8. The steel structure welding method without teaching based on BIM online analysis and real-time visual registration according to claim 1, characterized in that, The preset welding process parameters include welding voltage, welding current, and welding speed; The step of calculating the expected heat input to the area to be welded based on the preset welding process parameters, and converting the expected heat input into a thermal deformation perturbation term, includes: The expected heat input is calculated by dividing the product of the arc thermal efficiency, the welding voltage and the welding current by the welding speed. Based on the coupling relationship between the local inverse stiffness of the component and the linear energy per unit length, the thermal deformation perturbation term is calculated and generated in the Lie algebra increment field.

9. The steel structure welding method without teaching based on BIM online analysis and real-time visual registration according to claim 1, characterized in that, The step of introducing the thermal deformation perturbation term into the continuous deformation field for reverse pre-compensation, and generating a robot-less welding trajectory based on the reverse pre-compensated continuous deformation field, includes: The thermal deformation perturbation term is used as a compensation value and is inversely added to the Lie algebra pose increment after optimization. Using the Lie group exponential mapping rule, the Lie algebraic pose increment after inverse compensation is combined and transformed with the ideal theoretical pose extracted from the steel structure BIM model to calculate the final continuous pose of the center point of the welding robot end tool. The six-degree-of-freedom spatial trajectory sequence containing the final continuous pose is used as the robot's teach-free welding trajectory.

10. A steel structure teach-free welding system based on BIM online analysis and real-time visual registration, used to implement the method described in any one of claims 1-9, characterized in that, include: The online parsing and mechanical graph construction module is used to parse the steel structure BIM model online, extract physical manufacturing parameters, construct a theoretical weighted undirected graph containing theoretical nodes and mechanical semantic edge weights, and calculate the corresponding theoretical Laplace matrix. The visual skeleton extraction and actual graph construction module is used to acquire 3D point cloud data of actual steel structure workpieces through visual sensors, extract the global topological skeleton, construct an actual undirected graph containing actual nodes, and calculate the corresponding actual Laplacian matrix. The spectral space matching and coarse registration module is used to perform generalized eigenvalue decomposition on the two Laplacian matrices respectively, perform feature comparison in the Laplacian spectral space, establish a global topological mapping relationship, and obtain the coarse registration result. The physical constraint manifold optimization module is used to construct a non-rigid deformation energy functional in the Lie group manifold space, and use the mechanical semantic edge weights as constraints of the adaptive manifold regularization mechanism to participate in the optimization solution, generate a continuous deformation field, and complete real-time visual registration. The thermal compensation and trajectory generation module is used to calculate the expected heat input based on the extracted preset welding process parameters, convert the expected heat input into a thermal deformation perturbation term for reverse pre-compensation in the continuous deformation field, and generate a robot teachless welding trajectory.

Images