An intelligent splitting method based on data segmentation algorithm
By using data segmentation algorithms and multi-objective optimization techniques, the problem of insufficient human experience in steel structure module disassembly has been solved, achieving efficient and accurate steel structure module disassembly and improving disassembly quality and production efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2025-04-23
- Publication Date
- 2026-07-21
AI Technical Summary
Existing methods for disassembling steel structure modules rely on manual experience and lack precise data analysis and optimization strategies, resulting in low disassembly efficiency, serious waste of resources, and insufficient accuracy in feature recognition and segmentation, making it difficult to meet the needs of complex engineering projects.
An intelligent splitting method based on data segmentation algorithms is adopted, including point cloud data acquisition, BIM model fusion, finite element multi-scale stress analysis, adaptive DBSCAN clustering to mark feature regions, NSGA-II and PPO algorithm optimization of the segmentation scheme, and dynamic adjustment of cutting edge weights and processing parameters.
It improves the accuracy of feature area recognition and segmentation, reduces the discreteness of module division, enhances process adaptability, and achieves efficient and accurate steel structure module disassembly.
Smart Images
Figure CN120430168B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent data splitting technology, and more specifically, to an intelligent data splitting method based on a data segmentation algorithm. Background Technology
[0002] With the rapid development of the construction industry, steel structures are increasingly widely used in various building projects. Traditional methods for disassembling steel structure modules often rely on manual experience, lacking precise data analysis and optimization strategies, resulting in low disassembly efficiency, serious resource waste, and difficulty in meeting the needs of complex projects. Therefore, there is an urgent need for an intelligent, efficient, and precise method for optimizing the disassembly of steel structure modules to improve the production efficiency and disassembly quality of steel structures.
[0003] Existing intelligent steel structure dismantling methods acquire point clouds and BIM models, then manually select feature points to achieve initial alignment between the point cloud and the BIM model. Next, a classic iterative nearest-point algorithm is used for fine registration. After registration, single-scale finite element analysis is performed, and high-stress areas are marked. During steel structure segmentation, the steel structure is cut according to standard component dimensions. Finally, with cost minimization as the objective, a genetic algorithm is used to generate local optimal solutions, and cutting parameters are selected from a fixed process library. This approach, to a certain extent, ensures the quality and production efficiency of steel structure dismantling.
[0004] However, existing methods still have some problems: the error bandwidth of stress analysis using a single-scale model in feature recognition is high, manual marking of partitions is inefficient, and the recognition accuracy of transition zones is insufficient. The feature recognition process should be optimized. In terms of steel structure segmentation, existing methods use a single segmentation dimension, resulting in insufficient segmentation accuracy, excessive discreteness of module division, and low adaptability of static matching of processing parameters to the process. They can only ensure cost minimization. The segmentation process should be optimized. Summary of the Invention
[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent splitting method based on data segmentation algorithm to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent data splitting method based on a data segmentation algorithm, comprising the following steps: S1. Raw data acquisition: Scan the steel structure to be disassembled to obtain point cloud data of the steel structure surface and extract the BIM model data of the steel structure to be disassembled; S2. Steel structure data fusion: Denoise the point cloud data of the steel structure surface, and then register it with the steel structure BIM model data to obtain a fused model; S3, Feature Labeling: Perform finite element multi-scale stress analysis on the fusion model, map the stress data to the point cloud, label high stress areas, regular geometric areas and transition areas based on the adaptive DBSCAN clustering algorithm, and output the feature labeling results; minPts is dynamically set based on the local stress gradient ||∇σ||, with the specific formula as follows: , Where minPts0 is the minimum number of points, β P β is the adaptive adjustment coefficient. P ∈[0.1,1.0]; S4. Hierarchical Intelligent Segmentation: Perform graph segmentation on the feature partitions to obtain a preliminary module division scheme for the steel structure. Perform geometric segmentation on the preliminary module division scheme for the steel structure to obtain a refined module division scheme for the steel structure, a cutting edge weight table, and processing parameters. S5. Dynamic optimization of segmentation scheme: Based on the NSGA-II algorithm and PPO algorithm, the steel structure refined module segmentation scheme is dynamically optimized in multiple objectives, and the final Pareto front solution set and optimization scheme comparison matrix are output. The specific steps for dynamically optimizing the segmentation scheme are as follows: D1. Set optimization objectives and dynamic constraints: Set minimizing total cost, maximizing structural strength, and minimizing the number of modules as optimization objectives, and set transportation constraints and processing constraints as dynamic constraints; D2, NSGA-II Algorithm Application: The NSGA-II algorithm is used to obtain the initial Pareto front solution set and the comparison matrix of optimization schemes; D3, PPO Collaborative Optimization: Real-time acquisition of the current generation of NSGA-II's hypervolume growth rate, solution set diversity, and constraint violation rate; adjustment of crossover probability or target weight based on the current state and updating of NSGA-II parameters. D4. Termination judgment: Termination is determined when the Pareto front hypervolume growth rate of two adjacent generations is lower than the preset value and there is no improvement for 10 consecutive generations. Termination is also determined when the maximum generation is reached or the processing feasibility verification pass rate is higher than the preset value. D5. Output the final Pareto front solution set, the optimization scheme comparison matrix, and the dynamic parameter trajectory; S6. Closed-loop verification: Perform finite element simulation verification and digital twin feedback on the optimized scheme, and output finite element analysis report, digital twin verification report and tolerance compensation parameter table; S7. Dynamic Adjustment of Schemes: Based on the closed-loop verification results, the order of the optimized schemes is dynamically adjusted. The scheme ranked first is marked as the target splitting scheme, and the schemes ranked second and third are marked as the first alternative scheme and the second alternative scheme, respectively. The marked schemes are then uploaded to the steel structure splitting scheme management terminal.
[0007] The technical effects and advantages of this invention are as follows: This invention performs finite element multi-scale stress analysis on a fusion model, mapping stress data to a point cloud. Based on the adaptive DBSCAN clustering algorithm, it labels high-stress areas, regular geometric areas, and transition areas, outputting feature labeling results. This effectively reduces error bandwidth, improves labeling and partitioning efficiency, and enhances transition area identification accuracy. Furthermore, this invention performs graph segmentation on the feature partitions to obtain a preliminary steel structure module division scheme. Geometric segmentation of this preliminary scheme yields a refined steel structure module division scheme, a cutting edge weight table, and processing parameters. Based on the NSGA-II and PPO algorithms, it performs dynamic multi-objective optimization of the refined steel structure module division scheme, outputting the final Pareto front solution set and an optimization scheme comparison matrix. This improves segmentation accuracy, reduces module division dispersion, and enhances the adaptability to the process through dynamic process parameter matching. This achieves multi-objective collaborative optimization, improving the final segmentation quality and production efficiency. Attached Figure Description
[0008] Figure 1 This is a diagram illustrating the method steps of the present invention.
[0009] Figure 2 This is a system structure block diagram of the present invention. Detailed Implementation
[0010] 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.
[0011] like Figure 1 The embodiment shown provides an intelligent data splitting method based on a data segmentation algorithm, including the following steps: S1. Raw data acquisition: Scan the steel structure to be disassembled to obtain point cloud data of the steel structure surface and extract the BIM model data of the steel structure to be disassembled; Furthermore, the extracted BIM model data of the steel structure to be disassembled includes geometric parameters, connection relationships, material properties, non-geometric information, and metadata. Geometric parameters include component geometry, component dimensions, and component spatial coordinates; connection relationships include node types and connection methods between components; material properties include steel grade, material strength, and surface treatment requirements; non-geometric information includes component number, component type, and design load; and metadata includes BIM model version information and model coordinate system datum.
[0012] Specifically, in this embodiment, a 3D laser scanner can be used to acquire point cloud data of the steel structure surface, and BIM data such as geometric parameters and connection relationships of the steel structure model can be exported from the Revit platform.
[0013] S2. Steel structure data fusion: Denoise the point cloud data of the steel structure surface, and then register it with the steel structure BIM model data to obtain a fused model; Furthermore, the specific steps for steel structure data fusion are as follows: A1. Point cloud denoising based on nonlocal mean filtering of curvature; A2. Calculate the centroid of the point cloud and the BIM model, and then translate the point cloud so that its centroid coincides with the centroid of the BIM model. A3. Calculate the covariance matrix for the point cloud and BIM model respectively, then perform eigenvalue decomposition to obtain the eigenvectors of the point cloud and BIM model, construct a rotation matrix and apply the rotation matrix to align the principal axis direction of the point cloud with the BIM model. A4. After determining the initialization parameters, find the nearest point in the BIM model for each point in the point cloud. Solve for the optimal rotation matrix and optimal translation vector for each iteration by minimizing the sum of squared distances between corresponding points. Then update the point cloud coordinates and continue iterative processing until the iteration termination condition is met, and the steel structure data fusion is completed.
[0014] In this embodiment, the specific steps for point cloud denoising are as follows: A11. Neighborhood Search: Set the initial neighborhood radius r, and search for the point cloud coordinates p with the i-th point cloud coordinates. ai Points whose distance is less than or equal to the initial neighborhood radius constitute neighborhood D. ai ; A12. Calculate curvature: For the i-th point cloud coordinate p ai Calculate the eigenvalues λ of the covariance matrix within its initial neighborhood radius. a1 ≧λ a2 ≧λ a3 The coordinates of the i-th point cloud are p ai curvature value C (p) ai The ratio of the smallest eigenvalue to other eigenvalues is given by the following formula: ; A13, Calculate point p aj For point p ai weight w(p) ai ,p aj ): , p aj For neighborhood D ai The three-dimensional coordinates of the j-th original point, σd This is the spatial distance weighting parameter, whose value is equal to the initial neighborhood radius by default, σ c C(p) is the curvature difference weighting parameter, with a default value of 0.1. aj ) is the neighborhood D ai The curvature value of the j-th original point is calculated using the same method as the coordinates p of the i-th point cloud. ai curvature value C (p) ai The calculation method is the same; A14. For the i-th point p in the point cloud ai The weighted average of all points p in its neighborhood aj After obtaining the coordinates, the denoised 3D coordinates p are obtained. ai * The specific formula is as follows: .
[0015] In this embodiment, the specific steps for aligning the centroids of the point cloud and the BIM model are as follows: A21. Calculate the centroid μ of the denoised point cloud. a The specific formula is as follows: , N a Given the total number of points in the point cloud, calculate the centroid μ of the BIM model. d The specific formula is as follows: , N d q represents the number of vertices in the BIM model. dj Let j be the coordinates of the j-th vertex in the BIM model; A22. Translate the point cloud so that its centroid coincides with the centroid of the BIM model. Then, the coordinates p of the i-th denoised point cloud after centroid alignment are... ai 'satisfy: .
[0016] In this embodiment, the specific steps for aligning the point cloud with the main axis of the BIM model are as follows: A31. Calculate the point cloud covariance matrix M. a The specific formula is as follows: , (p) ai '-μ d ) T For (p) ai '-μ d Calculate the covariance matrix M of the BIM model from the transpose of the matrix. d The specific formula is as follows: , (q)dj '-μ d ) T For (q) dj '-μ d The transpose of the matrix; A32. Perform eigenvalue decomposition on the point cloud covariance matrix to obtain the eigenvector V. a V a =[v a1 ,v a2 ,v a3 The covariance matrix of the BIM model is decomposed into eigenvectors V. d V d =[v d1 ,v d2 ,v d3 ]; A33. Construct the rotation matrix R PCA The specific formula is as follows: , V a -1 For V a The inverse matrix; A34. Apply a rotation matrix to align the point cloud with the principal axis direction of the BIM model. The coordinates of the i-th point cloud after alignment with the principal axis direction are p. ai "satisfy: .
[0017] In this embodiment, the specific steps for A4 are as follows: A41. Determine the initialization parameters: the coordinates p of the i-th point cloud before iteration. ai (0) =p ai The initial rotation matrix R0 = I, where I is the identity matrix, representing a square matrix with 1s on the main diagonal and 0s on the rest, and the initial translation vector t0 = 0. A42, for each point p in the point cloud ai Find its nearest point q in the BIM model dj (i); A43. Solve for the optimal rotation matrix and optimal translation vector in the k-th iteration by minimizing the sum of squared distances between corresponding points in the k-th iteration. The mathematical expression for minimizing the sum of squared distances between corresponding points in the k-th iteration is: , V k-1 U represents the principal direction distribution of the decentralized coordinates of the BIM model in the (k-1)th iteration. k-1 This represents the principal direction distribution of the decentralized coordinates of the point cloud in the (k-1)th iteration. , μ q,k-1 μ represents the centroid of the set of nearest points in the BIM model for each point in the (k-1)th iteration point cloud. p,k-1 This represents the centroid of the point cloud in the (k-1)th iteration; A44. Update the point cloud coordinates p after the kth iteration. ai (k) The specific formula is as follows: ; A45. Iteration Termination Determination: when
[0018] Or k=T max The iteration terminates when the maximum number of iterations, T, is typically taken. max =100, convergence threshold ϵ=10 −5 .
[0019] S3, Feature Labeling: Perform finite element multi-scale stress analysis on the fusion model, map the stress data to the point cloud, label high stress areas, regular geometric areas and transition areas based on the adaptive DBSCAN clustering algorithm, and output the feature labeling results; Furthermore, the specific steps for performing finite element multi-scale stress analysis on the fusion model are as follows: B11. Macro-scale modeling: Select shell element or beam element type according to the steel structure material properties, load conditions and expected analysis objectives, and initially set the nonlinear options and convergence tolerance of the solver. Generate a sparse mesh for the overall structure, calculate the global stress distribution, and identify the initial high stress areas. B12. Microscale modeling: Establish a refined sub-model of solid elements in the high-stress region marked by macroscopicity. Use the macroscopic results as the boundary conditions of the sub-model through nodal displacement interpolation or stress field mapping. Refine the mesh size of the sub-model until it meets the stress gradient resolution requirements. B13 Solver Setup and Execution: Select a compatible solver based on the element type, configure the load step, substep, and output request. If the micro-analysis generates a new high-stress area, return to B11 to expand the macro-marked area and reanalyze. B14. Output Results: Outputs the overall Von Mises stress field σ of the macroscopic model. vm (x,y,z), a refined stress field in the microscopic sub-model including principal stress components, and displacement vectors U at each node. w [U wx U wy U wz [Location of maximum displacement, maximum stress value and its spatial coordinates (x)] max , y max , z max(Based on the safety factor of the specified yield strength and the geometric curvature distribution of the stress concentration region.)
[0020] Furthermore, the specific steps for mapping stress data to point clouds are as follows: B21. Stress interpolation: For point P in the point cloud bi (x,y,z), search for the number of finite element nodes N within its initial neighborhood radius. b Point P is then calculated based on the interpolation formula. bi stress value σ pbi The specific formula is as follows: , σ j Let P be the point bi The stress value of the j-th finite element node within the initial neighborhood radius, d i,j Let P be the point bi The distance to the j-th finite element node; B22. Global Normalization: Normalization is performed based on the stress range of all finite element nodes. The specific formula is as follows: , σ max σ min The values are, in order, the maximum and minimum stress values of the finite element nodes; B23. Outlier Truncation: When point P bi When the stress value is greater than the preset value, it is marked as a singularity and removed; B24. Data Output: Output a point cloud set with stress labels {p bi (x,y,z,σ pbi * )}.
[0021] Furthermore, the specific steps for labeling high-stress areas, regular geometric areas, and transition areas based on the adaptive DBSCAN clustering algorithm are as follows: B31, Calculate point p bi k-nearest neighbor average stress weighted density ρ pbi The specific formula is as follows: , N(p bi (k) represents point p bi The set of k-nearest neighbors, m pbj For the j-th nearest neighbor point p bj The number of neighborhood points under the initial neighborhood radius; B32. Calculate the dynamic adjustment value r of the neighborhood radius. c The specific formula is as follows: , r0 is the initial neighborhood radius, αs The attenuation coefficient; B33. Dynamically set minPts based on the local stress gradient ||∇σ||, using the following formula: , Where minPts0 is the minimum number of points, β P β is the adaptive adjustment coefficient. P ∈[0.1,1.0]; B34. Calculate the principal curvatures C1 and C2 using PCA, and then standardize the principal curvatures C1 and C2 to obtain C1. * C2 * The specific formula is as follows: , C e σ a The mean and standard deviation of the global curvature. ; B35. Construct the eigenvector F after balancing the dimensions of Mahalanobis distance. pbi F pbi =[σ pbi * C1 * C2 * ]; B36. Setting the upper limit threshold σ of stress based on the quantile of the global stress-curvature distribution. y1 upper limit threshold of curvature change C y1 Stress lower limit threshold σ y2 and the lower limit threshold C of curvature change y2 ; B37, σ pbi * >σ y1 And ||C1-C2||>C y1 The points are marked as high-stress areas, and σ pbi * <σ y2 And ||C1-C2|| <C y2 The points marked are designated as regular geometric regions, and the remaining points are designated as transition regions.
[0022] Specifically, in this embodiment, the feature labeling results include partition category labels, multi-level label distribution maps; average stress, maximum stress, stress standard deviation, area, point cloud proportion, high stress node coordinates, high stress node stress values, and average curvature change rate of the high stress zone; average stress, area, point cloud proportion, principal curvature mean, and curvature entropy of the regular geometric zone; average stress, stress gradient mean, area, point cloud proportion, and curvature change rate standard deviation of the transition zone; geometric boundary point coordinates and stress boundary point coordinates, boundary type labels, and closed boundary loops of the boundary point set.
[0023] In this embodiment, it should be specifically noted that the upper limit threshold of stress, the upper limit threshold of curvature change, the lower limit threshold of stress, and the lower limit threshold of curvature change are all selected based on actual needs, and no specific value limit is imposed here. In terms of numerical relationship, the upper limit threshold of stress is greater than the lower limit threshold of stress, and the upper limit threshold of curvature change is greater than the lower limit threshold of curvature change.
[0024] S4. Hierarchical Intelligent Segmentation: Perform graph segmentation on the feature partitions to obtain a preliminary module division scheme for the steel structure. Perform geometric segmentation on the preliminary module division scheme for the steel structure to obtain a refined module division scheme for the steel structure, a cutting edge weight table, and processing parameters. Furthermore, the specific steps of hierarchical intelligent segmentation are as follows: C1. Graph Model Abstraction and Edge Weight Calculation: Define the center point of the high-stress zone, the centroid of the geometrically regular zone, and the key points of the transition zone as graph nodes, and calculate the weight w of the edge connecting nodes i and j. bi,j The specific formula is as follows: , In the formula α y σ i σ i β h f i,j γ l θ i,j The following are, in order: stress weight coefficient, stress value of node i, stress value of node j, geometric complexity weight coefficient, geometric connection area ratio of edge (i,j), connection strength weight coefficient, and connection strength of edge (i,j). C2. Multi-constraint graph segmentation optimization: Determine the objective function, the specific mathematical expression of which is: , x i,j The value λ is 1 when edge (i,j) is cut, and 0 when edge (i,j) is not cut. a Imbalance (S) is the weighting adjustment coefficient. k The sign indicates a penalty for maintaining the balance of submodule size. S kIt is the k-th submodule obtained after graph partitioning, where Avgsize is the ratio of the total number of nodes to the number of submodules, |S k | is a submodule S k The number of nodes, setting process constraints and mechanical constraints; C3. Geometric Decomposition and Manufacturability Verification: Extract the outer contour of the module based on the Alpha Shape algorithm to avoid generating non-convex shapes, insert process segmentation surfaces for complex connection areas, and interact with the processing equipment database to verify whether the component dimensions match. C4. Priority Quantification of Cutting Edges: The priority score is expressed as the product of the stress gradient on both sides of the edge and the safety factor, and the matching degree of the bevel form. C5. Data Output: Outputs the geometric boundary description of the segmentation scheme, the refined sub-module CAD model, the cutting edge weight table, and the processing parameters. The geometric boundary description includes the dominant boundary type, boundary point coordinates, average boundary stress gradient, boundary curvature, and recommended connection method. The refined sub-module CAD model includes process features, assembly datum, and metadata. The cutting edge weight table includes the cutting edge ID, start coordinates, end coordinates, weight, priority score, recommended cutting angle, bevel type, and thermal deformation compensation. The processing parameters include temporary and final processing parameters. The final processing parameters include the sub-module ID, cutting sequence, welding sequence, bevel angle, welding current, and post-processing requirements.
[0025] S5. Dynamic optimization of segmentation scheme: Based on the NSGA-II algorithm and PPO algorithm, the steel structure refined module segmentation scheme is dynamically optimized in multiple objectives, and the final Pareto front solution set and optimization scheme comparison matrix are output. Furthermore, the specific steps for dynamic optimization of the segmentation scheme are as follows: D1. Set optimization objectives and dynamic constraints: Set minimizing total cost, maximizing structural strength, and minimizing the number of modules as optimization objectives, and set transportation constraints and processing constraints as dynamic constraints; In this embodiment, it should be specifically noted that the total cost is the sum of cutting cost, welding cost, and transportation cost; the structural strength is the minimum of two ratios: the ratio of maximum equivalent stress to material yield strength and the ratio of maximum displacement to allowable displacement.
[0026] D2, NSGA-II Algorithm Application: The NSGA-II algorithm is used to obtain the initial Pareto front solution set and the comparison matrix of optimization schemes; Furthermore, the specific steps for applying the NSGA-II algorithm are as follows: D21. Input the search space of the cutting edge selection sequence and the population size N, and then randomly generate N initial splitting schemes. Each scheme is encoded as a binary string. D22. After calculating three objective values for each scheme, the population is stratified by Pareto dominance relation. Rank 1 is the optimal non-dominated solution set, Rank 2 is the next best, and so on. D23. For individuals of the same Rank, sort them according to each objective function value and calculate the crowding distance between their adjacent solutions: fm is the m-th objective function value; D24. Randomly select k individuals from the population, and select individuals with higher Rank or greater crowding distance to enter the mating pool. Then, assign crossover probability p to the selected parent individuals. c Exchange some genes, with a mutation probability p m Flipped gene loci; D25. Merge the original population and offspring to obtain 2N individuals. Select the top N optimal individuals according to non-dominated sorting and crowding distance to form a new generation population. D26. If convergence has been achieved before the maximum number of generations is reached, the process will terminate directly. If the Pareto front hypervolume growth rate of two adjacent generations is lower than the preset value or convergence has not been achieved even after the maximum number of generations is reached, the process will be forcibly terminated.
[0027] D3, PPO Collaborative Optimization: Real-time acquisition of the current generation of NSGA-II's hypervolume growth rate, solution set diversity, and constraint violation rate; adjustment of crossover probability or target weight based on the current state and updating of NSGA-II parameters. D4. Termination judgment: Termination is determined when the Pareto front hypervolume growth rate of two adjacent generations is lower than the preset value and there is no improvement for 10 consecutive generations. Termination is also determined when the maximum generation is reached or the processing feasibility verification pass rate is higher than the preset value. D5. Output the final Pareto front solution set, the optimization scheme comparison matrix, and the dynamic parameter trajectory.
[0028] Furthermore, the final Pareto front solution set includes a unique identifier for each solution, a list of submodules, total cost, structural strength score, number of submodules, fabrication feasibility score, detailed information on the cutting edge, connection method, cutting priority, stress gradient, maximum stress, and maximum displacement. The optimization scheme comparison matrix includes scheme ID, cost weight, strength weight, number of modules weight, hypervolume contribution diversity score, cost-strength slope, and fabrication feasibility score. The dynamic parameter trajectory includes crossover probability, mutation probability, dynamically adjusted weights of PPO, and the dynamic change dataset of average hypervolume.
[0029] S6. Closed-loop verification: Perform finite element simulation verification and digital twin feedback on the optimized scheme, and output finite element analysis report, digital twin verification report and tolerance compensation parameter table; In this embodiment, the specific steps for verifying the optimization scheme using finite element simulation are as follows: E11. Import the CAD model containing process features and material properties into the finite element analysis software, and convert the cutting parameters into contact pairs or binding constraints in the finite element method. E12. Load and constraint design: Load conditions include static loads, dynamic loads and construction loads, and constraint conditions include hinged supports (releasing rotation), sliding supports (releasing translation) and foundation spring stiffness (set according to the geological report). E13. Intelligent Mesh Generation: Identify high-stress areas, apply second-order hexahedral meshes to high-stress areas, control the gradient change of mesh size in transition areas, and apply first-order tetrahedral meshes to regular areas. E14. Safety Verification Quantification: Set numerical safety standards for stress, displacement, and buckling; calculate the maximum Von Mises stress, displacement contour plot, and buckling factor; compare the calculation results with the safety standards to determine whether the optimized scheme meets the safety standards. E15. Anomaly Detection: Automatically marks areas exceeding limits and associates them with the corresponding components in the CAD model; E16. Data Output: Outputs a finite element analysis report and anomaly area markers. The finite element analysis report includes stress / strain distribution diagrams, safety factors, and critical load values. The anomaly area markers include the location of the anomaly area, the anomaly type, the cause of the anomaly, and the anomaly risk level.
[0030] In this embodiment, the specific steps for providing digital twin feedback on the optimization scheme are as follows: E21. Data Acquisition: Installation position deviation of acquisition module (Δx, Δy, Δz) and cutting size error; E22. Map the measured data to the digital twin model and update the geometric and mechanical properties; E23. Deviation Analysis: Calculate the installation error using the following formula: , The specific formula for calculating cutting accuracy is as follows: ; E24. Generate dynamic compensation instructions based on deviation analysis results; E25. Output a digital twin verification report and a tolerance compensation parameter table. The digital twin verification report includes module ID, design stress, measured stress, installation error, compensation instructions, processing feasibility score, and total cost. The tolerance compensation parameter table includes compensation items, theoretical values, measured errors, and compensation values.
[0031] S7. Dynamic Adjustment of Schemes: Based on the closed-loop verification results, the order of the optimized schemes is dynamically adjusted. The scheme ranked first is marked as the target splitting scheme, and the schemes ranked second and third are marked as the first alternative scheme and the second alternative scheme, respectively. The marked schemes are then uploaded to the steel structure splitting scheme management terminal.
[0032] In this embodiment, the specific steps for dynamically adjusting the scheme are as follows: F1. Retrieve the safety factor S for closed-loop verification of each scheme. qi Total cost W zi Number of abnormal regions N yi Processing feasibility score M xi and the mean value of tolerance compensation T ei ; F2, Regarding the safety factor S qi Total cost W zi Number of abnormal regions N yi Processing feasibility score M xi and the mean value of tolerance compensation T ei The normalization process is performed using the following formula: , , , , , S qmax S qmin W zmax W zmin N ymax M xmax M xmin T emax T emin The values are, in order: maximum safety factor, minimum safety factor, maximum total cost, minimum total cost, maximum number of abnormal areas, maximum processing feasibility score, minimum processing feasibility score, maximum average tolerance compensation value, and minimum average tolerance compensation value for closed-loop verification of the optimization scheme. F3, respectively using β Sb β Wb , λ Nbi β Mb β Tb This represents the dynamically adjusted safety weight, cost weight, anomaly penalty coefficient, process weight, and tolerance weight. The data value of each weight before closed-loop verification is obtained, and represented by β. Sa β Wa , λ Na βMa β Ta Indicates that λ Na =0; F4, N yi When β is greater than 0, Sb =β Sa +0.1, T ei When β is greater than the design compensation threshold, Wb =β Wa -0.05, M xi When the score is less than the minimum processing feasibility score, β Mb =β Ma +0.05, β Tb =β Ta , λ Nbi =N gi +0.2, N gi Let i be the number of high-risk anomalies in the i-th scheme; F5. Calculate the comprehensive score Z of the i-th scheme. i The specific formula is as follows: ; F6. Sort the optimized schemes from highest to lowest according to the comprehensive score value, mark the top-ranked scheme as the target splitting scheme, and mark the second and third ranked schemes as the first alternative scheme and the second alternative scheme respectively. Upload the marked schemes to the steel structure splitting scheme management terminal.
[0033] In this embodiment, it should be specifically noted that the preset values, expected values, set values, design values, thresholds, weight coefficients, and weight adjustment coefficients used are all selected based on actual needs, and no specific value restrictions are imposed here.
[0034] like Figure 2 This embodiment provides an intelligent data segmentation system based on a data segmentation algorithm, including a raw data acquisition module, a steel structure data fusion module, a feature processing module, an intelligent segmentation module, a dynamic optimization module for segmentation schemes, a closed-loop verification module, a dynamic adjustment module for the scheme, and a database. The raw data acquisition module, steel structure data fusion module, feature processing module, intelligent segmentation module, dynamic optimization module for segmentation schemes, closed-loop verification module, and dynamic adjustment module for the scheme are connected sequentially. The raw data acquisition module is connected to the feature processing module, intelligent segmentation module, and closed-loop verification module respectively. All modules in the system are connected to the database.
[0035] The original data acquisition module scans the steel structure to be disassembled to obtain point cloud data of the steel structure surface and extracts the BIM model data of the steel structure to be disassembled. The steel structure data fusion module denoises the point cloud data of the steel structure surface and then registers it with the steel structure BIM model data to obtain the fused model. The feature processing module performs finite element multi-scale stress analysis on the fusion model, maps the stress data to the point cloud, and marks high-stress areas, regular geometric areas and transition areas based on the adaptive DBSCAN clustering algorithm, and outputs feature marking results. The intelligent segmentation module performs graph segmentation on the feature partitions to obtain a preliminary module division scheme for the steel structure, and performs geometric segmentation on the preliminary module division scheme to obtain a refined module division scheme for the steel structure, a cutting edge weight table, and processing parameters. The dynamic optimization module of the segmentation scheme performs dynamic multi-objective optimization of the steel structure fine module segmentation scheme based on the NSGA-II algorithm and the PPO algorithm, and outputs the Pareto front solution set and the optimization scheme comparison matrix. The closed-loop verification module performs finite element simulation verification and digital twin feedback on the optimization scheme, and outputs a finite element analysis report, a digital twin verification report, and a tolerance compensation parameter table. The dynamic adjustment module dynamically adjusts the order of the optimized schemes based on the closed-loop verification results, marks the first scheme as the target splitting scheme, marks the second and third schemes as the first alternative scheme and the second alternative scheme respectively, and uploads the marked schemes to the steel structure splitting scheme management terminal. The database is used to store data information for all modules in the system.
[0036] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An intelligent data splitting method based on a data segmentation algorithm, characterized in that: Includes the following steps: S1. Raw data acquisition: Scan the steel structure to be disassembled to obtain point cloud data of the steel structure surface and extract the BIM model data of the steel structure to be disassembled; S2. Steel structure data fusion: Denoise the point cloud data of the steel structure surface, and then register it with the steel structure BIM model data to obtain a fused model; S3, Feature Markers: Finite element multi-scale stress analysis was performed on the fusion model, and the stress data was mapped to the point cloud. The high stress area, regular geometric area and transition area were marked based on the adaptive DBSCAN clustering algorithm, and the feature marking results were output. minPts is dynamically set based on the local stress gradient ||∇σ||, with the specific formula as follows: ， Where minPts0 is the minimum number of points, β P β is the adaptive adjustment coefficient. P ∈[0.1,1.0]; S4. Hierarchical Intelligent Segmentation: Perform graph segmentation on the feature partitions to obtain a preliminary module division scheme for the steel structure. Perform geometric segmentation on the preliminary module division scheme for the steel structure to obtain a refined module division scheme for the steel structure, a cutting edge weight table, and processing parameters. S5. Dynamic optimization of segmentation scheme: Based on the NSGA-II algorithm and PPO algorithm, the steel structure refined module segmentation scheme is dynamically optimized in multiple objectives, and the final Pareto front solution set and optimization scheme comparison matrix are output. The specific steps for dynamically optimizing the segmentation scheme are as follows: D1. Set optimization objectives and dynamic constraints: Set minimizing total cost, maximizing structural strength, and minimizing the number of modules as optimization objectives, and set transportation constraints and processing constraints as dynamic constraints; D2, NSGA-II Algorithm Application: The NSGA-II algorithm is used to obtain the initial Pareto front solution set and the comparison matrix of optimization schemes; D3, PPO Collaborative Optimization: Real-time acquisition of the current generation of NSGA-II's hypervolume growth rate, solution set diversity, and constraint violation rate; adjustment of crossover probability or target weight based on the current state and updating of NSGA-II parameters. D4. Termination judgment: Termination is determined when the Pareto front hypervolume growth rate of two adjacent generations is lower than the preset value and there is no improvement for 10 consecutive generations. Termination is also determined when the maximum generation is reached or the processing feasibility verification pass rate is higher than the preset value. D5. Output the final Pareto front solution set, the optimization scheme comparison matrix, and the dynamic parameter trajectory; S6. Closed-loop verification: Perform finite element simulation verification and digital twin feedback on the optimized scheme, and output finite element analysis report, digital twin verification report and tolerance compensation parameter table; S7. Dynamic Adjustment of Schemes: Based on the closed-loop verification results, the order of the optimized schemes is dynamically adjusted. The scheme ranked first is marked as the target splitting scheme, and the schemes ranked second and third are marked as the first alternative scheme and the second alternative scheme, respectively. The marked schemes are then uploaded to the steel structure splitting scheme management terminal.
2. The intelligent splitting method based on a data segmentation algorithm according to claim 1, characterized in that: The BIM model data of the steel structure to be disassembled extracted in step S1 includes geometric parameters, connection relationships, material properties, non-geometric information and metadata. The geometric parameters include the component geometry, component size and component spatial coordinates. Connection relationships include node types and connection methods between components; material properties include steel grade, material strength, and surface treatment requirements; non-geometric information includes component number, component type, and design load; metadata includes BIM model version information and model coordinate system datum.
3. The intelligent splitting method based on a data segmentation algorithm according to claim 1, characterized in that: The specific steps for steel structure data fusion in step S2 are as follows: A1. Point cloud denoising based on nonlocal mean filtering of curvature; A2. Calculate the centroid of the point cloud and the BIM model, and then translate the point cloud so that its centroid coincides with the centroid of the BIM model. A3. Calculate the covariance matrix for the point cloud and BIM model respectively, then perform eigenvalue decomposition to obtain the eigenvectors of the point cloud and BIM model, construct a rotation matrix and apply the rotation matrix to align the principal axis direction of the point cloud with the BIM model. A4. After determining the initialization parameters, find the nearest point in the BIM model for each point in the point cloud. Solve for the optimal rotation matrix and optimal translation vector for each iteration by minimizing the sum of squared distances between corresponding points. Then update the point cloud coordinates and continue iterative processing until the iteration termination condition is met, and the steel structure data fusion is completed.
4. The intelligent splitting method based on a data segmentation algorithm according to claim 1, characterized in that: The specific steps of step S3, which involves performing finite element multi-scale stress analysis on the fused model, are as follows: B11. Macro-scale modeling: Select shell element or beam element type according to the steel structure material properties, load conditions and expected analysis objectives, and initially set the nonlinear options and convergence tolerance of the solver. Generate a sparse mesh for the overall structure, calculate the global stress distribution, and identify the initial high stress areas. B12. Microscale modeling: Establish a refined sub-model of solid elements in the high-stress region marked by macroscopicity. Use the macroscopic results as the boundary conditions of the sub-model through nodal displacement interpolation or stress field mapping. Refine the mesh size of the sub-model until it meets the stress gradient resolution requirements. B13 Solver Setup and Execution: Select a compatible solver based on the element type, configure the load step, substep, and output request. If the micro-analysis generates a new high-stress area, return to B11 to expand the macro-marked area and reanalyze. B14. Output Results: Outputs the overall Von Mises stress field σ of the macroscopic model. vm (x,y,z), a refined stress field in the microscopic sub-model including principal stress components, and displacement vectors U at each node. w [U wx U wy U wz [Location of maximum displacement, maximum stress value and its spatial coordinates (x)] max , y max , z max (Based on the safety factor of the specified yield strength and the geometric curvature distribution of the stress concentration region.) 5. The intelligent splitting method based on a data segmentation algorithm according to claim 1, characterized in that: The specific steps in step S3, which maps the stress data to the point cloud, are as follows: B21. Stress interpolation: For point P in the point cloud bi (x,y,z), search for the number of finite element nodes N within its initial neighborhood radius. b Point P is then calculated based on the interpolation formula. bi stress value σ pbi The specific formula is as follows: ， σ j Let P be the point bi The stress value of the j-th finite element node within the initial neighborhood radius, d i,j Let P be the point bi The distance to the j-th finite element node; B22. Global Normalization: Normalization is performed based on the stress range of all finite element nodes. The specific formula is as follows: ， σ max σ min The values are, in order, the maximum and minimum stress values of the finite element nodes; B23. Outlier Truncation: When point P bi When the stress value is greater than the preset value, it is marked as a singularity and removed; B24. Data Output: Output a point cloud set with stress labels {p bi (x,y,z,σ pbi * )}.
6. The intelligent splitting method based on a data segmentation algorithm according to claim 1, characterized in that: The specific steps of step S3, which uses the adaptive DBSCAN clustering algorithm to label high-stress areas, regular geometric areas, and transition areas, are as follows: B31, Calculate point p bi k-nearest neighbor average stress weighted density ρ pbi The specific formula is as follows: ， N(p bi (k) represents point p bi The set of k-nearest neighbors, m pbj For the j-th nearest neighbor point p bj The number of neighborhood points under the initial neighborhood radius; B32. Calculate the dynamic adjustment value r of the neighborhood radius. c The specific formula is as follows: ， r0 is the initial neighborhood radius, α s The attenuation coefficient; B33. Calculate the principal curvatures C1 and C2 using PCA, and then standardize the principal curvatures C1 and C2 to obtain C1. * C2 * The specific formula is as follows: , C e σ a The mean and standard deviation of the global curvature. ； B34. Construct the eigenvector F after balancing the dimensions of Mahalanobis distance. pbi F pbi =[σ pbi * C1 * C2 * ]; B35. Setting the upper limit threshold σ of stress based on the quantile of the global stress-curvature distribution. y1 upper limit threshold of curvature change C y1 Stress lower limit threshold σ y2 and the lower limit threshold C of curvature change y2 ; B36, σ pbi * >σ y1 And ||C1-C2||>C y1 The points are marked as high-stress areas, and σ pbi * <σ y2 And ||C1-C2|| <C y2 The points marked are designated as regular geometric regions, and the remaining points are designated as transition regions.
7. The intelligent splitting method based on a data segmentation algorithm according to claim 1, characterized in that: The specific steps of step S4, hierarchical intelligent segmentation, are as follows: C1. Graph Model Abstraction and Edge Weight Calculation: Define the center point of the high-stress zone, the centroid of the geometrically regular zone, and the key points of the transition zone as graph nodes, and calculate the weight w of the edge connecting nodes i and j. bi,j The specific formula is as follows: ， In the formula α y σ i σ i β h f i,j γ l θ i,j The following are, in order: stress weight coefficient, stress value of node i, stress value of node j, geometric complexity weight coefficient, geometric connection area ratio of edge (i,j), connection strength weight coefficient, and connection strength of edge (i,j). C2. Multi-constraint graph segmentation optimization: Determine the objective function, the specific mathematical expression of which is: ， x i,j The value λ is 1 when edge (i,j) is cut, and 0 when edge (i,j) is not cut. a Imbalance (S) is the weighting adjustment coefficient. k () indicates a penalty term for submodule size balance. ， S k It is the k-th submodule obtained after graph partitioning, where Avgsize is the ratio of the total number of nodes to the number of submodules, |S k | is a submodule S k The number of nodes, setting process constraints and mechanical constraints; C3. Geometric Decomposition and Manufacturability Verification: Extract the outer contour of the module based on the Alpha Shape algorithm to avoid generating non-convex shapes, insert process segmentation surfaces for complex connection areas, and interact with the processing equipment database to verify whether the component dimensions match. C4. Priority Quantification of Cutting Edges: The priority score is expressed as the product of the stress gradient on both sides of the edge and the safety factor, and the matching degree of the bevel form. C5. Data Output: Outputs the geometric boundary description of the segmentation scheme, the refined sub-module CAD model, the cutting edge weight table, and the processing parameters.
8. The intelligent splitting method based on a data segmentation algorithm according to claim 1, characterized in that: The final Pareto front solution set output in step S5 includes a unique identifier for each solution, a list of submodules, total cost, structural strength score, number of submodules, fabrication feasibility score, detailed information on the cutting edge, connection method, cutting priority, stress gradient, maximum stress, and maximum displacement. The optimization scheme comparison matrix includes scheme ID, cost weight, strength weight, number of modules weight, hypervolume contribution diversity score, cost-strength slope, and fabrication feasibility score.