A large-scale BIM model optimization system based on intelligent loading

Abstract

The application relates to the technical field of building information model optimization, in particular to a large-volume BIM model optimization system based on intelligent loading, which comprises a model analysis unit, a function identification unit, a simplification execution unit and a model recombination unit. In the application, the model analysis unit reads an IFC file and separates component data, the function identification unit analyzes through a geometric and semantic double-channel analysis, outputs a decorative component (a first type), a non-load-bearing structure (a second type) and a load-bearing component (a third type) set, the simplification execution unit performs edge folding on the first type, performs vertex clustering on the second type and skips simplification on the third type, and the model recombination unit binds metadata and constructs a dynamic octree index to generate a lightweight model. The application solves the problems of single traditional simplification strategy and core information loss, improves model loading efficiency and application accuracy, and is suitable for complex building scenes.

Description

Technical Field

[0001] This invention relates to the field of building information modeling optimization technology, and more specifically, to a large-scale BIM model optimization system based on intelligent loading. Background Technology

[0002] Building Information Modeling (BIM) optimization is an important technology. As building size and complexity increase, the efficient loading and accurate application of large-scale BIM models become crucial. Lightweight optimization technology is of great significance for reducing hardware resource consumption and improving collaborative efficiency. Traditional model processing methods that rely on unified and simplified standards are no longer able to meet the differentiated needs of different functional components.

[0003] However, traditional large-scale BIM model optimization techniques suffer from core problems such as a single simplification strategy and loss of core information. Existing solutions use the same simplification algorithm for all components, failing to distinguish the functional differences between decorative components, non-load-bearing structures, and load-bearing core components. Oversimplification of decorative components leads to the loss of visual details and affects visualization effects, while insufficient simplification of non-load-bearing structures results in model redundancy and slow loading speed. Missimplification of load-bearing components can destroy mechanical topological features and jeopardize the accuracy of structural analysis. At the same time, the simplification process lacks collaborative verification of geometric shape and semantic attributes. When components with abrupt curvature changes or high-strength material components are blindly simplified, it not only reduces model accuracy but also increases subsequent modification costs and may even cause engineering risks due to distortion of core structural data. It is difficult to achieve efficient and lightweight large-scale BIM models while ensuring the integrity of key information, which restricts the in-depth application of BIM technology in complex building scenarios. To solve this technical problem, we provide a large-scale BIM model optimization system based on intelligent loading. Summary of the Invention

[0004] The purpose of this invention is to provide a large-scale BIM model optimization system based on intelligent loading to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, a large-scale BIM model optimization system based on intelligent loading is provided, comprising a model parsing unit for reading the IFC format file of the BIM model and separating the triangular mesh, component type code, and material strength parameters of each component in the IFC format file. The system is characterized by further comprising:

[0006] The functional identification unit inputs the component data output by the model parsing unit into the convolutional neural network model. The model contains parallel geometric analysis channels and semantic parsing channels, which are used to calculate and identify the component data and output a first type of component set, a second type of component set, and a third type of component set.

[0007] The simplified execution unit performs differentiated processing on the first type of component set, the second type of component set, and the third type of component set respectively:

[0008] The edge-folding algorithm is applied to the components in the first type of component set to simplify the patch density to meet the visual recognition threshold of the decorative component.

[0009] A vertex clustering algorithm is performed on the components in the second type of component set to preserve mechanical topological features and meet the simplification criteria for non-load-bearing structures.

[0010] Simplification processing is skipped for components in the third type of component set;

[0011] The model reorganization unit reorganizes the component data processed by the simplified execution unit into a lightweight BIM model.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0013] In a large-scale BIM model optimization system based on intelligent loading, the functional identification stage utilizes a dual-channel approach of geometric analysis and semantic parsing to accurately classify three types of component sets. The geometric analysis channel locates visually sensitive areas using features such as curvature and patch density, while the semantic parsing channel extracts structural attributes by combining material strength and type coding. A dynamic weight allocation module adjusts feature priorities according to scenario requirements, ensuring the accuracy of classification for decorative components, non-load-bearing structures, and load-bearing components. This provides a reliable basis for differentiated simplification, avoiding the simplification bias caused by traditional single classification. The simplification execution stage implements customized processing for different sets. For the first type of decorative components, an edge-folding algorithm reduces patches in flat areas, and the folding step size is dynamically adjusted by combining contour feature comparison, reducing patch density while preserving visual sensitivity. The model is designed with meticulous attention to detail, ensuring both the integrity of the decorative effect and reducing redundant data. For the second type of non-load-bearing structure, a dynamic protective zone is constructed around the key nodes of the force transmission path. Vertex clustering is performed outside the protective zone, and stiffness loss is verified using finite element micro-simulation to ensure that the mechanical topological characteristics are maintained after simplification, balancing lightweighting and structural stability. For the third type of load-bearing component, simplification is skipped by digital watermarking, fully preserving the original triangular mesh and material parameters to ensure the accuracy of structural analysis. The model reorganization unit is constructed by binding differentiated metadata tags, building a dynamic octree index and matching the LOD level to achieve intelligent model loading. When called, only the component data with the required precision for the current scene is loaded, greatly improving loading efficiency. At the same time, it supports reverse restoration of the original data of load-bearing components through the watermark interface to meet the needs of high-precision applications. Attached Figure Description

[0014] Figure 1 This is an overall block diagram of the present invention.

[0015] The meanings of the labels in the diagram are as follows:

[0016] 1. Model parsing unit; 2. Function identification unit; 3. Simplified execution unit; 4. Model reorganization unit. Detailed Implementation

[0017] 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.

[0018] This invention provides a large-scale BIM model optimization system based on intelligent loading. Please refer to [link / reference]. Figure 1 As shown, it includes a model parsing unit 1, used to read the IFC format file of the BIM model, and separate the triangular mesh, component type code, and material strength parameters of each component in the IFC format file. Its distinguishing feature is that it also includes:

[0019] Function recognition unit 2 inputs the component data output by model parsing unit 1 into the convolutional neural network model. This model contains parallel geometric analysis channels and semantic parsing channels, which are used to calculate and recognize the component data and output the first type of component set, the second type of component set, and the third type of component set.

[0020] To accurately distinguish components with different functions in the BIM model and provide a basis for subsequent differentiation and simplification, the function identification unit 2 generates three sets through the following steps, the specific implementation of which is as follows:

[0021] In the geometric analysis channel, the density distribution data of the triangular mesh of the component is calculated. The density gradient detection algorithm is used to identify the regions in the density distribution data whose density change rate exceeds the preset threshold. The geometric attribute set of the region is extracted and quantized into a geometric feature vector. The geometric attribute set includes the region area ratio, the number of curvature extreme points, and the boundary contour complexity.

[0022] The geometric analysis channel focuses on the spatial morphological characteristics of components. By quantifying the geometric properties of triangular meshes, it generates geometric feature vectors that reflect the complexity of the component's morphology. The specific steps are as follows:

[0023] For each component triangular mesh output by model analysis unit 1, the number of triangular faces per unit area, i.e., the face density, is counted, and the density distribution on the component surface is recorded. For example, the face density of the edge area of ​​a wall component is 50 faces / square meter, and the density of the middle area is 20 faces / square meter. The density gradient detection algorithm is used to analyze the density distribution data. When the density change rate of adjacent areas exceeds a preset threshold, the area is marked as a "shape-sensitive area". For example, at the connection between the column and the beam, the face density suddenly increases from 30 faces / square meter to 60 faces / square meter, with a change rate of 100%, which far exceeds the threshold and is marked as a shape-sensitive area. For the marked shape-sensitive areas, three core geometric attributes are extracted:

[0024] Area percentage: The ratio of the area of ​​this region to the total area of ​​the components, such as the connection part accounting for 15% of the total area of ​​the column;

[0025] Number of curvature extrema: The number of vertices with the maximum or minimum curvature in a region, such as the bending vertices of a curved surface component. The more vertices there are, the more complex the shape.

[0026] Boundary contour complexity: Measured by the number of turns in the contour lines. For example, the complexity of a straight contour is 1, and the complexity of a contour with 3 corners is 4.

[0027] These three attributes are converted into numerical values ​​according to preset rules, such as 0.15 for 15% area ratio, 0.5 for 5 curvature extrema, and 0.4 for contour complexity of 4, and combined to form a geometric feature vector (e.g., [0.15, 0.5, 0.4]).

[0028] In the semantic parsing channel, the structural attribute identifier of the component type code is parsed, the material strength parameter is extracted based on the structural attribute identifier, the material strength parameter is mapped to the safety level coefficient according to the building code, and a semantic feature vector is generated by combining the structural attribute identifier.

[0029] The semantic parsing channel extracts functional information from the component's type encoding and material parameters to generate semantic feature vectors that reflect the component's structural properties. The specific steps are as follows:

[0030] The component type code contains structured information reflecting its function. For example, in the code "C-02-03", "C" represents a load-bearing component and "02" represents the type of concrete. By parsing the identifier in the code, the basic functional category of the component is determined, such as load-bearing, decorative, or non-load-bearing. The material strength parameters are located according to the structural attribute identifier, such as the compressive strength of concrete cubes and the yield strength of steel. Then, according to the building code, the parameters are mapped to the safety level coefficient: the higher the strength, the larger the safety level coefficient (e.g., a coefficient of 0.8 for a strength of 30MPa and a coefficient of 1.2 for 50MPa). The functional category corresponding to the structural attribute identifier (e.g., load-bearing components are recorded as 1 and decorative components as 0) is combined with the safety level coefficient to form a semantic feature vector (e.g., for a load-bearing component with a strength coefficient of 1.2, the vector is [1, 1.2]).

[0031] The geometric and semantic feature vectors are input into the dynamic weight allocation module of the convolutional neural network model. The set to which the component belongs is determined through priority evaluation. The specific steps are as follows:

[0032] The priority allocation result based on the dynamic weight allocation module outputs three sets of components: a first set, a second set, and a third set. The module automatically adjusts the weights of geometric and semantic features according to the application scenario of each component. If the model is used for visualization, the weight of geometric features is increased (e.g., 0.6); if used for structural analysis, the weight of semantic features is increased (e.g., 0.7). For example, for decorative components, the weight of geometric features (morphological complexity) is 0.6, and the weight of semantic features (safety level) is 0.4; for load-bearing components, the weight of semantic features is 0.7, and the weight of geometric features is 0.3. A comprehensive score of the feature vectors is calculated based on these weights, and the sets are divided according to the score results.

[0033] The first category consists of components with high geometric feature scores (e.g., the proportion of morphologically sensitive areas exceeds 50%) and low semantic feature scores (e.g., the security level coefficient is <0.5), which are mostly decorative components (e.g., reliefs, patterns).

[0034] The second category consists of components with moderate geometric feature scores (20%-50% of the area is morphologically sensitive) and moderate semantic feature scores (0.5-0.8 safety level coefficient), which are mostly non-load-bearing structures (such as partition walls and secondary supports).

[0035] The third set consists of components with high semantic feature scores (safety level coefficient > 0.8), which are mostly load-bearing core components (such as main beams and columns). The three sets of components generated in the end accurately reflect the simplification requirements of different components, laying the foundation for subsequent differentiated optimization.

[0036] To accurately select decorative components whose core function is visual presentation, the generation of the first set is based on the collaborative approach of geometric analysis and semantic parsing, and is achieved through dual verification of visually sensitive area identification and functional attribute filtering. The specific implementation method is as follows:

[0037] Based on the patch density distribution data output from the geometric analysis channel, the visual key regions of the component are further located through curvature features, as follows:

[0038] The curvature clustering algorithm is used to divide the feature surface of a component into partitions, grouping areas with similar curvature values ​​into the same feature surface (e.g., planar areas with curvature close to 0, and curved areas with curvature in the range of 5-10). For example, the surface of a column capital component includes a planar base (curvature 0), an arc-shaped side (curvature 8), and a top carving (curvature 20-30). Using a clustering algorithm, it is divided into three feature surface partitions. A preset rate of change threshold is set (e.g., the rate of change of curvature between adjacent areas exceeds 50%). Areas with a rate of change of curvature exceeding the preset threshold are marked as visually sensitive areas. These areas are often visual focal points (e.g., at the junction of the carving and the arc-shaped side, the curvature suddenly increases from 8 to 20, a rate of change of 150%, far exceeding the threshold). For example, the carved area at the top of the column capital component and its boundary with the side are both marked as visually sensitive areas. The total area of ​​all visually sensitive areas is calculated and compared with the overall surface area of ​​the component to obtain the proportion (e.g., the proportion of the visually sensitive area of ​​the column capital component is 40%). Simultaneously, text matching rules are activated in the semantic parsing channel to block component type codes containing preset keywords and filter components whose material strength is higher than the standard value of decorative materials. Components whose proportion of visually sensitive areas exceeds the preset visual sensitivity threshold and passes the filter are included in the first set. The steps are as follows:

[0039] A pre-defined keyword list (such as "beam," "column," and "load-bearing," representing load-bearing or structural functions) is used. Component type codes are checked through text matching; if a code contains these keywords, it is directly excluded (e.g., the code "Z-01-load-bearing" is blocked because it contains the keyword "load-bearing"). Standard values ​​for decorative materials are set, and the material strength parameters corresponding to the component type codes are extracted. If the parameters are higher than the standard values, the component is judged as non-decorative and filtered. The results of geometric analysis and semantic parsing are combined to finally select components that meet the criteria and include them in the first category set.

[0040] Only retain components that satisfy both of the following conditions:

[0041] If the proportion of the visually sensitive area exceeds the preset visual sensitivity threshold (e.g., 30%, meaning the area of ​​the visually sensitive area must account for more than 30% of the total area of ​​the component), it will be filtered through the semantic parsing channel. That is, the code does not contain preset keywords, and the material strength does not exceed the standard value of decorative materials. For example, if the proportion of the visually sensitive area of ​​a certain carved bracket component is 50% (exceeding the 30% threshold), the type code is "Q-02-decoration" (without keywords), and the wood strength is 8MPa (below the standard value of 10MPa), then the component will be included in the first category set as a simplified object that focuses on visual preservation in the future.

[0042] To accurately select non-load-bearing structural components suitable for simplification using vertex clustering algorithms, the generation of the second set combines material strength filtering via semantic parsing with curvature feature recognition via geometric analysis. This ensures that the selected components have neither load-bearing function nor are feasible for morphological simplification. The specific implementation method is as follows:

[0043] In the semantic parsing channel, for components whose component type coding is clearly classified as non-load-bearing structures (such as those with codes containing identifiers like "partition wall," "fill," or "secondary support"), a dynamic filtering mechanism for material strength parameters is established, with the following steps:

[0044] From the attribute data associated with the component type code, the concrete strength grade and steel yield strength parameters are extracted. These parameters directly reflect the structural load-bearing potential of the component. Non-load-bearing structures usually do not require excessive strength. According to the material requirements for non-load-bearing structures in the building code, preset allowable values ​​are set. The logic for setting the allowable values ​​is: non-load-bearing structures only need to meet their own weight and slight additional loads, and do not need to reach the strength standards of load-bearing structures. When the component type code belongs to the non-load-bearing structure category, the concrete strength grade and steel yield strength parameters are extracted and compared with the preset allowable values ​​for non-load-bearing structures. Components with strength values ​​lower than the standard values ​​are screened out. The extracted strength parameters are compared with the allowable values, and only components with parameters lower than or equal to the allowable values ​​are retained. For example, if a non-load-bearing partition wall has a concrete strength grade of C20 (lower than C25) and a steel yield strength of 210 MPa (lower than 235 MPa), it passes the filter. If another non-load-bearing component has a concrete strength grade of C30 (higher than C25), it is excluded because the strength exceeds the standard. At the same time, the overall curvature distribution of the component is calculated in the geometric analysis channel. Components with smooth curvature changes and no abrupt changes are marked. Components that meet both of the above requirements are included in the second set. Simultaneously, in the geometric analysis channel, the morphological feature analysis of the components that pass the semantic filter is performed to determine whether they are suitable for simplification. The steps are as follows:

[0045] Based on the triangular mesh data of the components, the curvature value of each vertex is calculated (e.g., the curvature of a planar vertex is 0, and the curvature of a gently curved vertex is 5-10). The distribution of curvature values ​​on the component surface is statistically analyzed. The rate of curvature change is detected using a sliding window algorithm. If the rate of curvature change of any adjacent area on the component surface is lower than a preset abrupt change threshold, and there are no extreme curvature points (points where the curvature value suddenly increases to above 20), it is judged as having "gradual curvature change without abrupt changes." For example, a non-load-bearing partition wall is a planar structure with all curvature values ​​being 0 and a rate of change of 0, which meets the characteristics. However, partition walls with complex corners do not meet the criteria because of curvature abrupt changes (e.g., the curvature at the corner suddenly increases from 0 to 30). Components that meet the "gradual curvature change without abrupt changes" characteristic are marked, indicating that their form is simple and key features are not easily lost during simplification. Low-strength non-load-bearing components filtered by the semantic parsing channel are compared with components with gradual curvature marked by the geometric analysis channel. Only components that meet both conditions are retained and included in the second set. For example, a non-load-bearing infill wall:

[0046] Semantic analysis shows that its concrete strength grade is C20 (meeting the non-load-bearing strength standard), and geometric analysis shows that its surface is planar with a curvature change rate of 0 (meeting the smooth characteristic). Therefore, the component is included in the second set as the object of subsequent vertex clustering simplification. The semantic analysis channel ensures the compatibility of the component's non-load-bearing properties and material strength, while the geometric analysis channel ensures that the component's shape is feasible for simplification. The second set generated by the two together accurately identifies components suitable for simplification using the vertex clustering algorithm. This avoids oversimplification of high-strength non-load-bearing components and prevents distortion caused by simplification of complex-shaped components, providing a reliable basis for subsequent differentiated processing.

[0047] Simplified execution unit 3 performs differentiated processing on the first type of component set, the second type of component set, and the third type of component set respectively:

[0048] The edge-folding algorithm is applied to the components in the first type of component set to simplify the patch density to meet the visual recognition threshold of the decorative component.

[0049] To achieve model lightweighting while maintaining the visual recognizability of decorative components, when performing the edge-folding algorithm on components in the first type of component set, curvature analysis is used to locate simplified areas, and the folding strategy is dynamically adjusted by combining contour comparison. The specific implementation method is as follows:

[0050] The surface curvature distribution data of the component is calculated, and flat areas with curvature values ​​below the visual perception threshold are identified. Based on the component's triangular mesh model, the vertices of each triangular facet are traversed, and the curvature value at each vertex is calculated. The curvature value reflects the degree of surface curvature; the smaller the value, the closer it is to a plane. A visual perception threshold is set (e.g., curvature value ≤ 5, where the human eye cannot perceive the subtle curvature in the area). All areas with curvature values ​​below this threshold are marked as flat areas. These areas, due to their simple shape, are the main targets for edge folding simplification. For example, the background plane of a relief component has a curvature value of 2 (below the threshold of 5) and is identified as a flat area; while the relief pattern itself has a curvature value of 15, belonging to a visually sensitive area that cannot be simplified. Then, based on this flat area, candidate triangular facets for edge folding are located. Within the marked flat area, triangular facets with larger areas and a distance from the boundary of the visually sensitive area exceeding a preset safety distance (e.g., 5 cm, to avoid folding affecting the shape of the sensitive area) are selected as candidates. For example, in the background plane of the aforementioned relief component, triangular facets far from the relief pattern are selected as candidates to ensure that the folding operation only applies to areas that do not affect the visual effect. Then, after each folding operation, a real-time orthographic projection contour map of the component is generated. The current contour map and the original contour map are input into a convolutional neural network model for skeleton feature comparison. Edge folding merges the common edges of two adjacent triangular facets, reducing the number of facets (e.g., merging two adjacent triangles into a quadrilateral, reducing one facet). After each fold, a two-dimensional contour map is generated from the front of the component, extracting the skeleton features of the contour (such as contour turning points and line directions), and comparing them with the skeleton features of the original contour map before folding. For example, after folding triangular facets in a flat area, the overlap between the generated contour map and the skeleton lines of the original map needs to be maintained at a high level. When the feature similarity is lower than the similarity threshold, an adaptive reduction mechanism for the folding step size is triggered. The step size is the number of triangular facets processed in each folding operation. The reduction mechanism reduces the number of facets folded each time (e.g., from 5 facets per fold to 3), avoiding contour deformation due to excessively rapid folding. For example, if the similarity drops to 75% (below 80%) after a fold, the next fold will only process 3 facets to make the contour change smoother. Finally, the above folding and comparison process is iterated until the facet density of the component meets the visual recognition threshold of the decorative component. The visual recognition threshold refers to the lowest facet density (e.g., 30 facets per square meter) at which the human eye can clearly identify decorative details. When the facet density of the folded component reaches this threshold, the operation stops. For example, if the initial facet density of a component is 100 facets per square meter, after multiple folds it drops to 30 facets, and the contour skeleton similarity always remains above 80%, then simplification is complete.

[0051] To avoid excessive edge folding operations that could distort the contours of the first type of component (decorative component), an adaptive folding step size reduction mechanism adjusts the folding scale by calculating compensation coefficients in real time and dynamically updates the simplified region in conjunction with curvature re-detection. The specific implementation method is as follows:

[0052] The folding compensation coefficient is calculated based on the feature similarity difference between the current contour image and the original contour image. The feature similarity difference refers to the degree of difference between the original contour skeleton features and the current folded contour skeleton features (e.g., if the similarity drops from 90% to 70%, the difference is 20%). The compensation coefficient is positively correlated with this difference: the larger the difference, the smaller the compensation coefficient; the smaller the difference, the larger the compensation coefficient. For example, when the similarity difference is 10%, the compensation coefficient is set to 0.8; when the difference is 30%, the compensation coefficient drops to 0.3. The proportion of the total area of ​​the folded triangular facets in the current edge folding operation to the total area of ​​the flat area of ​​the component is extracted, i.e., the area percentage. For example, if 5 triangular facets are folded, the total area is 2 square meters, while the total area of ​​the flat area of ​​the component is 20 square meters, the area percentage is 10%. This area percentage is multiplied by the folding compensation coefficient to obtain the next folding step size, i.e., the area percentage of triangular facets that can be processed in the next folding operation. For example, if the area percentage is 10% multiplied by the compensation coefficient of 0.8, the next folding step size is 8%, i.e., the next folding step size is 8%. The total area of ​​each folded surface does not exceed 8% of the flat area. In this way, the folding step size is dynamically adjusted according to the degree of contour distortion: the greater the decrease in contour similarity, the smaller the step size, to avoid further contour deformation. The area percentage is multiplied by the folding compensation coefficient as the next folding step size. A preset period and a preset folding compensation coefficient threshold are set. If the compensation coefficient of each fold is less than the threshold within this period, it indicates that the folding of the current flat area is close to the boundary of the visually sensitive area. Continuing to fold according to the original area may affect the decorative details. At this time, the curvature re-detection module is activated to recalculate the curvature distribution data of the component surface and update the flat area distribution map.

[0053] Excessively folded areas are removed, and new flat areas are re-marked, such as areas that were not folded before and whose curvature is still below the visual perception threshold. If a component's compensation coefficient remains low after 5 folds, curvature re-detection reveals that the edge of the original flat area has reached the embossed pattern (visually sensitive area). The updated flat area distribution map then excludes this edge part, retaining only areas far from the sensitive area. This ensures that subsequent folding operations will not damage the decorative details. The adaptive reduction mechanism for folding step size dynamically controls the folding amplitude by multiplying the compensation coefficient and the area ratio to avoid contour distortion. It also updates and simplifies the area in real time through curvature re-detection to ensure that folding is always carried out within a safe range. Ultimately, this efficiently reduces the number of facets while maximizing the preservation of the visual characteristics of the decorative components.

[0054] A vertex clustering algorithm is performed on the components in the second type of component set to preserve mechanical topological features and meet the simplification criteria for non-load-bearing structures.

[0055] To simplify non-load-bearing structural components while preserving their mechanical and topological characteristics, a vertex clustering algorithm is applied to the second set of components. This simplification is achieved through a collaborative operation of key node protection and cluster verification, resulting in a safe and efficient simplification process. The specific implementation method is as follows:

[0056] The component mechanical topology features output by the loading function identification unit 2 include the force transmission path of the component, such as the main route of force transmission from one end of the component to the other and information on key stress points. Based on the component mechanical topology features, the coordinate set of key nodes in the force transmission path is identified. These nodes are turning points or concentration points of force transmission, such as the connection point between a non-load-bearing partition wall and the frame, or the endpoints of secondary supports, which are crucial for maintaining the basic mechanical performance of the component. For example, if the force transmission path of a non-load-bearing partition wall is from the top two ends to the bottom support, its top two ends and bottom support points are marked as key nodes, and their coordinates are recorded as a key node coordinate set. A protective zone with dynamically expanding radius is constructed around each node in the key node coordinate set. The protective zone is a three-dimensional region surrounding the nodes, used to prohibit vertex clustering operations and prevent the merging of key nodes from causing the loss of mechanical features. The radius of the protection zone employs a dynamic expansion mechanism: the initial radius is set based on the magnitude of the force exerted on the nodes. For example, nodes subjected to greater force have an initial radius of 10 cm, while those subjected to less force have an initial radius of 5 cm. If subsequent clustering operations approach the edge of the protection zone (e.g., the distance is less than 2 cm), the radius is automatically expanded (e.g., from 10 cm to 12 cm) to ensure that the mechanical topology around critical nodes is not disrupted. For instance, if a critical node initially has a protection zone radius of 10 cm, when a clustering operation involves a vertex 9 cm away, the radius expands to 12 cm to prevent that vertex from being merged and affecting the node's functionality.

[0057] A vertex clustering algorithm is performed in the area outside the protection zone to merge adjacent vertices that meet the simplification criteria for non-load-bearing components, reducing the number of vertices (e.g., merging a region with 100 vertices into 50) while preserving the overall shape of the component. After each clustering, the rate of change of the component's stiffness matrix is ​​verified through finite element micro-simulation. The stiffness matrix reflects the component's ability to resist deformation; a smaller rate of change indicates a smaller impact of clustering on mechanical performance. For example, a 5% change rate in the stiffness matrix after clustering indicates that simplification has a small impact on the component's stiffness; however, if the rate of change reaches 15%, it may affect the stability of the non-load-bearing structure. When the stiffness loss exceeds a preset stiffness loss threshold, the vertex clustering operation is rolled back to ensure that the component's mechanical performance meets the simplification criteria for non-load-bearing structures, meaning that it can still withstand its own weight and minor additional loads after simplification. If the stiffness loss reaches 12% after a certain clustering (exceeding the 10% threshold), that clustering is canceled, the merged vertices are restored, and vertices further away are re-selected for clustering to avoid structural risks caused by over-simplification.

[0058] To prevent insufficient stiffness in non-load-bearing components due to oversimplification during vertex clustering simplification, when finite element micro-simulation detects stiffness loss exceeding a preset threshold, it is necessary to accurately locate abnormal regions, roll back the clustering operation, and strengthen verification to ensure that the mechanical properties of the components meet the standards. The specific implementation method is as follows:

[0059] After the finite element micro-simulation detects that the stiffness loss exceeds the limit, the spatial coordinates of the component with abnormal stiffness matrix change rate are extracted. These coordinates are then matched with the vertex clustering operation log to locate the affected mesh region of the three most recent clustering operations.

[0060] Based on the spatial distribution characteristics of the abnormal coordinates in the affected mesh area, the range of the protection zone is expanded along the force transmission path of the component, and the vertex clustering operations that have been applied to the abnormal area in the most recent three times are canceled. At the same time, the protection zone expansion operation is performed, and the reverted mesh area and the expanded protection zone range are input into the vertex clustering algorithm for simplification. The reinforced finite element micro-simulation verification is then initiated on the re-simplified component.

[0061] In finite element micro-simulation, when the stiffness loss of a component exceeds a preset threshold, the spatial coordinates of the abnormal rate of change of the stiffness matrix are first extracted, i.e., the specific locations where the stiffness of the component has significantly decreased, such as the coordinates of the middle area of ​​a non-load-bearing partition wall. Then, the vertex clustering operation log is retrieved. This log records the vertex coordinates and merging time information involved in each clustering operation. The abnormal coordinates are matched with the clustering regions in the log to locate the mesh regions that have affected the abnormal region in the last three clustering operations. For example, if the last three clustering operations all involve vertices in the middle area of ​​the partition wall, indicating abnormal stiffness in the middle area, log matching reveals that the last three clustering operations merged adjacent vertices in this area. These clusterings are determined to be the cause of the stiffness loss. For the located affected mesh regions, the force transmission path of the component is expanded... Expanding the protection zone for critical nodes: The force transmission path is the main route through which force is transmitted in a component. For example, the force of a partition wall is transmitted from the top to the bottom support. When expanding the protection zone, it is necessary to cover the area around the critical nodes on the force transmission path. The radius of the protection zone for the top node of the partition wall is expanded from 10 cm to 15 cm to ensure that the mesh area on the force transmission path is not over-clustered. At the same time, the last three vertex clustering operations applied to the abnormal area are undoed, that is, the merged vertices are re-split and restored to the state before clustering. The area that was merged into 50 vertices is restored to 100 vertices. For example, if a partition wall has 50% fewer vertices in the middle area due to three clustering operations, after rollback, the number of vertices in the area is restored, and the stiffness is partially restored. The rolled-back mesh area and the expanded protection zone are input into the vertex clustering algorithm, and the simplification operation is re-executed.

[0062] In the area outside the new protective zone, vertices that are farther apart are selected for clustering, such as merging only vertices that are more than 5 cm apart to avoid excessively dense merging. The number of vertices in each cluster is reduced, for example, from 10 vertices to 5 vertices each time. After simplification, enhanced finite element micro-simulation verification is initiated. Compared with conventional verification, enhanced verification increases the load test scenario, such as applying 1.2 times the self-weight load to the surface of the component, and extends the simulation time to ensure that the simplified component can still meet the stiffness requirements under more stringent conditions. For example, after a partition wall is re-clustered, the enhanced verification confirms that the stiffness loss is 7%, which meets the simplification standard for non-load-bearing structures. The simplification is completed. Finally, while reducing the number of vertices, the mechanical topological characteristics of non-load-bearing components are not destroyed, meeting the dual requirements of lightweighting and structural stability.

[0063] Simplification processing is skipped for components in the third type of component set;

[0064] To ensure the integrity of the original data of the core load-bearing components in the third set and to avoid the impact of simplification on structural safety, a method of embedding digital watermark identifiers is used to skip simplification for these components. The specific implementation method is as follows:

[0065] While the model parsing unit 1 reads the IFC format file, special processing is simultaneously applied to components that have been marked as the third type of set by the function identification unit 2:

[0066] A hash value is generated from the component's type code, a fixed-length character sequence that uniquely identifies the code. For example, "ZL-01-load-bearing" is converted into a specific character combination. At the same time, the component's material strength parameters are extracted. The hash value and material strength parameters are concatenated in a fixed format (such as the string form "hash value + strength parameter") to form a digital watermark identifier. This identifier is unique; different components have different type codes or strength parameters, resulting in different generated identifiers. The generated digital watermark identifier is secretly embedded in the component's triangular mesh data, such as embedding it into the decimal places of the mesh vertex coordinates, without affecting normal data reading. This completes the marking of the third type of component. For example, if the type code hash value of a main beam component is "ABC123" and the strength parameter is "50MPa", the identifier "ABC123+50MPa" is generated and embedded in its mesh data, marking it as a component that needs to be skipped for simplification.

[0067] When the simplified execution unit 3 processes component data, it performs real-time detection of the digital watermark identifier for each component:

[0068] The system monitors the presence of digital watermark identifiers in real time, extracting potentially hidden watermark identifiers from component data using a dedicated algorithm. It verifies the correctness of the format, such as whether it conforms to the "hash value + strength parameter" concatenation rule, and whether the content matches the component type code and material parameters. For example, it checks if the strength parameter in the identifier matches the parameters extracted by the parsing unit. If a valid watermark identifier is detected, a skip procedure is triggered, stopping any simplification operations on the component, such as edge folding and vertex clustering, ensuring that the original triangular mesh data of the component, such as vertex coordinates and the number of faces, remains unchanged. For instance, if a column component is detected with a valid watermark, simplification execution unit 3 skips the process directly, preserving its original 1000 triangular face data intact. After the skip procedure is triggered, the original triangular mesh data of the skipped component is directly transmitted to model reconstruction unit 4, ensuring that the data is not modified during transmission. Simultaneously, the system automatically records the skipped component information, including the total number of skipped components and their spatial coordinates. These records are used for verification during subsequent model reconstruction, such as confirming whether all third-category components are loaded according to the original data and tracing model versions, such as checking the number and location of third-category components in a specific version of the model.

[0069] Model Reorganization Unit 4 reorganizes the component data processed by Simplified Execution Unit 3 into a lightweight BIM model.

[0070] To integrate the three types of components that have undergone differentiated processing into a complete lightweight BIM model, while retaining the processing characteristics and quick access capabilities of each component, model reorganization unit 4 achieves reorganization through a process of metadata binding, spatial index construction, and hierarchical matching. The specific implementation method is as follows:

[0071] The model reorganization unit 4 first receives three types of component sets output by the simplification execution unit 3, and binds differentiated metadata tags to each component in the three types of component sets. The content of the differentiated metadata tags is the processing features of the functional identification unit 2 and the simplification execution unit 3. The processing features of the functional identification unit 2 include: the component's set category (first type / second type / third type), the proportion of visually sensitive areas (for the first type), mechanical topological feature keywords (for the second and third types), and the material strength parameter level, etc. The processing features of the simplification execution unit 3 include: the number of edge folds and the final face density of the first type of component, the change in the number of vertices before and after vertex clustering and the stiffness loss rate of the second type of component, and the "unsimplified" mark and the original number of facets of the third type of component. For example, the metadata tag of a certain first type of carved component is: "Set category: first type; proportion of visually sensitive areas: 40%; number of edge folds: 8 times; final facet density: 30 pieces / square meter"; the tag of a certain third type of main beam component is: "Set category: third type; mechanical topological feature: main force transmission path; processing status: unsimplified; original number of facets: 1000 pieces". These labels are like "identity files" for the components, recording the processing steps and providing feature basis for subsequent model applications;

[0072] By recursively partitioning the space, components are assigned to leaf nodes, and precision control markers are added to the nodes. When building a lightweight encapsulated structure, the LOD level is automatically matched according to the leaf node markers of the octree. Finally, a lightweight BIM model carrying an octree index, data blocks and precision rule configuration file is output, which supports reverse restoration of the original data of load-bearing components through the watermark interface.

[0073] A dynamic octree index is constructed based on the spatial coordinates of the components. An octree index is a structure that recursively divides a three-dimensional space into eight sub-regions. The specific steps are as follows:

[0074] Using the entire spatial range of the model as the root node, it is evenly divided into 8 cubic sub-regions. The number of components in each sub-region is checked. If the number of components in a sub-region exceeds a preset threshold, the sub-region is further divided into 8 smaller sub-regions until the number of components in each sub-region (leaf node) does not exceed the threshold. For example, the root node range of a building model is X=0-100, Y=0-100, Z=0-50. After three recursive divisions, each leaf node corresponds to a spatial range of 5×5×3 meters, containing 10-15 components. Based on the spatial coordinates of each component, it is assigned to the corresponding leaf node. For example, if the coordinate range of a partition wall falls entirely within the leaf node with X=10-15, Y=20-25, Z=0-3, it is assigned to that node. A precision control mark is added to each leaf node. The mark content is the simplified precision level of the components within that node. For example, the first type of component corresponds to "visual precision level", the second type corresponds to "non-load-bearing precision level", and the third type corresponds to "original precision level". For example, leaf nodes containing carved components are labeled "visual precision level," nodes containing partition walls are labeled "non-load-bearing precision level," and nodes containing main beams are labeled "original precision level." The dynamic octree index enables rapid model retrieval and loading. For instance, when viewing a specific floor, only the components corresponding to the leaf nodes of that floor are loaded. Precision control tags provide a basis for model calling in different scenarios. Based on the dynamic octree index, a lightweight BIM model encapsulation structure is constructed. The specific steps are as follows:

[0075] Based on the precision control marker of the leaf node, the corresponding LOD level is automatically matched: "Original Precision Level" corresponds to LOD300 (high precision), "Non-load-bearing Precision Level" corresponds to LOD200 (medium precision), and "Visual Precision Level" corresponds to LOD100 (basic precision). For example, the leaf node where the main beam is located matches LOD300, the partition wall matches LOD200, and the carving matches LOD100. The component data (including geometric data and metadata tags) within each leaf node is packaged into an independent data block. The data block uses a compression algorithm to reduce its volume, but retains key information. The original triangular mesh data of the third type of component is not compressed. A precision rule configuration file is created to record the precision standard corresponding to each LOD level, the storage path of the data block, and the structural information of the octree index. The final output lightweight BIM model contains three parts: dynamic octree index, compressed component data blocks, and precision rule configuration file. At the same time, the model supports reverse restoration of the original data of the third type of component through the watermark interface. When it is necessary to call high-precision load-bearing component data, the unsimplified original triangular mesh data is extracted from the data block by reading the watermark identifier in the metadata tag, ensuring the accuracy of the structural analysis scenario. This allows the large-scale BIM model to significantly reduce storage and operating costs while ensuring core accuracy, meeting the application requirements of intelligent loading.

[0076] In this invention, the model parsing unit 1 reads the IFC file and separates the component data; the function identification unit 2 analyzes through geometric and semantic channels and outputs sets of decorative components (first type), non-load-bearing structures (second type), and load-bearing components (third type); the simplification execution unit 3 performs edge folding on the first type, vertex clustering on the second type, and skips simplification on the third type; and the model reorganization unit 4 binds metadata and constructs a dynamic octree index to generate a lightweight model. This solves the problems of traditional simplification strategies being too simplistic and losing core information, improves model loading efficiency and application accuracy, and is suitable for complex architectural scenarios.

[0077] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A large-scale BIM model optimization system based on intelligent loading, comprising a model parsing unit (1) for reading the IFC format file of the BIM model, separating the triangular mesh, component type code, and material strength parameters of each component in the IFC format file, characterized in that, Also includes: The functional identification unit (2) inputs the component data output by the model parsing unit (1) into the convolutional neural network model. The model contains parallel geometric analysis channels and semantic parsing channels, which are used to calculate and identify the component data and output the first type of component set, the second type of component set and the third type of component set. The functional identification unit (2) generates three sets through the following steps: In the geometric analysis channel, the density distribution data of the triangular mesh of the component is calculated. The density gradient detection algorithm is used to identify the region in the density distribution data whose density change rate exceeds a preset threshold. The geometric attribute set of the region is extracted and quantized into a geometric feature vector. The geometric attribute set includes the region area ratio, the number of curvature extreme points, and the boundary contour complexity. In the semantic parsing channel, the structural attribute identifier of the component type code is parsed, the material strength parameter is extracted based on the structural attribute identifier, the material strength parameter is mapped to the safety level coefficient according to the building code, and a semantic feature vector is generated by combining the structural attribute identifier. The geometric feature vector and semantic feature vector are input into the dynamic weight allocation module of the convolutional neural network model, and the first type of component set, the second type of component set and the third type of component set are output based on the priority allocation result of the dynamic weight allocation module. The simplified execution unit (3) performs differentiated processing on the first type of component set, the second type of component set, and the third type of component set respectively: The edge-folding algorithm is applied to the components in the first type of component set to simplify the patch density to meet the visual recognition threshold of the decorative component. A vertex clustering algorithm is performed on the components in the second type of component set to preserve mechanical topological features and meet the simplification criteria for non-load-bearing structures. Simplification processing is skipped for components in the third type of component set; The model reorganization unit (4) reorganizes the component data processed by the simplified execution unit (3) into a lightweight BIM model.

2. The large-scale BIM model optimization system based on intelligent loading according to claim 1, characterized in that, The generation of the first type of set further includes: Based on the patch density distribution data output from the geometric analysis channel, the characteristic surface partitions of the components are divided using a curvature clustering algorithm. Regions with curvature change rate exceeding a preset change rate threshold are marked as visually sensitive areas. Simultaneously, text matching rules are activated in the semantic parsing channel to block component type codes containing preset keywords and filter components whose material strength is higher than the standard value of decorative materials. Components with a visually sensitive area ratio exceeding a preset visual sensitivity threshold and passing the filter are included in the first set.

3. The large-scale BIM model optimization system based on intelligent loading according to claim 1, characterized in that, The generation of the second type of set further includes: A dynamic filtering mechanism for material strength parameters is established in the semantic parsing channel. When the component type code belongs to the non-load-bearing structure category, the concrete strength grade and steel yield strength parameters are extracted and compared with the preset allowable values ​​for non-load-bearing structures. Components with strength values ​​lower than the standard values ​​are selected. At the same time, the overall curvature distribution of the component is calculated in the geometric analysis channel. Components with smooth curvature changes and no abrupt changes are marked. Components that meet both of the above requirements are included in the second set.

4. The large-scale BIM model optimization system based on intelligent loading according to claim 1, characterized in that, The edge-folding algorithm is applied to the components in the first type of component set, as follows: The surface curvature distribution data of the component is calculated, and flat areas with curvature values ​​lower than the visual perception threshold are identified. Based on these flat areas, candidate triangular facets for edge folding are located. After each folding operation, a component orthographic projection contour map is generated in real time. The current contour map and the original contour map are input into a convolutional neural network model for skeleton feature comparison. When the feature similarity is lower than the similarity threshold, an adaptive reduction mechanism for the folding step size is triggered. This process is iterated until the visual recognition threshold of the decorative component is met.

5. A large-scale BIM model optimization system based on intelligent loading according to claim 4, characterized in that, The specific implementation of the adaptive reduction mechanism for the folding step size is as follows: The folding compensation coefficient is calculated based on the feature similarity difference between the current contour map and the original contour map. The area ratio of the triangular facets in the current edge folding operation is extracted, and the area ratio is multiplied by the folding compensation coefficient as the next folding step size. When the folding compensation coefficient of the folding operation within the preset period is less than the preset folding compensation coefficient threshold, the curvature re-detection module is activated to update the flat area distribution map.

6. The large-scale BIM model optimization system based on intelligent loading according to claim 1, characterized in that, The vertex clustering algorithm is performed on the components in the second type of component set, as follows: The component mechanical topology features output by the loading function identification unit (2) are used to identify the key node coordinate set in the force transmission path based on the component mechanical topology features. A protection zone with dynamically expanding radius is constructed with each node in the key node coordinate set as the center. A vertex clustering algorithm is executed in the area outside the protection zone to merge adjacent vertices that meet the simplification criteria for non-load-bearing components. After each clustering, the change rate of the component stiffness matrix is ​​verified by finite element micro-simulation. When the stiffness loss exceeds the preset stiffness loss threshold, the vertex clustering operation is rolled back.

7. A large-scale BIM model optimization system based on intelligent loading according to claim 6, characterized in that, The method for reverting to vertex clustering when the stiffness loss exceeds a preset stiffness loss threshold is as follows: After the finite element micro-simulation detects that the stiffness loss exceeds the limit, the spatial coordinates of the component with abnormal stiffness matrix change rate are extracted. These coordinates are then matched with the vertex clustering operation log to locate the affected mesh region of the three most recent clustering operations. Based on the spatial distribution characteristics of the abnormal coordinates in the affected mesh area, the range of the protection zone is expanded along the force transmission path of the component, and the vertex clustering operations that have been applied to the abnormal area in the most recent three times are canceled. At the same time, the protection zone expansion operation is performed, and the reverted mesh area and the expanded protection zone range are input into the vertex clustering algorithm for simplification. The reinforced finite element micro-simulation verification is then initiated on the re-simplified component.

8. A large-scale BIM model optimization system based on intelligent loading according to claim 1, characterized in that, The method for skipping simplification processing for components in the third type of component set is as follows: When the model parsing unit reads the IFC format file, it embeds a digital watermark identifier for the components marked as the third set. The identifier is generated by splicing the hash value of the component type code and the material strength parameter. When the simplified execution unit (3) processes the component data, it detects the existence status of the digital watermark identifier in real time. If a valid watermark identifier is detected, it triggers the skip procedure and directly transmits the original triangular mesh data of the skipped component to the model reorganization unit (4). At the same time, it records the total number of skipped components and their spatial coordinates.

9. A large-scale BIM model optimization system based on intelligent loading according to claim 1, characterized in that, The model reorganization unit (4) achieves lightweight BIM model reorganization through the following steps: Receive the three types of component sets output by the simplified execution unit (3), and bind a differentiated metadata tag to each component in the three types of component sets. The content of the differentiated metadata tag is the processing characteristics of the function identification unit (2) and the simplified execution unit (3). A dynamic octree index is constructed based on the spatial coordinates of the components. The components are assigned to leaf nodes through recursive spatial partitioning, and precision control markers are added to the nodes. When constructing a lightweight encapsulation structure, the LOD level is automatically matched according to the leaf node markers of the octree. Finally, a lightweight BIM model carrying the octree index, data blocks and precision rule configuration files is output. It supports reverse restoration of the original data of load-bearing components through the watermark interface.

Images