A lightweight processing method for a large complex building prefabricated component BIM model

Abstract

This invention provides a lightweight processing method for BIM models of large and complex prefabricated building components, belonging to the field of prefabricated component BIM technology. This invention eliminates geometric redundancy by using a generalized graph matching algorithm to identify and instantiate duplicate prefabricated components; it employs a curvature adaptive-semantic feature joint discretization model to dynamically and accurately discretize the parametric geometry; it uses octree spatial partitioning and virtual camera ray projection to identify and remove internal invisible meshes; it extracts attribute data to a distributed graph database, using lightweight globally unique identifiers to achieve decoupling and asynchronous loading of geometry and attributes; it utilizes artificial intelligence to output an optimized and simplified mesh; and finally, it encapsulates and outputs a lightweight BIM model file through geometry-attribute integration. This solves the technical problems of accuracy loss and computational efficiency imbalance caused by large geometric redundancy and the difficulty in coordinating discretization accuracy with computational resources in the lightweight processing of BIM models of large and complex prefabricated building components.

Description

Technical Field

[0001] This invention belongs to the field of prefabricated component BIM technology, specifically, it relates to a method for lightweighting BIM models of large and complex prefabricated building components. Background Technology

[0002] In the field of Building Information Modeling (BIM), lightweighting of prefabricated component BIM models is a crucial step in achieving visualization and collaborative management of large-scale building projects. Traditional lightweighting methods typically employ fixed-step mesh discretization and static mesh simplification algorithms to reduce the polygon count of component geometric data, storing geometric and attribute data in a single file structure, relying on manual setting of simplification scale and sampling density parameters. However, in large-scale building BIM scenarios containing tens of thousands of repetitive prefabricated components, the significant differences in geometric complexity among different components lead to oversparse sampling in high-curvature regions, resulting in loss of detail, and redundant polygons in flat regions, causing data bloat. Simultaneously, static simplification algorithms cannot perceive the semantic type of components or the current computing resource status, causing resource allocation and accuracy control to become decoupled. In other words, existing technologies suffer from the technical problem of large, complex prefabricated component BIM models having large amounts of geometric data redundancy, and the difficulty in coordinating discretization accuracy and computing resources, resulting in an imbalance between model accuracy loss and computational efficiency after lightweighting. Summary of the Invention

[0003] In view of this, the present invention provides a lightweight processing method for BIM models of large and complex prefabricated building components, which can solve the technical problem in the prior art that the large amount of geometric data redundancy and the difficulty in coordinating and adapting discretization accuracy and computing resources in the BIM models of large and complex prefabricated building components lead to a loss of model accuracy and an imbalance in computing efficiency after lightweight processing.

[0004] This invention is implemented as follows: This invention provides a method for lightweight processing of BIM models of large and complex prefabricated building components, including the following steps:

[0005] Traverse all prefabricated components in the BIM database, establish the attribute-geometric composite graph structure for each component, calculate the graph Laplacian matrix spectral characteristics, generate high-dimensional invariant hash key values, identify identical instantiated components through a generalized graph matching algorithm, retain the parent geometric data, and simplify the remaining duplicate components into affine transformation matrices and parent pointer mappings.

[0006] A curvature adaptive-semantic feature joint discretization model is used to dynamically discretize the parametric geometry of prefabricated components. Intensive sampling is performed in high curvature regions and key connection parts, while extremely sparse sampling is performed in flat regions and non-visible internal surfaces. The discretization template of high-frequency parametric components is stored in a caching mechanism.

[0007] An octree is used to divide the prefabricated components into multiple spatial layers. The absolutely invisible internal mesh is identified and marked from multiple angles by virtual camera ray projection. The internal steel reinforcement and pipeline mesh is removed by dynamic topology stripping technology, leaving only the outer envelope geometry and attribute links.

[0008] The attribute data of prefabricated components is extracted and stored in a distributed graph database. The geometric data is stored in a streaming compressed format and decoupled and associated with a lightweight globally unique identifier. The asynchronous on-demand loading of geometry and attributes is achieved through a microservice parallel processing mechanism.

[0009] Using the discretized mesh output by the curvature adaptive-semantic feature joint discretization model as input, the high-frequency geometric details and low-frequency macroscopic topological features of the prefabricated components are extracted by the complex domain-wavelet-multi-scale attention map convolutional simplification model. The number of internal iterations is adaptively determined according to the mesh complexity, and the optimal simplification mesh and vertex weight allocation scheme are output.

[0010] Based on the vertex weight allocation scheme, the lightweight BIM model of the whole scene is encapsulated by geometry-attribute integration. The final mapping relationship between geometric data and attribute data in the distributed graph database is established through a globally unique identifier, and the lightweight BIM model file is output.

[0011] Specifically, the attribute-geometric composite graph structure is a graph data structure constructed using the branch geometric features of prefabricated components as nodes and the spatial adjacency relationships and Boolean operation relationships between the various parts of the components as directed edges.

[0012] Specifically, the calculation of the spectral features of the graph Laplacian matrix involves constructing a normalized graph Laplacian matrix from the attribute-geometric composite graph. ,in It is an adjacency matrix. For the degree matrix, Perform eigenvalue decomposition and take the first eigenvalue. Each eigenvalue constitutes a spectral eigenvector.

[0013] The high-dimensional invariant hash key is generated by combining the following four types of quantities: the graph Laplacian matrix before... The four types of quantities—a sequence of feature values, the Euclidean distance between the coordinates of the three-dimensional geometric center of the precast component, the volume ratio of the bounding box aligned with the outer envelope axis, and the surface area of ​​the precast component—are normalized and concatenated into a high-dimensional vector, which is then mapped to a fixed-length key value using a locality-sensitive hash function.

[0014] The generalized graph matching algorithm refers to a graph theory-based generalized graph matching algorithm used to calculate the maximum common subgraph between two attribute-geometric composite graphs; when the node coverage of the maximum common subgraph of two components is higher than the similarity threshold, they are determined to be the same type of instantiated components.

[0015] The affine transformation matrix refers to a 4×4 homogeneous coordinate transformation matrix used to express translation, rotation, and scaling; the parent pointer mapping refers to a lightweight reference structure in the BIM scene database that points to the location of the parent geometric data storage with a unique index.

[0016] Specifically, the curvature adaptive-semantic feature joint discretization model is an algorithm model that dynamically adjusts the sampling density of the triangular mesh based on the local curvature distribution and the semantic type of the component during the parametric geometric discretization process; a smaller triangle step size is used in high curvature regions and key connection parts, while a larger step size is used in flat regions and non-visible internal surfaces.

[0017] The caching mechanism refers to establishing a hash cache table in memory with the parameterized component type identifier as the key and the corresponding discretized triangular mesh template as the value; when the same type of parameterized component enters the discretization process again, the mesh template in the cache is directly read and the parameters are scaled.

[0018] The octree multi-level spatial partitioning refers to recursively dividing the prefabricated component into eight equal parts by aligning the axis of the prefabricated component with the bounding box, until the leaf node size is less than the minimum voxel side length threshold; the virtual camera ray projection refers to uniformly distributing several virtual camera positions on the outer spherical surface of the prefabricated component, emitting a ray beam covering the entire component at each position, recording the first patch hit by each ray, and marking the patches not hit by any ray as absolutely invisible meshes.

[0019] The dynamic topology stripping technique refers to directly deleting the vertices, faces, and their index references of absolutely invisible meshes from the mesh data structure while maintaining the integrity of the outer envelope geometry topology, and updating the face adjacency list, while retaining the attribute link index between the removed internal structure and the outer envelope geometry.

[0020] Specifically, the simplified complex domain-wavelet-multi-scale attention graph convolutional model involves the input layer receiving vertex coordinate matrices, adjacency matrices, local curvature matrices, and high-dimensional semantic vectors of material properties. The input layer maps vertex spatial coordinate components to complex moduli and surface manifold trend angles to complex phase angles, resulting in a complex domain vertex feature matrix. Subsequently, it undergoes three-dimensional discrete wavelet transform decomposition into high-frequency geometric detail feature sub-bands and low-frequency macroscopic topological feature sub-bands. These two sub-bands are then fed into independent complex domain graph convolutional branches for feature extraction, and finally merged in the multi-scale attention fusion module to output vertex retention weights and CUDA flow allocation proposal vectors.

[0021] The multi-scale attention fusion module has a built-in residual iterative jump connection structure with a dynamic gating unit. The dynamic gating unit outputs gating weights based on the geometric complexity score of the current feature map. The gating weights determine whether the current feature map continues to enter the next iteration or is output directly.

[0022] Wherein, the vertex dynamic weight adjustment function Based on the current total number of facets of the component Global curvature mean Current available video memory Ratio of current iteration round to maximum iteration round Calculate, the expression is ,in , , , The weighting coefficients are and satisfy the following conditions: .

[0023] The high-dimensional semantic vector refers to the text encoding of the material name, component category name and manufacturer specification description of the prefabricated component into a dense real number vector of fixed dimension through a pre-trained language model, which is used to introduce semantic prior information in the complex domain-wavelet-multi-scale attention map convolution simplified model.

[0024] The similarity threshold ranges from 0.85 to 0.98; The value range is 8–32; the step size of the triangle in the high curvature region ranges from 0.5–5 mm; the step size of the flat region ranges from 20–200 mm; the minimum voxel side length threshold ranges from 1–10 mm; the number of virtual cameras ranges from 32–256; the number of convolution kernels in the complex domain graph convolutional layer ranges from 32–128; the maximum number of internal iterations ranges from 2–8; the geometric complexity score threshold ranges from 0.2–0.8; , , , The values ​​of each coefficient range from 0.1 to 0.5; the dimensions of the high-dimensional semantic vector range from 128 to 512.

[0025] This invention achieves instantiation and merging of repetitive components by constructing an attribute-geometric composite graph structure, employs a curvature adaptive-semantic feature joint discretization model for accuracy adaptive sampling, and combines a complex domain-wavelet-multi-scale attention graph convolution simplification model to achieve resource-aware dynamic mesh simplification. This solves the technical problems of accuracy loss and computational efficiency imbalance in BIM models of large and complex prefabricated building components, caused by large geometric data redundancy and the difficulty in coordinating discretization accuracy with computational resources. The geometric redundancy elimination of this invention allows subsequent discretization and simplification steps to process only a limited number of parent geometry types. Curvature adaptive sampling ensures geometric fidelity in key areas. Complex domain graph convolution captures both geometric position and surface orientation information through complex encoding. A dynamic gating iteration mechanism automatically adjusts the computational depth based on geometric complexity. In summary, this invention solves the technical problems of accuracy loss and computational efficiency imbalance in the lightweight processing of BIM models of large and complex prefabricated building components mentioned in the background art. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 A comparison chart of the total number of front panels before and after lightweighting different types of prefabricated components.

[0028] Figure 3 A diagram showing the vertex retention rate and mesh patch quantity distribution after simplification of various parent components. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0030] like Figure 1 The diagram shown is a flowchart of a method for lightweighting BIM models of large and complex prefabricated building components provided by this invention. This method includes the following steps:

[0031] S01. Traverse all prefabricated components in the BIM database, establish the attribute-geometric composite graph structure for each component, calculate the graph Laplacian matrix spectral characteristics, generate high-dimensional invariant hash key values, identify identical instantiated components through the generalized graph matching algorithm, retain the parent geometric data, and simplify the remaining duplicate components into affine transformation matrices and parent pointer mappings.

[0032] S02. The curvature adaptive-semantic feature joint discretization model is used to dynamically discretize the parametric geometry of the prefabricated component. The sampling is densified in the high curvature region and key connection parts, and the sampling is extremely sparse in the flat region and non-visible internal surface. The discretization template of the high-frequency parametric component is stored in the cache mechanism.

[0033] S03. Use octree to divide the prefabricated components into multiple spatial layers. Use virtual camera ray projection to identify and mark the absolutely invisible internal mesh from multiple angles. Use dynamic topology stripping technology to remove the internal steel reinforcement and pipeline mesh, and only retain the outer envelope geometry and attribute links.

[0034] S04. Extract and store the massive attribute data of prefabricated components in a distributed graph database. The geometric data is stored in a streaming compressed format and decoupled and associated with a lightweight globally unique identifier. Asynchronous on-demand loading of geometry and attributes is achieved through a microservice parallel processing mechanism.

[0035] S05. Using the discretized mesh output by the curvature adaptive-semantic feature joint discretization model as input, the high-frequency geometric details and low-frequency macroscopic topological features of the prefabricated components are extracted using the complex domain-wavelet-multi-scale attention map convolutional simplification model. The number of internal iterations is adaptively determined according to the mesh complexity, and the optimal simplification mesh and vertex weight allocation scheme are output.

[0036] S06. Based on the vertex weight allocation scheme obtained in S05, perform geometry-attribute integration and encapsulation on the full-scene lightweight BIM model, establish the final mapping relationship between geometric data and attribute data in the distributed graph database through a globally unique identifier, and output the lightweight BIM model file.

[0037] The attribute-geometry composite graph structure refers to a graph data structure constructed with the branch geometric features of prefabricated components as nodes and the spatial adjacency and Boolean operation relationships between the various parts of the components as directed edges, which is used to uniformly express the geometric shape and attribute dependencies of prefabricated components.

[0038] The calculation method for the spectral features of the graph Laplacian matrix is ​​as follows: a normalized graph Laplacian matrix is ​​constructed from the attribute-geometric composite graph. ,in It is an adjacency matrix. For the degree matrix, Perform eigenvalue decomposition and take the first eigenvalue. eigenvalues ​​( The value range is 8 to 32. The spectral feature vector is determined by conducting multiple rounds of offline experiments on BIM projects of different sizes and statistically analyzing the convergence interval of the feature values.

[0039] The high-dimensional invariant hash key is generated by combining the following four types of quantities: the graph Laplacian matrix before... The four types of quantities—a sequence of feature values, the Euclidean distance between the coordinates of the three-dimensional geometric centers of the precast components, the volume ratio of the bounding boxes aligned with the outer envelope axes, and the surface area of ​​the precast components—are normalized and concatenated into a high-dimensional vector. This vector is then mapped to a fixed-length key value using a locality-sensitive hash function, which is used to quickly retrieve suspected similar components.

[0040] The generalized graph matching algorithm refers to a graph theory-based generalized graph matching algorithm used to calculate the maximum common subgraph between two attribute-geometric composite graphs. When the node coverage of the maximum common subgraph of two components is higher than the similarity threshold, they are determined to be the same type of instantiated components. The similarity threshold ranges from 0.85 to 0.98, and the optimal value is determined by conducting receiver operating characteristic curve analysis experiments on manually labeled datasets of multiple typical prefabricated component families.

[0041] The affine transformation matrix refers to a 4×4 homogeneous coordinate transformation matrix used to express translation, rotation, and scaling; the parent pointer mapping refers to a lightweight reference structure in the BIM scene database that points to the parent geometric data storage location with a unique index. Non-parent instances do not store geometric data separately, but only store the affine transformation matrix and the parent pointer mapping, thereby eliminating scene-level geometric redundancy.

[0042] Technical effects of the generalized graph matching algorithm in the solution: This algorithm performs dual similarity measurement on prefabricated components in both the spectral and structural domains of the graph, grouping tens of thousands of repetitive prefabricated components in a large-scale BIM scene into a unique parent geometry and lightweight pointer system. Since prefabricated components are highly repetitive in actual engineering projects, this algorithm fundamentally eliminates redundant storage at the geometric level of the BIM scene. Subsequent steps such as discretization and ray casting only need to be performed once on a limited number of parent geometries, greatly reducing the total computational load of the entire process and providing a stable instance mapping foundation for subsequent asynchronous attribute loading.

[0043] The curvature adaptive-semantic feature joint discretization model refers to an algorithm model that dynamically adjusts the sampling density of the triangular mesh based on the local curvature distribution and the semantic type of the component during the parametric geometric discretization process. For curved surfaces with large curvature and key connection parts such as tenons, grooves, and reserved openings, a smaller triangle step size (step size range of 0.5 to 5 mm, determined through geometric error comparison experiments on prefabricated components of different sizes) is used. For flat areas and non-visible internal surfaces, a larger step size (step size range of 20 to 200 mm, also determined through geometric error comparison experiments) is used to retain key geometric details while controlling the number of faces.

[0044] The caching mechanism refers to establishing a hash cache table in memory with the parameterized component type identifier as the key and the corresponding discretized triangular mesh template as the value. When the same type of parameterized component enters the discretization process again, the mesh template in the cache is directly read and the parameters are scaled, without repeating the geometric analytical calculation.

[0045] The non-visible internal surface refers to a surface that is completely obscured by the outer geometry from all reasonable viewing angles and contributes nothing to any rendered view, including the surface mesh of the steel reinforcement skeleton inside the precast component and the surface mesh of the embedded pipeline.

[0046] The octree multi-layer spatial partitioning refers to recursively dividing the prefabricated component into eight equal parts by aligning the axis of the prefabricated component with the bounding box, until the leaf node size is smaller than the set minimum voxel side length (ranging from 1 to 10 mm, determined by conducting multiple sets of ray projection experiments on prefabricated components with different geometric accuracy requirements, and taking into account the optimal combination of internal mesh recognition rate and computation time); the number of partitioning layers ranges from 4 to 8 layers, determined by statistical analysis of the missed detection rate under different number of layers in pre-experiments.

[0047] The virtual camera ray projection refers to the uniform distribution of several virtual camera positions on the outer spherical surface of the prefabricated component, emitting a ray beam covering the entire component at each position, and recording the first facet hit by each ray; the facets not hit by any ray first are marked as absolutely invisible meshes and included in the dynamic topology stripping object; the number of virtual cameras ranges from 32 to 256, which is determined by statistical experiments on the missed detection rate of typical component families.

[0048] The dynamic topology stripping technique refers to directly deleting the vertices, faces, and their index references of absolutely invisible meshes from the mesh data structure while maintaining the integrity of the outer envelope geometry, and updating the face adjacency list. At the same time, it retains the attribute link index between the removed internal structure and the outer envelope geometry, so that attribute data is not lost during geometry stripping.

[0049] The distributed graph database refers to a distributed database system that stores attribute nodes and attribute relationships using a graph data model. It supports high-concurrency attribute reading and writing using globally unique identifiers as indexes and is suitable for storing various types of attribute data such as material, manufacturer, and mechanical parameters of prefabricated components.

[0050] The streaming compression format refers to an extended format that uses the Draco geometric compression protocol to perform lossy or lossless streaming encoding on the vertex coordinates and face connection relationships of the triangular mesh. It supports segment-by-segment decompression during transmission or loading without waiting for the full data to arrive.

[0051] The lightweight globally unique identifier refers to a 128-bit unique code generated according to the RFC 4122 standard. In this scheme, it serves as a unique decoupling association key between geometric data and corresponding attribute data in the distributed graph database. It has a fixed length, is generated without collisions, and is suitable for use in large-scale concurrent scenarios.

[0052] The microservice parallel processing mechanism refers to splitting independent tasks such as attribute data reading and writing, geometric data compression, and global unique identifier allocation into mutually decoupled microservice processes, which are asynchronously scheduled through message queues. Each microservice runs independently on different computing nodes, realizing the parallel execution of multiple tasks and eliminating single-point memory input and output blocking.

[0053] The simplified complex-domain-wavelet-multi-scale attention map convolutional model is a graph neural network model based on deep learning. The specific structure of this simplified model is as follows: The input layer receives four types of data: the original 3D mesh vertex coordinate matrix of the prefabricated component, the adjacency matrix of the patch connection graph, the local curvature matrix, and the high-dimensional semantic vector of the material. The input layer converts the vertex coordinates from the real domain to the complex domain, mapping the spatial coordinate components of each vertex to the modulus of a complex number, and mapping the manifold trend angle of the surface containing the vertex to the phase angle of a complex number, thus obtaining the complex-domain vertex feature matrix. Subsequently, the model enters a 3D discrete wavelet transform layer, applying a 3D discrete wavelet transform to the complex-domain vertex feature matrix. The graph signal is decomposed into high-frequency geometric detail feature sub-bands and low-frequency macroscopic topological feature sub-bands. These two sub-bands are then fed into independent complex-domain graph convolution branches for feature extraction. The complex-domain graph convolution operation employs complex multiplication rules to perform message passing and neighborhood aggregation in the complex spectral domain. The two feature sub-bands converge in a multi-scale attention fusion module, which incorporates a residual iterative jump connection structure with dynamic gating units. The dynamic gating units output gating weights based on the geometric complexity score of the current feature map (calculated jointly by the local variance of the curvature matrix and the semantic vector entropy). These gating weights determine whether the current feature map continues into the next iteration or is directly output. In each iteration, the attention mechanism dynamically calculates the weights of the high-frequency sub-bands. With low-frequency subband weights (satisfy The two features are weighted and fused; the maximum number of internal iterations of the model ranges from 2 to 8, adaptively determined by the geometric complexity score. The geometric complexity score threshold is determined through multiple rounds of training convergence experiments on different types of components such as irregular prefabricated stairs and standard rectangular wall panels, with a value range of 0.2 to 0.8; the model output layer outputs two types of results: the retention weight of each vertex (range of...). The model includes a CUDA stream allocation proposal vector; each component of the CUDA stream allocation proposal vector corresponds to the workload ratio of a CUDA stream, and the vector dimension is equal to the number of available CUDA streams. This vector is predicted by the model at the output layer through a fully connected subnetwork. The input of this subnetwork is the total number of faces of the current component, the average curvature, the semantic category encoding, and the current available memory. The output is the load allocation coefficient of each CUDA stream. The thread block size and thread bundle allocation of the CUDA kernel function are obtained by mapping the vertex retention weight sparsity statistics through a two-layer perceptron subnetwork, thus directly functionally relating to the dynamic discretization parameters in step S05 in both hierarchical resource scheduling and stream-level resource scheduling dimensions. The number of convolution kernels in the complex domain graph convolutional layer of the model ranges from 32 to 128, and is determined through accuracy-time trade-off experiments on multi-scale BIM scenarios.

[0054] The specific steps for establishing the training dataset for the complex domain-wavelet-multi-scale attention map convolution simplified model include: collecting original 3D models of prefabricated components from multiple real large-scale prefabricated building BIM projects, covering categories such as standard rectangular wall panels, composite floor slabs, irregular prefabricated stairs, frame columns, beams, and various connectors, with no less than 200 different parameter instances for each category; having professional BIM engineers manually annotate the vertices of each component to retain the ground truth weight (values ​​are 0 or 1), and recording the corresponding local curvature matrix, high-dimensional semantic vector of materials, and patch connection relationship diagram; controlling the ratio of positive and negative samples (retained vertices and vertices to be removed) in the dataset between 1:1 and 3:1, and balancing them through oversampling; dividing the dataset into training set, validation set, and test set, with a ratio of 7:1.5:1.5;

[0055] The training steps of the complex domain-wavelet-multi-scale attention map convolutional simplified model specifically include: using a weighted sum of the binary cross-entropy loss with vertex preservation weights and the mean squared error loss of the CUDA flow assignment proposal vector as the total loss function; the weight ratio is determined through multiple rounds of ablation experiments on the validation set, ranging from 1:0.1 to 1:0.5; using the Adam optimizer, with an initial learning rate ranging from... ~ The model is dynamically decayed using a cosine annealing strategy; the batch size ranges from 8 to 32, depending on the available GPU memory; the number of training rounds ranges from 50 to 300, with the early stopping condition being that the validation set loss does not decrease for 20 consecutive rounds; after each training round, the precision and recall of vertex retention are statistically analyzed on the validation set, and the harmonic mean of the two is used as the model selection metric.

[0056] The technical advantages of the simplified complex-domain-wavelet-multi-scale attention graph convolution model are as follows: This model encodes the spatial coordinates and manifold trends of the 3D mesh into the modulus and phase angle of the complex domain, enabling graph convolution to simultaneously capture geometric position information and surface topological orientation information during spectral computation. This overcomes the shortcomings of traditional real-domain graph convolution in representing non-manifold geometry. The 3D discrete wavelet transform decouples the graph signal into high-frequency detail and low-frequency macroscopic features, allowing the attention mechanism to independently evaluate the retention value of each region at different frequency scales, avoiding the problem of interference between detail and macroscopic features at a single scale. The dynamic gating iteration mechanism automatically reduces the number of iterations to save computational resources when processing geometrically simple components, and automatically increases the number of iterations to ensure accuracy when processing geometrically complex components. This achieves dual adaptive allocation of computational resources at both the component and surface levels, thus achieving an overall optimal balance between accuracy and efficiency throughout the lightweight process.

[0057] The vertex dynamic weight adjustment function W is calculated based on the following four input quantities: the total number of current facets of the component. Global curvature mean Current available video memory (unit The ratio of the current iteration round to the maximum iteration round The expression for the vertex dynamic weight adjustment function W is: ,in , , , The weighting coefficients were determined through grid search experiments on the validation set. Each coefficient ranges from 0.1 to 0.5 and satisfies the following conditions: The maximum number of faces in the dataset. Total system video memory (unit) );when When the model reduces the number of convolutional kernel channels to 50% of the base number for the current batch and compresses the number of current CUDA streams to a minimum (no less than 1), it also increases the gating threshold of the dynamic gating unit, causing the model to exit the iteration earlier; when At this time, the model maintains the basic number of channels and the standard number of CUDA streams unchanged, and the gating threshold remains at its default value; when At this time, the model expands the number of convolutional kernel channels to 150%–200% of the basic number of channels and activates all available CUDA streams for parallel computation. Simultaneously, it lowers the gating threshold to allow the model to execute the maximum number of iterations, ensuring that the geometric features of complex components are fully extracted. The vertex dynamic weight adjustment function W establishes a linear mapping relationship with the dynamic discretization step size parameter in step S05. When W is large, the sampling step size is reduced; when W is small, the sampling step size is increased, thereby unifying the model's resource allocation decisions and discretization accuracy control into the same adjustment framework.

[0058] The “outer envelope geometry” refers to the closed curved surface geometry of the outermost layer of the precast component that can be observed from an external perspective after removing the internal steel reinforcement skeleton and embedded structures such as pre-embedded pipelines. The relationship between the outer envelope geometry and the removed internal structure is preserved through attribute link index to support subsequent attribute queries.

[0059] The geometry-attribute integration encapsulation refers to the unified integration of the simplified lightweight geometric data file and the access interface address of the attribute data in the distributed graph database into the header index area of ​​the lightweight BIM model file through a globally unique identifier. This allows downstream applications to directly and asynchronously retrieve attribute data on demand based on the globally unique identifier when loading the lightweight BIM model file, without the need to preload all attributes.

[0060] The high-dimensional semantic vector refers to the text encoding of the material name, component category name, and manufacturer specification description of the prefabricated component into a dense real number vector of fixed dimension (the dimension range is 128 to 512, determined through multiple rounds of experiments on the material similarity retrieval task) through a pre-trained language model. This vector is used to introduce semantic prior information into the complex domain-wavelet-multi-scale attention map convolutional simplified model.

[0061] Furthermore, the present invention also provides a lightweight processing system for BIM models of large and complex prefabricated building components based on computer implementation. As an optional implementation of the above method, the system is provided with a readable storage medium in the computer, which stores program instructions. When the program instructions are run in the computer, they call a distributed graph database, a CUDA parallel computing unit, and a microservice parallel processing mechanism to execute all steps S01 to S06 of the above method.

[0062] The specific implementation of step S01 is as follows: For each prefabricated component in the BIM database, its branch geometric features (such as main slab surfaces, bosses, tenons, reserved openings, etc.) are used as nodes, and the spatial adjacency relationships and Boolean operation relationships (merge, subtract, intersect) between each branch geometry are used as directed edges to construct an attribute-geometry composite graph structure. The normalized graph Laplacian matrix is ​​calculated for the attribute-geometry composite graph of each component. And perform eigenvalue decomposition to extract the eigenvalues. eigenvalues ​​( The spectral feature vector is constructed using values ​​ranging from 8 to 32. The Laplacian spectral feature is invariant to node number permutations, allowing it to express global structural differences among components regardless of their storage order. The spectral feature vector is normalized by concatenating the component's 3D geometric center Euclidean distance, bounding box volume ratio aligned with its outer envelope axis, and surface area. This is then mapped to a fixed-length key value using a locality-sensitive hash function for rapid retrieval of suspected identical components. For component pairs with identical hash keys, a generalized graph matching algorithm is used to calculate the maximum common subgraph node coverage. Components with a coverage rate higher than the similarity threshold (reference value 0.92) are identified as identical instantiated components. For identical components, one parent geometric data is retained; other instances only store a 4×4 affine transformation matrix and a lightweight pointer to the parent, thus eliminating scene-level geometric redundancy.

[0063] The specific implementation of step S02 is as follows: For the instantiated and merged parent prefabricated components, a curvature adaptive-semantic feature joint discretization model is used to discretize the parametric geometry into triangular meshes. First, the principal curvature of each point on the component surface is calculated. For areas where the absolute curvature value exceeds the curvature threshold, as well as key connection parts such as tenons and reserved openings, a smaller triangle step size (reference range 0.5-5mm) is used for dense sampling. For flat areas where the absolute curvature value is below the curvature threshold, as well as non-visible internal surfaces, a larger step size (reference range 20-200mm) is used for extremely sparse sampling. Semantic type determination is based on the category label of the prefabricated components. For semantic categories such as frame columns and irregular staircases, the step size threshold of key areas is automatically reduced to ensure the geometric fidelity of key details. For each type of parametric component after processing, the discretized triangular mesh template is stored in an in-memory hash cache table with the component type identifier as the key. Subsequent components of the same type directly read the cache template and perform parameter scaling, without repeating geometric analysis calculations, significantly reducing the time consumption of batch processing.

[0064] The specific implementation of step S03 is as follows: For the discretized parent prefabricated component mesh, firstly calculate its axis-aligned bounding box, and recursively divide the bounding box into eight equal parts using an octree structure until the leaf node voxel side length is less than the minimum voxel side length threshold (reference range 1-10mm, recommended value 5mm), with the number of division layers controlled between 4 and 8. Subsequently, virtual cameras are evenly arranged on the outer sphere of the component (reference range 32-256, recommended value 128), and each virtual camera emits a ray beam covering the entire component, recording the first hit face. Traversing all faces, faces not first hit by any virtual camera ray are marked as absolutely invisible meshes. Dynamic topology stripping technology directly deletes the vertices, faces, and their index references of the above-mentioned absolutely invisible faces from the mesh data structure, and synchronously updates the face adjacency list to maintain the topological connectivity of the outer envelope geometry. The attribute link indexes corresponding to the removed internal steel reinforcement skeleton mesh and the embedded pipeline mesh are retained separately to ensure that the attribute data is not lost during geometric stripping. The relationship between the outer envelope geometry and the removed internal structure is maintained through the attribute link index.

[0065] The specific implementation of step S04 is as follows: All attribute data, such as material, manufacturer, and mechanical parameters, of the prefabricated components are extracted from the BIM model master file and written into a distributed graph database as a graph data model. Attribute nodes and attribute relationships are stored in a graph structure, supporting high-concurrency read and write operations. The outer envelope geometric mesh data after the stripping in step S03 is stream-compressed using an extended format based on the Draco geometric compression protocol, supporting segment-by-segment decompression and on-demand transmission. For each prefabricated component, a 128-bit globally unique identifier is generated according to the RFC 4122 standard. This identifier serves as the unique decoupling association key between the geometric compression file and the corresponding attribute data node in the distributed graph database. Tasks such as attribute data read / write, geometric data compression, and globally unique identifier allocation are split into independent microservice processes, asynchronously scheduled through message queues. Each microservice runs on different computing nodes, eliminating single-point memory input / output blocking and achieving asynchronous on-demand loading of geometry and attributes.

[0066] The specific implementation of step S05 is as follows: Taking the discretized triangular mesh output from step S02 as input, the three-dimensional coordinate components of the vertices are mapped to the modulus of complex numbers, and the manifold trend angle of the surface where the vertex is located is mapped to the phase angle of complex numbers, resulting in a complex domain vertex feature matrix. This allows the graph convolution to carry both geometric position and surface orientation information simultaneously in the spectral operation. A three-dimensional discrete wavelet transform is applied to the complex domain vertex feature matrix to decompose the graph signal into a high-frequency geometric detail feature sub-band and a low-frequency macroscopic topological feature sub-band. The two sub-bands enter independent complex domain graph convolution branches, with the number of convolution kernels ranging from 32 to 128. Complex multiplication rules complete message passing and neighborhood aggregation in the complex spectral domain. The two sub-bands converge in the multi-scale attention fusion module. The dynamic gating unit calculates the geometric complexity score based on the local variance of the curvature matrix and the semantic vector entropy, with a score threshold ranging from 0.2 to 0.8, determining whether the current round of feature map exits the iteration. The maximum number of iterations ranges from 2 to 8. The attention mechanism dynamically calculates the weights of the high-frequency sub-bands in each round. With low-frequency subband weights (satisfy And weighted fusion. Vertex dynamic weight adjustment function. Considering four dimensions—total number of faces, average global curvature, available GPU memory, and iteration progress—when When the value is below 0.3, the number of convolutional kernel channels is compressed and the iteration is terminated early. Maintain the baseline parameter when it is between 0.3 and 0.6. If the number of channels is not less than 0.6, expand the number of channels and activate all CUDA streams to perform the maximum iteration. Simultaneously, a linear mapping relationship is established with the discretization step size parameter in step S02, incorporating resource allocation decisions and discretization accuracy control into a unified adjustment framework. The model outputs the retention weights and CUDA flow allocation proposal vectors for each vertex.

[0067] The specific implementation of step S06 is as follows: Based on the vertex weight allocation scheme output in step S05, vertices and their related faces with retention weights lower than the retention weight threshold are removed from the mesh to obtain an optimized and simplified mesh. The simplified lightweight geometric data file and the access interface address of the attribute data in the distributed graph database are uniformly integrated and written into the header index area of ​​the lightweight BIM model file using a globally unique identifier. When downstream applications load the lightweight BIM model file, they can directly retrieve attribute data asynchronously on demand based on the globally unique identifier without preloading all attributes, thereby completing the geometry-attribute integration encapsulation and outputting the final lightweight BIM model file.

[0068] It should be noted that the key technologies of this invention include: Attribute-geometric composite graph structure combined with graph Laplacian spectrum features, enabling permutation invariance in the identification of identical instances, accurately merging massive amounts of duplicate components without being affected by storage order, thus reducing the computational scale of subsequent steps from the root; Curvature adaptive-semantic feature joint discretization combines curvature perception with semantic priors, automatically increasing sampling density in geometrically dense areas and decreasing it in redundant areas, preserving key details while controlling surface expansion; Complex domain-wavelet-multi-scale attention graph convolutional simplified model captures both geometric position and surface orientation information through complex encoding, and wavelet transform decouples the graph signal into different frequency subbands for feature extraction, avoiding interference between features at different scales; dynamic gating mechanism allows the computational depth to automatically scale with geometric complexity. These three key technologies work synergistically under the unified adjustment of the vertex dynamic weight adjustment function, allowing the geometric redundancy elimination effect to directly reduce the input scale of the simplified model, and the resource awareness capability of the simplified model to adjust the discretization accuracy. Together, they ensure the overall optimal balance between accuracy and efficiency in large-scale scenarios, a synergistic effect that no single technology can achieve independently.

[0069] It should be noted that in large-scale building BIM projects, the geometric complexity of prefabricated components is extremely unevenly distributed. Standard rectangular wall panels have simple geometry, while irregularly shaped prefabricated stairs or complex connectors contain a large number of free-form surfaces and small features. In this case, if a uniform and fixed mesh reduction ratio is used to reduce the number of faces of all components in the scene, a dilemma will arise: complex components will suffer from severe feature loss, while simple components will have too many redundant faces. The reason for this technical problem is that traditional mesh reduction algorithms (such as the quadratic error metric method) only rely on local geometric error scores when calculating the folding cost of each vertex. They cannot perceive the semantic type of the current component, the overall geometric complexity level, or the available computing resources of the current system. This results in a lack of adaptive adjustment capability in the whole-scene reduction decision. Complex components and simple components are allocated computing resources of the same granularity, which wastes the computing on simple components and is insufficient to support the fine reduction of complex components. The usual solution to the above technical problem is to manually group components according to complexity in advance, set different reduction ratio parameters for different groups, and then execute the reduction algorithm in batches. However, this approach relies on engineers' experience and judgment, resulting in coarse-grained grouping and an inability to dynamically respond to changes in system resources during processing. When available GPU memory drops sharply due to other tasks, continuing to execute with the original parameters can lead to memory overflow or computational blockage. This invention effectively solves this technical problem. The vertex dynamic weight adjustment function collects information in real-time from four dimensions in each iteration: the total number of faces, the global mean curvature, the current available GPU memory, and the iteration progress. It dynamically calculates the comprehensive weight value and adjusts the number of convolutional kernel channels, the number of CUDA streams, and the gating threshold accordingly. This allows complex components to automatically obtain more computational resources and deeper iterations, while simple components automatically reduce computational depth and exit early. Furthermore, this function links resource allocation decisions and discretization precision control through a linear mapping relationship and the discretization step size parameter, unifying them into a single adjustment framework. This ensures a real-time optimal balance between accuracy and efficiency in real-world engineering environments with dynamically changing system resources, without requiring manual intervention in grouping or parameter adjustments.

[0070] Specifically, the principle of this invention is as follows: This invention solves the aforementioned core technical problems by incorporating three key aspects—geometric redundancy elimination, precision adaptive discretization, and resource-aware grid simplification—into a unified, collaborative adjustment framework. First, the attribute-geometric composite graph structure uniformly expresses the geometric shape and attribute dependencies of components as a graph data structure. The spectral characteristics of the graph Laplacian matrix are permutationally invariant to the global structure of the graph. Therefore, without being affected by the component storage order, it can accurately identify identical instances in a large-scale scene using a generalized graph matching algorithm, merging duplicate components into a parent geometry and affine transformation pointer system. This eliminates redundant storage at the geometric level from the source, ensuring that the total computational load of subsequent steps is proportional to the number of component types in the scene, rather than the total number of components. Secondly, the curvature adaptive-semantic feature joint discretization model dynamically adjusts the triangular mesh step size based on the local curvature distribution. The logic is that regions with higher curvature have higher geometric information density, requiring denser sampling to preserve features, while flat regions have lower geometric information density, allowing for sparse sampling to ensure accuracy. Semantic type constraints further ensure sufficient sampling density for key engineering-significant connection parts such as tenons and pre-reserved openings, while a caching mechanism avoids redundant analytical calculations of similar components. Subsequently, octree multi-level spatial partitioning combined with virtual camera ray projection identifies internally invisible meshes from a visibility perspective. Dynamic topology stripping technology removes internal steel reinforcement and pipeline meshes while maintaining the integrity of the outer envelope topology, further reducing the input scale of subsequent simplified models. The core complex domain-wavelet-multi-scale attention graph convolution simplified model encodes vertex spatial coordinates as complex moduli and surface manifold trend angles as complex phase angles, enabling graph convolution to carry both geometric position and topological direction information in spectral operations, thus overcoming the expressive limitations of real domain graph convolution. The 3D discrete wavelet transform decouples the graph signal into a high-frequency detail subband and a low-frequency macroscopic subband, enabling the attention mechanism to independently evaluate retention value at different frequency scales and avoiding mutual interference between detail features and macroscopic features. The vertex dynamic weight adjustment function integrates four dimensions—total number of faces, global curvature mean, available memory, and iteration progress—and establishes a linear mapping relationship with the discretization step size parameter. This allows the model's resource allocation decisions and discretization accuracy control to operate collaboratively within the same adjustment framework, thereby achieving an overall optimal balance between accuracy and efficiency in large-scale scenes with significant differences in geometric complexity.

[0071] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0072] The specific implementation method of step S01 is as follows.

[0073] Traverse all prefabricated components in the BIM database, extract the set of branch geometric feature nodes and the set of directed edges with spatial adjacency and Boolean operations for each component, and construct an attribute-geometry composite graph. ,in For a set of nodes, Let be the set of directed edges. The normalized graph Laplacian matrix is ​​constructed for the composite graph, expressed by the following formula:

[0074] ;

[0075] In the formula, For the normalized graph Laplacian matrix, It is an adjacency matrix. For an angle matrix, its diagonal elements , representing a node The degree. To Perform eigenvalue decomposition and take the first eigenvalue. eigenvalues ​​( The value range is 8 to 32, determined by conducting multiple rounds of offline experiments on BIM projects of different sizes and statistically analyzing the convergence interval of the eigenvalues ​​to form the spectral eigenvector. ,in ( )for The Small eigenvalues. High-dimensional invariant hash keys are generated by concatenating four types of normalized quantities and then performing locality-sensitive hashing. The four types of quantities are spectral feature vectors. Euclidean distance between the three-dimensional geometric centers of the components Outer envelope axis aligned with bounding box volume ratio and component surface area The formulas for splicing and mapping are expressed as follows:

[0076] ;

[0077] In the formula, For fixed-length hash keys, , , , These are the results after max-min normalization for the corresponding four types of quantities. A locality-sensitive hash function maps high-dimensional vectors to fixed-length codes to support fast retrieval. For suspected homogeneous component pairs with similar hash keys, a generalized graph matching algorithm is used to calculate the maximum common subgraph. The node coverage formula is expressed as follows:

[0078] ;

[0079] In the formula, For node coverage, For the set of nodes of the maximum common subgraph of the two component attribute-geometric composite graphs, and These are the node sets of the two components, respectively. , , These represent the number of nodes in the corresponding set. When If the similarity exceeds a threshold (ranging from 0.85 to 0.98, determined through receiver operating characteristic (ROC) analysis on manually labeled datasets of multiple typical prefabricated component families), the two are considered identical instantiations. Among identical components, one parent geometry is retained, and the remaining instances are simplified to 4×4 affine transformation matrices. The affine transformation matrix formula for mapping to the parent pointer is expressed as follows:

[0080] ;

[0081] In the formula, ( ) represents the rotation and scaling components. , , For translation components along the three coordinate axes, non-parent instances only store... Instead of storing geometric data separately, the parent pointer eliminates scene-level geometric redundancy.

[0082] The specific implementation method of step S02 is as follows.

[0083] Perform curvature adaptive-semantic feature joint discretization on the parametric geometry of precast components, based on local curvature. Dynamically assign triangular mesh sampling step size with component semantic type The step size allocation formula is expressed as follows:

[0084] ;

[0085] In the formula, For the position of the curved surface Sampling step size (in mm) at the location. The minimum step size (unit: mm, range: 0.5–5 mm, determined through geometric error comparison experiments) is used. The maximum step size (unit: mm, range: 20–200 mm, determined through geometric error comparison experiments) is used. For position Local curvature at (units) ), The semantic key part indicator is 1 (for key connection parts such as tenons and grooves and reserved openings, take 1; for other areas, take 0). This is the curvature adjustment coefficient (unit: mm, empirical value: 10 mm, making the denominator dimensionless). This is a semantic adjustment coefficient (dimensionless, empirically valued at 5). For high-frequency parameterized component types, the discretized mesh template is stored in a hash cache table with the component type identifier as the key and the corresponding discretized triangular mesh template as the value. When the same type of component re-enters the process, it is directly read from the cache and scaled according to the parameters, avoiding repeated geometric analysis.

[0086] The specific implementation method of step S03 is as follows.

[0087] An octree was used to partition the prefabricated components into multiple layers. The axis-aligned bounding box was recursively divided into eight equal parts, with the number of layers ranging from 4 to 8. The minimum voxel side length of the leaf nodes ranged from 1 to 10 mm, determined by the optimal combination of internal mesh recognition rate and computation time through ray casting experiments. 32–256 virtual cameras were evenly distributed on the outer sphere (the number was determined through a miss rate statistical experiment). Ray beams covering the entire component were emitted from each camera position. The first facet hit by each ray was counted. Facets not hit by any ray were marked as absolutely invisible meshes. Subsequently, a dynamic topology stripping technique was used to delete the corresponding vertices, faces, and index references from the mesh data structure, and the face adjacency list was updated. Simultaneously, the attribute link indexes between the removed internal structure and the outer envelope geometry were retained, ensuring that attribute data was not lost during geometry stripping.

[0088] The specific implementation method of step S04 is as follows.

[0089] The massive attribute data of prefabricated components is extracted and stored in a distributed graph database. Geometric data is stored using a streaming compression format based on the Draco protocol extension. The geometric data and attribute data are decoupled and associated through a 128-bit globally unique identifier generated according to the RFC 4122 standard. Tasks such as attribute reading and writing, geometry compression, and identifier allocation are split into independent microservice processes, which are asynchronously scheduled through message queues to achieve parallel execution of multiple tasks and eliminate single-point memory input / output blocking.

[0090] The specific implementation method of step S05 is as follows.

[0091] The model input layer receives the vertex coordinate matrix. ( (Total number of vertices), adjacency matrix Local curvature matrix and material high-dimensional semantic vector (dimension) The dimension range is 128–512 (determined through multiple rounds of experiments on the material similarity retrieval task). The input layer converts vertex coordinates into a complex number domain representation, and the conversion formula is as follows:

[0092] ;

[0093] In the formula, For the first Complex field characteristics of vertices, For the first The three-dimensional coordinate vector of each vertex. for The length of the Euclidean mold, For the first The angle (in rad) of the manifold trend at each vertex. Using the imaginary unit, we obtain the characteristic matrix of the vertex of the complex field. Subsequently, on Applying a three-dimensional discrete wavelet transform decomposes the graph signal into high-frequency subbands. With low-frequency subband The two sub-bands enter independent complex-domain graph convolution branches respectively. The message passing formula for complex-domain graph convolution is expressed as follows:

[0094] ;

[0095] ;

[0096] In the formula, This is the current convolutional layer number. The adjacency matrix after adding self-loops, It is the identity matrix. for The corresponding angle matrix, and These are the high-frequency and low-frequency branches, respectively. Complex convolution kernel matrix, For the first The number of convolutional kernels ranges from 32 to 128 for layer feature dimensions (determined through accuracy-time trade-off experiments on multiple BIM scenarios). For complex domain activation functions, complex multiplication performs message passing and neighborhood aggregation in the complex spectral domain. The two feature streams are fused in the multi-scale attention fusion module, and the attention weight calculation formula is as follows:

[0097] ;

[0098] In the formula, and These are the attention weights for high-frequency and low-frequency sub-bands, respectively, satisfying... , and The scalar score, obtained after linear transformation of the corresponding feature map, is used to measure the relative importance of high-frequency and low-frequency features in the current iteration. The geometric complexity score formula is expressed as follows:

[0099] ;

[0100] In the formula, Scoring the geometric complexity (dimensionless). Curvature matrix The local variance, semantic vector Information entropy and These are weighting coefficients (empirical values ​​are 0.5 for each). The value range is 0.2 to 0.8. The threshold was determined through multiple rounds of training convergence experiments on different types of components such as irregularly shaped prefabricated stairs and standard rectangular wall panels. The dynamic gating unit is based on... This determines whether the current feature map should proceed to the next iteration. The formula for the vertex dynamic weight adjustment function is as follows:

[0101] ;

[0102] In the formula, This is the dynamic weight adjustment amount for the vertex (dimensionless). The mean global curvature of the component (normalized to dimensionless). The total number of faces in the current component. The maximum number of faces in the dataset. This represents the current available video memory (in GB). Total system video memory (in GB). It is the ratio of the current iteration round to the maximum iteration round (dimensionless, ranging from 0 to 1). , , , These are the weighting coefficients (each value ranges from 0.1 to 0.5, satisfying...). (Determined through a grid search experiment on the validation set). At this point, the number of convolutional kernel channels is reduced to 50% of the basic number of channels, the number of CUDA streams is compressed to a minimum (no less than 1), and the gating threshold is increased to facilitate early exit from the iteration; when At the same time, maintain the basic number of channels and the standard number of CUDA streams, and keep the gating threshold at its default value; when At this point, the number of convolutional kernel channels is expanded to 150%–200% of the base number, activating all available CUDA streams, and the gating threshold is lowered to execute the maximum number of iterations, ensuring that the geometric features of complex components are fully extracted. The maximum number of internal iterations ranges from 2 to 8. Adaptive decision. Dynamic discretization step size parameter and A linear mapping relationship is established, expressed by the following formula:

[0103] ;

[0104] In the formula, For the The adjusted sampling step size (in mm) is... When the sample size is large, reduce the sampling step size. When the sampling step size is small, the model resource allocation decision and discretization accuracy control are unified into the same adjustment framework. The model output layer retains the weights for each vertex. and CUDA flow allocation proposal vector ,in The number of available CUDA streams. By fully connected subnetwork , Semantic category encoding and The input is predicted, and each component of the vector corresponds to a workload proportion of a CUDA stream. The total loss function is expressed as follows:

[0105] ;

[0106] In the formula, For the total loss, Binary cross-entropy loss that preserves the weights of the vertices. Preserve the weight vectors for the vertices output by the model. Preserve the weight truth vector (with values ​​of 0 or 1) for vertices manually annotated by professional BIM engineers. Mean squared error loss for assigning proposal vectors to CUDA streams. For the corresponding truth vector, The balancing coefficient for the two types of losses (ranging from 0.1 to 0.5, determined through multiple rounds of ablation experiments on the validation set). Training uses the Adam optimizer, with an initial learning rate ranging from [value missing]. ~ The model is dynamically decayed using a cosine annealing strategy. The batch size ranges from 8 to 32, and the number of training rounds ranges from 50 to 300. The early stopping condition is that the validation set loss does not decrease for 20 consecutive rounds. The harmonic mean of precision and recall is used as the model selection metric.

[0107] The specific implementation method of step S06 is as follows.

[0108] Based on the vertex retention weight scheme output in step S05, the lightweight BIM model of the entire scene is subjected to geometry-attribute integration and encapsulation. The simplified lightweight geometric data file and the attribute data access interface address in the distributed graph database are uniformly integrated and written into the header index area of ​​the lightweight BIM model file through a globally unique identifier, establishing the final mapping relationship between geometric data and attribute data. This allows downstream applications to directly and asynchronously retrieve attribute data on demand based on the globally unique identifier when loading the lightweight BIM model file, and output the lightweight BIM model file.

[0109] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: This example takes the BIM model of a large-scale prefabricated residential building project as the processing object. The project includes standard rectangular exterior wall panels, composite floor slabs, irregular prefabricated stairs, frame columns, frame beams and various types of connectors. The total number of prefabricated components is approximately 18,000. The original BIM model file is huge, and there are obvious memory overruns and response delays when directly loaded into the engineering visualization platform. The lightweight processing method of this invention needs to be used for processing.

[0110] First, execute step S01, traversing the prefabricated components in the BIM database. For each component, establish an attribute-geometric composite graph structure, using branch geometric features (main slab surface, bosses, tenons, reserved openings) as nodes and spatial adjacency relationships and Boolean operations as directed edges. Calculate the normalized Laplacian matrix for each component's attribute-geometric composite graph, and take the first 16 eigenvalues ​​to form the spectral eigenvector. Taking 16, after normalizing the Euclidean distance of the 3D geometric center, the bounding box volume ratio, and the surface area, it is mapped to a 128-bit key value using a locality-sensitive hash function. For component pairs with the same key value, the maximum common subgraph node coverage rate is calculated using a generalized graph matching algorithm, with a similarity threshold set to 0.92. The recognition results show that there are 127 independent geometric parent types in the components, and the remaining duplicate instances are simplified to a 4×4 affine transformation matrix and a parent pointer mapping. After instantiation and merging, only the 127 parent geometry types need to be processed, significantly reducing the scale of geometric processing.

[0111] Then, step S02 is executed, where triangular mesh discretization is performed on the 127 types of parent components using a curvature adaptive-semantic feature joint discretization model. High curvature regions (where the absolute value of the principal curvature exceeds the curvature threshold of 0.5) are then considered. For critical connection points such as mortise and tenon joints and reserved openings, a 1.5mm step size is used for dense sampling. Flat areas and non-visible internal surfaces are sampled with an 80mm step size for extremely sparse sampling. Semantic type determination is based on component category labels, automatically reducing the step size to 1.0mm for critical areas of irregularly shaped prefabricated stairs. After processing, the discretized mesh templates of 127 parent types are written to an in-memory hash cache table. For components of the same type, the cache is directly read and parameters are scaled to avoid redundant geometric parsing.

[0112] In step S03, each type of parent component is partitioned into an octree space based on an axis-aligned bounding box, with a minimum voxel side length threshold of 5mm and a partitioning layer of 6 layers. 128 virtual cameras are evenly distributed on the outer spherical surface, each emitting a ray beam covering the entire component, and the ray hit rate of each facet is statistically analyzed. After ray projection identification by the virtual cameras, a large number of faces in the internal steel reinforcement skeleton mesh and the embedded pipeline mesh are marked as absolutely invisible meshes. These faces and their vertex indices are directly deleted from the mesh data structure using dynamic topology stripping technology, the facet adjacency table is updated, and the attribute link indexes of the removed internal structures are saved separately to ensure the integrity of the attribute data. The number of faces before and after stripping for each parent component is shown in Table 1.

[0113] Table 1 Comparison of the number of back pieces before and after dynamic topology stripping of typical parent components

[0114]

[0115] In step S04, the material, manufacturer, and mechanical parameters of all prefabricated components are extracted to a distributed graph database, where attribute nodes and relationships are stored using a graph data model. The outer envelope geometric mesh data is stream-compressed using an extended format based on the Draco geometric compression protocol. A 128-bit globally unique identifier is generated for each prefabricated component according to RFC 4122, serving as the unique association key between the geometric compression file and the attribute nodes in the distributed graph database. The three tasks—attribute reading / writing, geometric compression, and globally unique identifier allocation—are each split into independent microservice processes, asynchronously scheduled via message queues, with each process running on an independent computing node.

[0116] In step S05, the discretized mesh output from step S02 is used as input to enter the simplified model of complex domain-wavelet-multi-scale attention map convolution. Vertex coordinate components are mapped to the modulus of a complex number, and the trend angle of the curved manifold is mapped to the phase angle of a complex number. After three-dimensional discrete wavelet transform, it is decomposed into a high-frequency geometric detail feature sub-band and a low-frequency macroscopic topological feature sub-band. Features are extracted from these two sub-bands through complex domain graph convolution branches (64 convolution kernels), and then merged in the multi-scale attention fusion module. The geometric complexity score of the irregular prefabricated staircase, calculated by the dynamic gating unit, is 0.71, exceeding the gating threshold of 0.6. The maximum number of iterations is set to 7, and the model outputs vertex weights after fully extracting complex geometric features. The geometric complexity score of the standard rectangular exterior wall panel is 0.22, below the gating threshold of 0.3. The number of convolution kernel channels is compressed to 50% of the basic number of channels, and the model exits early after 2 iterations to save computational resources. Vertex dynamic weight adjustment function. The values ​​of each weight coefficient are: , , , , When dealing with complex connectors, a value of 0.67 is used to activate parallel computation of all available CUDA streams. The vertex retention rate of various parent components after simplification and the number of meshes after mesh simplification are as follows: Figure 3 As shown. This step outputs the optimized simplified mesh and vertex weight allocation scheme.

[0117] In step S06, vertices and related faces with retention weights below the retention weight threshold of 0.35 are removed according to the vertex weight allocation scheme. The optimized and simplified geometric data and the attribute data access interface address of the distributed graph database are uniformly written into the header index area of ​​the lightweight BIM model file through a globally unique identifier, completing the geometry-attribute integration encapsulation and outputting the final lightweight BIM model file. Figure 2 As shown, the comparison of the total number of facets in the original model and the lightweight model across different component types intuitively demonstrates the simplification effect of this invention on various types of prefabricated components. When the downstream platform loads the lightweight BIM model file, it can directly and asynchronously retrieve attribute data on demand based on the globally unique identifier, without needing to preload all attributes, significantly reducing memory usage during initial loading.

[0118] From the perspective of technological advancement, the progress of this invention compared to traditional methods is reflected in the following aspects: Traditional methods use a uniform and fixed simplification ratio for all components in the scene, failing to perceive the geometric complexity of the components and the status of system resources. In contrast, the vertex dynamic weight adjustment function of this invention incorporates the total number of faces, the average curvature, the available memory, and the iteration progress into a unified function framework, enabling real-time adaptive simplification decisions. Complex components obtain sufficient computational depth to retain details, while simple components exit early to release resources, eliminating the inherent defects of fixed parameter schemes from the perspective of computational resource allocation mechanisms. The introduction of complex domain encoding and three-dimensional discrete wavelet transform enables graph convolution to simultaneously capture geometric position, surface direction, and multi-scale frequency information. Compared to the information loss in non-manifold geometry caused by traditional real domain graph convolution, this invention improves geometric fidelity from the perspective of the mathematical expression of feature extraction. The attribute-geometry decoupled storage and the asynchronous loading mechanism of globally unique identifiers eliminate the memory bottleneck caused by full attribute preloading under the traditional single-file structure from the data architecture perspective, making on-demand access to large-scale prefabricated component BIM models feasible in engineering practice.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for lightweighting BIM models of large and complex prefabricated building components, characterized in that, Includes the following steps: Traverse all prefabricated components in the BIM database, establish the attribute-geometric composite graph structure for each component, calculate the graph Laplacian matrix spectral characteristics, generate high-dimensional invariant hash key values, identify identical instantiated components through a generalized graph matching algorithm, retain the parent geometric data, and simplify the remaining duplicate components into affine transformation matrices and parent pointer mappings. A curvature adaptive-semantic feature joint discretization model is used to dynamically discretize the parametric geometry of prefabricated components. Intensive sampling is performed in high curvature regions and key connection parts, while extremely sparse sampling is performed in flat regions and non-visible internal surfaces. The discretization template of high-frequency parametric components is stored in a caching mechanism. An octree is used to divide the prefabricated components into multiple spatial layers. The absolutely invisible internal mesh is identified and marked from multiple angles by virtual camera ray projection. The internal steel reinforcement and pipeline mesh is removed by dynamic topology stripping technology, leaving only the outer envelope geometry and attribute links. The attribute data of prefabricated components is extracted and stored in a distributed graph database. The geometric data is stored in a streaming compressed format and decoupled and associated with a lightweight globally unique identifier. The asynchronous on-demand loading of geometry and attributes is achieved through a microservice parallel processing mechanism. Using the discretized mesh output by the curvature adaptive-semantic feature joint discretization model as input, the high-frequency geometric details and low-frequency macroscopic topological features of the prefabricated components are extracted by the complex domain-wavelet-multi-scale attention map convolutional simplification model. The number of internal iterations is adaptively determined according to the mesh complexity, and the optimal simplification mesh and vertex weight allocation scheme are output. Based on the vertex weight allocation scheme, the lightweight BIM model of the whole scene is encapsulated by geometry-attribute integration. The final mapping relationship between geometric data and attribute data in the distributed graph database is established through a globally unique identifier, and the lightweight BIM model file is output.

2. The method of claim 1, wherein the method is characterized by: The attribute-geometric composite graph structure is specifically a graph data structure constructed using the branch geometric features of prefabricated components as nodes and the spatial adjacency relationships and Boolean operation relationships between the various parts of the components as directed edges.

3. The method of claim 2, wherein the method further comprises: The simplified complex domain-wavelet-multi-scale attention graph convolutional model specifically involves the input layer receiving vertex coordinate matrices, adjacency matrices, local curvature matrices, and high-dimensional semantic vectors of material properties. The input layer maps vertex spatial coordinate components to complex moduli and surface manifold trend angles to complex phase angles, resulting in a complex domain vertex feature matrix. Subsequently, it undergoes three-dimensional discrete wavelet transform decomposition into high-frequency geometric detail feature sub-bands and low-frequency macroscopic topological feature sub-bands. These two sub-bands are then fed into independent complex domain graph convolutional branches for feature extraction, and finally merged in the multi-scale attention fusion module to output vertex retention weights and CUDA flow allocation proposal vectors.

4. The method of claim 3, wherein the method further comprises: The high-dimensional invariant hash key is generated by a combination of the following four types of quantities: the graph Laplace matrix before... The four types of quantities—a sequence of feature values, the Euclidean distance between the coordinates of the three-dimensional geometric center of the precast component, the volume ratio of the bounding box aligned with the outer envelope axis, and the surface area of ​​the precast component—are normalized and concatenated into a high-dimensional vector, which is then mapped to a fixed-length key value using a locality-sensitive hash function.

5. The method of claim 4, wherein the method further comprises: The generalized graph matching algorithm refers to a graph theory-based generalized graph matching algorithm used to calculate the maximum common subgraph between two attribute-geometric composite graphs; when the node coverage of the maximum common subgraph of two components is higher than the similarity threshold, they are determined to be instantiated components of the same type.

6. The method of claim 5, wherein the method further comprises: The affine transformation matrix refers to a 4×4 homogeneous coordinate transformation matrix used to express translation, rotation, and scaling; the parent pointer mapping refers to a lightweight reference structure in the BIM scene database that points to the location of the parent geometric data storage with a unique index.

7. The method of claim 6, wherein the method further comprises: The curvature adaptive-semantic feature joint discretization model is specifically an algorithm model that dynamically adjusts the sampling density of the triangular mesh based on the local curvature distribution and the semantic type of the component during the parametric geometric discretization process; a smaller triangle step size is used in high curvature regions and key connection parts, while a larger step size is used in flat regions and non-visible internal surfaces. 8.The method of claim 7, wherein the method further comprises: The caching mechanism refers to establishing a hash cache table in memory with the parameterized component type identifier as the key and the corresponding discretized triangular mesh template as the value; when the same type of parameterized component enters the discretization process again, the mesh template in the cache is directly read and the parameters are scaled. 9.The method of claim 8, wherein the method further comprises: The octree multi-level spatial partitioning refers to recursively dividing the prefabricated component into eight equal parts by aligning the axis of the prefabricated component with the bounding box, until the leaf node size is less than the minimum voxel side length threshold; the virtual camera ray projection refers to uniformly distributing several virtual camera positions on the outer spherical surface of the prefabricated component, emitting a ray beam covering the entire component at each position, recording the first patch hit by each ray, and marking the patches not hit by any ray as absolutely invisible meshes. 10.The method of claim 9, wherein the method further comprises: The dynamic topology stripping technique refers to directly deleting the vertices, faces, and their index references of absolutely invisible meshes from the mesh data structure while maintaining the topological integrity of the outer envelope geometry, and updating the face adjacency list, while retaining the attribute link index between the removed internal structure and the outer envelope geometry.

Images