Layered quantification method and system for BIM lightweight engine

Abstract

The invention provides a hierarchical quantification method and system for a BIM lightweight engine, and belongs to the technical field of building information model design. Comprising the following steps: carrying out structured analysis on a BIM model, and dividing the BIM model into a plurality of sub-components according to an internal hierarchical structure and component semantic information of the BIM model; carrying out quantitative sensitivity evaluation on each sub-assembly, and dividing quantitative levels; performing layered mixed quantization according to the quantization hierarchy to which each sub-component belongs; and the quantized geometric data and non-geometric data are separately stored, inverse quantization operation is executed in the early stage of pipeline rendering by using GPU computing power of a client, vertex coordinates and normal vectors are recovered in real time, and then rendering is performed by using the recovered geometric data. According to the method, the efficiency and the computing power of the large BIM model for building design in the storage, transmission and real-time rendering processes can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of building information modeling design technology, and particularly relates to a layered quantization method and system for a lightweight BIM engine. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Building Information Modeling (BIM), a three-dimensional model-based approach to architectural design, construction, and management, has been widely adopted in the construction industry. The core of BIM technology lies in its "living, data-rich architectural representation," integrating information from the entire lifecycle of a building project into a single three-dimensional digital model. However, this massive data integration also presents a major obstacle to its application. Modern building projects are enormous, with individual BIM model files easily reaching several gigabytes in size and tens of millions of triangles. Such a massive data volume poses a severe challenge to the graphics processing capabilities of computers.

[0004] Despite extensive research conducted by technical personnel in this field, existing BIM lightweighting methods for architectural design still suffer from the following technical shortcomings: (1) Existing technologies cannot fundamentally and efficiently compress and represent the underlying geometric data that occupies the majority of storage and bandwidth in BIM models. For example, existing technologies focus more on how the data is organized and how it is loaded on demand, but do not change the way the geometric data is represented as high-precision floating-point numbers. This "preservation as is" approach has significant data redundancy, which is the main technical bottleneck restricting further reduction in model file size and will affect the efficiency of BIM models in storage, transmission and real-time rendering.

[0005] (2) The optimization effect of existing technologies mainly relies on the powerful graphics computing power of the client, which cannot achieve a high-quality and smooth experience on ordinary commercial hardware or mobile devices. Whether it is GPU acceleration or rendering technology, it only uses more efficient hardware resources to "hard-decode" massive data, but does not solve the problem of excessive computing burden on the client caused by the large model file at the source. Although it improves the loading and rendering speed, it still requires powerful hardware as support. This makes it difficult to achieve high-fidelity, real-time interaction and rendering of large-scale BIM models on popular devices such as laptops and tablets with limited computing power. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention provides a layered quantization method and system for a lightweight BIM engine, which can effectively improve the efficiency and computing power of large BIM models used in architectural design during storage, transmission and real-time rendering.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a hierarchical quantization method for a lightweight BIM engine.

[0008] A hierarchical quantization method for a lightweight BIM engine includes: The BIM model is structurally analyzed and divided into multiple sub-components based on its inherent hierarchical structure and component semantic information; a quantitative sensitivity assessment is performed on each sub-component, and quantitative levels are defined. Based on the quantization level of each sub-component, a layered hybrid quantization is performed: a high-precision quantization strategy is adopted for sub-components with high sensitivity, and the quantization operation is simulated and the quantization parameters are optimized to minimize the precision loss; a non-uniform quantization strategy is adopted for sub-components with medium sensitivity, and the quantization precision is dynamically allocated based on the local density of the vertices inside the sub-component; a unified quantization strategy is adopted for sub-components with low sensitivity, and the vertex coordinates and normal vectors are linearly mapped to low-precision integers. The quantized geometric data is stored separately from the non-geometric data. The client's GPU computing power is used to perform dequantization in the early stages of the rendering pipeline to restore vertex coordinates and normal vectors in real time. The restored geometric data is then used for rendering.

[0009] Furthermore, the quantitative sensitivity assessment is achieved by analyzing the geometric complexity and semantic importance of each sub-component, namely: parsing the BIM model to obtain the semantic information of each sub-component and calculating the geometric complexity index of each sub-component; combining the semantic information and the geometric complexity index, a weighted calculation method is used to obtain a comprehensive quantitative sensitivity score.

[0010] Furthermore, the semantic information includes component name, type, and function; the geometric complexity index includes at least one of the following: number of vertices, number of faces, average face curvature, and topological irregularity.

[0011] Furthermore, the high-precision quantization strategy includes: inserting a pseudo-quantization operation during forward propagation to simulate the impact of low-precision arithmetic on geometric data; defining a loss function to minimize the difference between full-precision geometric data and pseudo-quantized data; and optimizing quantization parameters through backpropagation and a pass-through estimator to preserve key geometric details.

[0012] Furthermore, the non-uniform quantization strategy employs an iterative Lloyd-Max algorithm to dynamically allocate quantization precision, including: optimizing the quantization interval boundary and quantization value based on the distribution characteristics of vertices within the sub-component; and minimizing the quantization error by alternately updating the quantization interval boundary and quantization value.

[0013] Furthermore, the Lloyd-Max algorithm is implemented based on two mutually synergistic conditions: the optimal core condition and the optimal boundary condition.

[0014] Furthermore, the unified quantization strategy includes: using a single linear scaling factor to map vertex coordinates and normal vectors to low-precision integers; and determining the quantization range based on the model's axis-aligned bounding box boundaries.

[0015] A second aspect of the present invention provides a hierarchical quantization system for a lightweight BIM engine.

[0016] A hierarchical quantization system for a lightweight BIM engine, comprising: The model acquisition module is configured to: acquire the original BIM file into the system; The model parsing module is configured to perform structured analysis on the BIM model and divide it into multiple sub-components based on its internal hierarchical structure and component semantic information. The quantization sensitivity assessment module is configured to: perform quantization sensitivity assessment on each sub-component and divide the quantization levels; The hierarchical hybrid quantization module is configured to perform hierarchical hybrid quantization based on the quantization level of each sub-component: a high-precision quantization strategy is used for sub-components at high sensitivity levels, minimizing precision loss by simulating quantization operations and optimizing quantization parameters; a non-uniform quantization strategy is used for sub-components at medium sensitivity levels, dynamically allocating quantization precision based on the local density of vertices within the sub-component; and a unified quantization strategy is used for sub-components at low sensitivity levels, linearly mapping vertex coordinates and normal vectors to low-precision integers. The data storage module is configured to store the quantized geometric data and non-geometric data in different table structures in the database. The rendering module is configured to utilize the client's GPU computing power to perform dequantization operations in the early stages of the rendering pipeline, recovering vertex coordinates and normal vectors in real time, and then using the recovered geometric data for rendering. A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of a layered quantization method for a lightweight BIM engine as described in the first aspect of the present invention.

[0017] The fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a layered quantization method for a lightweight BIM engine as described in the first aspect of the present invention.

[0018] The above one or more technical solutions have the following beneficial effects: (1) This invention performs structured analysis on the BIM model, dividing it into multiple sub-components based on its inherent hierarchical structure and component semantic information; it evaluates the quantization sensitivity of each sub-component and classifies the quantization levels; through a hierarchical hybrid quantization method, it applies adaptive quantization strategies (high-precision quantization, non-uniform quantization, and unified quantization) to different types of components, achieving targeted bit depth reduction of massive vertex coordinates and normal vector data. This can directly reduce the inherent redundancy of geometric data from the data source, thereby fundamentally reducing the model file size and significantly improving the efficiency of BIM models in storage and network transmission.

[0019] (2) This invention reduces the model file size and decreases the amount of data transmitted over the network and the memory usage by using a layered hybrid quantization method. Simultaneously, the client rendering engine only needs to perform highly computationally efficient integer operations and lightweight dequantization operations to recover geometric data early in the rendering pipeline, thereby freeing the GPU from the burden of complex calculations involving high-precision floating-point numbers. This makes real-time interaction and rendering of large-scale, high-fidelity BIM models possible on ordinary consumer hardware such as laptops and tablets, breaking the dependence of BIM applications on professional high-end graphics workstations.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of a layered quantization method for a lightweight BIM engine according to Embodiment 1 of the present invention. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] The overall idea proposed in this invention is as follows: This invention provides a hierarchical quantization method for BIM lightweight engines. This method proposes a hierarchical hybrid framework that can adaptively select the optimal quantization strategy based on the geometric complexity and semantic importance of the BIM model, thereby improving the compression rate of the BIM model and solving the problems of large model files and high rendering computing power requirements in existing BIM lightweight technologies.

[0027] Example 1 This embodiment discloses a hierarchical quantization method for a lightweight BIM engine.

[0028] like Figure 1 As shown, a hierarchical quantization method for a lightweight BIM engine includes: Step S1: Perform a structured analysis on the BIM model, dividing it into multiple sub-components based on its inherent hierarchical structure and component semantic information; conduct a quantitative sensitivity assessment on each sub-component and classify the quantitative levels. Step S2: Perform layered hybrid quantization according to the quantization level of each sub-component: For sub-components with high sensitivity, a high-precision quantization strategy is adopted, which minimizes the precision loss by simulating quantization operations and optimizing quantization parameters; for sub-components with medium sensitivity, a non-uniform quantization strategy is adopted, which dynamically allocates quantization precision based on the local density of vertices within the sub-component; for sub-components with low sensitivity, a unified quantization strategy is adopted, which linearly maps vertex coordinates and normal vectors to low-precision integers. Step S3: Separate and store the quantized geometric data from the non-geometric data. Utilize the client's GPU computing power to perform dequantization in the early stages of the rendering pipeline to restore vertex coordinates and normal vectors in real time. Then, use the restored geometric data for rendering.

[0029] Based on the above process, this invention can effectively improve the efficiency and computing power of large-scale BIM models used in architectural design during storage, transmission, and real-time rendering. To facilitate understanding of the technical solution of this invention, the specific implementation methods of this invention will be further explained and described below.

[0030] In step S1, the BIM model is subjected to structural analysis and divided into multiple sub-components based on its inherent hierarchical structure and component semantic information; a quantitative sensitivity assessment is performed on each sub-component, and a quantitative level is determined.

[0031] The assessment of quantitative sensitivity is achieved by analyzing the geometric complexity and semantic importance of each sub-component. Specifically, the BIM model is parsed to obtain the semantic information of each sub-component, and the geometric complexity index of each sub-component is calculated. Combining the semantic information and the geometric complexity index, a weighted calculation method is used to obtain a comprehensive quantitative sensitivity score.

[0032] When performing semantic analysis on components, the BIM model is first parsed to obtain semantic information such as the component name, type, and function of each component. For example, key semantic tags such as "ventilation and air conditioning system," "valve," "pipeline connector," and "high-precision equipment" are identified. During geometric complexity assessment, for each component, its geometric complexity indices are calculated, such as the number of vertices, the number of faces, the average face curvature, and topological irregularity. High vertex count and high curvature typically represent high geometric complexity.

[0033] Finally, combining the semantic importance and geometric complexity of the components, a comprehensive quantization sensitivity score is calculated. Quantization Sensitivity Score The calculation can be achieved through the following methods: Each BIM component is evaluated using a comprehensive scoring model, and the scoring results are used to categorize it into different quantitative levels. Quantitative Sensitivity Scoring Geometric complexity index and semantic importance weight It consists of two core parts. The scoring model is specifically represented as follows: ; in, This represents a semantic importance weighting factor, used to amplify the sensitivity of key components. Geometric complexity metric. It is used to measure the spatial complexity of a component and is obtained by a weighted sum of normalized indices such as the number of vertices, the number of faces, and the average curvature of the component. ; in, Indicates the number of vertices of the component. Indicates the number of facets of a component. This represents the average curvature of all surfaces in the component; , and Let these be the weighting factors corresponding to the number of vertices, the number of faces, and the average curvature, respectively, satisfying... . This represents the normalization function, used to normalize indicators. Mapped to the range of 0-1, represented as ;in, and These represent the maximum and minimum values ​​of the corresponding indicators in the entire model, respectively.

[0034] Table 1. Sensitivity Level Classification Based on IFC Component Type

[0035] Semantic importance weight It is a preset discrete score based on IFC component type, as shown in Table 1: when the IFC component type is a key device such as HVAC, pump, or structural connector, its semantic importance weight is... The value is 1.0, corresponding to a sensitivity level of "extremely high"; when the IFC component type is a functional component or safety component such as doors and windows, escape indicators, or important pipelines, its semantic importance weight is... The value is 0.6, corresponding to a sensitivity level of "medium-high"; when the IFC component type is a basic enclosure component such as a wall, floor slab, beam, or column, its semantic importance weight is... The value is 0.2, which corresponds to a sensitivity level of "low".

[0036] Furthermore, the component classification and quantification hierarchy includes: calculating sensitivity scores. Then, based on preset experience thresholds and The components are divided into three quantification levels: 1) High sensitivity (corresponding to high-precision quantization layer): If The components are categorized as high-precision quantization layers. These components are typically geometrically complex "geometric outliers" with key semantics. For example, a ventilation and air conditioning system with a high number of vertices and complex surfaces, or a component labeled as "critical equipment".

[0037] 2) Moderate sensitivity (corresponding to non-uniform quantization layer): If The components are classified as non-uniform quantization layers. These components are typically of medium geometric complexity, such as curved exterior walls and irregularly shaped windows. Their data distribution is non-uniform, requiring fine-grained processing.

[0038] 3) Low sensitivity (corresponding to unified quantization layer): If The components are categorized into a unified quantification layer. These components are typically geometrically simple, regular, and numerous, such as walls and floor slabs.

[0039] In step S2, hierarchical hybrid quantization is performed according to the quantization level of each sub-component: a high-precision quantization strategy is adopted for sub-components with high sensitivity, and the quantization parameters are optimized by simulating quantization operations to minimize the loss of precision; a non-uniform quantization strategy is adopted for sub-components with medium sensitivity, and the quantization precision is dynamically allocated based on the local density of vertices inside the sub-component; a unified quantization strategy is adopted for sub-components with low sensitivity, and the vertex coordinates and normal vectors are linearly mapped to low-precision integers.

[0040] 1) High-precision quantization.

[0041] For geometrically complex or semantically crucial components, such as electromechanical components, the number of vertices, triangles, and curvatures are far greater than their precision compared to other building structures—these are known as "geometric outliers." Applying a uniform quantization strategy can lead to verification distortion during reconstruction of these components. Therefore, a high-precision quantization strategy similar to Quantization-Aware Training (QAT) is employed, fine-tuning representative geometric data to minimize precision loss while preserving critical details. This high-precision quantization strategy includes: inserting pseudo-quantization operations during forward propagation to simulate the impact of low-precision arithmetic on geometric data; defining a loss function to minimize the difference between full-precision geometric data and pseudo-quantized data; and optimizing quantization parameters through backpropagation and a pass-through estimator to preserve critical geometric details.

[0042] The core idea of ​​high-precision quantization strategy is to simulate the impact of quantization on the model without changing the model's full-precision weights. Its quantization function follows the uniform quantization formula: ; in, This represents the pseudo-quantized value. Indicates the quantization step size. The floating-point input value represents the original full-precision value; The function is used to restrict the input value to a certain range. Within the range, parameter values and The range of geometric data of the components is dynamically determined.

[0043] This invention simulates the QAT method through a fine-tuning process, which includes forward propagation, loss function quantization, backpropagation, and STE. Quantization-Aware Training (QAT) is a compression technique for large models in AI training. It introduces pseudo-quantization operations (simulating low-precision computation of the target hardware, such as integer INT8) during the training and fine-tuning stages of a large model, while maintaining high-precision weights. By approximating the gradient of the quantization function through a STE, combined with backpropagation and optimization of model parameters (especially quantization range / step size), the model can adapt and minimize accuracy loss when deployed to low-precision environments.

[0044] Forward propagation: This process takes the geometric data of the component to be quantized as input and inserts fake quantization operations into the computation graph. These operations simulate the effects of low-precision arithmetic on vertices and normals, while the computation itself still uses full-precision floating-point numbers. In practice, forward propagation takes the geometric data of the component to be quantized as input, embeds a fake quantization module in the computation graph, and executes the uniform quantization formula described above. The output is the fake quantization result, providing the data foundation for subsequent processes.

[0045] Loss function quantization: Define a loss function that aims to minimize the difference between full-precision geometric data and pseudo-quantized data. This loss function can be measured using mean squared error. The goal is to find the optimal quantization parameters that minimize the loss function, i.e.: ; in, Indicates the quantization loss value. Indicates expectation or average. A floating-point input value representing the original full-precision value.

[0046] Backpropagation and STE: In the backpropagation process, in order to overcome... To address the non-differentiability of the function, this invention employs a Straight-Through Estimator (STE), which approximates the quantization function as an identity function with a derivative of 1 when calculating the gradient. This allows the gradient to flow smoothly through the quantization layer, thereby enabling the full-precision weights of the model to be updated smoothly.

[0047] In end-to-end training of neural networks, the backpropagation algorithm relies on gradient calculation to update parameters. However, some key operators (such as those in quantization)... In binarization The non-differentiable or zero-gradient property is addressed by STE (Skimming the Gradient Approximation). Its core idea is to perform the actual computation during forward propagation, strictly calculating the non-differentiable operation according to the original function. During backpropagation, the mathematical derivative of the non-differentiable operator is ignored, and the gradient is directly "penetrated" through this layer, approximating it as: ; in, Represents the loss function. This represents the pseudo-quantized geometric data; 1 represents scalar 1 (i.e., the assumed...). The derivative is 1). Pretend that the discretization operation did not occur (as if x were directly passed to the output), allowing the gradient to flow back without obstruction.

[0048] 2) Non-uniform quantization.

[0049] For moderately complex components (such as curved surfaces and decorative components), a non-uniform quantization technique is employed. This technique dynamically allocates different quantization precipitates based on the local density of vertices within the component, thereby achieving efficient compression while maintaining visual fidelity. Specifically, the non-uniform quantization strategy uses an iterative Lloyd-Max algorithm to dynamically allocate quantization precision: optimizing the quantization interval boundaries and quantization values ​​based on the distribution characteristics of vertices within the sub-component; and minimizing quantization error by alternately updating the quantization interval boundaries and quantization values. Unlike the fixed-step uniform quantization used in existing technologies, the Lloyd-Max quantizer can be optimized based on the statistical distribution characteristics of the input signal, thus more effectively reducing distortion when processing non-uniformly distributed data.

[0050] When using the iterative Lloyd-Max algorithm to find the optimal quantization parameters, the goal is to find the optimal quantization interval division and quantization value given a quantization level, so as to minimize the overall error (mean squared error) between the quantized data and the original data. This algorithm is based on two interdependent conditions: the optimal core condition and the optimal boundary condition.

[0051] Optimal quantization condition: Given a set of quantization boundaries, the optimal quantized value must be the average of all points falling within that interval, i.e.: ; in, It is the first quantized values ​​for each interval, and It is the boundary of the interval. These are the original data points that fall within this interval. This is the total number of data points falling within that interval. This condition ensures that each quantized value best represents all data points within its respective interval, thereby minimizing local errors.

[0052] Optimal boundary condition: Given a set of quantized values, the optimal quantization boundary L j It must be the midpoint of its two adjacent quantization values, that is: ; in, It is the first The boundaries of each interval, and These are two adjacent quantization values. This condition ensures that the spatial division of each quantization interval is the most reasonable, thus avoiding unnecessary errors.

[0053] By iteratively alternating between these two conditions, the algorithm gradually converges to a local optimum, thereby generating the most suitable quantization scheme for non-uniformly distributed geometric data.

[0054] 3) Standardized quantification.

[0055] For simple, large-scale, and geometrically uniform components (such as walls and floors), the simplest and most efficient unified quantization strategy, namely the standard post-trained quantization (PTQ) method, is adopted to linearly map vertex coordinates and normal vectors to low-precision integers, thereby achieving maximum file size compression. The unified quantization strategy involves: using a single linear scaling factor to map vertex coordinates and normal vectors to low-precision integers; and determining the quantization range based on the model's axis-aligned bounding box boundaries.

[0056] This method uses a single linear scaling factor to uniformly map all floating-point numbers to low-precision integers, specifically including vertex coordinate quantization and normal vector quantization.

[0057] Vertex coordinate quantization, i.e.: ; ; Normal vector quantization, i.e.: ; in, and These are the minimum and maximum boundaries of the axis-aligned bounding box (AABB) of the model, respectively. It is the number of quantization levels; It is the quantization range of the normal vector. The quantization step size represents the vertex coordinates. This represents the quantized vertex coordinates. This method is fast and efficient enough to handle elements where the loss of a few decimal places has a negligible impact on vision and function, thus achieving maximum model size compression.

[0058] This invention employs layered hybrid quantization to thoroughly reduce the bit depth of the most space-consuming geometric data in BIM models, and adopts differentiated quantization strategies based on the characteristics of different components, significantly reducing model file size. Simultaneously, it creatively applies the "outlier awareness" concept from large model quantization to BIM models, ensuring that critical and complex components maintain high accuracy after quantization, thus solving the detail loss problem caused by uniform quantization methods in existing technologies. The significant reduction in model file size directly reduces the data bandwidth required for streaming transmission, improving loading speed, especially in mobile scenarios with unstable network environments.

[0059] In step S3, the quantized geometric data and non-geometric data are stored separately, forming a proprietary, lightweight data format suitable for efficient streaming. The client's GPU computing power is used to perform dequantization in the early stages of the rendering pipeline, restoring vertex coordinates and normal vectors in real time. Subsequently, the restored geometric data is used for rendering. Thus, the client's rendering engine only needs to perform simple integer operations and dequantization, significantly reducing its reliance on memory bandwidth and CPU / GPU computing power, enabling smooth real-time rendering of large BIM models on low-end consumer hardware.

[0060] The quantized geometric data and non-geometric data are stored separately, forming a proprietary, lightweight data format suitable for efficient streaming transmission. In practice, the quantized geometric data (such as vertex coordinates and normal vectors in low-precision integer format) and non-geometric data (such as semantic attributes like component names, types, and functions) are stored in different table structures within the database. Simultaneously, the two types of data are encapsulated according to proprietary rules to form a lightweight format adapted for efficient streaming transmission, reducing transmission redundancy.

[0061] The client's GPU computing power is utilized to perform dequantization operations early in the rendering pipeline, restoring vertex coordinates and normal vectors in real time. In practice, the client GPU performs lightweight dequantization calculations early in the rendering pipeline (e.g., before vertex shading), based on the inverse logic of the corresponding quantization strategy, to restore low-precision integers to floating-point format vertex coordinates and normal vectors in real time. As an optional implementation, the following unified quantization form can be used as the inverse logic of a quantization strategy: ; in, Represents the original coordinates. Indicates quantized coordinates, Indicates the quantization step size. Indicates the lower limit of the value.

[0062] The recovered geometric data is used for rendering. Specifically, the floating-point geometric data obtained by dequantization is connected to subsequent stages of the rendering pipeline (such as vertex shading, rasterization, and fragment shading) to complete the real-time rendering of the model according to the interaction and display requirements of the BIM model.

[0063] This invention creatively applies the concepts of "outlier awareness" and "mixed precision" from deep learning model quantization to the geometric data of BIM models. Simultaneously, it treats the quantization processing of vertex coordinates and normal vectors as a unified process, but the choice of quantization strategy is based on the type and importance of the components. This effectively solves the problems of geometric data redundancy and rendering overhead in the field of lightweight BIM. By creatively applying the "quantization" principles from deep learning model compression, particularly the ideas of outlier handling and mixed precision, to BIM model geometry, this technology represents an incremental improvement. It transforms BIM from a tool requiring specialized high-end hardware into a universal enterprise-level asset management platform that can run on ordinary commercial hardware, thus fully unleashing the potential of BIM as an interactive model and promoting the application of digital twin technology in a wider range of fields.

[0064] Example 2 This embodiment discloses a hierarchical quantization system for a lightweight BIM engine.

[0065] A hierarchical quantization system for a lightweight BIM engine, comprising: The model acquisition module is configured to: acquire the original BIM file into the system; The model parsing module is configured to perform structured analysis on the BIM model and divide it into multiple sub-components based on its internal hierarchical structure and component semantic information. The quantization sensitivity assessment module is configured to: perform quantization sensitivity assessment on each sub-component and divide the quantization levels; The hierarchical hybrid quantization module is configured to perform hierarchical hybrid quantization based on the quantization level of each sub-component: a high-precision quantization strategy is used for sub-components at high sensitivity levels, minimizing precision loss by simulating quantization operations and optimizing quantization parameters; a non-uniform quantization strategy is used for sub-components at medium sensitivity levels, dynamically allocating quantization precision based on the local density of vertices within the sub-component; and a unified quantization strategy is used for sub-components at low sensitivity levels, linearly mapping vertex coordinates and normal vectors to low-precision integers. The data storage module is configured to store the quantized geometric data and non-geometric data in different table structures in the database. The rendering module is configured to utilize the client's GPU computing power to perform dequantization operations in the early stages of the rendering pipeline, recovering vertex coordinates and normal vectors in real time, and then using the recovered geometric data for rendering. Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0066] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps in a layered quantization method for a lightweight BIM engine as described in Embodiment 1 of this disclosure.

[0067] Example 4 The purpose of this embodiment is to provide an electronic device.

[0068] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a layered quantization method for a lightweight BIM engine as described in Embodiment 1 of this disclosure.

[0069] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0070] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0071] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A hierarchical quantization method for a lightweight BIM engine, characterized in that, include: The BIM model is structurally analyzed and divided into multiple sub-components based on its inherent hierarchical structure and component semantic information; Perform a quantization sensitivity assessment on each sub-component and divide it into quantization levels; Based on the quantization level of each sub-component, layered hybrid quantization is performed: a high-precision quantization strategy is adopted for sub-components with high sensitivity levels, and the precision loss is minimized by simulating quantization operations and optimizing quantization parameters. For sub-components with medium sensitivity, a non-uniform quantization strategy is adopted, which dynamically allocates quantization precision based on the local density of vertices within the sub-component; for sub-components with low sensitivity, a uniform quantization strategy is adopted, which linearly maps vertex coordinates and normal vectors to low-precision integers. The quantized geometric data is stored separately from the non-geometric data. The client's GPU computing power is used to perform dequantization in the early stages of the rendering pipeline to restore vertex coordinates and normal vectors in real time. The restored geometric data is then used for rendering.

2. The hierarchical quantization method for a lightweight BIM engine as described in claim 1, characterized in that, The quantitative sensitivity assessment is achieved by analyzing the geometric complexity and semantic importance of each sub-component, namely: parsing the BIM model to obtain the semantic information of each sub-component and calculating the geometric complexity index of each sub-component; combining the semantic information and the geometric complexity index, a weighted calculation method is used to obtain a comprehensive quantitative sensitivity score.

3. The layered quantization method for a lightweight BIM engine as described in claim 2, characterized in that, The semantic information includes component name, type, and function; the geometric complexity index includes at least one of the following: number of vertices, number of faces, average face curvature, and topological irregularity.

4. The hierarchical quantization method for a lightweight BIM engine as described in claim 1, characterized in that, The high-precision quantization strategy includes: inserting pseudo-quantization operations during forward propagation to simulate the impact of low-precision arithmetic on geometric data; defining a loss function to minimize the difference between full-precision geometric data and pseudo-quantized data; and optimizing quantization parameters through backpropagation and a pass-through estimator to preserve key geometric details.

5. The layered quantization method for a lightweight BIM engine as described in claim 1, characterized in that, The non-uniform quantization strategy employs an iterative Lloyd-Max algorithm to dynamically allocate quantization precision, including: optimizing the quantization interval boundary and quantization value based on the distribution characteristics of vertices within the sub-component; and minimizing quantization error by alternately updating the quantization interval boundary and quantization value.

6. The layered quantization method for a lightweight BIM engine as described in claim 5, characterized in that, The Lloyd-Max algorithm is based on two mutually reinforcing conditions: the optimal core condition and the optimal boundary condition.

7. The layered quantization method for a lightweight BIM engine as described in claim 1, characterized in that, The unified quantization strategy includes: using a single linear scaling factor to map vertex coordinates and normal vectors to low-precision integers; and determining the quantization range based on the model's axis-aligned bounding box boundaries.

8. A hierarchical quantification system for a lightweight BIM engine, characterized in that, include: The model acquisition module is configured to: acquire the original BIM file into the system; The model parsing module is configured to perform structured analysis on the BIM model and divide it into multiple sub-components based on its internal hierarchical structure and component semantic information. The quantization sensitivity assessment module is configured to: perform quantization sensitivity assessment on each sub-component and divide the quantization levels; The hierarchical hybrid quantization module is configured to perform hierarchical hybrid quantization according to the quantization level to which each sub-component belongs: a high-precision quantization strategy is adopted for sub-components with high sensitivity levels, and the accuracy loss is minimized by simulating quantization operations and optimizing quantization parameters. For sub-components with medium sensitivity, a non-uniform quantization strategy is adopted, which dynamically allocates quantization precision based on the local density of vertices within the sub-component; for sub-components with low sensitivity, a uniform quantization strategy is adopted, which linearly maps vertex coordinates and normal vectors to low-precision integers. The data storage module is configured to store the quantized geometric data and non-geometric data in different table structures in the database. The rendering module is configured to utilize the client's GPU computing power to perform dequantization operations in the early stages of the rendering pipeline, recovering vertex coordinates and normal vectors in real time, and then using the recovered geometric data for rendering.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the hierarchical quantization method for a lightweight BIM engine as described in any one of claims 1-7.

10. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the hierarchical quantization method for a lightweight BIM engine as described in any one of claims 1-7.

Images