Semantic-driven adaptive rendering method and system for industrial 3d models

Abstract

The application provides a semantic driving-based industrial three-dimensional model adaptive rendering method and system, and belongs to the field of computer graphics processing and industrial digitization technology.The method constructs a heterogeneous model data structure containing discrete grid data, parameterized feature data, feature semantic data and mapping relationship between features and local areas of discrete grids in an offline stage;in a rendering process, the rendering priority is dynamically determined according to the engineering semantic weight of the feature, the screen space projection amount and the sight distance, the feature meeting the condition is determined as a feature to be reconstructed; based on the parameterized feature data of the feature to be reconstructed, the parameter domain is sampled and corresponding vertex data is generated through parallel calculation; then, the corresponding local area in the original discrete grid is shielded, removed or replaced according to the mapping relationship, and the generated vertex data is fused and rendered with the remaining discrete grid data.The method realizes on-demand refinement reconstruction of key features of an industrial model.

Description

Technical Field

[0001] This invention relates to the fields of computer graphics and industrial digitalization technology, and in particular to a semantically driven adaptive rendering method and system for industrial 3D models. Background Technology

[0002] With the deepening of the Industrial Internet and digital transformation, the efficient display and interaction of 3D computer-aided design (CAD) models of complex industrial products (such as construction machinery, aerospace equipment, and automobiles) on the web has become a core requirement for application scenarios such as remote operation and maintenance, digital marketing, and collaborative design. However, industrial 3D models typically contain tens of thousands or even millions of parts, with massive data scale, complex structure, and a large amount of critical engineering semantic information, which poses a severe challenge to real-time rendering.

[0003] To address these challenges, existing technologies primarily explore two aspects: model lightweighting and rendering optimization.

[0004] On the one hand, model simplification techniques based on Level of Detail (LOD) are widely adopted. For example, Chinese patent application CN110362927A proposes a lightweight BIM model processing method, which separates attributes from mesh data and generates multiple meshes of different precision (coarse model, medium model, fine model) for components, loading the model of the corresponding precision according to the viewpoint distance. PCT application WO2023124842A1 further introduces the concept of Screen Space Error (SSE), determining when to switch to a finer model level by calculating the projection of geometric errors on the screen, and defining "incremental" and "replacement" node switching methods to achieve a smoother LOD transition. These methods alleviate rendering pressure to some extent, but their essence is still the scheduling and switching of pre-generated discrete mesh data. When users need to zoom in locally to observe key engineering features (such as positioning holes, mating surfaces, and sealing grooves), these features will still suffer from polygonal distortion due to mesh simplification, failing to meet the needs of accurate measurement and assembly analysis in industrial scenarios.

[0005] On the other hand, to improve data loading and parsing efficiency, the industry is committed to optimizing the organization format and transmission protocol of 3D data. For example, Chinese patent application CN112270756A proposes the idea of ​​component templates and instantiation, reducing data redundancy by reusing components with the same geometric shape. Chinese patent application CN116977523A describes a complete web-based rendering process, including data parsing, preprocessing, mesh transformation, progressive rendering, and viewport culling, and explores the possibility of improving performance using distributed rendering. However, these methods still focus on the compression, organization, and transmission of mesh data. After the model is restored on the browser side, its geometric accuracy depends entirely on the quality of the pre-generated and stored mesh. The loss of parametric features and boundary representation (Brep) information in the original computer-aided design (CAD) model causes the web-based model to become a "dumb graphic," unable to support feature-based intelligent interaction and downstream process applications.

[0006] Furthermore, existing technologies also include schemes that preset loading priorities based on application scenarios or user operations. For example, Chinese patent application CN110069733A sets static loading priorities for different components based on whole-machine or component-level application modes. This scheme has a coarse decision-making granularity, cannot respond to the user's dynamically changing observation intentions, and its priority determination is unrelated to the importance of the industrial features themselves, making it difficult to ensure the display accuracy of key process surfaces in complex scenarios.

[0007] In summary, existing technologies generally suffer from the following three interrelated defects when rendering ultra-large-scale industrial 3D models:

[0008] The inherent trade-off between accuracy and performance: Whether it's distance-based LOD switching or SSE-based HLOD refinement, the implementation of fine-grained levels relies on pre-stored high-precision grid data. To ensure smoothness, the overall data volume must be significantly reduced, leading to geometric distortion of key engineering features during scaling.

[0009] Complete loss of engineering semantics: Existing lightweight workflows generally simplify CAD models to pure triangular meshes, filtering out original geometric constraints, parametric equations, and topological relationships. This makes the models unable to support feature-level recognition, accurate measurement, and semantic-based analysis and simulation, greatly reducing the industrial application value of the models.

[0010] Insufficient intelligence in resource scheduling: Existing rendering scheduling strategies are mainly based on spatial distance, projection size, or preset application patterns, lacking consideration for the core dimension of "importance of industrial features." For key features such as mating surfaces and positioning holes, even if they are in the center of the user's field of view, they fail to receive priority accuracy guarantees.

[0011] To address the aforementioned issues, a novel technical solution is urgently needed that can fundamentally solve the problem of preserving the accuracy of key industrial features while ensuring rendering performance, and at the same time restore the engineering semantics of the model, thereby achieving truly intelligent and adaptive rendering. Summary of the Invention

[0012] The technical problem to be solved by this invention is that when rendering ultra-large-scale industrial 3D models on the Web, existing technologies cannot simultaneously take into account rendering performance and the display accuracy of key engineering features (such as assembly holes and positioning surfaces), resulting in the loss of industrial semantics and geometric distortion. This invention proposes a semantically driven adaptive rendering method and system for industrial 3D models.

[0013] The semantically driven adaptive rendering method for industrial 3D models according to embodiments of the present invention includes:

[0014] Obtain heterogeneous lightweight model data, including: discrete mesh data, parametric feature data for characterizing the geometry of the industrial model, feature semantic data, and the mapping relationship between the parametric feature data and the mesh regions in the discrete mesh data;

[0015] The rendering priority of the features is dynamically calculated based on the engineering semantic weights corresponding to the parameterized feature data and the view parameters of the features corresponding to the parameterized feature data in the current view state.

[0016] Features whose rendering priority meets preset conditions are identified as features to be reconstructed.

[0017] Based on the parameterized feature data of the features to be reconstructed, the parameter domain is sampled, and corresponding vertex data is generated through parallel computing; the local regions of the discrete mesh are masked, eliminated, or replaced according to the mapping relationship; the generated vertex data is fused and rendered with the remaining discrete mesh data, and then output to the display device.

[0018] According to some embodiments of the present invention, the parameterized feature data includes geometric parameters and spatial positioning information for describing the target features, which are obtained by parsing boundary representation data. The parameterized feature data also includes local precise geometric representation data corresponding to the target features.

[0019] In some embodiments of the present invention, the engineering semantic weights are pre-set based on the functional importance of industrial features in assembly, manufacturing or measurement. Different types of industrial features have different semantic weights, wherein the semantic weights of mating features and positioning features are higher than the semantic weights of non-functional decorative features.

[0020] According to some embodiments of the present invention, the view parameters include the projected area of ​​the feature in screen space and the distance between the feature and the observation point; the rendering priority is determined by a combination of the projected area, the engineering semantic weight, and the distance.

[0021] In some embodiments of the present invention, the calculation model for the rendering priority P is as follows:

[0022]

[0023] Where S is the projected area of ​​the feature in the screen space. denoted as the engineering semantic weight of the feature, D is the distance between the observation point and the feature, and f() is the calculation function.

[0024] According to some embodiments of the present invention, the parallel computing employs at least one of a central processing unit and a graphics processing unit, wherein the parallel computing unit includes at least one of a central processing unit and a graphics processing unit. When the parallel computing employs a WebGPU, generating vertex data via the WebGPU specifically includes:

[0025] Assign thread groups in the WebGPU compute shader that match the preset sampling density;

[0026] Each thread samples the parameter domain of the parameterized feature data in parallel and calculates the three-dimensional coordinates of the sampling points based on the corresponding geometric analytical equations.

[0027] The calculated 3D coordinates are directly written to the vertex buffer in the WebGPU memory to generate the vertex data.

[0028] In some embodiments of the present invention, the parameterized feature data includes cylindrical hole features, and the geometric analytical equations corresponding to the cylindrical hole features are reconstructed by polar coordinate sampling based on the axis starting point coordinates, axis vector, radius and height parameters.

[0029] According to some embodiments of the present invention, the method is executed in a browser environment, and the parallel computation is implemented using WebGPU compute shaders and / or vertex shaders.

[0030] In some embodiments of the present invention, the rendering priority has multiple levels, and the method further includes: dynamically adjusting the parameter sampling frequency during reconstruction through parallel computing according to the rendering priority to generate vertex data of different levels of refinement; masking, culling, or replacing local regions of discrete meshes according to the mapping relationship, and fusing the generated vertex data with the remaining discrete mesh data for rendering includes: locating the discrete mesh region corresponding to the feature to be reconstructed according to the mapping relationship, and performing rendering masking, triangle culling, or vertex data replacement on the discrete mesh region, replacing it with the generated vertex data for rendering.

[0031] An adaptive rendering system for industrial 3D models based on semantic driving according to an embodiment of the present invention includes:

[0032] The data processing module is used to acquire and parse heterogeneous lightweight model data, including discrete grid data, parameterized feature data, and the mapping relationship between the discrete grid data and the parameterized feature data;

[0033] The scheduling engine module is used to calculate the rendering priority based on the engineering semantic weight of the feature and the view parameters, and to determine the feature to be reconstructed based on the rendering priority;

[0034] The parallel computing module is used to sample the parameter domain and generate vertex data based on the parameterized feature data;

[0035] The rendering fusion module is used to mask, cull, or replace local areas of the discrete mesh according to the mapping relationship, and to fuse the generated vertex data with the remaining discrete mesh data for rendering.

[0036] The present invention has the following beneficial effects:

[0037] This invention fundamentally solves the contradiction between performance and accuracy when rendering large-scale industrial models by constructing a heterogeneous lightweight data architecture that includes discrete meshes and parametric features. By introducing engineering semantic weights to participate in the dynamic scheduling of rendering priorities, it ensures that key assembly features (such as positioning holes and mating surfaces) receive priority accuracy guarantees from any viewpoint, avoiding the geometric distortion of key features caused by traditional LOD technology. By utilizing parallel computing units to reconstruct high-precision geometry in real time on the browser side, it realizes an "on-demand refinement" rendering paradigm, significantly reducing the initial data transmission volume and video memory usage. This enables smooth interaction in a normal browser environment and supports accurate measurement based on industrial semantics, providing key technical support for remote operation and maintenance, collaborative design, and digital delivery of industrial equipment. Attached Figure Description

[0038] Figure 1This is a flowchart of a semantically driven adaptive rendering method for industrial 3D models according to an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of the structure of an adaptive rendering system for semantically driven industrial 3D models according to an embodiment of the present invention.

[0040] Figure label:

[0041] Data processing module 101, scheduling engine module 102, parallel computing module 103, rendering fusion module 104. Detailed Implementation

[0042] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0043] The steps described in the specification and the flowcharts in the accompanying drawings of this invention are not necessarily to be strictly followed according to the step numbers; the execution order of the steps can be changed. Furthermore, certain steps can be omitted, multiple steps can be combined into one step, and / or one step can be broken down into multiple steps.

[0044] As described in the background section, existing industrial 3D model rendering technologies suffer from the following problems:

[0045] 1. Precision distortion issue;

[0046] Traditional LOD strategies based on distance or hierarchy often result in significant polygonization of key engineering features (such as mating holes, sealing surfaces, and positioning surfaces) when zoomed in or magnified locally, making it difficult to meet the needs of industrial measurement and assembly analysis.

[0047] 2. The problem of loss of engineering semantics;

[0048] Lightweight 3D models typically retain only triangular geometric information, losing the engineering semantic attributes of the original CAD model. This results in web-based models being unable to support feature-level recognition, measurement, and downstream process or simulation applications.

[0049] 3. Performance and resource bottlenecks;

[0050] For complex industrial assemblies containing tens of thousands of parts, directly loading high-precision meshes or multi-level static LODs will lead to excessive video memory usage on the web client, and may even cause browser crashes.

[0051] To address the aforementioned issues, this invention proposes a feature-level adaptive rendering method and system for industrial 3D models based on semantics. In the offline phase, this invention extracts parameters, data, and features from the industrial CAD model, constructing a heterogeneous lightweight data structure containing discrete mesh data, Brep data, and parameterized feature data. During rendering, the rendering priority is dynamically determined based on industrial feature semantics and view state, and parallel computing is used to perform local real-time geometric reconstruction on high-priority features, thereby achieving adaptive high-precision rendering.

[0052] The present invention will now be described in detail with reference to the accompanying drawings, using specific embodiments of parallel computing implemented with WebGPU computing shaders. The described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0053] Example 1:

[0054] This embodiment provides a semantic-driven adaptive rendering method for industrial 3D models on the web. This method can achieve efficient and high-precision rendering of ultra-large-scale industrial 3D models in a browser environment. It is particularly suitable for web-based visualization of complex industrial assemblies (such as engineering machinery, aerospace equipment, automobiles, etc.) containing tens of thousands or even hundreds of thousands of parts.

[0055] Figure 1 This is a flowchart illustrating the method described in this embodiment. Figure 1 As shown, the method includes the following steps:

[0056] Step S1: Obtain heterogeneous lightweight model data.

[0057] In the offline processing stage, feature recognition and parameter extraction are performed on the original industrial CAD model. Specifically, parametric geometric features are identified from the industrial CAD model (such as STEP, IGES, etc.). Parametric geometric features include at least one or more of the following: cylindrical holes, stepped shafts, keyways, datum planes, and freeform surfaces. The corresponding geometric parameters (such as the axis start coordinates, axis vector, radius, height, etc.) and engineering semantic identifiers (such as "locating hole", "sealing surface", "mating surface", "decorative feature", etc.) are extracted from these features.

[0058] The extracted data is encapsulated into a heterogeneous lightweight data structure, which includes at least the following:

[0059] Discrete mesh data: Lightweight triangular mesh data used for the basic display;

[0060] Parametric feature data: Parametric equations and corresponding parameters used to characterize the key geometries of industrial models;

[0061] Feature semantic identifier: A semantic label corresponding to the parameterized feature data, representing the importance of the feature engineering.

[0062] Simultaneously, an indexable mapping relationship is established between the parameterized feature data and the discrete mesh data. This mapping relationship is used to quickly locate the spatial region in the discrete mesh data corresponding to a specific parameterized feature during subsequent rendering.

[0063] Optionally, the heterogeneous lightweight model data may also include boundary representation Brep data to enhance the model's topology analysis capabilities on the web.

[0064] Step S2: Calculate the rendering priority of features in real time.

[0065] During the rendering process on the web platform, the current view state is acquired in real time, including information such as camera position, view direction, and field of view. For each feature in the scene (i.e., the parametric features identified in step S1), its rendering priority is dynamically calculated based on the following factors:

[0066] Engineering semantic weight: A weight value preset based on the functional importance of a feature in industrial assembly, manufacturing, or measurement. For example, the semantic weight of mating features (such as bearing housing holes) and positioning features (such as locating pin holes) can be set to 0.9, while the semantic weight of decorative chamfers or non-functional decorative features can be set to 0.3. This weight value reflects the criticality of the feature in industrial applications.

[0067] Screen space projected area: The pixel area occupied by the bounding box of the calculated feature after being projected onto screen space in the current view. This parameter reflects the visual salience of the feature in the current viewpoint.

[0068] View distance information: Calculates the Euclidean distance between the observation point and the center point of the feature. This parameter reflects the proximity of the feature to the observer.

[0069] In a preferred embodiment, the calculation model for rendering priority P is as follows:

[0070]

[0071] Where S is the projected area of ​​the feature in the screen space. sem represents the engineering semantic weights of the features, where sem is an abbreviation for semantic, and D is the distance between the observation point and the feature. The function f can be a weighted sum, product, or other computational model that integrates these three factors.

[0072] It should be noted that engineering semantic weights These are pre-defined values ​​based on the functional importance of industrial characteristics in assembly, manufacturing, or measurement. The specific implementation method is as follows:

[0073] (1) Rules for setting weight values: For features that directly affect assembly accuracy, positioning accuracy or core functions of the product (such as reference holes, mating surfaces, locating pin holes, etc.), set higher weight values ​​(such as 0.8 to 1.0); for features that have a certain impact on function but are not critical (such as general connecting holes, non-precision mating surfaces), set medium weight values ​​(such as 0.4 to 0.7); for features that only play a decorative or auxiliary role (such as chamfers, fillets, non-functional bosses), set lower weight values ​​(such as 0.1 to 0.3).

[0074] (2) Methods for obtaining weight values: Weight values ​​can be obtained through one or more of the following methods:

[0075] Read the preset semantic attributes or annotation information directly from the original CAD model and convert them into weight values;

[0076] Weight values ​​are automatically calculated based on the geometric parameters of the features (such as dimensional tolerances and surface roughness requirements) and contextual relationships (such as whether it is a mating relationship);

[0077] By manually labeling key features, weight values ​​are preset for them during the offline processing stage.

[0078] (3) Storage of weight values: Engineering semantic weights and feature semantic data are associated and stored in heterogeneous lightweight model data for the scheduling engine module to read and use during rendering.

[0079] The above calculations ensure that features that are both important and prominent in the current view receive higher rendering priority.

[0080] Step S3: Identify the features that meet the preset conditions as the features to be reconstructed.

[0081] Set a rendering priority threshold T. When the feature rendering priority P calculated in step S2 is greater than or equal to the threshold T, the feature is identified as a "feature to be reconstructed," meaning that high-precision geometric reconstruction is required for this feature. If P is less than the threshold T, the original discrete mesh data is used for rendering.

[0082] In an optional embodiment, the rendering priority can be divided into multiple levels (such as high, medium, and low levels), each level corresponding to a different reconstruction precision. Step S3 determines the corresponding reconstruction precision level based on the priority level of the feature.

[0083] Step S4: Generate high-precision vertex data in parallel using WebGPU compute shaders.

[0084] In response to the features to be reconstructed and their corresponding reconstruction accuracy levels determined in step S3, the WebGPU computation shader is invoked on the browser side to generate high-precision vertex data in parallel based on the parameterized feature data of these features. This step specifically includes the following sub-steps:

[0085] Step S41: Thread Group Assignment. In the WebGPU compute shader, thread groups are assigned according to the preset sampling density and the number of sampling points. Each thread is responsible for computing the geometric data of one or more sampling points.

[0086] Step S42: Parallel Parameter Domain Sampling. Each thread samples the parameter domain of the parameterized feature data in parallel. For example, for a cylindrical hole feature, its geometric analytical equation is reconstructed using polar coordinate sampling based on the axis start coordinates, axis vector, radius, and height parameters. Each thread calculates the corresponding parameter domain coordinates (such as angle θ and height h) according to its assigned thread ID, and then substitutes them into the geometric analytical equation to calculate the three-dimensional spatial coordinates (x, y, z) of the sampling point.

[0087] Step S43: Vertex Normal and Color Calculation. Each thread calculates the normal vector at the sampling point based on the geometric analytical equations. If the feature contains color or texture information, the color value or texture coordinates of the sampling point can also be calculated simultaneously.

[0088] Step S44: Directly write to the vertex buffer. The calculated 3D coordinates, normal vectors, color values, and other data are directly written to the vertex buffer in the WebGPU's video memory. This process is completed entirely on the GPU side, without needing to send data back to the CPU, avoiding data round trips between the CPU and GPU, and significantly improving generation efficiency.

[0089] Step S45: Atomic operations ensure data consistency. When multiple threads write to the vertex buffer in parallel, atomic operations supported by WebGPU (such as atomicAdd) are used to ensure that each thread obtains a unique write position, avoiding data overwriting and conflicts.

[0090] Step S5: Fuse and render the generated high-precision vertex data with the discrete mesh data.

[0091] The high-precision vertex data generated in step S4 is fused with the original discrete mesh data to form a complete rendering scene. Specifically, this includes:

[0092] Step S51: Masking / Catasing. Using a masking mechanism or vertex replacement mechanism, the original low-precision discrete mesh data corresponding to the feature to be reconstructed is culled from the rendering pipeline. For example, based on the mapping relationship established in Step S1, the triangular regions corresponding to the feature to be reconstructed in the discrete mesh data can be located, and these triangles can be marked as invisible.

[0093] Step S52: Vertex Buffer Merging. The high-precision vertex data generated in step S4 is merged with the vertex buffer of the remaining discrete mesh data (i.e., the part that was not culled) to form a complete vertex buffer.

[0094] Step S53: Unified rasterization processing. The merged vertex buffer is sent to the rendering pipeline for standard rendering processes such as vertex processing, rasterization, and fragment processing, and finally output to the display device.

[0095] Through the above steps, when users interact with the model (such as rotating, scaling, and translating), the system can dynamically adjust the rendering precision of features in real time according to the view state: when a key feature enters the field of view and is magnified for observation, the system automatically triggers GPU reconstruction to present the feature with mathematical precision; when the feature is far from the field of view or in a secondary position, the system only uses a lightweight discrete mesh for rendering, thereby significantly reducing the occupation of video memory and computing resources while ensuring visual quality and smooth interaction.

[0096] Example 2:

[0097] This embodiment further refines the process of generating heterogeneous lightweight model data in step S1, based on embodiment 1.

[0098] The generation of heterogeneous lightweight model data includes:

[0099] Step S11: Feature Recognition. The original industrial CAD model is analyzed to identify its parametric geometric features. The recognition process can employ feature recognition algorithms based on graph matching, traces, or neural networks. The identified feature types include at least: cylindrical / hole features, conical features, planar features, spherical features, toroidal features, extruded features, revolved features, and swept features.

[0100] Step S12: Parameter Extraction. For each identified feature, extract its complete geometric definition parameters. Taking the cylindrical hole feature as an example, the extracted parameters include: axis start point coordinates (x0, y0, z0), axis direction vector (v... x v y v z ( ), radius R, height H. For complex features, it is also necessary to extract their boundary representation Brep data, including the topological relationships and geometric definitions of all faces, edges, and vertices that constitute the feature.

[0101] Step S13: Semantic annotation. Add engineering semantic identifiers to each feature. Semantic identifiers can be obtained in the following ways: (1) directly read existing semantic information (such as annotations, attributes, etc.) from the original CAD model; (2) automatically infer the semantics of the feature based on its geometric attributes (such as size, position, orientation) and contextual relationships (such as the relationship with other features); (3) preset the semantics of key features through manual annotation.

[0102] In a preferred embodiment, the semantic identifier adopts a hierarchical structure, including multi-dimensional information such as feature categories (e.g., "hole features"), feature subcategories (e.g., "through holes", "blind holes", "threaded holes"), functional categories (e.g., "positioning holes", "assembly holes", "weight reduction holes"), and importance levels (e.g., "critical", "important", "general").

[0103] Step S14: Data Encapsulation. The extracted data is encapsulated into a unified heterogeneous data structure. This data structure can be serialized using formats such as JSON, BSON, and Protocol Buffers. During the encapsulation process, an index mapping is established between features and grid regions, for example, recording the range of grid vertex indices or triangle indices covered by each feature.

[0104] Step S15: Data Compression and Encryption. The packaged heterogeneous data is compressed to reduce its size. General compression algorithms (such as gzip, zstd) or specialized compression algorithms for 3D geometric data can be used. For sensitive industrial data, encryption can also be applied to ensure data security during transmission and storage.

[0105] Example 3:

[0106] Based on Example 1, this embodiment further refines the calculation process of rendering priority in step S2 and provides multiple calculation models.

[0107] Model 1: A calculation model based on weighted summation;

[0108]

[0109] Among them, S max D is the maximum projected area of ​​all features in the current view. max This is the preset maximum effective line-of-sight distance (features exceeding this distance will no longer be included in the priority calculation). , , For the weighting coefficients, satisfying .

[0110] Model 2: Product-based computational model;

[0111]

[0112] in, S ref For reference projection area (e.g., 1% of screen area). D ref For reference viewing distance (e.g., 10% of the diagonal length of the model's bounding box). α , β , γ These are exponential parameters representing the projected area, semantic weight, and viewing distance, used to adjust the degree of non-linear influence of each factor on the priority. When α When the value is greater than 1, the larger the projected area, the more linearly the priority increases; when 0 < 1, the priority increases; when the value is less than ... α When the priority level is less than 1, the growth rate tends to level off. β and γ and α The adjustment principle is the same, so it will not be repeated here.

[0113] Model 3: Classification-based computational model;

[0114] Features are categorized into several classes (e.g., key features, important features, common features, decorative features) based on their engineering semantic weights. Then, the Level of Depth (LOD) level for each class is determined based on the viewing distance and projected area. Specifically, an LOD lookup table can be pre-built for each semantic class, with the viewing distance and projected area as inputs and the reconstruction accuracy level as output.

[0115] In one specific embodiment, the engineering semantic weight is a pre-set value, ranging from 0 to 1. Based on the magnitude of the engineering semantic weight, the features can be divided into multiple categories, for example:

[0116] Key features: Weight value ≥ 0.8, corresponding to features that directly affect the assembly accuracy or core function of the product, such as assembly reference surface, main positioning hole, and mating journal;

[0117] Key features: 0.5 ≤ weight value < 0.8, corresponding to secondary positioning surfaces, threaded holes, oil passage holes, and other features that affect the function but are not critical;

[0118] Common features: 0.2 ≤ weight value < 0.5, corresponding to general structural surfaces, stiffeners, etc.

[0119] Decorative features: Weight value < 0.2, corresponding to chamfers, rounded corners, and non-functional appearance features.

[0120] It should be noted that the above weighting values ​​can be adjusted according to specific industrial scenarios and product types.

[0121] The LOD lookup table is constructed as follows: For each type of feature (such as key features), a set of viewing distance intervals and projected area intervals are predefined, with each interval corresponding to a reconstruction accuracy level. For example, for key features, the following lookup table 1 can be set up:

[0122] Table 1 LOD Lookup Table

[0123]

[0124] The division of viewing distance and projected area intervals can be adjusted according to model size, display device resolution, and actual application scenario. During the actual rendering process, the reconstruction accuracy level of the feature can be determined by querying the corresponding LOD lookup table based on the viewing distance and projected area of ​​the current feature, thereby controlling the parameter sampling frequency of the parallel computing unit.

[0125] Model 4: Machine Learning-Based Learning Model;

[0126] Collect user interaction data in different scenarios (such as frequency of zooming in, dwell time, etc.) and train a machine learning model to predict feature rendering priority. Input features include: feature type, semantic weight, projected area, viewing distance, feature position on the screen, user history behavior, etc., and the output is a priority score.

[0127] Example 4:

[0128] This embodiment provides a semantically driven adaptive rendering system for industrial 3D models on the web, which is used to implement the method described in any one of embodiments 1 to 3.

[0129] Figure 2 This is a schematic diagram of the module structure of the system in this embodiment. Figure 2 As shown, the system includes:

[0130] The data processing module 101 is used to acquire and parse heterogeneous lightweight model data. The heterogeneous lightweight model data includes discrete mesh data, parameterized feature data, and the mapping relationship between the discrete mesh data and the parameterized feature data. The data processing module 101 is also used to load model data during system initialization and to respond to data requests from the scheduling engine module 102 during operation.

[0131] The scheduling engine module 102 is used to acquire the current view state in real time (including camera position, viewpoint, field of view, etc.), calculate the rendering priority of features based on the engineering semantic weight of features and view parameters (projected area, viewing distance, etc.), and determine the features to be reconstructed and their corresponding reconstruction accuracy levels based on the rendering priority. The scheduling engine module 102 is also responsible for managing a priority queue and dynamically updating the priority order of features in the queue according to changes in the viewpoint.

[0132] Parallel computing module 103, employing a WebGPU reconstruction kernel integrated into the browser, is used to respond to features to be reconstructed, calling the WebGPU computation shader to generate high-precision vertex data in parallel based on the corresponding parameterized feature data. The WebGPU reconstruction kernel includes:

[0133] A thread scheduling unit is used to allocate an appropriate number of thread groups based on the reconstruction precision level.

[0134] The parameter parsing unit is used to obtain parameterized data of the features to be reconstructed from the data processing module 101;

[0135] Parallel computing units are used to execute computational shaders and generate high-precision vertex data on the GPU.

[0136] The buffer management unit is used to manage the vertex buffer in video memory, ensuring efficient writing and updating of reconstructed data.

[0137] The rendering fusion module 104 is used to fuse the generated high-precision vertex data with discrete mesh data and drive the display device to perform rendering output. The rendering fusion module 104 includes:

[0138] The masking unit is used to remove the original discrete mesh data corresponding to the feature to be reconstructed from the rendering pipeline;

[0139] The buffer merging unit is used to merge the high-precision vertex data generated by reconstruction with the remaining discrete mesh data;

[0140] The rendering pipeline interface is used to send the merged data into the WebGPU rendering pipeline for rasterization processing.

[0141] Data interaction between the above modules is accomplished through API calls within the browser, without the need for additional network communication overhead.

[0142] Example 5:

[0143] This embodiment provides a lightweight data transmission format for industrial 3D models, which is particularly suitable for the storage and transmission of heterogeneous lightweight model data in the methods and systems described in Embodiments 1 to 4.

[0144] The data transmission format adopts a hierarchical nested structure, containing multiple data blocks within a single data stream. The format has the following characteristics:

[0145] Data separation: Discrete grid data and parameterized feature data are stored separately for easy loading on demand;

[0146] Feature Index: Each feature has a unique feature ID and contains a mapping pointer to a specific region of the grid;

[0147] Semantic embedding: Each feature contains a semantic type and semantic weight, which are used by the scheduling engine;

[0148] LOD presets: Each feature can be preset with LOD trigger conditions to provide a reference for scheduling decisions;

[0149] Binary encoding: All geometric data is encoded in binary format, reducing parsing overhead and supporting direct upload to the GPU.

[0150] Example 6:

[0151] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a semantically driven adaptive rendering method for industrial 3D models on a web-based platform, as described in any of embodiments 1 to 3.

[0152] Computer-readable storage media can be any medium capable of storing program code, such as ROM, RAM, magnetic disk, optical disk, or flash memory. When the program runs on the computer's processor, the computer performs the following operations:

[0153] Acquire heterogeneous lightweight model data, which includes discrete mesh data and parametric feature data used to characterize the geometry of the industrial model, and establish a mapping relationship between the parametric feature data and the discrete mesh data.

[0154] The rendering priority of features is dynamically calculated based on the engineering semantic weights corresponding to the parameterized feature data and the view parameters of the features corresponding to the parameterized feature data in the current view state.

[0155] Features whose rendering priority meets preset conditions are identified as features to be reconstructed.

[0156] In response to the features to be reconstructed, the WebGPU compute shader is invoked on the browser side. Based on the parameterized feature data of the features to be reconstructed, high-precision vertex data is generated in parallel by the WebGPU compute shader.

[0157] The generated high-precision vertex data is fused and rendered with the discrete mesh data, and then output to the display device.

[0158] In summary, this invention significantly reduces the initial data loading scale while ensuring the display accuracy of key industrial features, and avoids geometric distortion of key features caused by traditional mesh simplification. This invention supports high-precision measurement and interactive operation based on industrial semantics on the Web, and improves the rendering performance and stability of large-scale industrial assemblies in the browser environment.

[0159] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A semantically driven adaptive rendering method for industrial 3D models, characterized in that, include: Obtain heterogeneous lightweight model data, including: discrete mesh data, parametric feature data for characterizing the geometry of the industrial model, feature semantic data, and the mapping relationship between the parametric feature data and the mesh regions in the discrete mesh data; The rendering priority of the features is dynamically calculated based on the engineering semantic weights corresponding to the parameterized feature data and the view parameters of the features corresponding to the parameterized feature data in the current view state. Features whose rendering priority meets preset conditions are identified as features to be reconstructed. Based on the parameterized feature data of the features to be reconstructed, the parameter domain is sampled, and corresponding vertex data is generated through parallel computing; local regions of the discrete mesh are masked, culled, or replaced according to the mapping relationship; the generated vertex data is fused and rendered with the remaining discrete mesh data, and then output to the display device. The parameterized feature data includes geometric parameters and spatial positioning information used to describe the target features, which are obtained by parsing the boundary representation data. The parameterized feature data also includes local precise geometric representation data corresponding to the target features. The engineering semantic weights are pre-set based on the functional importance of industrial features in assembly, manufacturing, or measurement. Different types of industrial features have different semantic weights. Among them, the semantic weights of mating features and positioning features are higher than those of non-functional decorative features. The view parameters include the projected area of ​​the feature in screen space and the distance between the feature and the viewpoint; the rendering priority is determined by a combination of the projected area, the engineering semantic weight, and the distance. The calculation model for the rendering priority P is as follows: Where S is the projected area of ​​the feature in the screen space. SEM is the engineering semantic weight of the feature, where SEM is an abbreviation for the English word "semantic". D is the distance between the observation point and the feature, and f() is the calculation function, which includes a calculation function based on weighted summation and a calculation function based on product. When f() uses a weighted summation-based calculation function, the calculation model for the rendering priority P is: Among them, S max D is the maximum projected area of ​​all features in the current view. max The preset maximum effective viewing distance, , , For the weighting coefficients, satisfying ; When f() uses a product-based calculation function, the calculation model for the rendering priority P is as follows: in, S ref For reference projected area, D ref For reference viewing distance, α , β , γ These are exponential parameters representing the projected area, semantic weight, and viewing distance, used to adjust the degree of nonlinear influence of each factor on the priority.

2. The semantically driven adaptive rendering method for industrial 3D models according to claim 1, characterized in that, The parallel computing employs at least one of a central processing unit and a graphics processing unit. When the parallel computing employs a WebGPU, generating vertex data via the WebGPU specifically includes: Assign thread groups in the WebGPU compute shader that match the preset sampling density; Each thread samples the parameter domain of the parameterized feature data in parallel and calculates the three-dimensional coordinates of the sampling points based on the corresponding geometric analytical equations. The calculated 3D coordinates are directly written to the vertex buffer in the WebGPU memory to generate the vertex data.

3. The semantically driven adaptive rendering method for industrial 3D models according to claim 2, characterized in that, The parameterized feature data includes cylindrical hole features, and the geometric analytical equations corresponding to the cylindrical hole features are reconstructed by polar coordinate sampling based on the axis starting point coordinates, axis vector, radius and height parameters.

4. The semantically driven adaptive rendering method for industrial 3D models according to claim 1, characterized in that, The method is executed in a browser environment, and the parallel computation is implemented using WebGPU compute shaders and / or vertex shaders.

5. The semantically driven adaptive rendering method for industrial 3D models according to claim 1, characterized in that, The rendering priority has multiple levels, and the method further includes: dynamically adjusting the parameter sampling frequency during reconstruction through parallel computing according to the rendering priority to generate vertex data of different levels of refinement; masking, culling, or replacing local regions of discrete meshes according to the mapping relationship, and fusing the generated vertex data with the remaining discrete mesh data for rendering, including: locating the discrete mesh region corresponding to the feature to be reconstructed according to the mapping relationship, and performing rendering masking, triangle culling, or vertex data replacement on the discrete mesh region, replacing it with the generated vertex data for rendering.

6. A semantically driven adaptive rendering system for industrial 3D models, characterized in that, include: The data processing module is used to acquire and parse heterogeneous lightweight model data, including discrete grid data, parameterized feature data, and the mapping relationship between the discrete grid data and the parameterized feature data; The scheduling engine module is used to calculate the rendering priority based on the engineering semantic weight of the feature and the view parameters, and to determine the feature to be reconstructed based on the rendering priority; The parallel computing module is used to sample the parameter domain and generate vertex data based on the parameterized feature data; The rendering fusion module is used to mask, cull, or replace local areas of the discrete mesh according to the mapping relationship, and to fuse the generated vertex data with the remaining discrete mesh data for rendering. The parameterized feature data includes geometric parameters and spatial positioning information used to describe the target features, which are obtained by parsing the boundary representation data. The parameterized feature data also includes local precise geometric representation data corresponding to the target features. The engineering semantic weights are pre-set based on the functional importance of industrial features in assembly, manufacturing, or measurement. Different types of industrial features have different semantic weights. Among them, the semantic weights of mating features and positioning features are higher than those of non-functional decorative features. The view parameters include the projected area of ​​the feature in screen space and the distance between the feature and the viewpoint; the rendering priority is determined by a combination of the projected area, the engineering semantic weight, and the distance. The calculation model for the rendering priority P is as follows: Where S is the projected area of ​​the feature in the screen space. SEM is the engineering semantic weight of the feature, where SEM is an abbreviation for the English word "semantic". D is the distance between the observation point and the feature, and f() is the calculation function, which includes a calculation function based on weighted summation and a calculation function based on product. When f() uses a weighted summation-based calculation function, the calculation model for the rendering priority P is: Among them, S max D is the maximum projected area of ​​all features in the current view. max The preset maximum effective viewing distance, , , For the weighting coefficients, satisfying ; When f() uses a product-based calculation function, the calculation model for the rendering priority P is as follows: in, S ref For reference projected area, D ref For reference viewing distance, α , β , γ These are exponential parameters representing the projected area, semantic weight, and viewing distance, used to adjust the degree of nonlinear influence of each factor on the priority.

Images