PhysGaussian-based digital twin ancient building construction method and system

Abstract

The invention belongs to the field of digital twinning, and provides a PhysGaussian-based digital twinning historic building construction method and system, and the method comprises the steps: carrying out the fitting of geometric primitives through employing a differential optimization algorithm, constructing a scene model of a target historic building through employing a PhysGaussian modeling framework, and carrying out the dynamic simulation prediction of the scene model fusing physical parameters, thereby achieving the dynamic simulation prediction of the target historic building. And rendering the visual effect to generate a digital twin model. According to the method, the geometry, material characteristics and environment interaction data of the ancient building are fused through the Gaussian kernel, twinning is real, the method is suitable for complex structures and fragile materials of the ancient building, preventive protection and repair decision making are supported, and real-time updating of urban-level ancient building groups is supported.

Description

Technical Field

[0001] The present invention belongs to the field of digital twins, and specifically relates to a method and system for constructing digital twin ancient buildings based on PhysGaussian. Background Art

[0002] Digital twin technology is an advanced modeling method that integrates multi-source data acquisition, simulation calculation and intelligent analysis. Its core goal is to build a "twin" in virtual space that is highly consistent with the real physical object. This twin not only reproduces the geometric shape and material properties of the physical object, but can also dynamically respond to its operating status, environmental changes and even future behavior predictions through real-time data drive. As the core carrier of cultural heritage, ancient buildings face huge challenges in digital protection and dynamic monitoring. Although traditional digital twin technology is widely used in the industrial field, it is difficult to meet the stringent requirements of ancient buildings for high-precision geometric details, authenticity of material physical properties and dynamic behavior simulation.

[0003] Ancient buildings have complex geometric forms and delicate structural relationships. They often contain a large number of irregular components, such as curved tiles, brackets, hollow carvings, and mortise and tenon joints. Their structural characteristics are highly dependent on historical craftsmanship and cultural symbols. Traditional geometric modeling methods often use regular geometric bodies for approximation, which makes it difficult to restore the true form of these non-standard primitives, resulting in loss of details. In addition, the materials of ancient buildings are mostly wood, bricks, stones, and glazed tiles. Under the influence of long-term wind and rain erosion and temperature and humidity changes, they show complex physical behaviors such as weathering, aging, and cracking. Such nonlinear and time-varying characteristics are difficult to accurately capture and express in the current modeling framework. The optical properties of materials (such as the high light reflection of glazed tiles and the diffuse reflection roughness of wood) and mechanical responses (such as the elastic-plastic transition of mortise and tenon joints and the post-earthquake residual deformation of beam-column structures) lack effective analytical methods.

[0004] Moreover, the dynamic simulation capabilities of existing digital twin methods are also significantly insufficient. When faced with complex boundary conditions such as wind loads and earthquake shocks, most traditional methods can only perform static or simplified linear response analysis, which makes it difficult to capture the multi-stage evolution of real buildings under disasters. For objects such as ancient buildings that have been exposed to the natural environment for a long time, how to simulate the coupling process of chronic deterioration and sudden damage has always been an obstacle that digital twin systems cannot overcome. In addition, in actual urban environments, ancient buildings often interact with environmental factors such as vegetation, water bodies, and terrain. For example, humidity affects foundation stability through soil conduction, and tree shading affects ventilation and corrosion rates. These complex multi-physics field interactions lack unified modeling and simulation support in the current twin system. Summary of the invention

[0005] The present invention provides a method and system for constructing a digital twin ancient building based on PhysGaussian, so as to solve the problem that traditional digital twin technology is difficult to meet the stringent requirements of ancient buildings for high-precision geometric details, authenticity of material physical properties and dynamic behavior simulation. In order to solve the above technical problems, the embodiments of the present invention disclose the following technical solutions: One aspect of the present invention provides a method for constructing a digital twin ancient building based on PhysGaussian, comprising: Obtain point cloud data within the target ancient building area and characteristic images of the building appearance; Using a preset geometric segmentation network to identify geometric primitives from point cloud data; The geometric primitives are fitted using a differentiable optimization algorithm, and the PhysGaussian modeling framework is used to build a scene model of the target ancient building. Inputting the characteristic image of the building appearance and the data collected by the mechanical sensor on the target ancient building into a preset differentiable physical perception network to obtain the physical parameters of each geometric primitive, wherein the physical parameters include at least optical parameters and mechanical parameters; Merge each geometric primitive in the scene model with the corresponding physical parameters; The PhysGaussian modeling framework is used to perform dynamic simulation prediction on the scene model that integrates physical parameters, and render visual effects to generate a digital twin model.

[0006] Optionally, the step of acquiring point cloud data and building appearance feature images within the target ancient building area includes: Use drones equipped with high-precision laser radar to scan target ancient buildings and obtain basic point cloud data of basic structures and dense point cloud data of complex structures; A high-definition camera is used to collect an RGB-D depth image of the surface of the target ancient building, wherein the image records at least the wood surface texture, carved patterns, and coating cracks of the target ancient building; The microscopic data of the target ancient building is detected based on a multispectral imaging system, and the microscopic data at least includes wood aging, internal insect infestation, and water stain erosion.

[0007] Optionally, the step of using a preset geometric segmentation network to identify geometric primitives from point cloud data includes: The PointNet++ network or DGCNN network is used to process the point cloud data and automatically identify the geometric primitives that constitute the target ancient buildings.

[0008] Optionally, the step of fitting the geometric primitives using a differentiable optimization algorithm includes: A differentiable optimization algorithm is used to adjust the parameters of each geometric primitive with the fitting goal of minimizing the geometric error between the original point cloud data and the constructed model; Fit all geometric primitives into an initial structural model based on Boolean operations; The initial structural model is smoothed using deformation field interpolation technology to obtain the overall structural model of the target ancient building.

[0009] Optionally, the use of the PhysGaussian modeling framework to construct a scene model of the target ancient building includes: The incremental Gaussian evolution algorithm in the PhysGaussian modeling framework is used to discretize each geometric primitive in the overall structural model of the target ancient building into Gaussian particles, among which the preset key geometric primitives are discretized into fine-grained Gaussian particles, and other geometric primitives are discretized into coarse-grained Gaussian particles. In dynamic scenes, the Gaussian particle parameters of geometric primitives are adjusted to simulate deformation and generate dynamic scene models.

[0010] Optionally, the step of inputting the characteristic image of the building appearance and the data collected by the mechanical sensor on the target ancient building into a preset differentiable physical perception network to obtain the physical parameters of each geometric primitive includes: Collecting mechanical data of each mechanical sensor, wherein the mechanical sensor is arranged at a position corresponding to each geometric primitive on the target ancient building; Acquire a sub-image of each geometric primitive based on the building appearance feature image; The mechanical data and sub-image corresponding to each geometric primitive are input into a preset differentiable physical perception network to obtain the physical parameters of each geometric primitive, wherein the physical parameters include at least Young's modulus, Poisson's ratio, reflectivity, density and yield stress.

[0011] Optionally, the step of fusing each geometric primitive in the scene model with a corresponding physical parameter includes: Based on the particle-grid bidirectional coupling mechanism of PhysGaussian, the physical parameters corresponding to each geometric primitive are matched with Gaussian particles; The mechanical and optical characteristics corresponding to each Gaussian particle are determined according to the physical parameters.

[0012] Optionally, the use of the PhysGaussian modeling framework to perform dynamic simulation prediction on the scene model integrating physical parameters includes: Based on the explicit time integration in the Material Point Method physical simulation method and the local affine transformation assumption in the PhysGaussian modeling framework, each Gaussian particle in the scene model is dynamically evolved; A differentiable simulation framework is used to dynamically adjust the mechanical and optical characteristics of each Gaussian particle by back-propagating simulation errors. Dynamic simulation prediction is performed based on the updated Gaussian particles to obtain a simulation model.

[0013] Optionally, the rendering visual effect generates a digital twin model, including: PhysGaussian Splatting technology is used to render the visual effects of the simulation model, and key geometric primitives are highlighted to generate a digital twin model of the target ancient building.

[0014] Optionally, the method further includes: Adjusting the mechanical characteristics and optical characteristics of each Gaussian particle according to the simulation parameters input by the user, wherein the simulation parameters include at least wind speed and earthquake intensity; Generate a simulation model based on the adjusted Gaussian particles; Extract the mechanical indicators of each structure in the simulation model, and generate alarms and protection plans when the mechanical indicators exceed the preset strength threshold.

[0015] Optionally, the scene model includes at least a model of the target ancient building, and models of vegetation and water systems in the target ancient building area.

[0016] Another aspect of the present invention provides a digital twin ancient building construction system based on PhysGaussian, which executes the digital twin ancient building construction method based on PhysGaussian described in the above aspect.

[0017] The present invention discloses a method and system for constructing digital twin ancient buildings based on PhysGaussian, which uses a differentiable optimization algorithm to fit geometric primitives, uses a PhysGaussian modeling framework to construct a scene model of a target ancient building, and dynamically simulates and predicts the scene model integrating physical parameters, and renders visual effects to generate a digital twin model. The present invention integrates the geometry, material properties and environmental interaction data of ancient buildings through Gaussian kernels to achieve "twins are real", which is suitable for complex structures and fragile materials of ancient buildings, supports preventive protection and restoration decisions, and supports real-time updates of urban-level ancient building complexes.

[0018] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the disclosure, nor is it intended to limit the scope of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The above and other objects, features and advantages of the present disclosure will become more apparent through a more detailed description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present disclosure.

[0020] Figure 1 A schematic diagram of a process of constructing a digital twin ancient building based on PhysGaussian provided by an embodiment of the present invention; Figure 2 An implementation provided for an embodiment of the present invention Figure 1 Schematic diagram of the process of step S100; Figure 3 An implementation provided for an embodiment of the present invention Figure 1 Schematic diagram of the process of step S300; Figure 4 An implementation provided for an embodiment of the present invention Figure 1 Schematic diagram of the process of step S400; Figure 5 An implementation provided for an embodiment of the present invention Figure 1 Schematic diagram of the process of step S500; Figure 6 An implementation provided for an embodiment of the present invention Figure 1 Schematic diagram of the process of step S600; Figure 7 A schematic flow chart of another method for constructing a digital twin ancient building based on PhysGaussian provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0022] As used herein, the term "including" and its variations mean open inclusion, i.e., "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "based at least in part on". The terms "an example embodiment" and "an embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one additional embodiment". The terms "first", "second", etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0023] Figure 1A schematic diagram of a process of constructing a digital twin ancient building based on PhysGaussian provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method comprises the following steps: Step S100: Acquire point cloud data and building appearance feature images within the target ancient building area.

[0024] The embodiment of the present invention obtains multi-source data of the target ancient building area through multiple collection methods, including at least point cloud data, RGB-D depth image and multi-spectral microscopic detection data.

[0025] In one embodiment disclosed in the present invention, Figure 2 As shown, step S100 includes the following sub-steps: Step S101: Use a drone equipped with a high-precision laser radar to scan the target ancient building to obtain basic point cloud data of the basic structure and dense point cloud data of the complex structure.

[0026] The drone platform equipped with a high-precision laser radar system is used to conduct a full-scale aerial scan of the target ancient building to obtain the three-dimensional spatial distribution information of the building's external structure. The laser radar system used has centimeter-level spatial resolution and fast point cloud imaging capabilities, and can continuously scan the building and surrounding objects at multiple angles and multiple flight paths to generate basic point cloud data with wide coverage.

[0027] In order to obtain high-density 3D information of complex building structures (such as eaves, brackets, window lattices and other fine components), the specific area can be scanned multiple times in low-altitude hovering mode, and dense point cloud data can be constructed in combination with structured light or laser triangulation modules. The dense point cloud data is spatially registered and fused with the basic point cloud data, so that the overall point cloud data has both global structural coverage and detail accuracy.

[0028] During the data collection process, the system uses an inertial navigation unit and GPS to jointly locate and ensure the spatial consistency of the point cloud data. To improve the scanning accuracy, ground control points can also be used for later point cloud calibration. This method can fully record the overall form of the target ancient building and provide an accurate three-dimensional foundation for subsequent geometric segmentation and modeling.

[0029] Step S102: using a high-definition camera to collect an RGB-D depth image of the surface of the target ancient building.

[0030] A high-resolution RGB-D camera system is used to collect close-up images of the target ancient building surface. The RGB-D camera system can simultaneously obtain visible light images (RGB) and depth maps (Depth), where the RGB image records the real texture and color information of the building surface, and the depth map records the distance information from each pixel to the sensor, forming a color dot matrix surface data.

[0031] In actual acquisition, operators can hold or install RGB-D equipment on a ground mobile platform or lifting frame to perform a full-scale scan of the building's facade, eaves, roof edges and other areas in a close manner. To ensure image quality, multi-angle and multi-time acquisition methods can be used, and SLAM (simultaneous localization and mapping) technology can be combined for image sequence registration.

[0032] The RGB-D images collected in this step not only record the real texture and color of the wooden structure surface, but also accurately reflect its surface microstructural features such as carved patterns, paint peeling, weathering cracks, etc. This information will be used to assist in determining the material properties and surface state of each geometric element in the subsequent physical perception modeling step, thereby improving the authenticity of physical simulation and rendering.

[0033] Step S103: Detecting microscopic data of the target ancient building based on the multispectral imaging system.

[0034] Microscopic data includes at least wood aging, internal insect infestation, and water damage.

[0035] A multispectral imaging system is used to scan the target ancient buildings in detail to detect the material degradation and disease distribution. The imaging system used integrates infrared, ultraviolet, near-infrared and visible light band channels, and realizes the identification of different materials and defects through multi-band response analysis.

[0036] Multispectral imaging can reveal microscopic changes that cannot be directly observed on the surface of wood. For example, the degree of lignin degradation in wood can be determined through the light absorption characteristics of the ultraviolet band; structural abnormalities caused by internal cavities or insect infestations can be identified using the reflection characteristics of the near-infrared band; and wood water absorption areas can be identified through humidity response imaging in the infrared and short-wave infrared bands, and water seepage paths and possible structural hazards can be analyzed.

[0037] To improve the acquisition accuracy, it can be equipped with an automatic calibration lens and a stable gimbal, and combined with point cloud data for real-time positioning to ensure that the multispectral image is accurately aligned with the RGB-D image and point cloud in space, forming a complete "geometry-texture-material" three-dimensional information fusion.

[0038] Step S200: using a preset geometric segmentation network to identify geometric primitives from point cloud data.

[0039] Through deep learning-based geometric segmentation methods and differentiable optimization technology, intelligent structural analysis of ancient building point cloud data is performed, geometric primitives are identified and an accurate initial structural model is constructed, laying a geometric foundation for subsequent physical modeling and simulation.

[0040] In one embodiment disclosed in the present invention, this step can be implemented in the following manner: Use PointNet++ or DGCNN network structures to process input point cloud data. PointNet++ can extract multi-scale local features in point clouds by constructing hierarchical local regions, which is suitable for capturing local changes in building components such as tiles and brackets; DGCNN (Dynamic Graph CNN) can dynamically construct a k-nearest neighbor graph between points and perform convolution operations on the graph structure. It is suitable for expressing complex local relationships between points in a point cloud.

[0041] The network input is the multi-source fused point cloud obtained in the previous steps, and the output is the geometric primitive label to which each point belongs. Points of the same category are aggregated through post-processing (such as K-means clustering, region growing, etc.) to extract complete geometric primitives.

[0042] At the same time, a multi-scale feature fusion mechanism is introduced in the network training stage to jointly model the local scale (such as the curved features of the sharp corners of the brackets) and the global scale (such as the ridge, beams and columns) to improve the segmentation accuracy in complex ancient building environments.

[0043] Step S300: fitting geometric primitives using a differentiable optimization algorithm, and building a scene model of the target ancient building using a PhysGaussian modeling framework.

[0044] In one embodiment disclosed in the present invention, Figure 3 As shown, the geometric primitives are fitted using a differentiable optimization algorithm, which includes the following sub-steps: Step S301: using a differentiable optimization algorithm to adjust the parameters of each geometric primitive with the goal of minimizing the geometric error between the original point cloud data and the constructed model.

[0045] The segmented geometric primitive regions are further fitted using a differentiable optimization algorithm. A parametric geometric template (such as a cylinder with parameters including radius, height, and direction vector) is introduced, and the objective function is to minimize the geometric error (such as the Euclidean distance) between the template surface and the original point cloud.

[0046] For example, the distance from each point to the geometric primitive model is defined as the loss term, and the parameters of each primitive are dynamically optimized through the back-propagation algorithm, including the center coordinate position, size parameters (such as length, width, height, radius, etc.), spatial orientation (such as rotation angle), and optional offset and local deformation tensor.

[0047] Differentiable optimization algorithms can be seamlessly integrated with deep learning segmentation networks, allowing the construction of end-to-end structure reconstruction systems and achieving full-process automatic modeling from point clouds to parseable geometric models.

[0048] Step S302: Fit all geometric primitives into an initial structural model based on Boolean operations.

[0049] After fitting all the basic geometric bodies, they are spliced ​​into a continuous initial structure model through Boolean geometric operations.

[0050] Boolean operations include: 1. Union: connect adjacent components into a complete roof, corridor or other overall structure; 2. Intersection: extract overlapping areas to determine the nesting or connection of the structure; 3. Difference: remove interfering data or trim unnecessary component fragments.

[0051] This process ensures that the initial structural model is continuous and consistent in spatial structure, avoiding gaps, overlaps or unreasonable insertions between components.

[0052] A geometric topology consistency verification mechanism is also introduced in the Boolean combination to ensure that the model does not have topological loopholes (such as "hanging faces" and "self-intersecting edges" and other problems), thereby improving the usability of the model for subsequent physical simulation and rendering.

[0053] Step S303: Smoothing the initial structural model using deformation field interpolation technology to obtain the overall structural model of the target ancient building.

[0054] By constructing a continuous deformation field, the initial structural model is calibrated and smoothed in detail, so that it has good visual continuity and physical simulability while maintaining the accuracy of the geometric structure.

[0055] The deformation field construction methods include the following methods: 1. Local deformation interpolation based on sparse control points: Select several key control points on the initial model, calculate the offset vectors between these points and the original point cloud, and propagate the deformation to the overall model through RBF or MLS interpolation technology.

[0056] 2. Physically driven detail fitting: By constructing a local rigid / flexible transformation field, specific components such as tile edges and curved brackets can be elastically adjusted.

[0057] 3. Deep residual correction network: A lightweight neural network is introduced to perform learning correction on the model residuals to improve the accuracy and naturalness of fitting.

[0058] The final output is an overall structural model with complete topology, smooth details, and minimal error, which serves as the input for subsequent PhysGaussian modeling.

[0059] In one embodiment disclosed in the present invention, after completing the point cloud geometry modeling of the target ancient building, in order to support the structural evolution modeling of the building in a dynamic environment (such as wind, vibration, aging, earthquake and other factors), the PhysGaussian modeling framework is further introduced, and the geometric primitives are discretized into multi-scale Gaussian particles. The geometric state is dynamically updated through the incremental Gaussian evolution algorithm, thereby constructing a digital twin scene model that adapts to real-time changes.

[0060] The PhysGaussian modeling framework is used to build the scene model of the target ancient building. Figure 3 As shown, this can be achieved using the following sub-steps: Step S304: using the incremental Gaussian evolution algorithm in the PhysGaussian modeling framework to discretize each geometric primitive in the overall structural model of the target ancient building into Gaussian particles.

[0061] The preset key geometric primitives are discretized into fine-grained Gaussian particles, and other geometric primitives are discretized into coarse-grained Gaussian particles.

[0062] The overall structural model of the target ancient building outputted in the above steps is used as input, and the PhysGaussian modeling framework is introduced. Its incremental Gaussian evolution algorithm is used to convert all geometric primitives in the structural model into Gaussian particle representations that can be dynamically updated. The main process is as follows: 1. Particle Discretization of Geometric Primitives 1. For each geometric primitive in the structural model (such as beams, columns, roof ridges, tiles, etc.), construct its three-dimensional geometric parameters (center position, direction, scale).

[0063] 2. Granularity classification is performed according to the importance and complexity of the primitives. For example, key geometric primitives (such as load-bearing columns, bracket nodes, and roof interfaces) are discretized into fine-grained Gaussian particles (Gaussian Mixture with small covariance) to capture structural micro-variations; secondary geometric primitives (such as walls, platforms, and window sills) are discretized into coarse-grained Gaussian particles to reduce computational redundancy.

[0064] 3. Each Gaussian particle is represented as a three-dimensional position vector and a covariance matrix, with additional weights to represent the physical importance of the particle.

[0065] 2. Incremental Gaussian Evolution Initialization 1. Based on the initial state of the structural model and environmental response, a particle collection is established as the initial scene in a static state.

[0066] 2. Spatial constraint graphs (such as particle connection tension and damping) can be introduced between particles to model interactions.

[0067] 3. Introduce parameterized physical models (such as Young's modulus, density, material properties) to support physical simulation.

[0068] Step S305: In a dynamic scene, the Gaussian particle parameters of the geometric primitives are adjusted to simulate deformation and generate a dynamic scene model.

[0069] In order to support the continuous evolution of ancient building structures under changes in the external environment or human intervention, this step is based on the incremental Gaussian evolution algorithm to update the state of each Gaussian particle in real time in the dynamic scene and build a complete dynamic simulation model. The specific technical details are as follows: 1. External input perception mechanism Introduce a variety of external dynamic input sources, such as wind speed / direction sensors, temperature and humidity changes, earthquake simulation data, structural strain sensors, or time-series video reasoning, etc. Map the external input into a physical perturbation function acting on the particle collection to drive the particle deformation.

[0070] 2. Gaussian particle state update For each time step, the particle position and covariance are updated based on the perturbation function. The basis for the update may include the particle's elastic response, inertia terms, external force effects, etc., to simulate real physical changes such as deformation, shaking, and warping.

[0071] 3. Continuous modeling and backtracking mechanism All particle states are saved in the time series cache to form a time series model. The current scene state can be visualized in real time, and retrospective analysis or evolution prediction of a certain time period can also be supported. At the same time, the intermediate state can be output for application scenarios such as structural safety warning and material aging assessment.

[0072] 4. Reconversion from particle to structure model At any time step, the Gaussian field to mesh reconstruction algorithm (such as iso-surface extraction or voxel reconstruction) can be used to restore the deformed three-dimensional model of the current structure.

[0073] Step S400: Input the characteristic image of the building appearance and the data collected by the mechanical sensors on the target ancient building into a preset differentiable physical perception network to obtain the physical parameters of each geometric primitive.

[0074] In order to improve the physical simulability and simulation accuracy of the digital twin model of the target ancient building in real scenes, an embodiment of the present invention pre-constructs a differentiable physical perception network for automatically inferring the physical property parameters of each geometric element of the target ancient building from multimodal input data (including building appearance images, RGB-D images, mechanical sensor data, etc.), and the physical parameters include at least optical parameters and mechanical parameters.

[0075] In one embodiment disclosed in the present invention, Figure 4 As shown, step S400 includes the following sub-steps: Step S401: collecting mechanical data of each mechanical sensor, where the mechanical sensor is arranged at each geometric primitive on the target ancient building.

[0076] Mechanical sensor modules are installed at each key geometric element (such as columns, beams, and bracket nodes). For example, stress gauges are arranged on the beams to monitor bending stress; axial stress / compression stress gauges are arranged on the columns; and seismic response accelerometers are arranged at the column bases. Temperature and humidity / wet expansion strain gauges can also be added to some nodes to assist in modeling.

[0077] The data sampling frequency is not less than 100Hz, and mechanical indicators such as compressive stress, tensile stress, and displacement response can be continuously recorded. The collected content includes long-term stress data and short-term dynamic impact response under static load, and each sensor data records the physical response value in the three-axis direction. The collected data needs to be synchronized with the image data through time and space alignment, that is, all sensor data are aligned with the image data through timestamps, and the system automatically marks the geometric primitive number corresponding to each data.

[0078] Step S402: acquiring a sub-image of each geometric primitive based on the building appearance feature image.

[0079] The building appearance image is associated with the mechanical data collected in step S401, and the local image segments corresponding to each geometric primitive are extracted as part of the network input. For details, please refer to the following steps: 1. Image segmentation and annotation Use semantic segmentation networks (such as DeepLabV3+) to identify the regions of geometric primitives in the original RGB-D image. Automatically generate sub-image regions in the image and associate primitives. Use fusible 3D depth maps to supplement geometric contour information.

[0080] 2. Feature enhancement processing: Each sub-image is subjected to illumination normalization, edge enhancement, and crack texture high-pass filtering, and auxiliary labels (such as material type, historical restoration traces, etc.) are embedded in the input image as auxiliary input items of the network.

[0081] Step S403: Input the mechanical data and sub-image corresponding to each geometric primitive into a preset differentiable physical perception network to obtain the physical parameters of each geometric primitive.

[0082] The physical parameters include at least Young's modulus, Poisson's ratio, reflectivity, density and yield stress, among which Young's modulus can measure the stiffness of the material; Poisson's ratio can reflect the lateral deformation rate; optical reflectivity can be used for realistic rendering; density and yield stress can be used for physical simulation.

[0083] The mechanical data and sub-images in the previous steps are input into the preset differentiable physical perception network DPPN to complete the estimation of the physical properties of each geometric primitive.

[0084] Each primitive outputs a set of physical parameter vectors, among which Young's modulus and Poisson's ratio are used for structural finite element modeling, reflectivity is used for realistic rendering, and density / yield stress is used for dynamic response simulation.

[0085] All output physical properties will be synchronously mapped back to the physical simulation model to support real-time simulation based on physical properties (such as simulating the response of ancient buildings under earthquakes and wind loads). It can also be used for applications such as structural safety assessment and disaster warning.

[0086] Step S500: Fusing each geometric primitive in the scene model with the corresponding physical parameter.

[0087] The embodiment of the present invention introduces a particle-grid bidirectional coupling mechanism based on PhysGaussian. This mechanism can accurately bind the physical parameters of each geometric primitive obtained through the differentiable physical perception network to the Gaussian particles represented by PhysGaussian, realizing the unified expression of geometric structure and mechanical behavior. Ultimately, it can support the simulation of the real response of ancient building structures under wind loads and earthquakes, including plastic deformation, stress distribution and even local fracture prediction.

[0088] In one embodiment disclosed in the present invention, Figure 5 As shown, step S500 may include the following sub-steps: Step S501: Based on the particle-grid bidirectional coupling mechanism of PhysGaussian, the physical parameters corresponding to each geometric primitive are matched with Gaussian particles.

[0089] For the target ancient building structure model constructed by the Gaussian evolution algorithm (each geometric primitive has been discretized into Gaussian particles of a certain density), the embodiment of the present invention introduces a particle-grid bidirectional coupling mechanism to achieve the following functions: 1. The input data is a set of Gaussian particles, each of which contains its spatial position, covariance matrix, etc.; a set of physical parameters of the geometric primitives to which each particle belongs; and the external environmental load field (such as wind pressure distribution function, earthquake ground acceleration time history).

[0090] 2. Binding mechanism The physical parameters are assigned to the corresponding Gaussian particles, among which Young's modulus and Poisson's ratio determine the displacement stiffness of the particles during the elastic response process; the density determines the mass matrix term; the yield stress is used for plastic deformation threshold control; and the optical reflectivity is used for image rendering consistency.

[0091] The particle distribution information is mapped onto a continuous physical simulation grid (Grid-based FEM or MPM grid) using the weighted projection function in the PhysGaussian framework. The projection uses a Gaussian kernel function (such as RBF or SPH kernel), and the grid node attributes can be obtained by particle weighted averaging to obtain the local stiffness, mass, and damping matrix.

[0092] Particle-to-grid mapping (P2G) is used to initialize the physical field; and grid-to-particle feedback (G2P) is used to return the physical response results after the load is applied to update the particle state (position, deformation, and damage mark).

[0093] 3. During the binding process, boundary constraints (such as fixing the root of the building) and contact stiffness constraints of adjacent particles are introduced to avoid particle penetration or non-physical movement during the simulation.

[0094] Step S502: Determine the mechanical characteristics and optical characteristics corresponding to each Gaussian particle according to the physical parameters.

[0095] According to the physical parameters corresponding to each Gaussian particle, its mechanical characteristic parameters (e.g., mass density, inertia tensor, connection stiffness) and optical characteristic parameters (e.g., color, transparency, luminescence factor) are determined.

[0096] Step S600: Use the PhysGaussian modeling framework to perform dynamic simulation prediction on the scene model that integrates physical parameters, and render the visual effects to generate a digital twin model.

[0097] In one embodiment disclosed in the present invention, Figure 6 This step includes the following sub-steps: Step S601: Based on the explicit time integration in the Material Point Method physical simulation method and the local affine transformation assumption in the PhysGaussian modeling framework, each Gaussian particle in the scene model is dynamically evolved.

[0098] The dynamic simulation prediction is first based on the PhysGaussian modeling framework, integrating the explicit time integration strategy in Material PointMethod (MPM) to gradually evolve each Gaussian particle representing the structure of the ancient building. This step not only regards each Gaussian particle as a geometric representation of the three-dimensional shape, but also gives it physical properties (including mass, stress tensor, deformation gradient, etc.), realizing a unified representation of geometry and physics.

[0099] In each frame of time advancement, the system performs the following processing on particles according to the MPM method: 1. Transfer the physical quantities (mass, momentum) on the particles to the background grid nodes.

[0100] 2. Perform mechanical calculations on the mesh nodes and update the speed.

[0101] 3. Map the result back to the particle and update its position and deformation gradient.

[0102] 4. Synchronously update the particle covariance matrix to support subsequent rendering.

[0103] This step supports modeling of complex physical behaviors such as elasticity, plasticity, and fracture. For example, in fracture simulation, the Drucker-Prager plasticity model is used to describe the nonlinear mechanical behavior of masonry materials, and the yield surface and plastic return mapping are defined. When the difference in the accumulated plastic strain of the particles exceeds the set threshold, the particle separation mechanism is triggered to simulate the crushing of bricks.

[0104] In order to improve the computational efficiency, a multi-resolution particle strategy is adopted: high-resolution particles (such as resolution r p <0.01m, while in the static background area coarse resolution particles (such as r p >0.1m).

[0105] The entire simulation is executed in parallel in a GPU-accelerated environment. CUDA is used to implement parallel computing of key MPM modules (including particle-mesh mapping, stress update, feedback, etc.). Combined with an explicit time advancement strategy, real-time simulation updates at 30 frames per second are achieved, supporting user interactive operations and maintaining stable graphics response.

[0106] Step S602: using a differentiable simulation framework, dynamically adjusting the mechanical and optical characteristics of each Gaussian particle by back-propagating simulation errors.

[0107] In order to further improve the simulation accuracy of digital twin ancient buildings, a differentiable physical simulation framework is introduced, and the mechanical and optical parameters are incorporated into the joint optimization process to achieve dynamic matching of simulation behavior and real data.

[0108] In terms of mechanical parameters, the system supports adaptive calibration of elastic modulus (such as Young's modulus of wood E=8.5GPa), density, yield stress, friction angle and cohesion.

[0109] 1. Apply external loads or interactive forces and use motion capture to obtain real response data (such as beam and column deformation and tile sliding trajectory).

[0110] 2. Construct a loss function to evaluate the deviation between simulation and reality.

[0111] 3. Use automatic differentiation tools to back-propagate the loss function and update the physical parameters so that the simulation results gradually approach historical observations.

[0112] 4. In the fracture simulation scenario, the crack path and fragment distribution generated in the simulation are compared with the actual post-earthquake photos, and the Drucker-Prager parameter group is fine-tuned through differentiable optimization.

[0113] In terms of optical features, photometric stereo vision technology is used to invert the parameters of the object surface: 1. After the multi-angle RGB image is input, the reflectivity and normal direction are reversed.

[0114] 2. The optical inversion results are used to optimize the coefficients of the spherical harmonic basis functions and reconstruct the illumination response of the Gaussian basis elements.

[0115] 3. Using differentiable rendering technology, the difference between the real image and the rendered image is used as the loss source, and the optical properties of Gaussian particles, such as roughness and metallicity, are optimized by backpropagation.

[0116] Mechanical simulation and visual reconstruction form a closed-loop optimization process to achieve high-fidelity digital twin modeling driven by historical data.

[0117] Step S603: Perform dynamic simulation prediction based on the updated Gaussian particles to obtain a simulation model.

[0118] Based on the optimized Gaussian particle set, the system constructs a simulation model with high realism, and performs dynamic prediction and visualization.

[0119] In one embodiment disclosed in the present invention, rendering visual effects to generate a digital twin model can be implemented in the following manner: PhysGaussian Splatting technology is used to render the visual effects of the simulation model, and key geometric primitives are highlighted to generate a digital twin model of the target ancient building.

[0120] After the dynamic simulation is completed, the PhysGaussian Splatting technology is further used to perform high-quality visual rendering of the simulation results, and to realize the construction and presentation of the digital twin model of the target ancient building.

[0121] The PhysGaussian Splatting technology can accurately present the complex optical effects of ancient buildings, such as high light reflection and sub-surface scattering, and combine spherical harmonic lighting and real-time shadow functions to vividly restore the historical style of ancient buildings. At the same time, through multi-dimensional data mapping, physical information such as stress field and displacement field are encoded into the transparency and color gradient of the Gaussian kernel, so as to intuitively display the mechanical state of key structures of ancient buildings (such as beams and columns), so as to facilitate the timely discovery of potential risk points.

[0122] PhysGaussian Splatting is a real-time differentiable rendering method based on three-dimensional Gaussian primitives. Its core is to project each physically simulated Gaussian particle directly onto the image plane, and accumulate and synthesize color and depth in the form of Gaussian kernel to generate a continuous, smooth, and mesh artifact-free visual image.

[0123] In the rendering process, each Gaussian particle carries its position, covariance matrix, color, normal, roughness and other attributes, which are encoded into a rendering parameter vector with more than six dimensions. The system performs perspective projection on the particles according to the camera's perspective and calculates the screen space shape (elliptical outline) and transparency weight of each particle.

[0124] The GPU-accelerated Alpha Blending algorithm is used for forward synthesis to generate realistic simulation images. At the same time, physical information such as the stress state and deformation gradient of the particles is visually mapped. For example, high-stress areas are highlighted in red and fracture edges are faded with transparency, allowing users to identify weak structural locations at a glance.

[0125] For geometric components with important significance (such as brackets, eaves, load-bearing columns, etc.), the system uses visual labels and high-resolution particles to focus on display, setting their Gaussian primitives to fine granularity (such as resolution r p <0.01m), enhancing rendering resolution and interactive experience.

[0126] Through the above rendering and display process, the system finally generates a digital twin model with real geometric appearance, dynamic physical response, and visual interactive control. This model can not only be used to display the structure and form of ancient buildings, but also support multiple application scenarios such as structural stability analysis, disaster warning assessment, historical state restoration, and virtual protection teaching.

[0127] The digital twin model has the following key features: 1. Real-time response and interactive control Users can directly apply disturbances (such as wind load, earthquake intensity adjustment, dragging objects, etc.) through interface controls, and convert these inputs into force terms in the simulation domain in real time to trigger physical response simulation. The updated particle position and deformation parameters will be updated in each frame to ensure that the feedback delay is controlled within 33ms (above 30FPS).

[0128] 2. Fault and disaster prediction Combined with the fracture mechanism, the simulation model can simulate the sliding of roof tiles, the plastic flow of beams and columns, and the crushing of bricks under impact during earthquakes. The prediction results are compared with historical damage patterns, and a risk analysis report is automatically generated, with recommended structural reinforcement solutions.

[0129] 3. Multimodal data fusion and display The geometric, optical, and mechanical information are stored in a unified manner through the Gaussian kernel and mapped into visual variables such as color, transparency, normal maps, etc. in the rendering pipeline, achieving a seamless transition from physical simulation to visual expression.

[0130] 4. Intelligent protection suggestions Combined with dynamic prediction results, it can automatically analyze the safety threshold of ancient buildings according to different disaster conditions, and put forward protection and restoration suggestions (such as reinforcement of specific nodes), providing technical support for cultural relics restoration work.

[0131] Through the digital twin model with Gaussian particles as the core, not only dynamic evolution and predictive simulation are realized, but also high-precision rendering and interactive feedback are integrated to build a digital twin system of ancient buildings that can be operated in real time, sustainably optimized, and visually analyzed.

[0132] The following is a specific embodiment of the simulation process: 1. Calibration of wood elastic modulus The elastic modulus of wood is E=8.5GPa, which simulates the slippage of roof tiles and the plastic deformation of beams and columns under earthquake action to optimize the structural reinforcement plan.

[0133] 2. Optical parameter analysis Through photometric stereo vision technology, the reflectivity of the object surface is inferred from the multi-light source RGB image. With normal direction , optimizing the spherical harmonic coefficients of Gaussian kernel combined with differentiable rendering.

[0134] 3. Mechanical parameter calibration: Apply known external forces to the elastic body and invert the Young's modulus from the motion capture data The strain-stress relationship is calculated using the incremental deformation of PhysGaussian, and the iterative optimization is performed until the error is minimized.

[0135] The purpose of calculating the reflectivity again here is to further verify and optimize the existing calculation results and ensure that the actual simulation is highly consistent with the surface characteristics of the building. In addition, by comparing the reflectivity output by manual calibration with that by the physical perception network, the error of the model in processing complex surface reflections can be effectively reduced, thereby improving the overall simulation accuracy.

[0136] 4. Parameter prediction and simulation (1) Fracture simulation example: Application of the Drucker-Prager plasticity model Technical principles and processes: a. Model initialization Model the bricks of the ancient building as a dense collection of Gaussian particles, each particle has a bound mass , initial deformation gradient (initial identity matrix) and Drucker-Prager plasticity parameters (friction angle , cohesion ).

[0137] Plasticity returns mapping according to definition , calculate the plastic strain evolution of the particles under yield conditions.

[0138] b. Impact load application Set dynamic boundary conditions in the simulation domain to apply instantaneous impact force to the brick surface , simulating external shocks (such as impact or seismic loads).

[0139] Calculate the velocity updates of the mesh nodes through the explicit time integration of the MPM , and passed to the particle to update its position and deformation gradient .

[0140] c. Fracture process simulation When a particle undergoes a large deformation, the Drucker-Prager yield condition is triggered and the elastic deformation gradient is updated. And mark the plastic strain accumulation.

[0141] Through the particle separation algorithm, when the difference in plastic strain between adjacent particles exceeds a threshold, the fracture effect is triggered, splitting the particle cluster into independent groups, simulating the brick-breaking effect.

[0142] d. Parameter optimization and historical data matching Based on the actual fracture mode (such as crack path, fragment distribution), construct a loss function Simulation Results - Historical Data .

[0143] Back-propagating gradients to plasticity parameters via a differentiable simulation framework and , using the gradient descent method to iteratively optimize until the error between the simulation results and the historical data is minimized.

[0144] (2) Real-time interactive simulation: GPU-accelerated MPM and Gaussian dynamic update Technical implementation steps: a. GPU parallel architecture design The grid calculation and particle state update of MPM are decomposed into CUDA kernel functions, and the parallel computing capability of GPU is used to accelerate key steps (such as mass / momentum transfer from particles to grids and stress calculation).

[0145] Through memory optimization (such as shared memory cache, asynchronous data transfer), the latency between GPU cores is reduced to ensure that the simulation frame rate is stable at 30 FPS.

[0146] b. Real-time deformation update of Gaussian primitives At each time step , according to the deformation gradient updated by the particle , dynamically calculate the covariance matrix of the Gaussian kernel .

[0147] Through incremental evolution, the global deformation gradient accumulation error is avoided, supporting the numerical stability of long-term simulation.

[0148] c. User interaction and dynamic response Interactive force input: Capture the external force applied by the user through UI controls (such as mouse dragging, touch screen gestures) , which is mapped into external force terms in the simulation domain .

[0149] Real-time feedback: In the GPU pipeline, the particle position and deformation parameters It is passed to the rendering pipeline in real time and the current frame image is generated through 3D Gaussian Splatting, ensuring that the delay between user operation and visual feedback is less than 33ms (30 FPS).

[0150] The core role of PhysGaussian in simulation is reflected in the following aspects: 1. Unified physical-geometric representation: Gaussian kernels carry both geometric shapes (covariance matrices ) and physical properties (stress , deformation gradient ), eliminating the “simulation grid Rendering Mesh" data conversion.

[0151] 2. Efficient dynamic evolution: Through incremental deformation updates and GPU-accelerated MPM, real-time simulation of complex physical phenomena (fracture, plastic flow) is achieved.

[0152] 3. Closed-loop parameter optimization: The differentiable framework supports end-to-end optimization from data to parameters, ensuring the high fidelity and predictive ability of the digital twin model.

[0153] In one embodiment disclosed in the present invention, Figure 7 As shown, the digital twin ancient building construction method based on PhysGaussian also includes the following steps: Step S700: adjusting the mechanical characteristics and optical characteristics of each Gaussian particle according to the simulation parameters input by the user.

[0154] The simulation parameters include at least wind speed and earthquake intensity.

[0155] The embodiment of the present invention provides a parameter input module based on a human-computer interaction interface, allowing users to set disaster simulation conditions in real time. A graphical interactive panel is used to provide controls such as sliders, numeric input boxes, selectors, etc., to support users to input parameters such as wind speed level, earthquake intensity, and flood water level. The input parameters will be mapped to corresponding physical simulation boundary conditions in real time. For example: the wind speed parameter will be converted into a wind pressure distribution that varies with height; the earthquake intensity is mapped to a ground acceleration time-history function and loaded to the boundary between the foundation and the bottom of the structure; the flood water level will affect the buoyancy field of the particles and the impact surface pressure distribution.

[0156] After loading the physical field, the system dynamically adjusts the mechanical characteristic parameters (including mass density, inertia tensor, connection stiffness) and optical characteristic parameters (such as color, transparency, and luminescence factor) of each Gaussian particle according to the intensity of the disaster. For example, when the structure enters a disaster state, the color of the corresponding particle gradually changes from green to yellow or even red, so as to provide real-time feedback on the stress state in the visual rendering; at the same time, the transparency will increase according to the local strain to assist in identifying the critical damage area.

[0157] Step S800: Generate a simulation model based on the adjusted Gaussian particles.

[0158] The embodiment of the present invention performs real-time dynamic simulation and builds a complete digital simulation model based on the incremental evolution simulation engine of PhysGaussian.

[0159] The fusion algorithm combining Material Point Method (MPM) and Gaussian Splatting is used to construct the particle field evolution equation, and the GPU parallel architecture is used to calculate the movement and interaction of particles under external forces. The deformation, crack development, node detachment and other processes of the structure in the particle system can be simulated in real time, and the simulation frame rate can reach more than 30FPS to ensure interactive response.

[0160] For areas where key structures (such as brackets, beams, columns, and eaves) are located, the simulation accuracy is enhanced by increasing the particle density (reducing the particle radius) and precision parameters to reflect the development of detailed-level damage.

[0161] During the physical simulation, the optical characteristics of particles (such as color) change in real time, reflecting the evolution of the mechanical state. At the same time, PhysGaussian Splatting is used for continuous real-time rendering to achieve dynamic disaster demonstration and interactive analysis in a three-dimensional view.

[0162] Step S900: extracting mechanical indicators of each structure in the simulation model, and generating an alarm and a protection plan after the mechanical indicators exceed a preset strength threshold.

[0163] 1. Mechanical index extraction and safety threshold determination: Continuously extract key mechanical indicators of each structural area in the simulation model, including maximum stress, maximum displacement, local strain, deformation gradient tensor, etc.

[0164] According to traditional material mechanics standards and ancient building protection specifications, failure criteria for different materials (such as wood, masonry, and concrete) are preset. For example, the Tresca yield criterion is used for wooden components; the Mohr-Coulomb fracture model is used for stone components; and the node connection is set to have a disengagement condition based on the limit displacement and friction slip model.

[0165] When the structural indicators reach 80% of the elastic limit, the system generates a potential risk warning; when the indicators exceed the ultimate strength, the system immediately triggers a red alarm and automatically enters the protection plan recommendation process.

[0166] A visual feedback mechanism is adopted. For example, in the 3D view, high stress areas are marked with color gradients (green→yellow→red) and V-shaped danger labels are superimposed to help users intuitively judge risky areas.

[0167] (2) Automatically generate protection scheme: Data clustering and deformation pattern recognition algorithms are used to classify and determine typical damage types that occur in the structure, such as bending buckling, shear cracks, loose mortise and tenon joints, and beam-column displacement.

[0168] Calling the ancient building protection knowledge base and rule reasoning engine, automatically matching the optimal protection suggestions under the current damage mode. For example, when an earthquake causes the mortise and tenon to come loose, it is recommended to install customized iron parts or carbon fiber coverings; when wind pressure causes the tiles to slide, it is recommended to repair with traditional mortar or antique polymer glue; when floods erode wooden columns, it is recommended to install waterproof isolation layers and wood sealing reinforcement agents.

[0169] The recommended measures are loaded back into the simulation model in a parameterized form, automatically simulating the response effect of the reinforced building under the same disaster scenario, and iterative optimization of the plan is achieved through indicator comparison, such as reducing the amount of reinforcement materials, adjusting the reinforcement location, etc.

[0170] Finally, a comprehensive report containing graphic descriptions, 3D models and budget estimates is generated, including: reinforcement locations and material specifications, repair construction steps and precautions, project cost estimates, and a "post-disaster performance preview" that supports interactive browsing with VR devices.

[0171] The following is a specific example of generating an early warning and protection plan: 1. Implementation mechanism of real-time prediction of safety threshold (1) Dynamic parameter input and simulation triggering a. User interface control design: Provide visual interactive panels (such as sliders, numerical input boxes), allowing users to dynamically adjust disaster parameters (such as wind speed level, earthquake intensity, flood level) and trigger simulation calculations in real time.

[0172] b. Parameter-physical field mapping: Convert input parameters into boundary conditions of the physical simulation engine. For example, wind speed is mapped into dynamic wind pressure distribution, and earthquake intensity is converted into ground acceleration time history curve, which is loaded into the building structure through the particle-field coupling model of PhysGaussian.

[0173] (2) Real-time physical simulation and index extraction a. GPU accelerated computing: Based on Material Point Method (MPM) and PhysGaussian incremental evolution algorithm, GPU parallel computing architecture is used to achieve real-time dynamic simulation (30FPS or more) of large-scale particle systems (such as bricks and tiles of ancient buildings, beams and columns).

[0174] b. Key safety indicator monitoring: Mechanical indicators: real-time extraction of maximum structural stress , displacement , local strain And other data.

[0175] Stability threshold: preset the failure criteria of ancient building materials (such as wood and stone) (such as Tresca yield condition and Mohr-Coulomb fracture criterion) and dynamically compare them with simulation data.

[0176] (3) Safety threshold determination and early warning Multi-level early warning mechanism: Yellow warning: When the key indicator reaches the elastic limit of the material (like ), indicating potential risks.

[0177] Red warning: When the indicator exceeds the material strength threshold (such as ), determining that the structure is about to fail, triggering an emergency protection recommendation.

[0178] Visual feedback: In the 3D Gaussian rendering interface, the color gradient (green yellow Red) High stress areas are marked in real time and superimposed Straight label.

[0179] 2. Technical logic for automatically generating protection solutions (1) Knowledge base and rule engine a. Engineering protection knowledge base: integrates cultural heritage protection standards (such as the Technical Specifications for the Maintenance and Reinforcement of Wooden Structures of Ancient Buildings), material mechanics databases (such as wood aging models, stone weathering coefficients), and historical restoration case libraries.

[0180] b. Rule reasoning engine: Based on the expert system and machine learning model, establish the mapping rules of "disaster type-structural damage-protection measures". For example: Earthquake causes mortise and tenon joints to come loose It is recommended to add iron reinforcement or carbon fiber cloth wrapping.

[0181] Wind load induced tile slippage It is recommended to use traditional mortar bonding or new antique waterproof glue for fixing.

[0182] (2) Protection plan generation process a. Damage pattern recognition: Identify the main damage types (such as bending deformation, shear cracks, and node loosening) through cluster analysis of simulation data.

[0183] b. Solution matching and optimization: Primary plan: Match the protection measures that match the current damage pattern from the knowledge base and sort them by priority (such as lowest cost and least impact on the original appearance).

[0184] Parametric optimization: Combine the geometric model of the ancient building with the physical parameters and verify the effectiveness of the solution through differentiable simulation. For example, simulate the stress distribution of the reinforced structure under the same disaster and iteratively optimize the reinforcement location and material usage.

[0185] c. Solution output: Generate graphic reports and 3D visualization instructions, including: Construction drawings: mark reinforcement points and material specifications (such as carbon fiber cloth thickness, anchor bolt model).

[0186] Budget assessment: Automatically estimate costs based on material unit prices and project quantities.

[0187] VR simulation: supports users to preview the performance of reinforced buildings in disaster scenarios through virtual reality devices.

[0188] In one embodiment disclosed in the present invention, the scene model at least includes a model of the target ancient building, and models of vegetation and water systems in the target ancient building area.

[0189] The embodiment of the present invention builds a highly coordinated digital twin platform by deeply integrating multi-source data and physical simulation technology, realizing unified modeling, dynamic visualization and interactive analysis of all elements of the city, including ancient buildings, vegetation and water systems. First, through multimodal data acquisition methods such as laser scanning, drone photography and hydrological sensors, the millimeter-level geometric details of ancient buildings, the three-dimensional root structure of vegetation and the dynamic flow information of the water system are accurately obtained, and these heterogeneous data are uniformly converted into a Gaussian primitive model based on PhysGaussian technology. Each Gaussian primitive not only encodes spatial distribution and geometric characteristics, but also embeds material properties (such as the elastic modulus of ancient building wood and the optical reflectivity of glazed tiles) and physical interaction parameters (such as the fixation of vegetation roots on the soil and the scouring intensity of floods on building foundations), thereby realizing the deep integration of physics and geometry under a unified framework.

[0190] At the visualization level, a multi-resolution rendering strategy is adopted: the carved brackets and weathered bricks and tiles of ancient buildings are realistically presented through high-precision Gaussian particles (resolution less than 0.01 meters), and the vegetation canopy and water surface are rendered with coarse-grained particles (resolution greater than 0.1 meters) to balance efficiency and realism. Through layered overlay technology, users can freely switch between geometric detail layers, physical field data layers (such as ancient building stress thermal maps, vegetation root stress distribution, flood velocity fields) and dynamic behavior layers (such as the swing of ancient trees in the wind, earthquake-induced building vibrations), and visually observe the complex interactions between multiple elements. For example, in flood simulations, water particles collide with the Gaussian primitives of the ancient bridge piers in real time, the system dynamically calculates the impact of scouring force on the bridge foundation, and maps the erosion risk level from blue to red through a color gradient, while marking the additional stress value of the roots of neighboring ancient trees on the foundation to form a comprehensive risk warning.

[0191] The interactive analysis function gives users the ability to dynamically control and support decisions. By dragging the controls, you can adjust the wind speed, water level or earthquake intensity in real time. The system is based on the GPU-accelerated Material Point Method and PhysGaussian incremental evolution algorithm to update the physical state of the entire scene in millisecond response. For example, when the user raises the flood level to 2 meters, the system not only simulates the impact of water on the ancient bridge piers, but also simultaneously calculates the attenuation of the root grip of the ancient trees along the coast due to soil saturation, and then predicts the risk of tilting of the ancient bridge. In response to these risks, the platform's built-in intelligent decision-making engine can automatically generate multiple sets of protection plans - such as strengthening bridge piers, transplanting some ancient trees or building diversion channels - and verify the effectiveness of each plan through differentiable simulation. Users can immersively preview the scenarios after the implementation of different plans in virtual reality: the peak stress of the bridge piers reinforced with carbon fiber decreases by 30% during floods, the stress distribution of the roots of transplanted ancient trees tends to be balanced, and the diversion channel diverts the flood away from the core area of ​​the building. All data are fed back in real time in the form of dynamic charts and three-dimensional heat maps, supplemented by cost estimates and original appearance protection assessments, providing managers with a decision-making basis that is both scientific and culturally sensitive.

[0192] It not only provides precise tools for the preventive protection of cultural heritage, but also reveals the deep connection between the natural and human environment through the collaborative simulation of all elements of the city. For example, the thousand-year stability of an ancient bridge not only depends on its own structure, but also is closely related to the soil-fixing effect of the ancient trees along the coast and the dredging history of the river. Through the time and space comparison function, the impact of vegetation changes and water diversion on ancient buildings over the past century can be traced back, injecting historical wisdom into future urban planning. Ultimately, this platform transforms the complexity of the physical world into a computable, interactive, and predictable digital mirror.

[0193] An embodiment of the present invention further provides a digital twin ancient building construction system based on PhysGaussian, which is used to execute the digital twin ancient building construction method based on PhysGaussian disclosed in the above embodiments.

[0194] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for constructing digital twin ancient buildings based on PhysGaussian, characterized in that: include: Obtain point cloud data within the target ancient building area and characteristic images of the building appearance; Using a preset geometric segmentation network to identify geometric primitives from point cloud data; The geometric primitives are fitted using a differentiable optimization algorithm, and the PhysGaussian modeling framework is used to build a scene model of the target ancient building. Inputting the characteristic image of the building appearance and the data collected by the mechanical sensor on the target ancient building into a preset differentiable physical perception network to obtain the physical parameters of each geometric primitive, wherein the physical parameters include at least optical parameters and mechanical parameters; Merge each geometric primitive in the scene model with the corresponding physical parameters; The PhysGaussian modeling framework is used to perform dynamic simulation prediction on the scene model that integrates physical parameters, and render visual effects to generate a digital twin model.

2. The method according to claim 1, characterized in that The step of obtaining point cloud data and building appearance feature images within the target ancient building area includes: Use drones equipped with high-precision laser radar to scan target ancient buildings and obtain basic point cloud data of basic structures and dense point cloud data of complex structures; A high-definition camera is used to collect an RGB-D depth image of the surface of the target ancient building, wherein the image records at least the wood surface texture, carved patterns, and coating cracks of the target ancient building; The microscopic data of the target ancient building is detected based on a multispectral imaging system, and the microscopic data at least includes wood aging, internal insect infestation, and water stain erosion.

3. The method according to claim 1, characterized in that The method of using a preset geometric segmentation network to identify geometric primitives from point cloud data includes: The PointNet++ network or DGCNN network is used to process the point cloud data and automatically identify the geometric primitives that constitute the target ancient buildings.

4. The method according to claim 1, characterized in that The method of fitting the geometric primitives using a differentiable optimization algorithm comprises: A differentiable optimization algorithm is used to adjust the parameters of each geometric primitive with the fitting goal of minimizing the geometric error between the original point cloud data and the constructed model; Fit all geometric primitives into an initial structural model based on Boolean operations; The initial structural model is smoothed using deformation field interpolation technology to obtain the overall structural model of the target ancient building.

5. The method according to claim 4, characterized in that The scene model of the target ancient building is constructed using the PhysGaussian modeling framework, including: The incremental Gaussian evolution algorithm in the PhysGaussian modeling framework is used to discretize each geometric primitive in the overall structural model of the target ancient building into Gaussian particles, among which the preset key geometric primitives are discretized into fine-grained Gaussian particles, and other geometric primitives are discretized into coarse-grained Gaussian particles. In dynamic scenes, the Gaussian particle parameters of geometric primitives are adjusted to simulate deformation and generate dynamic scene models.

6. The method according to claim 5, characterized in that The method of inputting the characteristic image of the building appearance and the data collected by the mechanical sensors on the target ancient building into a preset differentiable physical perception network to obtain the physical parameters of each geometric primitive includes: Collecting mechanical data of each mechanical sensor, wherein the mechanical sensor is arranged at a position corresponding to each geometric primitive on the target ancient building; Acquire a sub-image of each geometric primitive based on the building appearance feature image; The mechanical data and sub-image corresponding to each geometric primitive are input into a preset differentiable physical perception network to obtain the physical parameters of each geometric primitive, wherein the physical parameters include at least Young's modulus, Poisson's ratio, reflectivity, density and yield stress.

7. The method according to claim 6, characterized in that The step of fusing each geometric primitive in the scene model with the corresponding physical parameter includes: Based on the particle-grid bidirectional coupling mechanism of PhysGaussian, the physical parameters corresponding to each geometric primitive are matched with Gaussian particles; The mechanical and optical characteristics corresponding to each Gaussian particle are determined according to the physical parameters.

8. The method according to claim 7, characterized in that The method of using the PhysGaussian modeling framework to dynamically simulate and predict the scene model integrating physical parameters includes: Based on the explicit time integration in the Material Point Method physical simulation method and the local affine transformation assumption in the PhysGaussian modeling framework, each Gaussian particle in the scene model is dynamically evolved; A differentiable simulation framework is used to dynamically adjust the mechanical and optical characteristics of each Gaussian particle by back-propagating simulation errors. Dynamic simulation prediction is performed based on the updated Gaussian particles to obtain a simulation model.

9. The method according to claim 8, characterized in that The rendering visual effect generates a digital twin model, including: PhysGaussian Splatting technology is used to render the visual effects of the simulation model, and key geometric primitives are highlighted to generate a digital twin model of the target ancient building.

10. The method according to claim 9, characterized in that The method further comprises: Adjusting the mechanical characteristics and optical characteristics of each Gaussian particle according to the simulation parameters input by the user, wherein the simulation parameters include at least wind speed and earthquake intensity; Generate a simulation model based on the adjusted Gaussian particles; Extract the mechanical indicators of each structure in the simulation model, and generate alarms and protection plans when the mechanical indicators exceed the preset strength threshold.

11. The method according to claim 1, characterized in that: The scene model at least includes a model of the target ancient building, and models of vegetation and water systems in the target ancient building area.

12. A digital twin ancient building construction system based on PhysGaussian, characterized in that: The system executes the PhysGaussian-based digital twin ancient building construction method according to any one of claims 1-11.

Images