Three-dimensional gaussian modeling and adaptive optimization method for multi-granularity scenarios of power transmission projects

Abstract

The application discloses a power transmission and transformation engineering multi-granularity scene three-dimensional Gaussian modeling and adaptive optimization method and belongs to the technical field of three-dimensional scene reconstruction. In view of the problems that the modeling precision and efficiency of the prior art are difficult to be considered and the dynamic adaptation capability is poor, the application collects multi-source heterogeneous data, constructs a three-level multi-granularity hierarchical system of a global region-local, and based on structured Gaussian anchor point layered construction and representation modeling, generates a basic scene Gaussian representation after precision verification; the final optimized Gaussian model is generated through view adaptive optimization, anchor point growth and redundancy pruning; combined with static and dynamic region decoupling and incremental updating, local retraining strategy, the whole life cycle dynamic updating and lightweight interaction of the model are realized. The application improves the overall representation efficiency and local detail fidelity of a complex engineering scene, enhances the multi-view robustness, reduces the Gaussian point redundancy, supports mobile terminal adaptation, and has remarkable engineering practical value.

Description

Technical Field

[0001] This invention relates to the field of 3D scene reconstruction technology, specifically a method for 3D Gaussian modeling and adaptive optimization of multi-granularity scenes in power transmission and transformation engineering. Background Technology

[0002] Current power transmission and transformation engineering construction faces severe challenges due to complex geographical environments, diverse facility structures, and dynamically intertwined construction processes. These scenarios are characterized by large scale, multiple structures, and dynamic evolution. Traditional modeling technologies, primarily relied upon in the industry, such as 2D drawings, BIM, and oblique photogrammetry, have significant shortcomings in addressing the complex scenarios of power transmission and transformation projects: 2D drawings are not intuitive and easily lead to design conflicts; while BIM models are accurate, they cannot accurately represent large-scale real-world scenes; and oblique photogrammetry data is bloated and lacks interactivity. These technologies generally suffer from core defects such as difficulty in balancing model accuracy and rendering efficiency, insufficient ability to express dynamic scenes, and challenges in integrating multi-source data, severely restricting the improvement of digitalization and refined management of power transmission and transformation projects.

[0003] While 3D Gaussian Splatting (3D-GS) technology boasts advantages such as fast training speed, high real-time rendering efficiency, and excellent visual fidelity, its current applications are mostly limited to small-scale objects or indoor scenes. In complex industrial scenarios like power transmission and transformation projects, it faces challenges such as insufficient scene representation adaptability, difficulty in balancing multi-granularity details, mismatch between Gaussian point distribution and engineering features, high Gaussian point redundancy, and poor robustness under view changes. It cannot simultaneously meet the overall representation efficiency requirements of large-scale power transmission and transformation projects and the accuracy requirements of local details such as equipment connections and structure connections. Furthermore, it lacks adaptability to multi-view observation scenarios such as power transmission and transformation project inspections and construction simulations, making it difficult to achieve efficient representation and optimization of complex scenes, thus hindering its application expansion in the field of power transmission and transformation engineering.

[0004] Existing Scaffold-GS (Structured 3D Gaussian) technology effectively alleviates problems such as Gaussian redundancy and poor view robustness in traditional 3D-GS by constructing Gaussian distributions through anchor point layering, adaptive generation of neural Gaussian distributions through view adaptation, and optimization of distribution through anchor point growth and pruning. However, this technology has not been specifically adapted to the industrial scene characteristics of power transmission and transformation projects. It still has significant shortcomings in multi-granularity scene representation, engineering structure adaptation, and large-scale scene optimization, and cannot directly meet the actual engineering needs of power transmission and transformation projects for high-precision representation, efficient optimization, and stable reconstruction of large-scale complex scenes. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a three-dimensional Gaussian modeling and adaptive optimization method for multi-granularity scenarios in power transmission and transformation engineering, aiming to solve the problems in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering, comprising: We conducted multi-source data acquisition and multi-granularity feature layered extraction for power transmission and transformation engineering scenarios. We collected multi-source heterogeneous data of scenario 3D, images, and engineering semantics. We completed scenario structure classification and semantic binding through the PointNet++ model and constructed a three-level multi-granularity hierarchical system of global-regional-local. Based on a three-level multi-granularity hierarchical system, we carried out multi-granularity structured Gaussian anchor point hierarchical construction and representation modeling, completed the hierarchical initialization of the three-level granularity anchor points, bound geometric, semantic features and engineering space constraints to the anchor points, and completed the parameterization of the three-dimensional Gaussian distribution and the fusion of the multi-granularity Gaussian set with the anchor points as the core. After accuracy verification and iterative optimization, we generated the Gaussian representation of the basic scene. Using the Gaussian representation of the basic scene as input, the view adaptive Gaussian optimization is completed. View information is collected through multi-view sampling, multi-dimensional view features are extracted, the core parameters of Gaussian are dynamically optimized, the Gaussian set is refined through anchor point growth and redundancy pruning, and the final optimized Gaussian model is generated after multi-view consistency verification. Based on the final optimized Gaussian model, dynamic updates and interactive optimizations are performed; differentiated management of static and dynamic regions is achieved through semantic decoupling; for scene changes at different scales, incremental updates based on anchor point displacement and local retraining strategies based on joint optimization of anchor point-3D Gaussian distribution are adopted for updates; view cone activation and lightweight mechanisms are introduced to ensure interactive efficiency; and finally, the model is maintained throughout its entire lifecycle through verification.

[0007] Furthermore, the specific process of constructing a three-level multi-granularity hierarchical system of global-regional-local is as follows: Collect 3D point cloud data, image data, and engineering semantic data of power transmission and transformation engineering scenarios, and establish the spatial correspondence between engineering semantic data and 3D point cloud data and image data through coordinate registration; Using the PointNet++ model, the structure of a power transmission and transformation engineering scenario is divided into high-precision core structures, medium-precision conventional structures, and low-precision dynamic structures. The classification results are then spatially registered with the BIM model, and each 3D point cloud is assigned a corresponding BIM component ID and engineering semantic label. Among them, the high-precision core structures include transformers and power distribution equipment; the medium-precision conventional structures include supports and pipelines; and the low-precision dynamic structures include construction areas and temporary facilities. Based on 3D point clouds with BIM component IDs and engineering semantic tags, a three-level multi-granularity hierarchical system of global-region-local granularity is constructed sequentially from global to local. Differentiated data preprocessing operations are performed on point cloud data of different granularities in a three-level multi-granularity hierarchical system. Simultaneously, perspective transformation, histogram equalization, and SIFT feature point extraction are performed on image data. A multi-granularity semantic mapping table is established for semantic data to clarify the priority of semantic labels for different levels of engineering.

[0008] Furthermore, based on the 3D point cloud with BIM component IDs and engineering semantic tags, the specific process of constructing a global-regional-local three-level multi-granularity hierarchical system from global to local granularity is as follows: Based on the overall spatial range of the power transmission and transformation engineering scene contained in the labeled 3D point cloud with semantic annotation, a global granularity is constructed through sparse voxel mesh to bind global semantic labels and carry the overall representation of the low-precision dynamic structure. Based on the global granularity, a regional granularity is constructed using a sparse voxel mesh according to the structure type and precision level, which binds semantic labels of the structure type and carries the representation of the medium-precision conventional structure. Based on the regional granularity, a local granularity is constructed for the high-precision core structure region by using sparse voxel meshes to bind individual BIM component IDs and semantic tags for precision requirements, focusing on the geometric shape, surface features and connection details of individual high-precision core structures.

[0009] Furthermore, the specific process for generating the Gaussian representation of the basic scene is as follows: Based on a three-level multi-granularity hierarchical system of global-regional-local, layered anchor points are constructed using sparse voxel meshes. Preprocessed point cloud data of different granularities and SIFT feature points are bound to corresponding anchor points to generate fusion features and spatial constraint features for the anchor points. Using the bound hierarchical anchor points as the core, and combining the three-dimensional Gaussian distribution function, targeted representation modeling of different granularity scenes is performed to generate multi-granularity Gaussian sets; Accuracy verification and iterative adjustment are performed on global granularity Gaussian sets, regional granularity Gaussian sets, and local granularity Gaussian sets; The Gaussian sets of each granularity after accuracy verification and iterative adjustment are re-fused into an optimized overall scene Gaussian, i.e., the basic scene Gaussian representation, according to preset weights.

[0010] Furthermore, based on a three-level multi-granularity hierarchical system of global-regional-local, the specific process of constructing layered anchor points using sparse voxel meshes is as follows: For scene-level global granularity, a sparse voxel mesh is used to divide the label 3D point cloud into multiple spatial regions, and all point clouds in a single spatial region are used as global anchor points to bind the semantic tags of the functional partition project. For the structural group-level region granularity, a sparse voxel mesh is used to perform secondary spatial partitioning within the global anchor point range to generate region anchor points and bind structural type semantic tags. For component-level local granularity, a sparse voxel mesh is used to perform fine spatial division within the coverage area of ​​the regional anchor point, generating local anchor points and binding semantic tags for accuracy requirements and component material attributes.

[0011] Furthermore, the specific process of binding preprocessed point cloud data at different granularity levels and SIFT feature points to corresponding level anchor points, and generating fused features and spatial constraint features for the anchor points, is as follows: For global anchor points, the overall 3D coordinate statistical features of the point cloud within the corresponding spatial range are extracted, and the extracted SIFT feature points are combined to complete feature matching of the global scene from multiple perspectives. The overall 3D coordinate statistical features of the point cloud within the corresponding spatial range of the global anchor point and the geometric information of SIFT feature point matching are fused to generate the geometric description of the global anchor point. The semantic tags of the functional zoning project are embedded to generate global semantic features, and the global semantic features are concatenated with the geometric description of the global anchor point to form the global anchor point fusion feature. Global spatial constraint features are added based on the overall engineering boundary parameters of the BIM model. For regional anchor points, the 3D coordinate statistical features of the point cloud within the corresponding spatial range of the regional anchor point are extracted. Combined with the extracted SIFT feature points, feature matching of the multi-view structural group region is completed. The 3D coordinate statistical features of the point cloud within the corresponding spatial range of the regional anchor point and the geometric information of SIFT feature point matching are fused to generate the geometric description of the regional anchor point. The semantic labels of the structure type are embedded to generate regional semantic features. The regional semantic features are concatenated with the geometric description of the regional anchor point to form the regional anchor point fusion feature. Regional spatial constraint features are added based on the structural group size parameters of the BIM model. For local anchor points, the 3D coordinate statistical features of the point cloud within the corresponding area of ​​the local anchor point are extracted, and the extracted SIFT feature points are combined to complete feature matching of local details from multiple perspectives. The 3D point cloud within the neighborhood of the local anchor point is selected as a subset of the point cloud, and the 3D coordinate covariance matrix of this subset is constructed. Principal component analysis eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is used as the normal vector of the corresponding point cloud to obtain the point cloud normal vector feature. The point cloud normal vector feature is fused with the 3D coordinate statistical features of the point cloud within the corresponding area of ​​the local anchor point and the geometric information of SIFT feature point matching to generate the geometric description of the local anchor point. The accuracy requirement semantic label and component material attributes are embedded to generate local semantic features, and the local semantic features are concatenated with the geometric description of the local anchor point to form the local anchor point fusion feature. Local spatial constraint features are added based on the individual component size parameters of the BIM model. The output consists of a set of global anchors, regional anchors, and local anchors bound to corresponding hierarchical anchor fusion features and spatial constraint features.

[0012] Furthermore, the specific process for generating the final optimized Gaussian model is as follows: Collect view information from power transmission and transformation engineering scenarios under multiple poses and lighting conditions, and extract local texture features and multi-resolution view data. Figure 1 Consistency features, illumination features, and engineering semantic consistency features; The covariance matrix, color, and opacity of the 3D Gaussian distribution of anchor points in the Gaussian set corresponding to the Gaussian representation of the basic scene are dynamically optimized. Anchor point growth and Gaussian pruning optimization: Add new anchor points based on view errors and remove redundant 3D Gaussian distributions with contributions below the threshold; Visual processing of Gaussian sets after Gaussian pruning Figure 1 Consistency testing; See Figure 1 After consistency testing, the three-dimensional Gaussian distribution of the anchor points in the Gaussian set is weighted and integrated, and an index is built using a hierarchical storage method to output the final optimized Gaussian model.

[0013] Furthermore, the specific process of dynamically optimizing the covariance matrix of the 3D Gaussian distribution of anchor points in the Gaussian set corresponding to the Gaussian representation of the basic scene is as follows: Obtain the camera rotation matrix and translation matrix for each viewpoint, and calculate the covariance matrix of the 3D Gaussian distribution of each anchor point in the Gaussian set corresponding to the Gaussian representation of the basic scene for each viewpoint. Target location is calculated based on local texture features, based on multi-resolution view. Figure 1 The confidence weights for each viewpoint are calculated using the consistency features. The view dependency loss is constructed and minimized to obtain the optimal covariance matrix of the three-dimensional Gaussian distribution of each anchor point under different viewpoints. The covariance loss is constructed based on the optimal covariance matrix and the target covariance matrix, and the Gaussian geometric consistency total loss is constructed by combining the view dependency loss. The covariance matrix of the three-dimensional Gaussian distribution of the anchor points is updated by the gradient descent algorithm.

[0014] Furthermore, the specific process of dynamically optimizing the color and opacity of the 3D Gaussian distribution of anchor points in the Gaussian set corresponding to the Gaussian representation of the basic scene is as follows: For each anchor point in the Gaussian set corresponding to the Gaussian representation of the basic scene, its 3D position is projected onto a multi-view image, and its multi-dimensional features are sampled or aggregated. These multi-dimensional features include local texture features, multi-resolution view features, and other features. Figure 1 Consistency features, illumination features, and engineering semantic consistency features; The multi-dimensional features are input into the MLP decoder to predict the color and opacity of the three-dimensional Gaussian distribution corresponding to the anchor point, respectively. The color loss constraint ensures that the predicted color matches the true color, and the opacity loss constraint ensures that the predicted opacity matches the true opacity. The color loss and opacity loss are combined to construct a joint optimization total loss. The color and opacity of the anchor point corresponding to the three-dimensional Gaussian distribution are updated by the gradient descent algorithm.

[0015] Furthermore, the specific process of adding new anchor points based on view errors and removing redundant 3D Gaussian distributions with contributions below a threshold is as follows: Calculate the total rendering error, which is the sum of geometric error and visual rendering error; The error gradient is calculated based on the first partial derivative of the three-dimensional position coordinates of the three-dimensional Gaussian distribution corresponding to the anchor point with respect to the total rendering error. A new anchor point is added at the position with the largest error gradient, and the corresponding Gaussian distribution parameters are initialized to obtain the Gaussian set after the anchor point is grown. For each anchor point in the Gaussian set after anchor point growth, calculate the contribution of its corresponding Gaussian distribution under different views. The contribution is determined by the product of visibility weight, depth weight and view-covariance matching weight. Remove anchor points whose contribution is below a preset threshold to obtain a Gaussian set after Gaussian pruning.

[0016] Compared with existing technologies, the present invention has the following advantages: (1) This invention integrates three-dimensional Gaussian sputtering technology with three-level multi-granularity features of power transmission and transformation engineering, namely global, regional and local. By constructing structured anchor points in a hierarchical manner, it configures Gaussian point density, covariance matrix and engineering semantic constraints for different granularity scenes. Combined with multi-granularity Gaussian set weighted fusion and accuracy verification iterative adjustment, and then through anchor point growth and redundancy pruning strategies in view adaptive optimization, a complete technical chain from multi-granularity modeling to view optimization is formed. This ensures efficient rendering of coarse-grained scenes while achieving millimeter-level accurate modeling of key equipment-level structures. It breaks through the bottleneck of traditional methods in large-scale industrial scenes where accuracy and efficiency are difficult to balance, significantly reduces Gaussian point redundancy and improves overall representation efficiency.

[0017] (2) This invention addresses the multi-view observation needs of power transmission and transformation engineering inspection and construction simulation. In the process of view adaptive Gaussian optimization, it integrates BIM semantic attributes to construct a multi-constraint collaborative optimization strategy. Through view-dependent covariance matrix optimization, color and opacity joint optimization, and anchor point growth to complete details based on error gradient, combined with multi-view consistency verification and local anchor point illumination adaptive factor, the robustness of the engineering scene under different observation views is significantly improved. The model can still maintain stable geometric and visual performance under the conditions of view and illumination changes, and achieves the optimal balance between computing cost and reconstruction accuracy.

[0018] (3) This invention achieves efficient local updates of dynamic scenarios such as construction areas and temporary equipment without affecting the global static model by decoupling static and dynamic regional features and incremental updates, local retraining, and anchor point reconstruction through a three-level differentiated update strategy. At the same time, combined with the view cone activation strategy and lightweight interaction mode, it supports on-demand loading and smooth interaction of low computing power devices such as mobile devices. Furthermore, it can ensure the geometric accuracy and semantic consistency of the model throughout its entire life cycle through lightweight verification, thus promoting the practical application of three-dimensional Gaussian representation technology in all scenarios of power transmission and transformation engineering and significantly improving the dynamic adaptability and engineering value of the model. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] like Figure 1 As shown, the present invention provides a technical solution: a three-dimensional Gaussian modeling and adaptive optimization method for multi-granularity scenarios in power transmission and transformation engineering, comprising the following steps: Step S1: Conduct multi-source data acquisition and multi-granularity feature layering extraction for power transmission and transformation engineering scenarios. Collect multi-source heterogeneous data of scene 3D, images, and engineering semantics. Complete scene structure classification and semantic binding through PointNet++ model, and construct a three-level multi-granularity hierarchical system of global-regional-local; provide standardized data and structured framework for subsequent modeling.

[0021] Step S2: Based on the three-level multi-granularity hierarchical system, carry out multi-granularity structured Gaussian anchor point layer construction and representation modeling, complete the layer initialization of the three-level granularity anchor points, bind geometric, semantic features and engineering space constraints to the anchor points, complete the three-dimensional Gaussian distribution parameterization and multi-granularity Gaussian set fusion with the anchor points as the core, and generate the basic scene Gaussian representation through accuracy verification and iterative optimization.

[0022] Step S3: Using the basic scene Gaussian representation as input, complete the view adaptive Gaussian optimization. Collect view information through multi-view sampling, extract multi-dimensional view features, dynamically optimize the core parameters of Gaussian, refine the Gaussian set through anchor point growth and redundancy pruning, and generate the final optimized Gaussian model after multi-view consistency verification.

[0023] Step S4: Based on the final optimized Gaussian model, perform dynamic updates and interactive optimizations. Semantic decoupling enables differentiated management of static and dynamic regions; for scene changes at different scales, incremental updates based on anchor point displacement and local retraining strategies based on joint optimization of anchor points and 3D Gaussian distributions are used for updates; view cone activation and lightweight mechanisms are introduced to ensure interactive efficiency; finally, validation is used to achieve full lifecycle maintenance of the model.

[0024] The specific process of step S1 is as follows: Step S1.1, Multi-source Data Acquisition and Fusion: Based on the characteristics of different structures of equipment, buildings, and construction areas in power transmission and transformation engineering scenarios, three types of multi-source heterogeneous data are collected: 3D point cloud data, image data, and engineering semantic data. Spatial correspondence between engineering semantic data and 3D point cloud data and image data is established through coordinate registration, constructing a multi-source associated basic data source; specifically: Collect 3D point cloud data: Use LiDAR scanning equipment (such as ground 3D laser scanner or drone equipped with LiDAR) to acquire 3D point clouds of power transmission and transformation engineering scenes, ensuring coverage of all equipment, structures and construction areas, and adding scanning stations for severely obscured areas; Image data acquisition: Use a high-resolution camera (such as a DSLR or industrial camera) to capture images of the power transmission and transformation project scene from multiple perspectives; at the same time, record the camera intrinsic parameters (focal length, principal point) and extrinsic parameters (calibrated using COLMAP tool) for each image; Collect semantic data for the project: Import the BIM model of the power transmission and transformation project scenario, and extract the geometric parameters (length, width, height, spatial location), component type (transformer, busbar, frame, etc.), and material properties (steel, concrete, etc.) of the BIM components in the BIM model. Establish data association: Establish preliminary spatial correspondence between BIM components and 3D point cloud data and image data through coordinate registration.

[0025] Step S1.2, Scene Structure Classification and Semantic Labeling: Based on multi-source associated basic data sources, the PointNet++ model is used to classify the structure of the power transmission and transformation engineering scene into high-precision core structure, medium-precision conventional structure, and low-precision dynamic structure; the classification results are spatially registered with the BIM model, and each 3D point cloud is assigned a corresponding BIM component ID and engineering semantic label (such as "Transformer 3" and "Steel Frame"), and the output is a labeled 3D point cloud containing the overall spatial range of the semantically labeled power transmission and transformation engineering scene; thus realizing the binding of geometric data and engineering semantics; Among them, the high-precision core structure includes equipment or facilities that are crucial to functionality and safety, such as transformers and power distribution devices. These need to be represented very finely and accurately in the model to ensure that the geometric shape and functional requirements are reflected with high precision. Medium-precision conventional structures refer to some common equipment or facilities, such as supports and pipes. Their precision requirements are lower than those of the core structure, but they still need to ensure sufficient detail to support engineering applications and management. Low-precision dynamic structures refer to elements that can change over time, such as construction areas and temporary facilities. Because they change significantly during the engineering process, the accuracy requirements for their representation are relatively low.

[0026] Specifically, the PointNet++ model is trained using a structure classification loss function. Defined as: ; In the formula, This represents the total number of three-dimensional point clouds sampled in the power transmission and transformation engineering scenario; Indicates the number of categories; Indicates the first Point cloud about category The true label; Indicates the prediction of the first The point cloud belongs to the category The probability of; Represents the regularization coefficient; Indicates the regularization loss; Represents a logarithmic function; Regularization loss Defined as: ; In the formula, Indicates the first A point cloud Nearest neighbor set; Indicates the first The predicted probability vector of a point cloud; Indicates the first The predicted probability vector of the nearest neighbors; The L2 distance represents the predicted probability of neighboring points, used to ensure that the classification results of continuous structures such as transformer bodies and tower bodies in substations are not fragmented.

[0027] Step S1.3, Multi-granularity hierarchical division: Based on labeled 3D point clouds, and according to the principle of matching granularity to accuracy requirements, a three-level multi-granularity hierarchical system is constructed sequentially from global to local: scene-level global granularity, structure group-level regional granularity, and component-level local granularity. Specifically: Step S1.31: Constructing Global Granularity: Based on the labeled 3D point cloud, according to the principle of matching granularity with accuracy requirements, and taking the overall spatial range of the power transmission and transformation engineering scene contained in the labeled 3D point cloud as the benchmark, a global granularity is constructed by binding global semantic labels (such as "substation area") and carrying the overall representation of the low-precision dynamic structure through a large-scale sparse voxel mesh, generating global granularity point cloud data; the global granularity level mainly focuses on the overall spatial relationship of the scene, such as the spatial distribution of equipment and structures, functional zoning, etc.

[0028] Step S1.32: Constructing regional granularity: Based on the global granularity, construct regional granularity according to the structure type and accuracy level through a secondary sparse voxel mesh, binding semantic labels of structure type (such as "transformer group" and "support frame"), carrying the medium-precision conventional structure representation and serving as a high-precision core structure grouping container, and generating regional granularity point cloud data; this level of granularity is mainly used to represent the relative position and shape between structure groups.

[0029] Step S1.33: Constructing Local Granularity: Based on the regional granularity, for the high-precision core structure region, a local granularity is constructed by using a small-scale sparse voxel mesh to bind individual BIM component IDs and precision requirement semantic tags (such as "Transformer #3", "Support #5"), focusing on the geometric shape, surface features, and connection details of individual high-precision core structures, generating local granularity point cloud data; defining local granularity anchor points generated by dividing the regional granularity point cloud data into a small-scale sparse voxel mesh, each anchor point corresponding to the point cloud within a local detail area of ​​a component divided by the small-scale sparse voxel mesh; this level ensures the high-precision representation of key components in power transmission and transformation projects.

[0030] Step S1.4, Multi-granularity data preprocessing: For point cloud data of different granularity levels in the global-regional-local three-level multi-granularity hierarchy, differentiated data preprocessing operations are performed. Simultaneously, perspective transformation is performed on image data to correct viewpoint distortion, histogram equalization is used to optimize illumination consistency, and SIFT feature points are extracted. A multi-granularity semantic mapping table is established for semantic data to clarify the priority of semantic labels at different levels for subsequent Gaussian representation and modeling. Specifically: Preprocessing of global granular point cloud data: A filtering algorithm is used to remove noise points from the global granular point cloud data, ensuring data accuracy and continuity over a large area; large-scale voxel downsampling is used to reduce the data volume of the global granular point cloud data, improving computational efficiency; large-scale scene features of the global granular point cloud data are preserved; this process ensures that the global granular point cloud data can effectively represent the overall layout and spatial relationships of the substation, while avoiding excessive redundant data; Preprocessing of regional granular point cloud data: The regional granular point cloud data is simplified by mesoscale voxel downsampling to balance accuracy and computational cost, accurately express the relative position and morphology between each structural group, and avoid excessive computational cost due to too much detailed information. Preprocessing of local granular point cloud data: Fine voxel downsampling is used to preserve high-precision information of local granular point cloud data, and noise removal and point cloud smoothing are performed simultaneously to ensure the accuracy of details of key components.

[0031] The image data is corrected for perspective distortion using perspective transformation, and the consistency of illumination is optimized by histogram equalization. SIFT feature points are extracted for subsequent view matching.

[0032] Semantic data uses a pre-defined multi-granularity semantic mapping table to clarify the priority of semantic labels at different strength levels, providing a basis for the subsequent anchor point layering of Gaussian representation.

[0033] The multi-granularity semantic mapping table is built using BIM component list, PointNet++ classification results, and manually confirmed engineering area attributes as inputs. Each item includes at least: tag source, BIM component type, BIM component ID, and PointNet++ classification confidence score. Candidate granularity level, granularity priority Accuracy level Material properties and tag templates ; The specific mapping rule is as follows: When the BIM component type belongs to transformer, circuit breaker, disconnector, instrument transformer, busbar connector, insulator, terminal box, tower node, or equipment connection point, the candidate granularity level is mapped to local granularity, with granularity priority... Tag template The format is "BIM Component ID + Component Type + Accuracy Level + Material Attribute"; when the BIM component type belongs to a structural group such as support, pipe, cable trench, steel frame, foundation cap, or wall, the candidate granularity level is mapped to the regional granularity, with granularity priority... Tag template The format is "Structure Group Type + Area Number"; when the category belongs to construction area, temporary facilities, ground, road, green space, material storage area, or background area, the candidate granularity level is mapped to the global granularity, with granularity priority. Tag template It consists of "functional partitions + dynamic attributes".

[0034] The specific process of step S2 is as follows: Step S2.1: Based on a three-level multi-granularity hierarchical system of global-regional-local, and combined with a sparse voxel mesh, construct layered anchor points; specifically: For global granularity, a scale of is adopted. The large-scale sparse voxel mesh divides the labeled 3D point cloud into multiple spatial regions, using all point clouds within a single spatial region as global anchor points. Each global anchor point corresponds to a functional area or a large area of ​​the scene, and is bound to a "functional partition" engineering semantic tag. The initial variance of the 3D Gaussian distribution corresponding to the global anchor point is set to... (Used to provide a reference scale for subsequent Gaussian distribution initialization), constructing an overall layout representation that adapts to the global scene; For regional granularity, a scale of [scale value] is adopted within the global anchor point range. The mesoscale sparse voxel mesh is used for secondary spatial partitioning to generate region anchor points, which are then bound with "structure type" semantic tags. The initial variance of the 3D Gaussian distribution is set to... (Used to provide a reference scale for subsequent Gaussian distribution initialization), balancing region rendering efficiency and structural integrity; For local granularity, a scale of [scale value] is adopted within the coverage area of ​​the regional anchor point. A small-scale sparse voxel mesh is used for fine spatial partitioning, generating local anchor points, binding semantic tags of "accuracy requirements" and component material properties, and setting the initial variance of the three-dimensional Gaussian distribution to... (Used to provide a reference scale for subsequent Gaussian distribution initialization), thereby ensuring the accuracy of the representation of core details.

[0035] Step S2.2: Bind the preprocessed point cloud data of different granularities and SIFT feature points from step S1.4 to the corresponding level anchor points; specifically: For global anchor points, the overall 3D coordinate statistical features of the point cloud within the corresponding spatial range are extracted, and feature matching of the extracted SIFT feature points is completed for the global scene from multiple perspectives. The overall 3D coordinate statistical features of the point cloud within the corresponding spatial range of the global anchor point and the geometric information of SIFT feature point matching are fused to generate the geometric description of the global anchor point. The semantic tags of the "functional partition" project are used to generate global semantic features through the Embedding Layer. Geometric description of global anchor points splicing to form global anchor point fusion features Add global spatial constraint features to the overall engineering boundary parameters based on the BIM model. Limit the overall range of the three-dimensional Gaussian distribution corresponding to the global anchor point; For regional anchor points, the 3D coordinate statistical features of the point cloud within the corresponding spatial range are extracted. These features are then combined with the extracted SIFT feature points to perform feature matching for the multi-view structure group region. Finally, the 3D coordinate statistical features of the point cloud within the corresponding spatial range and the geometric information from the SIFT feature point matching are fused to generate a geometric description of the regional anchor points. The semantic labels of "structure type" are used to generate region semantic features through the Embedding Layer. Geometric description of region anchor points splicing forms regional anchor point fusion characteristics Adding area spatial constraint features to the structural group dimensional parameters based on the BIM model. The range of the three-dimensional Gaussian distribution corresponding to the anchor point in the defined region; For local anchor points, the 3D coordinate statistical features of the point cloud within the corresponding region are extracted, and combined with the extracted SIFT feature points to complete feature matching of local details from multiple perspectives. The 3D point cloud within the neighborhood of the local anchor point is selected as a subset of the point cloud, and the 3D coordinate covariance matrix of this subset is constructed. Principal component analysis (PCA) eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is used as the normal vector of the corresponding point cloud to obtain the point cloud normal vector features, which are then adapted to the curved / angular structures of the device. Finally, the point cloud normal vector features are fused with the 3D coordinate statistical features of the point cloud within the corresponding region of the local anchor point and the geometric information from SIFT feature point matching to generate a geometric description of the local anchor point. The semantic tags for "precision requirements" and component material attributes are used to generate local semantic features through an Embedding Layer. Geometric description of local anchor points splicing to form local anchor point fusion features Add local spatial constraint features to the dimensional parameters of individual components based on the BIM model. This limits the range of the three-dimensional Gaussian distribution corresponding to the local anchor point, avoiding representation distortion caused by overflow of the three-dimensional Gaussian distribution.

[0036] The output consists of a set of global anchor points, regional anchor points, and local anchor points bound to corresponding hierarchical anchor point fusion features (global anchor point fusion features, regional anchor point fusion features, and local anchor point fusion features) and spatial constraint features (global spatial constraint features, regional spatial constraint features, and local spatial constraint features). The set of anchor points retains the three-level hierarchical nesting relationship and spatial position information for subsequent 3D Gaussian modeling.

[0037] Step S2.3: Using the hierarchical anchor points bound in Step S2.2 as the core, and combining them with a 3D Gaussian distribution function, perform targeted representation modeling for different granularity scenarios to generate a multi-granularity Gaussian set; specifically: Step S2.31, 3D Gaussian distribution parameterization: From the set of anchor points consisting of global anchor points, regional anchor points, and local anchor points bound with corresponding hierarchical anchor point fusion features and spatial constraint features, one anchor point is selected in sequence. Get anchor point Coordinates as Gaussian mean Based on anchor points Corresponding anchor point fusion features Generate covariance matrix ; This refers to small neural networks, such as multilayer perceptrons (MLPs). Represent the initial covariance matrix; collect anchor points The pixel colors of multi-view images within the neighborhood are weighted and averaged to obtain the texture color (the weights are inversely proportional to the view distance); anchor points are then used to... The coordinates are input into a lightweight Neural Radiation Field Encoding Network (NeRF), which outputs radiation color; the texture color and radiation color are fused to obtain the color feature. Based on Gaussian mean Covariance matrix and color features , as anchor point Parameterizing a three-dimensional Gaussian distribution Traverse all anchor points in the set of anchor points consisting of global anchor points, regional anchor points, and local anchor points that are bound with corresponding hierarchical anchor point fusion features and spatial constraint features, and obtain the three-dimensional Gaussian distribution of each anchor point.

[0038] Specifically, for local anchors within the anchor set consisting of global anchors, regional anchors, and local anchors bound with corresponding hierarchical anchor fusion features and spatial constraint features, an additional illumination adaptation factor is introduced, which is then coupled with the corresponding color feature. Multiplication improves view robustness. Specifically, for each local anchor point, the pixel brightness values ​​of its neighborhood (e.g., a 5×5 pixel area around the projected location) are collected from all viewpoints, and the average brightness is calculated. Based on average brightness Calculate the illumination adaptation factor , This indicates the reference brightness; this makes the color of local anchor points more robust under different viewing angles and lighting conditions, reducing inconsistencies in rendered colors caused by changes in lighting.

[0039] Step S2.32, Generation of multi-granularity Gaussian sets: All global anchor points and their corresponding 3D Gaussian distributions are incorporated into a global granularity Gaussian set. The 3D Gaussian distribution density is set to 50 points / m³ to characterize the overall layout and spatial relationships of the scene; all region anchor points and their corresponding 3D Gaussian distributions are grouped into a region-granularity Gaussian set. The three-dimensional Gaussian distribution density is set to 200 points / m³ to characterize the morphology and relative positions of the structure group; all local anchor points and their corresponding three-dimensional Gaussian distributions are grouped into a local granular Gaussian set. The three-dimensional Gaussian distribution density is set to 1000 units / m³ to characterize the fine structure of key details; Preset global granularity weight factor Regional granularity weighting factor Local granularity weighting factor Global granular Gaussian set Regional granularity Gaussian set and local granular Gaussian sets Blend into a unified scene Gaussian , means as follows: ; Ultimately, a balanced representation of multiple granular details is achieved.

[0040] To ensure the stability of the project implementation, a reference range constraint is imposed on the weights: Take a value between 0.10 and 0.35. Take a value between 0.25 and 0.45. Take values ​​between 0.30 and 0.60; satisfy the following conditions. .

[0041] Step S2.4: Perform accuracy verification and iterative adjustment on the global granularity Gaussian set, the regional granularity Gaussian set, and the local granularity Gaussian set; specifically: Global granularity verification: Calculate the 3D Gaussian distribution of global anchor points in the global granularity Gaussian set and the Chamfer distance between the corresponding point cloud in the 3D point cloud data of the power transmission and transformation engineering scenario. If the Chamfer distance exceeds the preset distance threshold, the global granularity verification fails. The global anchor point density is automatically increased, and the anchor point construction to multi-granularity Gaussian set fusion step is re-executed until the accuracy requirements are met. Regional granularity verification: Randomly sample several regional points from the real-world 3D data of the power transmission and transformation project scenario, obtain the 3D Gaussian distribution of the regional anchor points at the corresponding positions in the corresponding regional Gaussian set, calculate the Chamfer distance between the 3D Gaussian distribution of the regional anchor points in the regional Gaussian set and the corresponding position point cloud in the 3D point cloud data of the power transmission and transformation project scenario. If the Chamfer distance exceeds the preset distance threshold, the regional granularity verification fails. The regional anchor point density of the regional granularity is automatically increased, and the anchor point construction to multi-granularity Gaussian set fusion step is re-executed until the engineering accuracy requirements are met. Local granularity verification: Select several points in local areas from the real-world 3D data of the power transmission and transformation project scenario. Calculate the 3D Gaussian distribution of local anchor points in the local granularity Gaussian set and the Chamfer distance between the corresponding point cloud locations in the 3D point cloud data of the power transmission and transformation project scenario. If the Chamfer distance exceeds a preset threshold, the local granularity verification fails. Increase the local anchor point density in the local area and re-execute the anchor point construction to multi-granularity Gaussian set fusion step until the accuracy requirements are met. At the same time, check the matching degree between the local granularity 3D Gaussian distribution and semantic tags in the local area to ensure that the 3D Gaussian distribution accurately reflects the structure and function of the detailed area.

[0042] Step S2.5: The global granularity Gaussian set, regional granularity Gaussian set, and local granularity Gaussian set after accuracy verification and iterative adjustment in step S2.4 are re-fused into an optimized overall scene Gaussian set according to preset weights. This optimized overall scene Gaussian set is the basic scene Gaussian representation.

[0043] The specific process of step S3 is as follows: Step S3.1, Multi-view observation information acquisition and view feature extraction: To meet the multi-view observation needs of power transmission and transformation engineering inspections and construction simulations, view information under multiple poses and lighting conditions of power transmission and transformation engineering scenarios is acquired, and view features adapted to the engineering scenario are extracted; specifically: Multi-view sampling: Based on the spatial range and structural distribution of the power transmission and transformation engineering scenario, a sampling method of key view plus supplementary view is adopted. Twelve fixed key views are set to cover the main observation directions of the power transmission and transformation engineering scenario. A spherical uniform sampling algorithm is used to generate 36 dynamic supplementary views around the key views to cover areas that are prone to view distortion, such as equipment connection points and structure joints. View feature extraction: Multi-dimensional feature extraction is performed on the image data from each viewpoint, specifically extracting four types of features: local texture features, multi-resolution view features, and multi-dimensional features. Figure 1 Consistency features, illumination features, and engineering semantic consistency features; the extraction process is as follows: Local texture features: Extract texture feature points such as corners and edges from each viewpoint image and calculate a 128-dimensional descriptor; Multi-resolution vision Figure 1Consistency Features: For each viewpoint image, based on the Scaffold-GS method, a three-level resolution feature set is constructed: "original - 1 / 2 downsampled - 1 / 4 downsampled". The feature matching relationship between adjacent viewpoint images at each resolution in the three-level resolution feature set is calculated using optical flow, generating a multi-scale viewpoint feature set. Figure 1 Consistency matrix; Illumination features: Calculate the mean brightness, contrast, and color saturation of the image from each viewpoint to generate a three-dimensional illumination feature vector; Engineering semantic consistency feature: Based on the corresponding BIM component ID and engineering semantic label assigned to each 3D point cloud in step S1.2, calculate the matching degree of semantic labels from different perspectives. This is to supplement the semantic constraints unique to engineering scenarios.

[0044] Step S3.2, Dynamic Optimization of 3D Gaussian Distribution Based on View Features: Based on the view features output in Step S3.1 and the Gaussian sets corresponding to the basic scene Gaussian representation generated in Step S2 (the global granularity Gaussian set, regional granularity Gaussian set, and local granularity Gaussian set after accuracy verification and iterative adjustment), the 3D Gaussian distribution (Gaussian points) of the anchor points is dynamically optimized to improve the robustness of the engineering scene view; specifically: Step S3.21, Covariance Matrix View Dependency Optimization: Based on the Scaffold-GS method, optimize the covariance matrix of the 3D Gaussian distribution of anchor points in the Gaussian set corresponding to the Gaussian representation of the basic scene; specifically: Anchor-guided Gaussian property prediction: Obtaining the first Camera rotation matrix from multiple perspectives Translation matrix ; The Gaussian set corresponding to the Gaussian representation of the basic scene. The three-dimensional Gaussian distribution position of each anchor point and the Camera rotation matrix from multiple perspectives Translation matrix Calculate the first From the perspective of the first The covariance matrix of the three-dimensional Gaussian distribution of each anchor point : ; In the formula, The view depends on the covariance prediction function, which can be represented by a multilayer perceptron (MLP). View-dependent parameter optimization: Calculate the target position based on the local texture features in step S3.1. Based on multi-resolution vision Figure 1 Calculation of consistency characteristics Confidence weights for each perspective Based on target location , No. Weight of each perspective The three-dimensional Gaussian distribution of each anchor point in the Gaussian set corresponding to the Gaussian representation of the basic scene. Calculate view dependency loss: ; In the formula, Indicates view dependency loss; Indicates the first The projection position of the three-dimensional Gaussian distribution from each viewpoint; Indicates the total number of viewpoints; Minimize the view dependency loss to obtain the optimal covariance matrix of the 3D Gaussian distribution of each anchor point under different viewpoints. Ultimately, this achieves the optimal state of geometric and visual consistency. Covariance matrix optimization objective: To optimize the covariance matrix based on the 3D Gaussian distribution of each anchor point under different viewpoints. and the target covariance matrix (Usually derived from reference data or geometric features of real objects), calculate the covariance loss: ; Comprehensive optimization: Constructing a Gaussian geometric consistency total loss based on view dependency loss and covariance loss. : ; Minimize the total loss of Gaussian geometric consistency using the gradient descent algorithm. Update the covariance matrix of the 3D Gaussian distribution of the anchor point, and output the updated covariance matrix of the 3D Gaussian distribution.

[0045] Step S3.22, Joint Optimization of Color and Opacity: An MLP decoder is used to predict the color and opacity of the 3D Gaussian distribution of each anchor point in the Gaussian set corresponding to the Gaussian representation of the base scene, ensuring consistent visual performance and geometric uniformity across multiple viewpoints; specifically: Color optimization: For each anchor point in the Gaussian set corresponding to the Gaussian representation of the basic scene, its 3D position is projected onto the multi-view image of step S3.1 to obtain the projection coordinates under each viewpoint; multi-dimensional features of the image are sampled or aggregated at the projection coordinates to form the multi-dimensional features of the anchor point; the multi-dimensional features include: local texture features of the anchor point, multi-resolution viewpoints, etc. Figure 1 Consistency features, illumination features, and engineering semantic consistency features; input the multi-dimensional features of the anchor point into the MLP decoder to predict the color of the corresponding three-dimensional Gaussian distribution. And constrain its consistency with the real color through color loss; Among them, color loss , represented as: ; In the formula, Indicates the target color (derived from the pixel values ​​of the actual image); This represents the color loss weighting coefficient; Opacity optimization: Input the multi-dimensional features of the anchor point into the MLP decoder to predict the opacity of the three-dimensional Gaussian distribution corresponding to the anchor point. And by constraining its consistency with the true opacity through opacity loss; The opacity loss is expressed as: ; In the formula, Indicates loss of opacity; Indicates the target opacity (which can be derived from the actual image transparency or semantic priors); This represents the opacity loss weighting coefficient; Joint optimization: Constructing the joint optimization total loss based on color loss and opacity loss. : ; Minimize the total loss of the joint optimization using the gradient descent algorithm. Update the color and opacity of the 3D Gaussian distribution corresponding to the anchor point, and output the updated color and opacity of the 3D Gaussian distribution.

[0046] Step S3.3, Anchor Point Growth and Gaussian Pruning Optimization: The dynamically optimized Gaussian set from Step S3.2 (i.e., the Gaussian set composed of updated 3D Gaussian distributions with updated covariance matrices, colors, and opacities) is subjected to anchor point growth and Gaussian pruning strategies to further optimize the 3D Gaussian distribution in the scene. Anchor point growth refers to adding new anchor points based on areas with large view errors, while Gaussian pruning refers to removing redundant 3D Gaussian distributions to improve rendering efficiency; specifically: Perform anchor point growth on the dynamically optimized Gaussian set: Calculate total rendering error Defined as the sum of geometric error and visual rendering error: ; In the formula, The geometric error is composed of the previous view dependency loss and covariance loss, reflecting the deviation between the projection position of the Gaussian distribution, the covariance matrix and the engineering geometric constraints; It represents visual rendering error, consisting of color loss and opacity loss, reflecting the deviation of Gaussian-distributed color and opacity from the actual image observation; Based on total rendering error Calculate the error gradient using the first partial derivatives of the three-dimensional position coordinates of the anchor points corresponding to the three-dimensional Gaussian distribution in the dynamically optimized Gaussian set. : ; Based on error gradient and preset gradient step size Calculate the position offset of the new anchor point : ; Based on the offset of the new anchor point Add a new anchor point at the location with the largest error gradient, initialize the corresponding Gaussian distribution parameters, update the dynamically optimized Gaussian set, and obtain the Gaussian set after anchor point growth.

[0047] Wherein, gradient step size The following settings are applied based on different granularity levels: global granularity ranges from 0.2m to 1.25m, regional granularity ranges from 0.05m to 0.375m, and local granularity ranges from 0.005m to 0.05m.

[0048] Perform Gaussian pruning on the Gaussian set after anchor point growth: calculate the contribution of the Gaussian distribution corresponding to each anchor point in the Gaussian set under different viewpoints. : ; In the formula, It is the first The three-dimensional Gaussian distribution of the anchor point and the... Matching weights for each perspective; It is the first Transparency from various perspectives; Among them, the The three-dimensional Gaussian distribution of the anchor point and the... Weight of each perspective Represented as: ; In the formula, Indicates visibility weight; Indicates depth weight; Indicates view-covariance matching weights; Visibility weight The Gaussian distribution 3D coordinates in the Gaussian set after anchor point growth are processed by the camera extrinsic parameters (camera rotation matrix). Translation matrix Transform to the first In a camera coordinate system with multiple viewpoints, if the depth value of a point is positive (located in front of the camera) and its projection onto the image plane falls within the effective pixel range, then... Otherwise, it is 0. Used to remove three-dimensional Gaussian distributions that are facing away from or outside the field of view; Depth weights The calculation is based on the depth normalization index (softmin), which gives higher weights to 3D Gaussian distributions with shallower depths (closer to the camera). The calculation formula is as follows: ; In the formula, This represents the depth of the current three-dimensional Gaussian distribution; Indicates the first A three-dimensional Gaussian distribution at the current viewpoint; Indicates the temperature coefficient; View - Covariance Matching Weights The degree of matching between the covariance matrix of the three-dimensional Gaussian distribution and the ideal point projection is evaluated based on the area or eigenvalue of the projected ellipse by transforming it to the camera coordinate system and projecting it onto the image plane.

[0049] Set contribution threshold Remove Find the anchor point, update the Gaussian set after the anchor point growth, and obtain the Gaussian set after Gaussian pruning.

[0050] Among them, contribution threshold The contribution distribution is adaptively determined by all current three-dimensional Gaussian distributions. ,in This represents the set of contributions consisting of all contributions from three-dimensional Gaussian distributions. The first of the contribution set Quantiles, lower threshold Take values ​​from 0.005 to 0.010. Take 5 to 15.

[0051] Step S3.4: Perform visual processing on the Gaussian set after Gaussian pruning. Figure 1 Consistency checks are performed to ensure that the 3D Gaussian distribution representation under different viewpoints can effectively reproduce the geometric and texture details of the real scene; specifically: See Figure 1 Consistency measurement: Gaussian rendered images of the substation engineering scene are obtained by rendering Gaussian sets after Gaussian pruning. The similarity between the Gaussian rendered images and the multi-view images of the power transmission and transformation engineering scene in step S3.1 is evaluated by PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity) indicators. The PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity) indices of Gaussian-rendered images and original images of the substation engineering scene from different viewpoints were calculated to ensure that these indices met preset quality requirements. If the performance of a certain viewpoint did not meet the requirements, the parameters were optimized to improve the viewpoint. Figure 1Consistency (return to step S3.2 for adjustment).

[0052] Step S3.5, View Figure 1 Storage of Gaussian sets after consistency checks: View Figure 1 After consistency testing, the three-dimensional Gaussian distributions of anchor points in the Gaussian set are weighted and integrated to obtain the final weighted Gaussian set. To improve computational and storage efficiency, a compact storage strategy is adopted, storing the three-dimensional Gaussian distribution of anchor points corresponding to different granularities in the final Gaussian set in a hierarchical manner. The design employs a hierarchical storage approach, loading the corresponding 3D Gaussian distributions at different granular levels as needed. An efficient index structure is established to enable fast loading and querying, outputting an indexed Gaussian set, which is ultimately an optimized Gaussian model, for fast loading, rendering, and visualization in power transmission and transformation engineering scenarios.

[0053] The specific process of step S4 is as follows: Step S4.1, Static / Dynamic Region Feature Decoupling: Based on the engineering semantic tags in the BIM model, the 3D Gaussian distribution of all anchor points in the final optimized Gaussian model obtained in Step S3 is divided into regions. Anchor points and corresponding 3D Gaussian distributions of fixed structures such as transformers, busbars, and frames are identified as static regions, maintaining their geometry and appearance while preserving high precision. Anchor points and corresponding 3D Gaussian distributions of facilities that change over time, such as construction areas and temporary equipment, are divided into dynamic regions according to update frequency (hours / days / weeks) and preset precision thresholds, providing a basis for subsequent differentiated updates.

[0054] Step S4.2, Dynamic Regional Differentiation Update Strategy: For incremental updates targeting minor changes: When only minor adjustments to the device position are detected within a dynamic region, a strategy of "moving anchor points to drive a 3D Gaussian distribution" is adopted. First, one or more local anchor points corresponding to the changed device are located using engineering semantic tags; then, the displacement vectors of these local anchor points due to the device displacement are calculated, and their spatial coordinates are updated; finally, based on the new anchor point coordinates, the position parameters of their corresponding 3D Gaussian distribution are recalculated and fine-tuned to complete the incremental update.

[0055] For local retraining with significant changes: When substantial scene changes such as equipment replacement, structural addition, or demolition are detected within a dynamic region, the original anchor point structure becomes inapplicable, requiring "joint reconstruction of anchor points and 3D Gaussian distribution." This operation first fixes the global model parameters, keeping the anchor point parameters, 3D Gaussian distribution parameters, engineering space constraint parameters, preset granularity weighting factors, and global camera calibration parameters (camera intrinsic and extrinsic parameters) unchanged at both the global and regional granularity levels. Then, only the dynamic region where changes have occurred is delineated, and the anchor points and their corresponding 3D Gaussian distributions within this dynamic region are locally retrained: for old structures that no longer exist, their corresponding anchor points and 3D Gaussian distributions are removed; for newly added or replaced structures, small-scale voxel mesh generation is performed again to generate new anchor points, and the 3D Gaussian distribution parameters of the new anchor points are initialized and optimized.

[0056] Step S4.3: Perform anchor point growth and Gaussian pruning on the incrementally updated anchor points: Both incremental updates and local retraining may lead to insufficient Gaussian representation density or redundancy in the dynamic region. Therefore, in areas of significant change (i.e., local areas in the dynamic region where equipment is replaced, structures are added / removed, or the construction scope is significantly shifted), further optimization is performed: by calculating the error gradient, a new anchor point is added at the location with the largest error and its Gaussian distribution is initialized, which is anchor point growth; at the same time, the cross-view contribution of all three-dimensional Gaussian distributions in the region is calculated, and redundant three-dimensional Gaussian distributions with a contribution below the threshold are removed, which is Gaussian pruning (using the same steps as step S3.3).

[0057] Step S4.4, View Cone Activation and Lightweight Interaction: To support efficient interaction, a view cone activation strategy is adopted, directly applied to the final optimized Gaussian model after anchor point growth and Gaussian pruning in Step S4.3. Based on the current interaction perspective, anchor points located within the view cone are quickly filtered by index, and only the corresponding 3D Gaussian distributions of these anchor points are activated and rendered. When the interaction perspective zooms in to a detailed magnified area of ​​a device, the local anchor points corresponding to that detailed magnified area are automatically identified, and their generation density is increased to ensure the fine presentation of details. For devices with low computing power, such as mobile devices, a lightweight interaction mode is enabled: by reducing the parameter dimension and complexity of the 3D Gaussian distribution, and performing on-demand loading and local caching of anchor points outside the current view cone and their corresponding 3D Gaussian distributions, smooth interaction is ensured.

[0058] Step S4.5, Lightweight Verification: After each update, lightweight verification is used to check the geometric accuracy and semantic consistency of the final optimized Gaussian model.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering, characterized in that, include: We conducted multi-source data acquisition and multi-granularity feature layered extraction for power transmission and transformation engineering scenarios. We collected multi-source heterogeneous data of scenario 3D, images, and engineering semantics. We completed scenario structure classification and semantic binding through the PointNet++ model and constructed a three-level multi-granularity hierarchical system of global-regional-local. Based on a three-level multi-granularity hierarchical system, we carried out multi-granularity structured Gaussian anchor point hierarchical construction and representation modeling, completed the hierarchical initialization of the three-level granularity anchor points, bound geometric, semantic features and engineering space constraints to the anchor points, and completed the parameterization of the three-dimensional Gaussian distribution and the fusion of the multi-granularity Gaussian set with the anchor points as the core. After accuracy verification and iterative optimization, we generated the Gaussian representation of the basic scene. Using the Gaussian representation of the basic scene as input, the view adaptive Gaussian optimization is completed. View information is collected through multi-view sampling, multi-dimensional view features are extracted, the core parameters of Gaussian are dynamically optimized, the Gaussian set is refined through anchor point growth and redundancy pruning, and the final optimized Gaussian model is generated after multi-view consistency verification. Based on the final optimized Gaussian model, dynamic updates and interactive optimizations are performed; differentiated management of static and dynamic regions is achieved through semantic decoupling; for scene changes at different scales, incremental updates based on anchor point displacement and local retraining strategies based on joint optimization of anchor point-3D Gaussian distribution are adopted for updates respectively. View cone activation and lightweight mechanisms are introduced to ensure interaction efficiency; finally, full lifecycle maintenance of the model is achieved through validation. Introducing view cone activation includes: A view cone activation strategy is adopted, which directly affects the final optimized Gaussian model: based on the current interactive viewpoint, anchor points located within the view cone are selected, and only the 3D Gaussian distribution corresponding to these anchor points is activated and rendered. When the interactive viewpoint is zoomed in to the detailed zoom area, the local anchor points corresponding to the detailed zoom area are automatically identified and their generation density is increased.

2. The method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering according to claim 1, characterized in that: The specific process of constructing a three-level multi-granularity hierarchical system of global-regional-local is as follows: Collect 3D point cloud data, image data, and engineering semantic data of power transmission and transformation engineering scenarios, and establish the spatial correspondence between engineering semantic data and 3D point cloud data and image data through coordinate registration; Using the PointNet++ model, the structure of a power transmission and transformation engineering scenario is divided into high-precision core structures, medium-precision conventional structures, and low-precision dynamic structures. The classification results are then spatially registered with the BIM model, and each 3D point cloud is assigned a corresponding BIM component ID and engineering semantic label. Among them, the high-precision core structures include transformers and power distribution equipment; the medium-precision conventional structures include supports and pipelines; and the low-precision dynamic structures include construction areas and temporary facilities. Based on 3D point clouds with BIM component IDs and engineering semantic tags, a three-level multi-granularity hierarchical system of global-region-local granularity is constructed sequentially from global to local. Differentiated data preprocessing operations are performed on point cloud data of different granularities in a three-level multi-granularity hierarchical system. Simultaneously, perspective transformation, histogram equalization, and SIFT feature point extraction are performed on image data. A multi-granularity semantic mapping table is established for semantic data to clarify the priority of semantic labels for different levels of engineering.

3. The method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering according to claim 2, characterized in that: Based on 3D point clouds with BIM component IDs and engineering semantic tags, the specific process of constructing a three-level multi-granularity hierarchical system (global-regional-local) sequentially from global to local granularity is as follows: Based on the overall spatial range of the power transmission and transformation engineering scene contained in the labeled 3D point cloud with semantic annotation, a global granularity is constructed through sparse voxel mesh to bind global semantic labels and carry the overall representation of the low-precision dynamic structure. Based on the global granularity, a regional granularity is constructed using a sparse voxel mesh according to the structure type and precision level, which binds semantic labels of the structure type and carries the representation of the medium-precision conventional structure. Based on the regional granularity, a local granularity is constructed for the high-precision core structure region by using sparse voxel meshes to bind individual BIM component IDs and semantic tags for precision requirements, focusing on the geometric shape, surface features and connection details of individual high-precision core structures.

4. The method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering according to claim 3, characterized in that: The specific process of generating the Gaussian representation of the basic scene is as follows: Based on a three-level multi-granularity hierarchical system of global-regional-local, layered anchor points are constructed using sparse voxel meshes. Preprocessed point cloud data of different granularities and SIFT feature points are bound to corresponding anchor points to generate fusion features and spatial constraint features for the anchor points. Using the bound hierarchical anchor points as the core, and combining the three-dimensional Gaussian distribution function, targeted representation modeling of different granularity scenes is performed to generate multi-granularity Gaussian sets; Accuracy verification and iterative adjustment are performed on global granularity Gaussian sets, regional granularity Gaussian sets, and local granularity Gaussian sets; The Gaussian sets of each granularity after accuracy verification and iterative adjustment are re-fused into an optimized overall scene Gaussian, i.e., the basic scene Gaussian representation, according to preset weights.

5. The method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering according to claim 4, characterized in that: The specific process of constructing layered anchor points based on a three-level multi-granularity hierarchical system of global-regional-local and combined with sparse voxel meshes is as follows: For scene-level global granularity, a sparse voxel mesh is used to divide the label 3D point cloud into multiple spatial regions, and all point clouds in a single spatial region are used as global anchor points to bind the semantic tags of the functional partition project. For the structural group-level region granularity, a sparse voxel mesh is used to perform secondary spatial partitioning within the global anchor point range to generate region anchor points and bind structural type semantic tags. For component-level local granularity, a sparse voxel mesh is used to perform fine spatial division within the coverage area of ​​the regional anchor point, generating local anchor points and binding semantic tags for accuracy requirements and component material attributes.

6. The method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering according to claim 5, characterized in that: The specific process of binding preprocessed point cloud data of different granularities and SIFT feature points to corresponding anchor points, and generating fused features and spatial constraint features for the anchor points, is as follows: For global anchor points, the overall 3D coordinate statistical features of the point cloud within the corresponding spatial range of the global anchor point are extracted, and the extracted SIFT feature points are combined to complete feature matching of the global scene from multiple perspectives. The overall 3D coordinate statistical features of the point cloud within the corresponding spatial range of the global anchor point and the geometric information of SIFT feature point matching are fused to generate the geometric description of the global anchor point. The semantic tags of the functional partition project are embedded to generate global semantic features, and the global semantic features are concatenated with the geometric description of the global anchor point to form the global anchor point fusion feature. Add global spatial constraint features to the overall engineering boundary parameters based on the BIM model; For regional anchor points, the three-dimensional coordinate statistical features of the point cloud within the corresponding spatial range of the regional anchor point are extracted. Combined with the extracted SIFT feature points, feature matching of the multi-view structure group region is completed. The three-dimensional coordinate statistical features of the point cloud within the corresponding spatial range of the regional anchor point and the geometric information of SIFT feature point matching are fused to generate the geometric description of the regional anchor point. The semantic label of the structure type is used to generate regional semantic features through the embedding layer. The regional semantic features are then concatenated with the geometric description of the regional anchor point to form the regional anchor point fusion feature. Add regional spatial constraint features to the structural group size parameters based on the BIM model; For local anchor points, the 3D coordinate statistical features of the point cloud within the corresponding region are extracted, and the extracted SIFT feature points are combined to complete feature matching of local details from multiple perspectives. The 3D point cloud in the neighborhood of the local anchor point is selected as a subset of the point cloud, and the 3D coordinate covariance matrix of this subset of the point cloud is constructed. Principal component analysis is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is used as the normal vector of the corresponding point cloud to obtain the point cloud normal vector feature. The point cloud normal vector feature is fused with the 3D coordinate statistical features of the point cloud within the corresponding region of the local anchor point and the geometric information of SIFT feature point matching to generate the geometric description of the local anchor point. The semantic label of accuracy requirement and the material attribute of the component are used to generate local semantic features through an embedding layer. The local semantic features are then concatenated with the geometric description of the local anchor point to form the local anchor point fusion feature. Add local spatial constraint features to the dimensional parameters of individual components based on the BIM model; The output consists of a set of global anchors, regional anchors, and local anchors bound to corresponding hierarchical anchor fusion features and spatial constraint features.

7. The method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering according to claim 6, characterized in that: The specific process of generating the final optimized Gaussian model is as follows: Collect view information under multiple poses and lighting conditions in power transmission and transformation engineering scenarios, and extract local texture features, multi-resolution view consistency features, lighting features and engineering semantic consistency features; The covariance matrix, color, and opacity of the 3D Gaussian distribution of anchor points in the Gaussian set corresponding to the Gaussian representation of the basic scene are dynamically optimized. Anchor point growth and Gaussian pruning optimization: Add new anchor points based on view errors and remove redundant 3D Gaussian distributions with contributions below the threshold; Perform a view consistency check on the Gaussian set after Gaussian pruning; After the view consistency check, the three-dimensional Gaussian distribution of the anchor points in the Gaussian set is weighted and integrated, and an index is built using a hierarchical storage method to output the final optimized Gaussian model.

8. The method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering according to claim 7, characterized in that: The specific process of dynamically optimizing the covariance matrix of the 3D Gaussian distribution of anchor points in the Gaussian set corresponding to the Gaussian representation of the basic scene is as follows: Obtain the camera rotation matrix and translation matrix for each viewpoint, and calculate the covariance matrix of the 3D Gaussian distribution of each anchor point in the Gaussian set corresponding to the Gaussian representation of the basic scene for each viewpoint. The target position is calculated based on local texture features, the confidence weight of each view is calculated based on multi-resolution view consistency features, the view dependency loss is constructed and minimized, and the optimal covariance matrix of the three-dimensional Gaussian distribution of each anchor point under different views is obtained. The covariance loss is constructed based on the optimal covariance matrix and the target covariance matrix, and the Gaussian geometric consistency total loss is constructed by combining the view dependency loss. The covariance matrix of the three-dimensional Gaussian distribution of the anchor points is updated by the gradient descent algorithm.

9. The method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering according to claim 8, characterized in that: The specific process of dynamically optimizing the color and opacity of the 3D Gaussian distribution of anchor points in the Gaussian set corresponding to the Gaussian representation of the base scene is as follows: For each anchor point in the Gaussian set corresponding to the Gaussian representation of the basic scene, the multi-dimensional features of the anchor point are sampled or aggregated according to its three-dimensional position projected onto the multi-view image. The multi-dimensional features include local texture features, multi-resolution view consistency features, lighting features, and engineering semantic consistency features. The multi-dimensional features are input into the MLP decoder to predict the color and opacity of the three-dimensional Gaussian distribution corresponding to the anchor point, respectively. The color loss constraint ensures that the predicted color matches the true color, and the opacity loss constraint ensures that the predicted opacity matches the true opacity. The color loss and opacity loss are combined to construct a joint optimization total loss. The color and opacity of the anchor point corresponding to the three-dimensional Gaussian distribution are updated by the gradient descent algorithm.

10. The method for three-dimensional Gaussian modeling and adaptive optimization of multi-granularity scenarios in power transmission and transformation engineering according to claim 9, characterized in that: The specific process of adding new anchor points based on view errors and removing redundant 3D Gaussian distributions with contributions below a threshold is as follows: Calculate the total rendering error, which is the sum of geometric error and visual rendering error; The error gradient is calculated based on the first partial derivative of the three-dimensional position coordinates of the three-dimensional Gaussian distribution corresponding to the anchor point with respect to the total rendering error. A new anchor point is added at the position with the largest error gradient, and the corresponding Gaussian distribution parameters are initialized to obtain the Gaussian set after the anchor point is grown. For each anchor point in the Gaussian set after anchor point growth, calculate the contribution of its corresponding Gaussian distribution under different views. The contribution is determined by the product of visibility weight, depth weight and view-covariance matching weight. Remove anchor points whose contribution is below a preset threshold to obtain a Gaussian set after Gaussian pruning.

Images