An industrial equipment 3D visualization method based on key feature recognition

Abstract

This invention provides a 3D visualization method for industrial equipment based on key feature recognition, belonging to the field of industrial equipment digitization technology. The method achieves accurate lightweighting and efficient visualization of industrial equipment 3D models through a five-step process: constructing a standardized knowledge base for functional components of power distribution equipment, identifying key regions based on a knowledge-enhanced graph convolutional neural network, hierarchical lightweight reduction of polygons, ASTC texture hierarchical compression, and a line-of-sight adaptive LOD mechanism. This invention uses a hierarchical polygon reduction technique that preserves key features, dividing the power distribution equipment into key and non-key feature areas, and selectively allocating polygon reduction rates of 20% and 80%. Combined with an ASTC near-lossless compression strategy for key areas, the recognition accuracy of key information such as circuit breaker opening / closing indicators and busbar T-type connection structures remains above 95%. Simultaneously, a topology protection algorithm ensures the integrity of the overall model structure and connection relationships, balancing model simplification with industrial practicality.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment digitization technology, and in particular to a method for 3D visualization of industrial equipment based on key feature recognition. Background Technology

[0002] In the process of industrial digital transformation, 3D visualization technology has become a core support means for equipment operation, maintenance, inspection and fault diagnosis in fields such as power. Among them, the accuracy and practicality of 3D modeling and visualization of the power distribution room, as a key node of the power system, directly affects the operation and maintenance efficiency and safety level of the industrial site.

[0003] Current technologies in 3D power distribution room modeling primarily employ the traditional approach of "full-detail modeling + general lightweight design." In the modeling phase, this approach uses "full-detail modeling" as its core strategy. Taking a power distribution room containing 120 devices as an example, it creates a high-precision 3D model for each device, containing 8,000 to 15,000 polygons, ultimately forming a complex mesh structure with approximately 1 million polygons. For textures, each device is configured with an independent, uncompressed PNG texture map with a resolution of 2048×2048 pixels, with a single texture file size of approximately 5MB, resulting in a total texture data size of around 200MB for the entire power distribution room. This technical implementation process involves five core steps: manual modeling, UV mapping, engine parameter settings, scene assembly, and format export. It relies on professional modeling software such as 3DS Max and Maya, as well as game engines like Unity and Unreal Engine, creating 3D scenes through parametric configuration. While it boasts advantages such as high development efficiency and a mature technology stack, it is essentially "general-purpose" and lacks targeted optimization for industrial operation and maintenance scenarios.

[0004] To adapt to the performance limitations of mobile devices, existing technologies also attempt to apply general lightweight methods: model simplification relies on the automatic polygon reduction tool built into the modeling software, using a globally uniform polygon reduction rate of 50% to 70%, and automatically merging adjacent faces and deleting redundant vertices through algorithms; texture optimization uses general compression format conversion or batch processing with online tools, which can only reduce texture size by 30% to 50%; LOD technology only sets two levels of switching, switching between the original model and the simplified model with 50% polygon reduction based on the straight-line distance from the device to the camera.

[0005] However, the aforementioned existing technologies have drawbacks:

[0006] First, there is a significant performance bottleneck. The 3D model of the power distribution room containing 120 devices takes 15 to 20 seconds to load in a mobile 4G network environment, with an operation frame rate of only 15 frames, which is far from meeting the rapid response requirements of industrial sites. This may lead to delays in fault handling and affect normal production and life.

[0007] Second, there is a fundamental contradiction between preserving details and making the model lightweight. A globally uniform "one-size-fits-all" simplification strategy will lead to the loss of key details such as unclear markings on operating devices, unclear structure of equipment connection points, and unrecognizable text on cabinet markings, which in turn will cause safety hazards.

[0008] Third, the application scenarios are poorly adapted. The general 3D modeling thinking does not fully consider the needs of industrial operation and maintenance for accurate visual feedback and rapid spatial positioning. The mobile terminal usage rate is only 40%, and the technology is seriously out of touch with the actual application scenarios. Summary of the Invention

[0009] In view of the shortcomings of the prior art, the purpose of this invention is to provide a 3D visualization method for industrial equipment based on key feature recognition. This method can solve the problems of existing technologies, such as prominent performance bottlenecks on mobile devices, long 3D model loading times, low operation frame rates, and inability to meet the rapid response needs of industrial sites; a fundamental contradiction between detail preservation and model lightweighting, with a globally uniform one-size-fits-all simplification strategy leading to the loss of key details such as operation labels and connection structures, causing safety hazards; poor adaptability to application scenarios, with general 3D modeling thinking not fully combined with the operational habits and environmental requirements of industrial operation and maintenance, resulting in low mobile device usage and difficulty in supporting actual on-site technical problems.

[0010] A first aspect of this invention proposes a method for 3D visualization of industrial equipment based on key feature recognition, comprising the following steps:

[0011] S1: Construct a standardized knowledge base for functional components of power distribution equipment. The knowledge base includes at least a functional layer, a geometric layer, and a topology layer. Calculate the comprehensive weight of each component based on safety criticality, operational importance, identification necessity, and maintenance frequency.

[0012] S2: Obtain the 3D mesh model of the industrial equipment to be visualized and construct graph structure data. Based on the functional semantic rules and topological constraint rules obtained in step S1, use a knowledge-enhanced graph convolutional neural network to perform feature learning on the graph structure data, output component function classification results, semantic confidence and importance scores, and generate regional classification results and component weights accordingly.

[0013] S3: Based on the region classification results, the component comprehensive weight and / or the importance score, perform hierarchical lightweight reduction processing on the three-dimensional mesh model. During the reduction process, use the semantic confidence and the topological constraint rules to constrain or penalize candidate simplification operations to obtain the three-dimensional geometric model after reduction.

[0014] S4: Based on the region classification results and / or the component weights, the texture data corresponding to the three-dimensional geometric model is subjected to hierarchical compression using adaptive scalable texture compression ASTC to obtain compressed textures and generate three-dimensional visualization resources.

[0015] S5: Obtain camera viewpoint and viewing distance parameters at the rendering end, construct a multi-level LOD switching mechanism with adaptive viewing distance, and adjust the LOD switching by delay or advance according to the component weights. At the same time, adopt a smooth transition strategy to realize continuous switching between adjacent LOD models, so as to output the three-dimensional visualization results of industrial equipment.

[0016] In a further preferred embodiment, in step S1:

[0017] The functional layer establishes component entries according to the hierarchical structure of "equipment type → functional category → specific function". The equipment type includes at least circuit breakers, busbar systems, distribution cabinets, disconnect switches, current transformers, voltage transformers, surge arresters, capacitors, reactors, transformers and switch cabinets. The functional category includes at least operation, connection, protection, transmission, insulation, identification, structure and auxiliary categories.

[0018] The geometry layer establishes a standardized geometric template for each component and configures its shape features and size parameters. The geometric template includes at least an operating handle template, a connection terminal template, an identification area template, and a structural support template.

[0019] The topology layer defines the electrical connection relationships, mechanical connection relationships, and spatial constraint relationships between components, and divides the topology rules into strong constraints, medium constraints, and weak constraints. Among them, strong constraints include at least electrical connection continuity and mechanical transmission relationships.

[0020] More preferably, the component comprehensive weight W in step S1 is calculated according to the following formula: W = 0.40W1 + 0.30W2 + 0.20W3 + 0.10W4;

[0021] Among them, W1 is the safety criticality weight, W2 is the operational importance weight, W3 is the identification necessity weight, and W4 is the maintenance frequency weight; the normalized values ​​of the above weights range from 0.1 to 1.0, and the values ​​of W1 and W2 for critical functional components are preferably not lower than 0.8, and the values ​​of W1 and W2 for non-critical components are preferably not higher than 0.3.

[0022] More preferably, the graph structure data in step S2 uses grid vertices and / or patches as graph nodes and grid adjacency relationships as graph edges;

[0023] The node features include at least geometric attributes and topological attributes, and further include the predicted functional label features obtained by geometric template matching in step S1;

[0024] Edge features include at least geometric relationship features, and further include constraint features such as electrical connection legality, mechanical dependency relationship and / or structural support relationship.

[0025] More preferably, the knowledge-enhanced graph convolutional neural network in step S2 includes: an input encoding layer, at least one graph convolutional layer, a knowledge enhancement layer, a global feature aggregation layer, and a dual-head output layer;

[0026] The input encoding layer standardizes and normalizes node and edge features to unify their dimensions.

[0027] Graph convolutional layers aggregate neighborhood features through a message passing mechanism and pass features through residual connections to avoid gradient vanishing.

[0028] The knowledge enhancement layer is used to fuse domain knowledge with geometric features;

[0029] The global feature aggregation layer uses attention pooling to aggregate vertex features into graph-level features;

[0030] The dual-head output layer is used to output the component function classification results and the component importance score, respectively.

[0031] 6. The industrial equipment 3D visualization method based on key feature recognition according to claim 5, characterized in that: the knowledge enhancement layer fuses the knowledge vector encoded by functional semantic rules and topological constraint rules with the output features of the graph convolutional layer, and the fusion method satisfies: Fen h = Fgcn + Fattn + Fk;

[0032] Wherein, Fgcn is the output feature of the graph convolutional layer, Fattn is the feature obtained based on the multi-head self-attention mechanism, with an optimal number of heads of 8 and a dropout rate of 0.1, Fk is the feature after linear mapping of the knowledge vector, and Fenh is the enhanced feature after fusion.

[0033] In a further preferred embodiment, in step S2, the region is divided into at least five levels based on its importance score, with different levels of regions corresponding to different reduction rates and geometric error thresholds:

[0034] Preferably, the key functional areas (importance score 0.8 to 1.0) correspond to a reduction rate of ≤20% and a geometric error of <1%;

[0035] Important identification regions (importance score 0.6–0.8) correspond to a reduction rate of ≤40% and a geometric error of <2%;

[0036] Functionally important regions (importance score 0.4–0.6) correspond to a reduction rate of ≤60% and a geometric error of <5%;

[0037] The structural support area (importance score 0.2-0.4) corresponds to a reduction rate of ≤80% and a geometric error of <10%;

[0038] Non-critical areas (importance score 0-0.2) correspond to a reduction rate of ≥95% and an allowable error of ≤20%.

[0039] In a further preferred embodiment, in step S3, a semantically aware surface reduction cost function is defined for the candidate simplification operation a:

[0040] C(a)=E(a)·(1+λ·W·conf)+μ·Ptopo(a);

[0041] Where E(a) is the geometric error term (calculated from the changes in edge length and curvature), W is the component comprehensive weight obtained in step S1, conf is the semantic confidence or importance score output in step S2, Ptopo(a) is the penalty term for violating topological constraint rules, preferably ≥1000 for strong constraint violations, preferably 500~1000 for medium constraint violations, and preferably 100~500 for weak constraint violations; λ and μ are weight coefficients, preferably λ=1000 and μ=500;

[0042] Candidate simplification operations are sorted according to the size of C(a), and operations with lower costs are performed first.

[0043] More preferably, in step S4, different ASTC compression block sizes are selected for different texture regions based on the region classification results and / or component weights:

[0044] Key texture areas preferably use ASTC 2x2 or 4x4 format, important texture areas preferably use ASTC 6×6 or 8×8 format, general texture areas preferably use ASTC 10×10 format, and minor texture areas preferably use ASTC 12×12 or 16×16 format.

[0045] Establish a standardized component-shared texture library, extract texture resources from common components, enable texture reuse for components of the same type, and reduce data redundancy.

[0046] In a further preferred embodiment, in step S5:

[0047] Construct at least a three-level LOD model, with different polygon retention ratios corresponding to different model levels;

[0048] The multi-level LOD model is switched according to the line-of-sight parameters of the camera and the device or component, and a smooth transition factor is calculated based on the line-of-sight parameters. The smooth transition factor is preferably: transition_factor = max(0.1, 1.0 - normalized_distance·0.3), where normalized_distance is the distance ratio after the line-of-sight parameters are normalized.

[0049] The LOD switching logic is adjusted based on the overall weight of the components, with the preferred method being to delay the switching by one level when W≥0.8 and advance the switching by one level when W≤0.3;

[0050] Linear interpolation fusion is performed on the current LOD model and the target LOD model based on the transition factor. The preferred transition time is 0.5 seconds to ensure that the frame rate of the mobile device is stable at more than 30fps, so as to achieve continuous switching between adjacent LOD models.

[0051] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following:

[0052] I. This invention uses a layered surface reduction technique that preserves key features to divide power distribution equipment into key feature areas and non-key feature areas, and allocates surface reduction rates of 20% and 80% accordingly. Combined with the near-lossless compression strategy of ASTC key areas, the recognition rate of key information such as circuit breaker opening and closing indicators and bus T-type connection structures is maintained at over 95%. At the same time, the topology protection algorithm ensures the integrity of the overall structure and connection relationship of the model, completely solving the pain point of "lightweighting inevitably loses details" in the prior art, and taking into account both model simplification and industrial practicality.

[0053] II. This invention achieves a performance leap through three levels of technical optimization: First, the layered polygon reduction technology accurately reduces the number of polygons in the model from 1 million to 200,000, reducing the GPU vertex processing load by 80%; second, the ASTC hierarchical compression technology compresses the texture size from 200MB to 25MB, achieving a compression rate of 87.5%. Combined with GPU hardware decoding acceleration, this shortens the mobile loading time from 15-20 seconds to 2-3 seconds, increasing the loading speed by 86.7%; third, the view distance adaptive three-level LOD mechanism dynamically allocates rendering resources according to different view distances in the operation and maintenance scenario, increasing the mobile frame rate from 15 frames to 60 frames and reducing the operation response latency from 66 milliseconds to 16.7 milliseconds, fully meeting the real-time requirements of rapid on-site inspection and fault handling.

[0054] Third, this invention significantly improves adaptability through scenario-oriented technical design: the LOD mechanism accurately matches three core operation and maintenance scenarios—detailed operation, equipment identification, and overall positioning—and dynamically adjusts the model's detail accuracy; high-precision preservation of key feature areas ensures that operation and maintenance personnel can quickly identify equipment status and locate faulty components; optimized interaction logic eliminates the need for global model updates for operations such as scaling and rotation, further improving operational smoothness. Ultimately, mobile usage increased from 40% to 70%, and on-site operation and maintenance efficiency improved by 50%, effectively supporting the actual work needs of industrial sites.

[0055] Fourth, this invention reduces overall costs through standardization and reusable design: First, it constructs a standardized knowledge base for functional components of power distribution equipment. New project development only requires parameter configuration to automatically generate models, achieving a code reuse rate of over 80% and improving development efficiency by 50%. Second, the significant reduction in model and texture data volume reduces server bandwidth costs by 30%, while the smaller file format improves CDN cache hit rate, further reducing redundant transmission overhead. Third, the technical solution adopts a modular architecture, which can be seamlessly extended to other industrial equipment scenarios such as substations and factory workshops, forming a replicable lightweight methodology for industrial 3D models, possessing significant value for large-scale promotion. Attached Figure Description

[0056] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0057] Figure 1 This is a flowchart of steps S1 to S5 provided in an embodiment of the present invention;

[0058] Figure 2 A flowchart of the topology constraint protection mechanism provided in an embodiment of the present invention;

[0059] Figure 3 This is a flowchart of the ASTC texture hierarchical compression process provided in an embodiment of the present invention. Detailed Implementation

[0060] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0061] The large-scale safe solidification and storage method for fly ash provided by the present invention will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0062] Reference manual attached Figure 1-3 This paper illustrates a method for 3D visualization of industrial equipment based on key feature recognition, provided by an embodiment of the present invention, comprising the following steps:

[0063] Step S1: Construction of a standardized knowledge base for functional components of power distribution equipment

[0064] (1) Functional layer construction

[0065] Component entries are established strictly according to a three-level hierarchical structure of "equipment type → function category → specific function" to ensure that the functional positioning of each component is clear and traceable.

[0066] Equipment types: Comprehensive coverage of core equipment in power distribution rooms, including circuit breakers, busbar systems, distribution cabinets, disconnect switches, current transformers, voltage transformers, surge arresters, capacitors, reactors, transformers and switchgear, totaling 11 categories;

[0067] Functional Categories: Based on the needs of power operation and maintenance, they are divided into eight categories: operation, connection, protection, transmission, insulation, identification, structure, and auxiliary.

[0068] Specific function mapping: Taking a circuit breaker as an example, the operation components include the operation handle and the opening and closing indicators; the connection components include the main contact system and the connection terminals; the protection components include the arc-extinguishing chamber, the insulating tie rod, the outer shell body, and the base support; and the auxiliary components include the auxiliary contacts and the fastening bolts.

[0069] (2) Geometric layer construction

[0070] Standardized geometric templates are established for each type of component. Their shape characteristics, size range, and material properties are clearly defined through parametric descriptions to ensure the algorithmic recognizability of geometric features. The geometric template parameters for core components are as follows:

[0071] Operating handle: The geometry is a lever mechanism, the shape is a long bar lever, the length-to-width ratio is 3:1, the length range is 200-400mm, the thickness range is 20-40mm, the material is high-strength engineering plastic or metal, and the location is on the outside of the equipment.

[0072] Connection terminal: The geometric type is circular terminal, the shape is disc-shaped, the diameter range is 8-25mm, the height range is 2-8mm, the material is electrolytic copper plated with silver, and the location is the external connection area of ​​the equipment;

[0073] Opening and closing indicator: The geometric type is a sign, the shape is rectangular, the standard size is 30×15mm, the thickness range is 1-5mm, the material is high-contrast plastic, and the location is at the front of the equipment.

[0074] Main busbar conductor: The geometric type is a rectangular conductor, the shape is a cuboid, the cross-sectional size is 40×10mm, the length is cut as needed, the material is T2 electrolytic copper, and the location is the main busbar trunk.

[0075] (3) Construction of topology layer

[0076] The system defines the connection relationships, spatial relationships, and dependencies between components, and classifies them into three levels of rules—strong constraints, medium constraints, and weak constraints—according to their impact, to ensure that critical topological structures are not disrupted during subsequent processing.

[0077] Electrical connection relationships: The series connection between the main contact system and the arc-extinguishing chamber (strong constraint) requires an electrical continuity error of ≤0.1mm; the bolted connection between the bus conductor and the connection terminal (strong constraint) must meet the requirements for current carrying capacity and insulation distance.

[0078] Mechanical connection relationships: Shaft connection between the operating handle and the transmission mechanism (strong constraint), coaxiality deviation ≤ 0.5mm; Bolt connection between the T-type connector and the bus conductor (medium constraint), connection strength and anti-loosening measures must be guaranteed;

[0079] Spatial relationships: The proximity of the opening and closing indicators to the operating handle (medium constraint), with a distance range of 0.2-0.4m; the horizontal alignment of the instrument display and the operating buttons (weak constraint), ensuring ease of operation and reasonable layout.

[0080] (4) Calculation of component comprehensive weight

[0081] The weighted summation formula is used to calculate the overall basic weight of the components. The weight allocation ratio fully reflects the safety priority principle of power operation and maintenance. The formula is as follows: W=0.40W1+0.30W2+0.20W3+0.10W4, where W1 is the safety criticality weight, W2 is the operational importance weight, W3 is the identification necessity weight, and W4 is the maintenance frequency weight. After normalization, the value of each weight is in the range of 0.1-1.0.

[0082] Step S2: Key Region Identification Based on Knowledge-Enhanced GCN

[0083] By converting 3D mesh models into graph-structured data, and utilizing knowledge-enhanced graph convolutional neural networks to achieve deep fusion of functional semantics and geometric features, accurate regional hierarchical results are output. The specific implementation process is as follows:

[0084] (1) Construction of graph structure data

[0085] Node definition: Using the vertices of the 3D mesh model as graph nodes, for the original model with 1 million polygons, approximately 800,000 valid nodes are extracted, and duplicate and isolated vertices are removed;

[0086] Edge definition: The adjacency relationship between grid vertices is used as the graph edge. Approximately 2.4 million edges are automatically generated based on the grid topology to ensure that the edge connection relationship is consistent with the original model.

[0087] Feature extraction:

[0088] Node features include geometric attributes (3D position coordinates, principal curvature, Gaussian curvature, mean curvature, unit normal vector), topological attributes (degree, clustering coefficient, betweenness centrality, eigenvector centrality), and predictive function labels (matching score based on S1 geometric template, ranging from 0 to 1).

[0089] Edge features: include geometric relationships (edge ​​length, direction vector, dihedral angle, local curvature) and constraint features (electrical connection legality, mechanical dependency relationship, structural support relationship; legal state is marked as 1, illegal state is marked as 0).

[0090] (2) Knowledge-enhanced GCN model construction and training

[0091] The knowledge-enhanced GCN model adopts an architecture of "input encoding - graph convolution - knowledge enhancement - global aggregation - dual-head output", which improves the accuracy of feature recognition by fusing domain knowledge.

[0092] Model Structure: The input encoding layer standardizes and normalizes node and edge features, unifying them into 128-dimensional vectors. The graph convolutional layer uses a three-layer stacked structure, aggregating neighborhood features through a message passing mechanism. Each layer's output dimension remains 128, and gradient vanishing is avoided through residual connections. The knowledge enhancement layer encodes the functional semantic rules and topological constraint rules of S1 into 64-dimensional knowledge vectors, which are then linearly mapped to 128 dimensions and fused with graph convolutional features and multi-head self-attention outputs. The global feature aggregation layer uses attention pooling to aggregate vertex-level features into graph-level features. The dual-head output layer outputs the functional classification results (5 functional labels) and importance scores (0-1), respectively.

[0093] Training parameters: Input feature dimension is 64 (32-dimensional geometric attributes + 16-dimensional topological attributes + 16-dimensional predicted function labels), hidden layer dimension is 128, functional classification output dimension is 5, and knowledge vector dimension is 64; learning rate is set to 0.001, batch size is 32, training iterations are 100 rounds, and the cross-entropy loss function (classification task) and MSE loss function (regression task) are jointly optimized.

[0094] Training results: The model achieved a functional classification accuracy of 94% and a key region recognition rate of 92% on the test set. The MAE (mean absolute error) of importance score prediction was 0.03, which meets the requirements of practical applications.

[0095] (3) Generation of regional classification results

[0096] Based on the importance score (0-1) output by the model, the device model is divided into five levels of regions;

[0097] Step S3: Layered lightweight surface reduction process

[0098] Step S3, based on the region classification results of S2, selects the optimal simplification operation through a semantically aware surface reduction cost function, and applies differentiated surface reduction algorithms to regions of different levels. This achieves model lightweighting while preserving key features and topological constraints. The specific implementation process is as follows:

[0099] (1) Execution of Differentiated Surface Reduction Algorithm

[0100] To address the characteristics of different regions, a dedicated polygon reduction algorithm is employed, and the polygon reduction parameters are strictly controlled.

[0101] Level 5 region: A conservative polygon reduction algorithm is employed, performing key feature identification, conservative edge folding, topological constraint verification, and functional integrity checks. By identifying key features such as the operating handle and main contact system, an edge folding priority queue is generated, and simplification operations are performed from low to high cost, with the maximum polygon reduction rate controlled within 20%. Ultimately, the polygon reduction rate for this region is 18%, retaining 82% of the original geometric details, with a geometric error ≤0.8%.

[0102] Level 4 area: Employing a shape-preserving and surface-reduction algorithm, this area prioritizes the protection of the contour, angular, and proportional features of components such as equipment markings and instrument displays. Visual fidelity is ensured by calculating shape similarity, resulting in a final surface reduction rate of 35% and a visual fidelity of 92%.

[0103] Level 3 region: A balanced face reduction algorithm is adopted, which combines edge folding and vertex clustering techniques to seek a balance between computational efficiency and visual quality, and finally the face reduction rate is 55%;

[0104] Level 2 area: Using a structural reduction algorithm, the relative positions of key structural points of components such as the cabinet frame and support base are preserved to ensure overall structural stability, resulting in a final reduction rate of 78%;

[0105] Level 1 region: An aggressive face reduction algorithm is adopted, which maximizes simplification through vertex clustering and large-area edge folding, resulting in a face reduction rate of 96%.

[0106] (2) Application of semantic-aware surface cost function

[0107] To ensure the priority retention of key areas, a semantically aware face reduction cost function is defined as follows: C(a)=E(a)⋅(1+λ⋅W⋅conf)+μ⋅Ptopo(a) where E(a) is the geometric error term of the candidate simplification operation a (calculated from the changes in edge length and curvature), W is the component comprehensive weight, conf is the semantic confidence, Ptopo(a) is the penalty term for violating topological constraints, and λ=1000 and μ=500 are weight coefficients.

[0108] Penalty settings: Ptopo(a) = 1000 for strong constraint violation, Ptopo(a) = 800 for medium constraint violation, and Ptopo(a) = 300 for weak constraint violation;

[0109] Execution logic: Sort by cost function value in ascending order, prioritize simplification operations with lower cost, and avoid oversimplification of critical areas.

[0110] (3) Verification and optimization of topological constraints

[0111] During the reduction of the surface area, the strong constraint rules defined in S1 are verified in real time, and the integrity of the connection and spatial relationships is verified through topological graph traversal and geometric calculations.

[0112] Electrical connection constraint verification: By traversing the electrical connection diagram between components, the continuity of the connection between the main contact and the arc-extinguishing chamber, and between the busbar and the terminal was checked. Two potential connection breaks were found, triggering operation rollback and adjustment of the reduction parameters.

[0113] Mechanical transmission constraint verification: Verify the coaxiality of the operating handle and the transmission mechanism, and reject any folding operations with errors exceeding the standard in 3 places;

[0114] Final output: The number of polygons in the 3D geometric model after reduction is approximately 180,000, which is 82% less than the original model. The topological constraint satisfaction rate is 100%, and the key feature recognition rate is 96%.

[0115] (4) Quality Control Closed-Loop Execution: To ensure that key features are not lost, the topological structure is intact, and the geometric error meets the threshold requirements during the surface reduction process, this scheme designs a multi-dimensional quality control closed-loop mechanism, which performs quality assessment and dynamic correction in the two core stages of surface reduction at the regional level and before global fusion, respectively:

[0116] Regional-level quality assessment and correction (triggered after each level of regional surface reduction): A multi-dimensional quality assessment system is adopted to calculate the quality score Q of the current region. The assessment indicators include: geometric error (30% after Hausdorf distance normalization), topological constraint satisfaction rate (30%), key feature retention rate (25%), and visual fidelity (15%).

[0117] Preset quality thresholds for each region level: Level 5 region Q≥0.95, Level 4 region Q≥0.90, Level 3 region Q≥0.85, Level 2 region Q≥0.80, Level 1 region Q≥0.75. If the quality score falls below the corresponding threshold, the system automatically triggers a correction mechanism.

[0118] Reduce the area reduction target for this region (e.g., reduce the reduction rate from 20% to 15% for level 5 regions and from 40% to 30% for level 4 regions).

[0119] Switch to a more conservative surface reduction algorithm strategy (such as switching from balanced surface reduction to shape-preserving surface reduction for level 3 regions, and switching from structural surface reduction to balanced surface reduction for level 2 regions).

[0120] Increase the protection weight of key features (in the semantic-aware surface cost function, increase the coefficient of λ from 1000 to 1500 to strengthen the constraint penalty of key components).

[0121] Before global fusion, quality assessment and correction are triggered after all regions have reduced their surface area and before the final model is output: Calculate the global quality score Qtotal (the weighted sum of the quality scores of each region according to its proportion), with a preset global threshold of Qtotal≥0.88; if it does not meet the standard, the system will first perform a secondary correction on the region with the lowest quality score. If it still does not meet the standard after the secondary correction, the surface area reduction rate will be uniformly reduced by 5%-10% for all regions of all levels, and the surface area reduction process will be re-executed until the global quality score meets the threshold requirement.

[0122] Step S4: ASTC Texture Hierarchical Compression Optimization

[0123] Step S4, based on the region classification results and component weights, performs differentiated ASTC compression on the texture data. By sharing a texture library, redundancy is reduced, and the amount of texture data is reduced while ensuring the readability of key information. The specific implementation process is as follows:

[0124] (1) Texture region classification

[0125] Combining the region classification results of S2 with the component weights of S1, the texture data is divided into four regions:

[0126] Key texture areas (W≥0.8): These include operating handle markings, opening and closing indicator text, equipment nameplates, etc., and require extremely high readability.

[0127] Important texture areas (0.5 ≤ <): including instrument display interface, bus phase indicators, status indicator lights, etc., must be clearly marked;

[0128] General texture areas (0.3≤<): including cabinet surface material, insulator texture, ventilation details, etc., the basic material texture must be preserved;

[0129] Secondary texture areas (<): These include shell decorative textures, auxiliary component surfaces, fastener textures, etc., and have lower requirements for visual quality.

[0130] (2) Hierarchical compression execution

[0131] Differential compression is performed using the ASTC format;

[0132] (3) Construction of shared texture library

[0133] Common component textures (such as standard connection terminals, operation buttons, phase indicators, etc.) of 11 types of power distribution equipment were extracted to establish a standardized shared texture library, containing 28 types of common texture resources. Components of the same type can directly reuse textures to reduce redundant data. Through texture feature extraction and similarity matching, textures with a repetition rate of more than 85% are grouped into one category, assigned a unique identifier, and stored in the shared library. When the model is loaded, the corresponding texture is called by the identifier.

[0134] (4) Texture detail enhancement processing

[0135] For critical texture areas, normal map baking technology is used to help preserve details: the surface bump features and material grain details of the high-resolution original textures in the shared texture library are baked into normal maps and bound to the compressed base texture for storage; during mobile rendering, the normal map information is superimposed onto the base texture through GPU hardware acceleration to restore the realistic texture of the device surface. This processing method does not increase the amount of additional texture loading. The normal map is compressed in ASTC8×8 format, with a single image size of ≤500KB, and can improve the material detail recognition of critical areas by more than 15%, meeting the observation needs of industrial operation and maintenance for the surface condition of equipment.

[0136] Compression result verification

[0137] The final total texture data is approximately 18MB, a 91% reduction from the original 200MB. The text recognition accuracy in key texture areas is ≥98%, and the recognition rate of important symbols is ≥95%, meeting the visual quality requirements for mobile visualization.

[0138] Key texture areas (W≥0.8): These include operating handle markings, circuit breaker indicator text, equipment nameplates, etc., which have extremely high readability requirements; ASTC2x2 or 4x4 format, compression ratio 8:1-10:1, quality loss ≤5%; At the same time, a dedicated detail layer is superimposed to extract high-frequency details of the original texture, such as text edge sharpening information and color contrast enhancement information of the markings. The detail layer adopts YUV color space optimized encoding, retaining only the high-frequency data of the luminance channel, with a volume ratio ≤10% of the base texture. Through the superposition and rendering of the base texture and the detail layer, the text recognition accuracy is ensured to be ≥98%, and the marking lines are clear and free of jagged edges.

[0139] Step S5: Deployment of the line-of-sight adaptive LOD mechanism

[0140] Step S5 constructs a multi-level LOD model and an adaptive switching mechanism, dynamically adjusting model details based on camera viewing distance and component weights. A smooth transition strategy ensures visual continuity and enhances the mobile interactive experience. The specific implementation process is as follows:

[0141] (1) Construction of multi-level LOD model

[0142] A 5-level LOD model is constructed based on the 3D geometric model after surface reduction and compressed texture.

[0143] (2) Implementation of LOD switching logic

[0144] Distance calculation: The real-time distance between the camera and the center of the component is calculated using the Euclidean distance formula: d=(x1−x2)2+(y1−y2)2+(z1−z2)2 where (x1,y1,z1) are the camera position coordinates and (x2,y2,z2) are the coordinates of the center of the component.

[0145] Weighting adjustment strategy: When the overall weight of a component is W≥0.8, the switching is delayed by one level (e.g., LOD0 is retained at 3 meters); when W≤0.3, the switching is advanced by one level (e.g., switching to LOD1 at 3 meters).

[0146] Smooth transition is achieved by using cross-fading technology when switching between adjacent LODs. The transition factor is calculated as follows: transitionfactor = max(0.1, 1.0 − normalized_distance ⋅ 0.3), where normalized_distance is the view distance normalization ratio (current view distance / switching threshold). Vertex and texture data of the current LOD and the target LOD are fused through linear interpolation, and the transition duration is set to 0.5 seconds to avoid visual jumps.

[0147] The results of the actual test conducted in a mobile 4G network environment are as follows:

[0148] Loading time: 2.2 seconds, an 87.8% reduction compared to the original model's 18 seconds;

[0149] Operating frame rate: average 38fps, peak 45fps, minimum 32fps, meeting the target of ≥30fps;

[0150] Visual continuity: There are no obvious jumps when switching between LODs, and key areas remain clearly distinguishable at all viewing distances;

[0151] Interactive response: The response latency for click, rotation, scaling, and other operations is ≤16ms, ensuring smooth operation without lag.

[0152] This embodiment achieves the following technical effects through the complete execution of S1-S5:

[0153] The number of polygons in the model was reduced from 1 million to 180,000, a reduction of 82%; texture data was compressed from 200MB to 18MB, a reduction of 91%, which greatly reduced data transmission and storage costs.

[0154] Mobile loading time ≤ 2.5 seconds, frame rate stable at 30fps or higher, operation response latency ≤ 16ms, solving the performance bottleneck problem of existing technologies;

[0155] The identification rate of key functional areas is ≥95%, the geometric error is ≤0.8%, and the topological constraint satisfaction rate is 100%, thus avoiding security risks caused by the loss of key details;

[0156] On-site maintenance personnel can quickly identify equipment status and locate faulty components through mobile devices, reducing the average fault handling response time by 12 minutes and increasing the mobile device usage rate from 40% to 75%, effectively improving industrial maintenance efficiency.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A method for 3D visualization of industrial equipment based on key feature recognition, characterized in that, Includes the following steps: S1: Construct a standardized knowledge base for functional components of power distribution equipment. The knowledge base includes at least a functional layer, a geometric layer, and a topology layer. Calculate the comprehensive weight of each component based on safety criticality, operational importance, identification necessity, and maintenance frequency. S2: Obtain the 3D mesh model of the industrial equipment to be visualized and construct graph structure data. Based on the functional semantic rules and topological constraint rules obtained in step S1, use a knowledge-enhanced graph convolutional neural network to perform feature learning on the graph structure data, output component function classification results, semantic confidence and importance scores, and generate regional classification results and component weights accordingly. S3: Based on the region classification results, the component comprehensive weight and / or the importance score, perform hierarchical lightweight reduction processing on the three-dimensional mesh model. During the reduction process, use the semantic confidence and the topological constraint rules to constrain or penalize candidate simplification operations to obtain the three-dimensional geometric model after reduction. S4: Based on the region classification results and / or the component weights, the texture data corresponding to the three-dimensional geometric model is subjected to hierarchical compression using adaptive scalable texture compression ASTC to obtain compressed textures and generate three-dimensional visualization resources. S5: Obtain camera viewpoint and viewing distance parameters at the rendering end, construct a multi-level LOD switching mechanism with adaptive viewing distance, and adjust the LOD switching by delay or advance according to the component weights. At the same time, adopt a smooth transition strategy to realize continuous switching between adjacent LOD models, so as to output the three-dimensional visualization results of industrial equipment.

2. The industrial equipment 3D visualization method based on key feature recognition according to claim 1, characterized in that: In step S1: The functional layer establishes component entries according to the hierarchical structure of "equipment type → functional category → specific function". The equipment type includes at least circuit breakers, busbar systems, distribution cabinets, disconnect switches, current transformers, voltage transformers, surge arresters, capacitors, reactors, transformers and switch cabinets. The functional category includes at least operation, connection, protection, transmission, insulation, identification, structure and auxiliary categories. The geometry layer establishes a standardized geometric template for each component and configures its shape features and size parameters. The geometric template includes at least an operating handle template, a connection terminal template, an identification area template, and a structural support template. The topology layer defines the electrical connection relationships, mechanical connection relationships, and spatial constraint relationships between components, and divides the topology rules into strong constraints, medium constraints, and weak constraints. Among them, strong constraints include at least electrical connection continuity and mechanical transmission relationships.

3. The industrial equipment 3D visualization method based on key feature recognition according to claim 1, characterized in that: In step S1, the overall weight W of the components is calculated according to the following formula: W = 0.40W1 + 0.30W2 + 0.20W3 + 0.10W4; Among them, W1 is the safety criticality weight, W2 is the operational importance weight, W3 is the identification necessity weight, and W4 is the maintenance frequency weight; the normalized values ​​of the above weights range from 0.1 to 1.0, and the values ​​of W1 and W2 for critical functional components are preferably not lower than 0.8, and the values ​​of W1 and W2 for non-critical components are preferably not higher than 0.

3.

4. The industrial equipment 3D visualization method based on key feature recognition according to claim 1, characterized in that: The graph structure data described in step S2 uses grid vertices and / or patches as graph nodes and grid adjacency relationships as graph edges. The node features include at least geometric attributes and topological attributes, and further include the predicted functional label features obtained by geometric template matching in step S1; Edge features include at least geometric relationship features, and further include constraint features such as electrical connection legality, mechanical dependency relationship and / or structural support relationship.

5. The industrial equipment 3D visualization method based on key feature recognition according to claim 1, characterized in that: The knowledge-enhanced graph convolutional neural network described in step S2 includes: an input encoding layer, at least one graph convolutional layer, a knowledge enhancement layer, a global feature aggregation layer, and a dual-head output layer; The input encoding layer standardizes and normalizes node and edge features to unify their dimensions. Graph convolutional layers aggregate neighborhood features through a message passing mechanism and pass features through residual connections to avoid gradient vanishing. The knowledge enhancement layer is used to fuse domain knowledge with geometric features; The global feature aggregation layer uses attention pooling to aggregate vertex features into graph-level features; The dual-head output layer is used to output the component function classification results and the component importance score, respectively.

6. The industrial equipment 3D visualization method based on key feature recognition according to claim 5, characterized in that: The knowledge enhancement layer fuses the knowledge vectors encoded by functional semantic rules and topological constraint rules with the output features of the graph convolutional layer, and the fusion method satisfies: Fen h = Fgcn + Fattn + Fk; Wherein, Fgcn is the output feature of the graph convolutional layer, Fattn is the feature obtained based on the multi-head self-attention mechanism, with an optimal number of heads of 8 and a dropout rate of 0.1, Fk is the feature after linear mapping of the knowledge vector, and Fenh is the enhanced feature after fusion.

7. The industrial equipment 3D visualization method based on key feature recognition according to claim 1, characterized in that: In step S2, the region is divided into at least five levels based on its importance score. Different levels of regions correspond to different reduction rates and geometric error thresholds. Preferably, the key functional areas (importance score 0.8 to 1.0) correspond to a reduction rate of ≤20% and a geometric error of <1%; Important identification regions (importance score 0.6–0.8) correspond to a reduction rate of ≤40% and a geometric error of <2%; Functionally important regions (importance score 0.4–0.6) correspond to a reduction rate of ≤60% and a geometric error of <5%; The structural support area (importance score 0.2-0.4) corresponds to a reduction rate of ≤80% and a geometric error of <10%; Non-critical areas (importance score 0-0.2) correspond to a reduction rate of ≥95% and an allowable error of ≤20%.

8. The industrial equipment 3D visualization method based on key feature recognition according to claim 1, characterized in that: In step S3, a semantically aware surface reduction cost function is defined for the candidate simplification operation a: C(a)=E(a)·(1+λ·W·conf)+μ·Ptopo(a); Where E(a) is the geometric error term (calculated from the changes in edge length and curvature), W is the component comprehensive weight obtained in step S1, conf is the semantic confidence or importance score output in step S2, Ptopo(a) is the penalty term for violating topological constraint rules, preferably ≥1000 for strong constraint violations, preferably 500~1000 for medium constraint violations, and preferably 100~500 for weak constraint violations; λ and μ are weight coefficients, preferably λ=1000 and μ=500; Candidate simplification operations are sorted according to the size of C(a), and operations with lower costs are performed first.

9. The industrial equipment 3D visualization method based on key feature recognition according to claim 1, characterized in that: In step S4, different ASTC compression block sizes are selected for different texture regions based on the region classification results and / or component weights: Key texture areas preferably use ASTC 2x2 or 4x4 format, important texture areas preferably use ASTC 6×6 or 8×8 format, general texture areas preferably use ASTC 10×10 format, and minor texture areas preferably use ASTC 12×12 or 16×16 format. Establish a standardized component-shared texture library, extract texture resources from common components, enable texture reuse for components of the same type, and reduce data redundancy.

10. The industrial equipment 3D visualization method based on key feature recognition according to claim 1, characterized in that: In step S5: Construct at least a three-level LOD model, with different polygon retention ratios corresponding to different model levels; The multi-level LOD model is switched according to the line-of-sight parameters of the camera and the device or component, and a smooth transition factor is calculated based on the line-of-sight parameters. The smooth transition factor is preferably: transition_factor = max(0.1, 1.0 - normalized_distance·0.3), where normalized_distance is the distance ratio after the line-of-sight parameters are normalized. The LOD switching logic is adjusted based on the overall weight of the components, with the preferred method being to delay the switching by one level when W≥0.8 and advance the switching by one level when W≤0.3; Linear interpolation fusion is performed on the current LOD model and the target LOD model based on the transition factor. The preferred transition time is 0.5 seconds to ensure that the frame rate of the mobile device is stable at more than 30fps, so as to achieve continuous switching between adjacent LOD models.

Images