A multi-source heterogeneous STP model lightweight method and system based on photon collision and recombination characteristics

Abstract

The present application relates to the technical field of engineering data processing and deep learning application, and particularly relates to a multi-source heterogeneous STP model lightweight method and system based on photon collision and composite feature construction. The method acquires multi-source heterogeneous STP data in batches through interface adaptation and classifies and preprocesses; multi-directional photon beams are used to simulate space collision, accurately capture external contours and physically strip hidden redundant geometric features; geometric and engineering semantic multi-dimensional features are extracted to construct a composite vector, key weights are inferred by a hybrid deep network, non-key components in the assembly tree are intelligently pruned, and the topology is adaptively updated; finally, redundant grids are centralized and simplified, and the bottom layer metadata is inherited, and a lightweight standard STP model is generated by hybrid compression encoding. The present application realizes the physical limit reduction of the model, while high-fidelity retains the core engineering semantics and topology, greatly improving the computing efficiency of industrial large models in cross-platform collaboration and visualization.

Description

Technical Field

[0001] This invention relates to the fields of engineering data processing and deep learning application technology, and in particular to a lightweight method and system for constructing a multi-source heterogeneous STP model based on photon collision and composite features. Background Technology

[0002] With the advancement of digitalization in large-scale engineering projects such as nuclear power plants, large chemical plants, and super high-rise buildings, and the in-depth application of Building Information Modeling (BIM) technology, multi-source heterogeneous 3D models have become a common feature of engineering lifecycle management. In the process of collaborative design by multiple teams, different disciplines often use different 3D industrial design software (such as PDMS, E3D, Revit, etc.), resulting in inconsistent model formats and huge differences in data structures.

[0003] To achieve collaborative review and browsing of the entire plant area or building, these heterogeneous models must be uniformly converted into the international standard STP (ISO10303-21) format and lightweighted. However, existing technologies have the following serious technical shortcomings when handling such large heterogeneous models: First, traditional geometric simplification algorithms, such as edge folding or blind face reduction, reduce the number of faces based on the overall grid geometry and do not further determine spatial visibility. This results in the complete preservation of a large number of completely invisible hidden geometric elements located inside equipment, pipes, or other structures, such as valve cavities and internal liners, thus occupying a huge amount of data. Meanwhile, key external contour features are severely distorted in the blind reduction of surfaces, failing to meet the accuracy requirements of actual engineering standards such as ASME Y14.5. Secondly, existing lightweight engineering heterogeneous models lack effective identification of the importance of engineering semantics and assembly features. When pruning complex component assembly trees, manual deletion or crude filtering based on a single volume size is still used, which easily leads to the accidental deletion of critical functional components such as small but extremely important instrument panels and connectors, or the retention of a large number of redundant temporary supports, compromising the engineering usability of the model. Finally, existing technologies often suffer from serious metadata loss during cross-software conversion. In particular, after heterogeneous models are converted to lightweight standard formats, problems such as the breakage or loss of color, material, and original assembly tree topology often occur, making it difficult to meet the high-quality technical requirements for lossless sharing of multi-source heterogeneous models.

[0004] Therefore, how to construct a lightweight solution that can physically eliminate internal invisible redundancy, intelligently identify engineering semantic keyness, and achieve high-fidelity output across software is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To address the technical problems that existing lightweight methods for engineering heterogeneous models often fail to balance geometric simplification efficiency and accuracy, lack effective identification of the importance of engineering semantics and assembly features, and are prone to data loss during cross-model conversion, resulting in poor engineering usability of the models and difficulty in meeting the high-quality technical requirements for lossless sharing of multi-source heterogeneous models, this invention provides a lightweight method and system for multi-source heterogeneous STP models based on photon collision and composite features.

[0006] One of the lightweight methods for constructing a multi-source heterogeneous STP model based on photon collision and recombination features includes the following steps: S1. Obtain the native STP output interface information of each 3D industrial design software, call the native STP output interface of each 3D industrial design software in batches through the interface adapter plugin, export the first STP dataset in STEP file format and store it in the distributed cache pool. S2. Analyze the first STP dataset, determine whether each 3D model in the first STP dataset is a solid model or a surface model, and according to the determination result, perform geometric feature enhancement extraction processing on the solid model and surface smoothness regularization processing on the surface model to obtain the second STP dataset. S3. By simulating the collision process between multi-directional photon beams and the surfaces of various three-dimensional models in the second STP dataset, a photon collision detection matrix is ​​constructed to obtain the external contour data of the three-dimensional model. Based on the preset collision intensity threshold, the external contour data of the three-dimensional model is marked as visible geometric element data and internal hidden redundant geometric element data, respectively. S4. Extract the geometric features, semantic features, and assembly relationship features of each 3D model in the second STP dataset and construct a composite feature vector. Input the vector into a pre-trained lightweight network model for inference to obtain key feature weights. Based on the key feature weights, delete non-key level 3D models and update the topological hierarchy of the assembly tree simultaneously to obtain the third STP dataset. S5. Based on the third STP dataset, perform centralized geometric simplification on the STP data marked with internally hidden redundant geometric elements. By inheriting the STP metadata through the attribute definition fields in the STEP file format, and generating a lightweight STP model through boundary representation optimization algorithm and hybrid compression coding.

[0007] Furthermore, step S1 includes the following sub-steps: S101. Based on the underlying STP output interface characteristic parameters of different 3D industrial design software, configure dedicated interface plug-ins for each 3D industrial design software, and establish a connection with each 3D industrial design software through a process communication mechanism, so as to control each 3D industrial design software to synchronously execute native STP export by sending a unified format call command. S102. Establish a distributed data cache pool to receive the first STP data exported by various 3D industrial design software in parallel. S103. Verify the file header format, EXPRESS language syntax compliance, and geometric data integrity of the first STP data respectively. If any verification fails, repeat steps S101-S102 until the verification passes and the first STP data is obtained.

[0008] Furthermore, step S2 includes the following sub-steps: S201. Based on the first STP dataset, the first STP dataset is parsed using the OCC system to extract geometric topology information, component functional classification semantic labels, and metadata; the metadata includes model color data, model material data, and model assembly relationship data; S202. Based on the parsing and extraction results, determine the model type. If the parsing and extraction results contain a solid component field, it is determined to be a solid model; if the parsing and extraction results contain a surface geometry field, it is determined to be a surface model. S203. Based on the judgment result, if it is judged to be a solid model, then enhance the extraction of hole and chamfer geometric features and mark the feature priority; if it is judged to be a curved surface model, then use the curved surface smoothing algorithm to remove burrs and repair the damaged curved surface. S204. Based on the model results obtained in step S203, construct the second STP dataset.

[0009] Furthermore, step S3 includes the following sub-steps: S301. Based on the second STP dataset, construct a three-dimensional spatial mesh coordinate system with the geometric center points of each three-dimensional model, and emit photon beams to the surface of each three-dimensional model within the second STP dataset according to the six-orthogonal, four-tilted multi-directional photon emission mechanism. The formula for calculating the photon beam density is expressed as: ; in, Indicates the photon beam density. This represents the preset photon beam reference density. This represents the preset reference volume of the 3D model. This represents the volume of the bounding box of a 3D model. S302. Based on the three-dimensional spatial grid coordinate system, assign a unique ID to each photon within the photon beam, and perform collision point error compensation for the oblique photon beam. Record the emission coordinate data, direction vector data, collision state data, collision point coordinate data, collision surface normal vector data, and collision intensity data corresponding to each photon, and construct a collision detection matrix. The error compensation formula is expressed as: ; in, This indicates the collision point error compensation. This represents the material correction factor for the 3D model. Indicates the angle of incidence of the photon; S303. Filter valid collision points in the collision detection matrix with a collision intensity greater than or equal to a set threshold, mark them as visible geometric element data, use the Delaunay triangulation algorithm to connect all valid collision points, obtain the continuous outer contour of each reconstructed 3D model, and mark the spatial regions where no collision occurred or where the collision intensity is less than the set threshold as internal hidden redundant geometric element data.

[0010] Furthermore, step S4 includes the following sub-steps: S401. Based on the second STP dataset, extract multi-dimensional features for each 3D model. These multi-dimensional features include geometric features, semantic features, and assembly relationship features. The geometric features include the number of triangular faces, vertex density, rate of curvature change, and volume percentage. The semantic features include component functional classification and engineering importance level based on STEP AP214 semantic tags. The assembly relationship features include the parent-child component association degree and motion constraint relationship. After normalizing the multi-dimensional features, construct a high-dimensional composite feature vector. S402. Construct a lightweight network model based on CNN network and Transformer structure, use historical engineering model samples as training set, construct a joint optimization objective based on lightweight rate loss function and key feature retention rate loss function, and pre-train the lightweight network model; S403. Based on the trained lightweight network model, the composite feature vector is input into the lightweight network model, the CNN structure in the lightweight network model is used to extract the local correlation of geometric features, the Transformer structure is used to extract the global dependency of semantic features and assembly relationship features, and the key feature weight values ​​of a single component are output. S404. Compare the key feature weight values ​​with preset weight thresholds. Determine the 3D models with weight values ​​greater than or equal to the preset weight thresholds as key level models. Determine the 3D models with weight values ​​less than the preset weight thresholds as non-key level models and delete them. After deletion, automatically traverse the assembly information of each 3D model in the second STP dataset and update the association relationship and geometric topology connection information of parent and child components to obtain the third STP dataset.

[0011] Furthermore, step S5 includes the following sub-steps: S501. Based on the third STP dataset, delete the STP data marked with internal hidden redundant geometric elements, perform an edge folding algorithm based on quadratic error metric on non-critical features, and perform order reduction optimization on NURBS surfaces to reduce the number of control points. S502. Inherit the color, material, and function label metadata from the first STP data by defining the attribute fields in the STEP file format; S503. The boundary representation method is used to simplify the geometric description statement, and LZ77 encoding is used to compress continuous geometric coordinate data. Huffman encoding is used to compress high-frequency semantic fields to generate the final lightweight STP model.

[0012] Furthermore, the method also includes constructing a smart filter based on regular expressions and a random forest classification model to identify and remove duplicate components and temporary labeled data in the first STP data.

[0013] Furthermore, when there are multiple STP model lightweighting tasks, the priority of each STP model lightweighting task is configured through the batch processing task manager based on the actual working conditions. The parallel lightweighting computing resources are allocated to multiple STP model lightweighting tasks according to the task priority, and the task progress is recorded in real time through the local database.

[0014] This invention also provides a lightweight system for multi-source heterogeneous STP models constructed based on photon collision and recombination features. This system is implemented based on any of the aforementioned lightweight methods for multi-source heterogeneous STP models constructed based on photon collision and recombination features, and includes: The interface adaptation and export module is used to obtain the native STP output interface information of each 3D industrial design software, and to call the native STP output interface of each 3D industrial design software in batches through the interface adaptation plugin, export the first STP dataset in STEP file format and store it in a distributed cache pool. The model preprocessing module is used to parse the first STP dataset, determine whether each 3D model in the initial STP dataset is a solid model or a surface model, and perform geometric feature enhancement extraction processing on the solid model and surface smoothness regularization processing on the surface model to obtain the second STP dataset. The external contour capture module is used to construct a photon collision detection matrix by simulating the collision process between multi-directional photon beams and the surfaces of various three-dimensional models in the second STP dataset, obtain the external contour data of the three-dimensional model, and mark the external contour data of the three-dimensional model as visible geometric element data and internal hidden redundant geometric element data based on a preset collision intensity threshold. The feature network optimization module is used to extract the geometric features, semantic features, and assembly relationship features of each 3D model in the second STP dataset and construct a composite feature vector. This vector is then input into a pre-trained lightweight network model for inference to obtain key feature weights. Based on these key feature weights, non-key level 3D models are deleted, and the topological hierarchy of the assembly tree is updated synchronously to obtain the third STP dataset. The optimization output and scheduling module is used to perform centralized geometric simplification on STP data marked with redundant geometric elements based on the third STP dataset. It inherits STP metadata through STP attribute definition fields and generates a lightweight STP model through boundary representation optimization algorithm and hybrid compression coding.

[0015] Furthermore, it also includes an intelligent filtering and batch processing module, comprising an intelligent filtering unit and a batch processing resource allocation unit. The intelligent filtering unit is used to construct an intelligent filter based on regular expressions and a random forest classification model to identify and remove duplicate components and temporary labeled data in the first STP data. The batch processing resource allocation unit is used to configure the priority of multiple STP model lightweight tasks according to actual working conditions through a batch processing task manager, allocate parallel lightweight computing resources for multiple STP model lightweight tasks according to task priority, and record task progress in real time through a local database.

[0016] A computer-readable storage medium for storing a computer program that, when run on a computer, causes the computer to execute any of the above-mentioned lightweight method for a multi-source heterogeneous STP model constructed based on photon collision and recombination features.

[0017] An electronic device includes: a memory for storing a computer program; and a processor for executing the computer program to implement a lightweight method for a multi-source heterogeneous STP model based on photon collision and recombination features as described above.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention, based on photon collision theory, employs a ten-directional photon beam scanning model ("six orthogonal + four tilted") combined with a dynamic photon density adjustment formula and material oblique deviation compensation to accurately capture the external contour data of a 3D model. It also removes invisible geometric redundancy hidden within the model, eliminating invalid data at its physical source and significantly reducing the overall size while ensuring external assembly and project progress. Furthermore, this invention utilizes a hybrid CNN and Transformer neural network to extract geometric features, STEP AP214 semantic label features, and assembly relationship features including motion constraints from multiple dimensions. It precisely quantifies the function and hierarchical weight of each component, accurately deletes non-critical components, and automatically reconstructs assembly relationships, eliminating the technical risks of losing core engineering features or hierarchical breaks. On the other hand, the solution uses a unified standardized interface adapter, uniformly outputting data compliant with ISO standards. The STP model based on the 10303-21 standard achieves high-fidelity inheritance of metadata such as materials, colors, and assembly torque through attribute definition fields in the STEP file format. Combined with boundary representation and LZ77 Huffman hybrid coding, it effectively improves the compatibility of lightweight models and overcomes the shortcomings of poor compatibility in cross-platform multi-team collaboration in existing technologies. Attached Figure Description

[0019] Figure 1 This is a flowchart of a lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features, as proposed in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of a lightweight system for a multi-source heterogeneous STP model based on photon collision and recombination features proposed in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the terminal device structure of a lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features proposed in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of a computer-readable storage medium structure for a lightweight method of multi-source heterogeneous STP model based on photon collision and recombination features proposed in an embodiment of the present invention.

[0023] In the diagram, 200 is the terminal device, 210 is the memory, 211 is the RAM, 212 is the cache memory, 213 is the ROM, 214 is the program / utility, 215 is the program module, 220 is the processor, 230 is the bus, 240 is the external device, 250 is the I / O interface, 260 is the network adapter, and 300 is the program product. Detailed Implementation

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, and not all of them. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention. It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0027] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or machine that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or machine. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or machine that includes said element.

[0028] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0029] Example 1: like Figure 1 As shown, this embodiment provides a lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features, including the following steps: S1. Obtain the native STP output interface information of each 3D industrial design software, call the native STP output interface of each 3D industrial design software in batches through the interface adapter plugin, export the first STP dataset in STEP file format and store it in the distributed cache pool. Furthermore, step S1 includes the following sub-steps: S101. Based on the underlying STP output interface characteristic parameters of different 3D industrial design software, configure dedicated interface plug-ins for each 3D industrial design software, and establish a connection with each 3D industrial design software through a process communication mechanism, so as to control each 3D industrial design software to synchronously execute native STP export by sending a unified format call command. S102. Establish a distributed data cache pool to receive the first STP data exported by various 3D industrial design software in parallel. S103. Verify the file header format, EXPRESS language syntax compliance, and geometric data integrity of the first STP data respectively. If any verification fails, repeat steps S101-S102 until the verification passes and the first STP data is obtained.

[0030] Specifically, this step addresses the technical pain points of closed formats and difficulties in cross-disciplinary collaboration for multi-source heterogeneous models. It deploys an interface adaptation matrix composed of multiple low-level APIs in the target design environment, each taking over the low-level output of heterogeneous software such as PDMS and E3D. Existing technologies often rely on third-party transfer tools or downgraded intermediate formats (such as OBJ and STL) for multi-format model conversion. Various geometric attributes become isolated during secondary or even tertiary translations, leading to broken topology trees and significant data loss. This method, through direct cross-software calls from the entire source, fuses and unifies heterogeneous data within different industrial software, constructing a standardized, lossless, and natively coherent first-level STP dataset for subsequent geometric beam and physical calculations.

[0031] Specifically, the implementation principle described above is as follows: First, through process communication scripts deployed on the servers of various 3D industrial design software, unified format commands containing export paths and accuracy parameters are sent in parallel to different 3D industrial design software around the clock. This forces the invocation of their native STEP file driver kernel, uniformly integrating and converting the acquired internal private primitive features into a common language under the ISO10303-21 standard protocol. Simultaneously, a distributed cache pool is used to handle massive, high-concurrency, high-density mesh data streams. Furthermore, it should be noted that the topological relationships of large-scale engineering design models are extremely complex, making them highly susceptible to memory overflow during export. By introducing the EXPRESS syntax and a boundary loop integrity verification mechanism, not only are incomplete geometries disguised as STP objects effectively intercepted, but scattered fields reflecting assembly constraints and material definitions are also forcibly solidified in a unified data pool.

[0032] S2. Analyze the first STP dataset, determine whether each 3D model in the first STP dataset is a solid model or a surface model, and according to the determination result, perform geometric feature enhancement extraction processing on the solid model and surface smoothness regularization processing on the surface model to obtain the second STP dataset. Furthermore, step S2 includes the following sub-steps: S201. Based on the first STP dataset, the first STP dataset is parsed using the OCC system to extract geometric topology information, component functional classification semantic labels, and metadata; the metadata includes model color data, model material data, and model assembly relationship data; S202. Based on the parsing and extraction results, determine the model type. If the parsing and extraction results contain a solid component field, it is determined to be a solid model; if the parsing and extraction results contain a surface geometry field, it is determined to be a surface model. S203. Based on the judgment result, if it is judged to be a solid model, then enhance the extraction of hole and chamfer geometric features and mark the feature priority; if it is judged to be a curved surface model, then use the curved surface smoothing algorithm to remove burrs and repair the damaged curved surface. S204. Based on the model results obtained in step S203, construct the second STP dataset.

[0033] Specifically, this step addresses the problem of inconsistent modeling accuracy and extremely chaotic topology in heterogeneous models. During the data preprocessing stage, the STEP interface based on the OpenCASCADE 7.5.0 (OCC 7.5.0) platform is introduced to parse the first STP data, capturing and organizing information on holes, chamfers, and broken surfaces in the model. Existing techniques, when processing massive models, often rely on a one-size-fits-all approach of indiscriminate mesh discretization, failing to distinguish between subtle features like threads on solids and burr boundaries on surfaces. This leads to severe geometric noise or chaotic topology during upper-level mesh reconstruction. This method, through a decoupled "solid-surface" dual-regulation channel, eliminates geometrically dirty data caused by differences in modeling techniques among professionals, constructing a smooth, clean second STP dataset with extremely high geometric continuity for subsequent high-frequency photon ray capture.

[0034] Specifically, the implementation principle described above is as follows: First, the first STP data stream is parsed using the STEP interface of OpenCASCADE 7.5.0, extracting geometric topology information (vertices, faces, assembly structures), semantic tags (component functional classification), and metadata. Based on whether the extraction results declare low-level language tags such as SOLID_VOLUME or NURBS_SURFACE, if the data contains key fields such as SOLID and BOOLEAN_OPERATION, it is determined to be a solid model; if SURFACE and CURVE are the main geometric descriptions, it is determined to be a surface model, and the solid model and surface model are decoupled in multiple dimensions. For solid models, such as gears and bearings in mechanical assemblies, key geometric features such as threads and keyways are extracted, and feature priorities are marked. For surface models, such as the exterior walls and roof surfaces of building BIM models, a surface smoothing algorithm is used to remove burrs and repair damaged surfaces. Furthermore, it should be noted that the physical photon collision algorithm has extremely stringent requirements for the orientation of the spatial mesh's normals and the continuity of its boundaries. Without pre-normalization of the original rough surface, when photon rays hit mesh "burrs" or penetrate "cracks," it is easy to induce diffuse reflection disorder and overflow of the collision matrix. Through this pre-geometric cleaning step, the discrete and poor model patches are reshaped into highly consistent and topologically complete 3D manifold boundaries, providing a zero-interference physical constraint environment for the spatial collision of photon beams.

[0035] S3. By simulating the collision process between multi-directional photon beams and the surfaces of various three-dimensional models in the second STP dataset, a photon collision detection matrix is ​​constructed to obtain the external contour data of the three-dimensional model. Based on the preset collision intensity threshold, the external contour data of the three-dimensional model is marked as visible geometric element data and internal hidden redundant geometric element data, respectively. Furthermore, step S3 includes the following sub-steps: S301. Based on the second STP dataset, construct a three-dimensional spatial mesh coordinate system with the geometric center points of each three-dimensional model, and emit photon beams to the surface of each three-dimensional model within the second STP dataset according to the six-orthogonal, four-tilted multi-directional photon emission mechanism. The formula for calculating the photon beam density is expressed as: ; in, Indicates the photon beam density. This represents the preset photon beam reference density. This represents the preset reference volume of the 3D model. This represents the volume of the bounding box of a 3D model. S302. Based on the three-dimensional spatial grid coordinate system, assign a unique ID to each photon within the photon beam, and perform collision point error compensation for the oblique photon beam. Record the emission coordinate data, direction vector data, collision state data, collision point coordinate data, collision surface normal vector data, and collision intensity data corresponding to each photon, and construct a collision detection matrix. The error compensation formula is expressed as: ; in, This indicates the collision point error compensation. This represents the material correction factor for the 3D model. Indicates the angle of incidence of the photon; S303. Filter valid collision points in the collision detection matrix with a collision intensity greater than or equal to a set threshold, mark them as visible geometric element data, use the Delaunay triangulation algorithm to connect all valid collision points, obtain the continuous outer contour of each reconstructed 3D model, and mark the spatial regions where no collision occurred or where the collision intensity is less than the set threshold as internal hidden redundant geometric element data.

[0036] Specifically, this step addresses the problem of massive internal geometry in complex, large models, which consumes a significant amount of rendering computation. It deploys dynamic density photon beams in a virtual 3D coordinate space to capture and record the physical collision intensity and normal response data of the model's outer surface. Existing techniques for geometric reduction often rely on vertex-based blind sampling simplification algorithms like QEM, resulting in simultaneous random collapse of the outer shell and internal core components. This leads to severe distortion of the external contour and an inability to remove deeply hidden debris. This method constructs a six-orthogonal, four-tilted multi-dimensional photon beam scan, creating highly accurate external boundary contour features for subsequent internal mesh removal.

[0037] Specifically, the implementation principle described above is as follows: First, an adaptively scaling spatial grid coordinate system is established with the geometric centroid of the 3D model as the origin. Virtual photon beams are emitted to each geometric plane of the model according to a dynamic density formula based on the model's volume gradient. Simultaneously, the spatial coordinates, incident deviation angle, and kinetic energy decay after the first physical collision of each photon are recorded in real time. This process maps and fuses massive amounts of 3D spatial collision states to construct a high-dimensional sparse collision matrix. Furthermore, it should be noted that large-scale equipment such as enclosed pumps and heat exchangers contain numerous complex pipes and impeller grids. These geometric structures have no engineering value for the final collaborative visual browsing. Utilizing the rectilinear propagation and occlusion properties of light, any grid that can reflect photon beams and generate energy exceeding a preset threshold is characterized by its external visible shell. All deep grids not covered by light spots or subjected to secondary refraction are precisely marked as redundant data. This eliminates polygons that are of no value to the visual display, completely removing the ineffective occupation of system memory by redundant grids.

[0038] S4. Extract the geometric features, semantic features, and assembly relationship features of each 3D model in the second STP dataset and construct a composite feature vector. Input the vector into a pre-trained lightweight network model for inference to obtain key feature weights. Based on the key feature weights, delete non-key level 3D models and update the topological hierarchy of the assembly tree simultaneously to obtain the third STP dataset. Furthermore, step S4 includes the following sub-steps: S401. Based on the second STP dataset, extract multi-dimensional features for each 3D model. These multi-dimensional features include geometric features, semantic features, and assembly relationship features. The geometric features include the number of triangular faces, vertex density, rate of curvature change, and volume percentage. The semantic features include component functional classification and engineering importance level based on STEP AP214 semantic tags. The assembly relationship features include the parent-child component association degree and motion constraint relationship. After normalizing the multi-dimensional features, construct a high-dimensional composite feature vector. S402. Construct a lightweight network model based on CNN network and Transformer structure, use historical engineering model samples as training set, construct a joint optimization objective based on lightweight rate loss function and key feature retention rate loss function, and pre-train the lightweight network model; S403. Based on the trained lightweight network model, the composite feature vector is input into the lightweight network model, the CNN structure in the lightweight network model is used to extract the local correlation of geometric features, the Transformer structure is used to extract the global dependency of semantic features and assembly relationship features, and the key feature weight values ​​of a single component are output. S404. Compare the key feature weight values ​​with preset weight thresholds. Determine the 3D models with weight values ​​greater than or equal to the preset weight thresholds as key level models. Determine the 3D models with weight values ​​less than the preset weight thresholds as non-key level models and delete them. After deletion, automatically traverse the assembly information of each 3D model in the second STP dataset and update the association relationship and geometric topology connection information of parent and child components to obtain the third STP dataset.

[0039] Specifically, this step addresses the issue of core components being easily mistakenly removed during lightweighting of 3D engineering models. A deep cognitive map, fused with a CNN network and a Transformer structure, is constructed within the data feature space to evaluate the volume percentage, engineering importance level, and assembly motion relationships of a vast number of components in the system in parallel. Existing lightweighting removal strategies often rely on preset absolute values ​​of bounding box dimensions, leading to the mistaken removal of small but crucial kinematic components such as key transmission gears and fasteners, rendering the model worthless for engineering logic analysis. This method, through multi-dimensional, multi-modal AI semantic mapping, overcomes the blindness of relying solely on visual size to determine retention in 3D lightweighting, providing a robust intelligent decision-making mechanism for the final STP assembly tree simplification.

[0040] Specifically, the implementation principle described above is as follows: First, the geometric features, semantic features, and assembly relationship features from the second STP dataset are multi-dimensionally fused and packaged, mapped into a 128-dimensional high-density composite feature vector. Deployed on the PyTorch 1.12 framework, a lightweight network model is constructed using a CNN+Transformer hybrid structure. Using historical engineering model samples as the training set, a joint optimization objective based on a lightweight rate loss function and a key feature retention rate loss function is built to pre-train the lightweight network model. After inputting the high-density composite feature vector into the pre-defined hybrid model, convolutional layers are used to efficiently extract the topological compactness of local micro-grids, while the Transformer's self-attention mechanism is used to extract the strong correlation and coupling effects between parts in different branches of the assembly tree. Furthermore, it should be noted that in extremely complex electromechanical assembly environments, isolating a single part to determine its survival carries a very high risk of damage; for example, a tiny limit pin determines the movement boundary of the entire transmission arm. By calculating the dynamic key feature weights of the output through a hybrid neural network, the system can accurately eliminate low-weight auxiliary parts such as transport fixing brackets and temporary solder pads. After non-critical layers are physically stripped away, the system's underlying adaptive reconstruction traversal is immediately initiated to automatically repair the parent-child assembly tree connection boundaries, fundamentally ensuring that the simplified lightweight model still possesses complete and rigorous engineering kinematic properties and review logic chains.

[0041] S5. Based on the third STP dataset, perform centralized geometric simplification on the STP data marked with internally hidden redundant geometric elements. By inheriting the STP metadata through the attribute definition fields in the STEP file format, and generating a lightweight STP model through boundary representation optimization algorithm and hybrid compression coding.

[0042] Furthermore, step S5 includes the following sub-steps: S501. Based on the third STP dataset, delete the STP data marked with internal hidden redundant geometric elements, perform an edge folding algorithm based on quadratic error metric on non-critical features, and perform order reduction optimization on NURBS surfaces to reduce the number of control points. S502. Inherit the color, material, and function label metadata from the first STP data by defining the attribute fields in the STEP file format; S503. The boundary representation method is used to simplify the geometric description statement, and LZ77 encoding is used to compress continuous geometric coordinate data. Huffman encoding is used to compress high-frequency semantic fields to generate the final lightweight STP model.

[0043] Specifically, this step addresses the issues of bloated data reorganization and loss of engineering attributes in the later stages of lightweight transformation. In the final stage of model data generation, a fusion output pipeline consisting of redundancy removal, semantic inheritance, and hybrid encoding is deployed to reorganize and optimize the spatial geometric matrix and non-entity attached attributes. Existing technologies often focus on simple geometric compression output during final file generation, permanently stripping away model colors, origin assembly attributes, and motion parameters, making them unrecognizable by downstream process reuse and simulation systems. This method overcomes the technical deficiency of lightweighting being equivalent to merely retaining a geometric shell through multi-level lossless data mapping and encoding, providing a high-fidelity, compact, and strictly internationally compliant lightweight engineering model for cross-enterprise collaborative environments.

[0044] Specifically, the implementation principle described above is as follows: First, based on the third STP dataset, STP data marked with internally hidden redundant geometric elements is removed, and secondary error reduction and smoothing calculations are performed on the remaining external non-critical surfaces. Then, the metadata is re-injected intact into each simplified sub-component node using the PROPERTY_DEFINITION attribute definition field in the STEP file format. Furthermore, it should be noted that simply reducing the number of faces cannot compress the file storage volume to its physical limit. Since the underlying STP file is a text stream recording point clouds and topological declarations in plain ASCII, it inevitably contains a large number of repeated coordinates and long characters in standard declarations. This step feeds the spatial digital array reflecting the vertex drift of the continuous surface into LZ77 encoding to compress the continuous geometric data. Simultaneously, commonly used descriptive long texts in engineering are fed into frequency-based Huffman coding for encoding reconstruction. Before output, the STP model is verified using the ISO10303-21 standard verification tool to ensure that it can be natively opened by various industrial software. This secondary deep compression, from the geometric physical layer to the text plane layer, transforms the originally scattered massive features into high-density data encoding. This not only effectively compresses the storage size of the final lightweight model, but also completely solves the problem of three-dimensional collaboration between the web and mobile terminals caused by bandwidth limitations.

[0045] Furthermore, it also includes building intelligent filters based on regular expressions and random forest classification models to identify and remove duplicate components and temporary labeled data in the first STP data; when there are multiple STP model lightweighting tasks, the priority of each STP model lightweighting task is configured through the batch processing task manager according to the actual working conditions, and the parallel lightweighting computing resources are allocated to multiple STP model lightweighting tasks according to the task priority, and the task progress is recorded in real time through the local database.

[0046] Specifically, this step addresses the robustness and efficiency requirements of processing massive numbers of models in large-scale engineering scenarios. Existing processing methods are mostly single-threaded serial processing, which is prone to crashing due to memory overflow or individual model anomalies when dealing with tens of thousands of model components. This method purifies the data source through an intelligent filtering mechanism and combines it with industrial-grade scheduling strategies to build a highly available and monitorable automated processing pipeline.

[0047] Specifically, the implementation principle described above is as follows: First, a pre-scanned random forest model is used to pre-scan the first STP dataset, automatically identifying and removing construction noise data such as temporary baselines and placeholder virtual components. Then, the batch processing scheduler dynamically distributes computational tasks to the CPU / GPU cluster based on the geometric complexity of each model and the user-defined priority. Furthermore, by introducing a local database-based status monitoring mechanism throughout the entire processing flow, the system can record the completion percentage of each subtask in real time. When encountering storage space exhaustion or network interruptions, it supports resuming interrupted downloads without recalculating completed lightweight stages. This engineered scheduling design significantly improves the system's throughput and robustness, ensuring efficient processing of large-scale heterogeneous model lightweighting tasks. Example 2 like Figure 2 As shown, as a preferred embodiment of the above embodiments, a lightweight system for multi-source heterogeneous STP models based on photon collision and recombination features is provided. This system is implemented based on any of the lightweight methods for multi-source heterogeneous STP models based on photon collision and recombination features described above, and includes: The interface adaptation and export module is used to obtain the native STP output interface information of each 3D industrial design software, and to call the native STP output interface of each 3D industrial design software in batches through the interface adaptation plugin, export the first STP dataset in STEP file format and store it in a distributed cache pool. The model preprocessing module is used to parse the first STP dataset, determine whether each 3D model in the initial STP dataset is a solid model or a surface model, and perform geometric feature enhancement extraction processing on the solid model and surface smoothness regularization processing on the surface model to obtain the second STP dataset. The external contour capture module is used to construct a photon collision detection matrix by simulating the collision process between multi-directional photon beams and the surfaces of various three-dimensional models in the second STP dataset, obtain the external contour data of the three-dimensional model, and mark the external contour data of the three-dimensional model as visible geometric element data and internal hidden redundant geometric element data based on a preset collision intensity threshold. The feature network optimization module is used to extract the geometric features, semantic features, and assembly relationship features of each 3D model in the second STP dataset and construct a composite feature vector. This vector is then input into a pre-trained lightweight network model for inference to obtain key feature weights. Based on these key feature weights, non-key level 3D models are deleted, and the topological hierarchy of the assembly tree is updated synchronously to obtain the third STP dataset. The optimization output and scheduling module is used to perform centralized geometric simplification on STP data marked with redundant geometric elements based on the third STP dataset. It inherits STP metadata through STP attribute definition fields and generates a lightweight STP model through boundary representation optimization algorithm and hybrid compression coding.

[0048] Furthermore, it also includes an intelligent filtering and batch processing module, comprising an intelligent filtering unit and a batch processing resource allocation unit. The intelligent filtering unit is used to construct an intelligent filter based on regular expressions and a random forest classification model to identify and remove duplicate components and temporary labeled data in the first STP data. The batch processing resource allocation unit is used to configure the priority of multiple STP model lightweight tasks according to actual working conditions through a batch processing task manager, allocate parallel lightweight computing resources for multiple STP model lightweight tasks according to task priority, and record task progress in real time through a local database.

[0049] Specifically, the implementation principle of the above embodiments is as follows: First, the interface adaptation and export module deployed at the front end, together with the intelligent filtering and batch processing module, obtains the underlying raw data (including primitive features in private format, assembly relationships, and engineering metadata) of multi-source heterogeneous 3D industrial design software in real time through process communication mechanism and underlying API calls. To ensure the standardization and effectiveness of massive concurrent data in the initial stage of conversion, the intelligent filtering unit uses regular expressions and random forest classification models to identify and remove duplicate components and temporary tags in the input data stream in real time. At the same time, the batch processing resource allocation unit, based on actual working conditions, issues concurrent call instructions with priority identifiers to each interface adaptation plugin, controlling each software to securely and uniformly convert the cleaned private data into the first STP dataset conforming to the ISO 10303-21 standard and centrally store it in a distributed cache pool. Then, after receiving the first STP dataset, the model preprocessing module performs multi-dimensional decoupling parsing of the underlying topology and semantic labels on the standard data stream based on the built-in graph kernel (OCC). Based on the entity or surface fields extracted from the parsed text, the model type is determined. For the holes and chamfer features of the entity model and the burrs and broken boundaries of the surface model, feature priority marking and targeted stitching smoothing and regularization are performed in parallel, thereby constructing a second STP dataset with eliminated initial geometric topological noise and a regular structure. Next, the external contour capture module extracts the normalized second STP dataset. In the three-dimensional mesh coordinate system adaptively constructed with the geometric center of the model, dynamic density photon beams are projected onto the model surface in multiple directions using a preset "six orthogonal and four tilted" mechanism. The geometric plane coordinates and energy attenuation data during the reflection and interference of the beams are analyzed to construct a photon collision detection matrix. Valid collision points are screened in three-dimensional space using a preset collision intensity threshold, and the continuous boundary contour of the physical shell of the model is reconstructed. At the same time, the deep internal space regions that are occluded by the shell are forcibly marked as hidden redundant geometric element data to be deleted. Furthermore, the feature network optimization module comprehensively extracts the geometric compactness, STEPAP214 functional classification labels, and hierarchical motion constraint data of each model in the second STP dataset, normalizes them to construct a high-dimensional composite feature vector, and inputs it into a lightweight network model composed of CNN and Transformer to perform deep inference of local geometry and global assembly dependencies. It accurately quantifies and outputs the key feature weights of each component, and performs physical stripping of non-critical hierarchical auxiliary components based on the weights. It also automatically traverses and updates the broken assembly tree connection relationship, and outputs the third STP dataset with lossless core engineering logic. Finally, the optimization output and scheduling module receives the third STP dataset after feature pruning, uses centralized computing power to destroy the internal hidden redundant meshes previously marked by the external contour capture module, and performs surface reduction and edge folding calculations based on quadratic error metric. At the same time, it forces the inheritance of low-level metadata such as color and material in the initial dataset through the attribute definition field hook built into the STEP protocol. Finally, it uses the boundary representation optimization algorithm and LZ77 and Huffman hybrid reconstruction encoder to package and compile the processed topological matrix and semantic high-frequency text to generate a standardized and lightweight STP final model with extremely high compression ratio. Furthermore, throughout the entire process of multi-source heterogeneous model transformation and pruning, the resource allocation and breakpoint recording control logic of the intelligent filtering and batch processing module runs through each hardware processing node. It uses a local database to record the progress of all lightweight phase tasks of all models in real time, and dynamically schedules multiple cross-disciplinary model processing processes in parallel based on the current remaining multi-core CPU and GPU memory of the system. This ensures high concurrency disaster recovery capability and optimal resource coordination of underlying computing power throughout the entire process from cleaning massive multi-source heterogeneous dirty data to photon collision adjudication and AI intelligent feature extraction in terms of physical computing architecture.

[0050] Example 3

[0051] Building upon Examples 1 and 2, there exists a practical application scenario for lightweighting multi-source heterogeneous engineering models. This scenario employs a lightweighting method and system for multi-source heterogeneous STP models based on photon collision and recombination features, as described in the aforementioned examples. Specifically: S1. Heterogeneous Model Interface Adaptation and Initial STP Export: Develop multi-software underlying STP interface adaptation plugins, batch call the native STP output function of general industrial 3D design software, export initial STP data and centrally summarize and verify it. S2. Model Preprocessing and Data Warming: Parse the initial STP data after parsing, extract geometric, semantic and metadata, determine the model type and perform targeted warping operations to generate a standardized processed dataset; S3. External contour capture based on photon collision theory: Configure multi-directional photon beam parameters, construct a collision detection matrix, reconstruct the external contour of the model and mark internal redundant hidden elements; S4. Feature-based lightweight network model optimization: Extract multi-dimensional features from the model, output feature weights through the trained lightweight network model, accurately delete non-critical level components and update topological relationships; S5. Centralized Lightweighting and STP Optimization Output: Centralized lightweighting processing is performed on the marked redundant STP data, metadata is integrated and hybrid compression encoding is used to generate a lightweight STP model that conforms to the ISO10303-21 standard. S6. Intelligent Filtering and Batch Processing Scheduling: Valid data is filtered through intelligent filters, and batch processing queue managers are used to achieve automated processing of multiple models throughout the entire process. Task priority configuration and breakpoint resume are supported.

[0052] Furthermore, step S1 includes: A dedicated adapter plugin is developed for general industrial 3D design software, unifying the command format; the initial STP export parameters are configured according to the ISO10303-21 standard, with an accuracy level set to 0.01mm, and the metadata includes color, material, and assembly relationship; the initial STP data is centrally stored using a Redis distributed cache pool, and format compliance verification is performed through the STEPcode parsing library, with a verification time of ≤10s / GB.

[0053] Furthermore, step S2 includes: The initial STP data is parsed using OpenCASCADE 7.5.0, eliminating the need to parse software-specific data. Solid models are enhanced by extracting geometric features such as holes and chamfers, while surface models undergo smoothness regularization. The maximum gap size of the repaired damaged surface is ≤0.1mm. Data is labeled with model ID, software source, and original format type tags to create a structured index.

[0054] Furthermore, step S3 includes: Employing a "six-orthogonal + four-tilted" photon emission strategy, the photon beam density is expressed as: Reference photon density =100 collisions / mm³, dynamically adjusted according to model volume; collision intensity threshold. The effective collision points are reconstructed using Delaunay triangulation, with a contour capture accuracy of ≤0.05mm; an error compensation formula is introduced. (k=0.003) Corrects the oblique photon collision deviation.

[0055] Furthermore, step S4 includes: Geometric features (4D), semantic features (3D), and assembly relationship features (5D) are extracted to construct a 128-dimensional multi-dimensional feature vector. The lightweight network model adopts a CNN+Transformer hybrid structure and is trained with 12,000 engineering model samples, achieving a feature weight prediction accuracy of ≥96%. The weight threshold θ is configurable (0.3-0.6). After deleting non-critical components, the parent-child relationship of the assembly tree is automatically updated to avoid hierarchical breakage.

[0056] Furthermore, step S5 includes: The algorithm employs a hybrid geometric simplification method combining edge folding and NURBS order reduction, with an edge folding threshold set to 0.05mm, reducing the surface order to 3-5. Metadata is inherited through STP native attribute fields, achieving a retention rate of ≥99%. LZ77+Huffman hybrid compression coding is used, resulting in an overall data compression ratio of 3:1-5:1.

[0057] Furthermore, step S6 includes: The intelligent filter integrates regular expressions and a random forest classification model (100 decision trees), achieving an effective data recognition accuracy of ≥98.5%. The batch processing queue manager supports parallel processing of up to 100 models, supports high, medium, and low priority settings, and records task progress through a local SQLite database.

[0058] Specifically: S1: Heterogeneous Model Interface Adaptation and Initial STP Export S11: Development and Implementation of Interface Adapter Plugins The PDMS adapter plugin is based on TCL scripts written using the Scripting API. It allows models to be selected by region using the CREATESTEPFILE command and supports batch export of PDMS version 12.1. S12: Batch execution operation A batch call controller was written in Python 3.8. It communicates with PDMS via Socket and sends call commands in a unified format: plaintext{"export_path":"D: / cache / gearbox_initial.stp",}. It can control 20 software instances to export in parallel, with an export rate of 500MB / min. The export time of a single model is reduced by 85% compared with manual operation.

[0059] S13: Initial STP Data Summary and Verification

[0060] Set up a Redis cluster as a distributed cache pool, supporting data storage of over 100GB, and receive initial STP data exported from various software; verify the EXPRESS syntax compliance and geometric data integrity of STP files through the STEPcode parsing library, and automatically remove corrupted files (99.8% pass rate); classify and store valid data according to model type, and create an index table to record information such as model ID, software source, initial size, and export time.

[0061] S2: Model Preprocessing and Data Warming

[0062] S21: Initial STP Data Parsing

[0063] Based on the STEP interface of OpenCASCADE 7.5.0, the initial STP data is parsed to extract geometric topology information (vertices, faces, assembly structure), semantic tags (component functional classification) and metadata, with a parsing time of ≤15s / GB. For example, the initial STP data (5.8GB) of a complex factory PDMS model was parsed, and the geometric and semantic information of more than 3200 components was successfully extracted without data loss.

[0064] S22: Model Type Determination

[0065] The model type is determined by analyzing the results: if the data contains key fields such as SOLID and BOOLEAN_OPERATION, it is determined to be a solid model; if SURFACE and CURVE are the main geometric descriptions, it is determined to be a surface model; in this embodiment, mechanical assembly and complex factory PDMS models are determined to be solid models, and building BIM models are determined to be surface models.

[0066] S23: Feature Enhancement and Regularization

[0067] Solid model processing: For core components of mechanical assemblies such as gears and bearings, enhance the extraction of key geometric features such as threads and keyways, and mark feature priorities; Surface model processing: For the exterior walls and roof surfaces of the building BIM model, a surface smoothing algorithm was used to remove burrs (maximum burr size 0.08mm) and repair two damaged surfaces (gap size 0.09mm). S3: External contour capture based on photon collision theory S31: Photon parameter configuration Taking a mechanical assembly (initial STP size 9.2GB, bounding box volume V=500000mm³) as an example, according to the formula Calculate photon density: The system employs a "six orthogonal (±X-axis, ±Y-axis, ±Z-axis) + four tilted (XY, YZ, XZ plane 45°)" launch strategy, with a total of 10 launch directions, to ensure full coverage of the model surface.

[0068] S32: Collision Detection Matrix Construction

[0069] A three-dimensional spatial grid coordinate system was established (grid accuracy 0.01 mm), and a unique ID was assigned to each photon (total 3.97 × 10⁻⁶). 9 (1 photon), record the following information to construct a collision detection matrix: Photon ID emission coordinates (x0, y0, z0) Direction vector (α, β, γ) Collision state Collision point coordinates (x, y, z) Collision surface normal vector (n) x ,n γ Collision intensity (n_z); P001(100,200,300)(0.707,0,0.707) is (105.2,200,305.2)(0.6,0.2,0.8)0.0012; P002(100,200,300)(-0.707,0,0.707)No—0; The matrix is ​​stored using a two-dimensional array in C++, which supports fast querying and filtering. The matrix construction takes 28 minutes (under GPU acceleration).

[0070] S33: Contour Extraction and Redundancy Marking

[0071] Screening collision intensity Valid collision points (total 2.86 × 10) 9 (Number of collision points), using the Delaunay triangulation algorithm based on the PCL library to connect valid collision points and reconstruct the external contour of the mechanical assembly, with a contour smoothness error ≤0.02mm; for non-collision (such as internal cavities) or collision intensity... The area is marked with the corresponding geometric elements (total 1.12 × 10). 6 Each piece of dough, 3.56×10 6 (1 vertex) is redundant and needs to be deleted.

[0072] S34: Precision Optimization

[0073] For collision point deviations in oblique photon beams (e.g., 45° in the XY plane), the error compensation formula is applied. Correction (θ is the incident angle of the photon, in this embodiment θ=45°, After correction, the external contour capture accuracy reached 0.045mm, meeting the requirements of ASME Y14.5 engineering standard.

[0074] S4: Feature-based lightweight network model optimization

[0075] S41: Multidimensional Feature Extraction

[0076] Taking the bearing assembly of a mechanical assembly as an example, we extract multi-dimensional feature vectors: Geometric features: 1200 facets, 80 vertices / mm², curvature change rate 0.35, volume percentage 5% (4D); Semantic features: Functional classification "transmission components", engineering importance level "core components", material "bearing steel" (3D); Assembly relationship characteristics: parent-child component correlation degree 0.9 (12 connecting bolts), motion constraint relationship "rotational pair", disassembly and assembly difficulty level "high" (5 dimensions); The 12-dimensional features are normalized to the [0,1] interval to construct a 128-dimensional extended feature vector (including the model's global feature mapping).

[0077] S42: Network Model Inference

[0078] The lightweight network model is deployed based on the PyTorch 1.12 framework and adopts a CNN+Transformer hybrid structure (3 convolutional layers and 2 encoder layers). The feature vector of the bearing component is input into the model, and the output feature weight value is 0.82 (≥ threshold θ=0.6), which is determined to be a critical component. For components such as oil seals and dust covers, the output weight value is 0.28 (< θ=0.6), which is determined to be a non-critical component.

[0079] S43: Precise Removal of Non-Critical Components

[0080] Based on the OpenCASCADE topology editing interface, non-critical components with a weight value <0.6 (32 in total, accounting for 12% of the total number of components) in the mechanical assembly were deleted; for example, components such as oil seals, dust covers, and temporary fixing brackets were deleted to ensure that the core transmission components (gears, bearings, shafts) were completely preserved.

[0081] S44: Topology Update

[0082] After deleting non-critical components, the STP assembly tree is automatically traversed to update the parent-child component relationships: the original "shaft-oil seal-bearing" relationship is adjusted to a direct "shaft-bearing" relationship, and the motion constraints and assembly constraints are updated synchronously; after the update, the assembly tree hierarchy is complete, with no broken or isolated components.

[0083] S5: Centralized Lightweight Design and STP Optimized Output

[0084] S51: Redundancy Removal and Geometric Simplification

[0085] Batch deletion of marked redundant geometric elements reduces the number of facets in the mechanical assembly from 8.6 × 10⁻⁶. 7 The number decreased to 2.1 × 10 7 For non-critical features (such as the chamfer of the non-mating surface of the bearing outer ring), an edge folding algorithm (based on a quadratic error metric) is performed, resulting in 3.2 × 10⁻⁶ edge folds. 6 One; the NURBS tooth surface of the gear was optimized by reducing the order from 7 to 5, reducing the number of control points by 40% while maintaining the tooth surface contact accuracy.

[0086] S52: Metadata Integration

[0087] The STP standard PROPERTY_DEFINITION field inherits metadata from the initial STP data: the bearing assembly material "bearing steel", color "silver gray", and assembly torque "25N". Information such as "m" is fully preserved; the metadata retention rate reaches 99.2%, with no fields missing or mismatched.

[0088] S53: Hybrid Compression and Standard Verification

[0089] The zlib library was used to implement LZ77 encoding (compressing continuous geometric coordinate data with a compression ratio of 3.1:1). A Huffman coding dictionary was constructed based on the frequency statistics of engineering STP fields (compression ratio of 5.3:1 for high-frequency semantic fields), achieving an overall compression ratio of 4.2:1. The ISO10303-21 compliance of the output STP model was verified using STEPToolsValidationToolkit, with a 100% pass rate. Finally, a lightweight STP model was generated, reducing the size from the initial 9.2GB to 1.1GB, a reduction rate of 88%.

[0090] S6: Intelligent Filtering and Batch Processing Scheduling

[0091] S61: Intelligent Filtering Execution

[0092] The intelligent filter uses regular expressions to match key fields such as "temporary markers" and "duplicate components" and combines them with a random forest classification model to determine the validity of the data. In the initial STP data of the mechanical assembly, 2 duplicate components and 3 temporary marker records were removed, and the accuracy of valid data identification was 98.5%.

[0093] S62: Batch Task Scheduling

[0094] Add three model processing tasks to the batch processing queue manager: Complex Factory PDMS Model (high priority), Building BIM Model (medium priority), and Mechanical Assembly (high priority); set the number of CPU cores to 16 and the GPU memory usage limit to 24GB, execute the tasks in parallel, and the total processing time is 120 minutes (40 minutes per model on average, which is 60% shorter than serial processing).

[0095] S63: Fully Automated Operation

[0096] The entire process of "interface call - data aggregation - preprocessing - lightweighting - optimized output" is automated using Python scripts, requiring no manual intervention. During processing, if a model is interrupted due to a storage anomaly, the process can be resumed using the breakpoint resume function (the recovery point is the S52 metadata integration stage), avoiding redundant calculations. Finally, three lightweight STP models are output, each corresponding to a different application scenario: a complex factory PDMS model (collaborative review), a building BIM model (VR walkthrough), and a mechanical assembly (process simulation). The lightweighting effects of these three engineering models (complex factory PDMS model, building BIM model, and mechanical assembly) are shown in Table 1.

[0097] Table 1. Lightweighting effect parameters of complex factory PDMS model, building BIM model and mechanical assembly

[0098] Furthermore, a large-scale chemical project used this embodiment to process a 6.9GB PDMS native format model. The initial STP data (5.8GB) was exported through the PDMS underlying STP interface and reduced to 1.3GB after processing by the method of this invention (lightweighting rate of 77.6%). The lightweight STP model can be opened natively in various general-purpose industrial 3D design software without geometric distortion and with complete metadata. During collaborative review, the network transmission time was shortened from 2 hours to 22 minutes, and the model loading frame rate was maintained at 60fps during the review process, improving the efficiency of multi-team collaboration by 3 times.

[0099] Example 4

[0100] like Figure 3 As shown, this embodiment proposes a terminal device based on a lightweight method for constructing a multi-source heterogeneous STP model using photon collision and recombination features. The terminal device 200 includes at least one memory 210, at least one processor 220, and a bus 230 connecting different platform systems.

[0101] The memory 210 may include a readable medium in the form of volatile memory, such as random access memory (RAM) 211 and / or cache memory 212, and may further include read-only memory (ROM) 213.

[0102] The memory 210 also stores a computer program that can be executed by the processor 220, causing the processor 220 to execute any of the above-mentioned lightweight method for multi-source heterogeneous STP models based on photon collision and recombination features in the embodiments of this application. The specific implementation method and the technical effects achieved are consistent with those described in the embodiments of the above method, and some details will not be repeated. The memory 210 may also include a program / utility 214 having a set (at least one) of program modules 215. Such program modules include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment.

[0103] Accordingly, processor 220 can execute the aforementioned computer program, as well as executable program / utility 214.

[0104] Bus 230 can represent one or more of several types of bus structures, including a memory bus or memory controller, peripheral bus, graphics acceleration port, processor, or a local bus using any of the various bus structures.

[0105] Terminal device 200 can also communicate with one or more external devices 240, such as keyboards, pointing devices, Bluetooth devices, etc., and with one or more devices capable of interacting with it, and / or with any device that enables it to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be performed via input / output (I / O) interface 250. Furthermore, terminal device 200 can also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapter 260. Network adapter 260 can communicate with other modules of terminal device 200 via bus 230. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with terminal device 200, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0106] Example 5

[0107] like Figure 4 As shown in Example 1, this example proposes a computer-readable storage medium for a lightweight method of multi-source heterogeneous STP model constructed based on photon collision and recombination features. The computer-readable storage medium stores instructions that, when executed by a processor, implement any of the aforementioned lightweight methods of multi-source heterogeneous STP model constructed based on photon collision and recombination features. The specific implementation method and the achieved technical effects are consistent with those described in the examples above, and some details will not be repeated.

[0108] This embodiment provides a program product 300 for implementing the above-described method, which may employ a portable compact disc read-only memory (CD-ROM) and include program code, and can run on a terminal device, such as a personal computer. However, the program product 300 of the present invention is not limited thereto. In this embodiment, the readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The program product 300 may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0109] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on a user computing device, partially on a user device, as a standalone software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to user computing devices via any type of network, including local area networks (LANs) or wide area networks (WANs), or they can be connected to external computing devices (e.g., via the Internet using an Internet service provider).

[0110] This invention is described from the perspectives of its intended use, effectiveness, progress, and novelty. Its practical and progressive features meet the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings are merely preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc., that are similar to or identical to those of this application, i.e., all equivalent substitutions or modifications made in accordance with the scope of this patent application, shall fall within the scope of protection of this patent application.

[0111] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features, characterized in that, Includes the following steps: S1. Obtain the native STP output interface information of each 3D industrial design software, call the native STP output interface of each 3D industrial design software in batches through the interface adapter plugin, export the first STP dataset in STEP file format and store it in the distributed cache pool. S2. Analyze the first STP dataset, determine whether each 3D model in the first STP dataset is a solid model or a surface model, and according to the determination result, perform geometric feature enhancement extraction processing on the solid model and surface smoothness regularization processing on the surface model to obtain the second STP dataset. S3. By simulating the collision process between multi-directional photon beams and the surfaces of various three-dimensional models in the second STP dataset, a photon collision detection matrix is ​​constructed to obtain the external contour data of the three-dimensional model. Based on the preset collision intensity threshold, the external contour data of the three-dimensional model is marked as visible geometric element data and internal hidden redundant geometric element data, respectively. S4. Extract the geometric features, semantic features, and assembly relationship features of each 3D model in the second STP dataset and construct a composite feature vector. Input the vector into a pre-trained lightweight network model for inference to obtain key feature weights. Based on the key feature weights, delete non-key level 3D models and update the topological hierarchy of the assembly tree simultaneously to obtain the third STP dataset. S5. Based on the third STP dataset, perform centralized geometric simplification on the STP data marked with internally hidden redundant geometric elements. By inheriting the STP metadata through the attribute definition fields in the STEP file format, and generating a lightweight STP model through boundary representation optimization algorithm and hybrid compression coding.

2. The lightweight method for a multi-source heterogeneous STP model based on photon collision and recombination features as described in claim 1, characterized in that, Step S1 includes the following sub-steps: S101. Based on the underlying STP output interface characteristic parameters of different 3D industrial design software, configure dedicated interface plug-ins for each 3D industrial design software, and establish a connection with each 3D industrial design software through a process communication mechanism, so as to control each 3D industrial design software to synchronously execute native STP export by sending a unified format call command. S102. Establish a distributed data cache pool to receive the first STP data exported by various 3D industrial design software in parallel. S103. Verify the file header format, EXPRESS language syntax compliance, and geometric data integrity of the first STP data respectively. If any verification fails, repeat steps S101-S102 until the verification passes and the first STP data is obtained.

3. The lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features according to claim 1, characterized in that, Step S2 includes the following sub-steps: S201. Based on the first STP dataset, the first STP dataset is parsed using the OCC system to extract geometric topology information, component functional classification semantic labels, and metadata; the metadata includes model color data, model material data, and model assembly relationship data; S202. Based on the parsing and extraction results, determine the model type. If the parsing and extraction results contain a solid component field, it is determined to be a solid model; if the parsing and extraction results contain a surface geometry field, it is determined to be a surface model. S203. Based on the judgment result, if it is determined to be a solid model, then enhance the extraction of hole and chamfer geometric features and mark the feature priority; If the model is determined to be a curved surface, a surface smoothing algorithm is used to remove burrs and repair the damaged surface. S204. Based on the model results obtained in step S203, construct the second STP dataset.

4. The lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features according to claim 1, characterized in that, Step S3 includes the following sub-steps: S301. Based on the second STP dataset, construct a three-dimensional spatial mesh coordinate system with the geometric center points of each three-dimensional model, and emit photon beams to the surface of each three-dimensional model within the second STP dataset according to the six-orthogonal, four-tilted multi-directional photon emission mechanism. The formula for calculating the photon beam density is expressed as: ； in, Indicates the photon beam density. This represents the preset photon beam reference density. This represents the preset reference volume of the 3D model. This represents the volume of the bounding box of a 3D model. S302. Based on the three-dimensional spatial grid coordinate system, assign a unique ID to each photon within the photon beam, and perform collision point error compensation for the oblique photon beam. Record the emission coordinate data, direction vector data, collision state data, collision point coordinate data, collision surface normal vector data, and collision intensity data corresponding to each photon, and construct a collision detection matrix. The error compensation formula is expressed as: ； in, This indicates the collision point error compensation. This represents the material correction factor for the 3D model. Indicates the angle of incidence of the photon; S303. Filter valid collision points in the collision detection matrix with a collision intensity greater than or equal to a set threshold, mark them as visible geometric element data, use the Delaunay triangulation algorithm to connect all valid collision points, obtain the continuous outer contour of each reconstructed 3D model, and mark the spatial regions where no collision occurred or where the collision intensity is less than the set threshold as internal hidden redundant geometric element data.

5. A lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features according to claim 1, characterized in that, Step S4 includes the following sub-steps: S401. Based on the second STP dataset, extract multi-dimensional features for each 3D model. These multi-dimensional features include geometric features, semantic features, and assembly relationship features. The geometric features include the number of triangular faces, vertex density, rate of curvature change, and volume percentage. The semantic features include component functional classification and engineering importance level based on STEP AP214 semantic tags. The assembly relationship features include the parent-child component association degree and motion constraint relationship. After normalizing the multi-dimensional features, construct a high-dimensional composite feature vector. S402. Construct a lightweight network model based on CNN network and Transformer structure, use historical engineering model samples as training set, construct a joint optimization objective based on lightweight rate loss function and key feature retention rate loss function, and pre-train the lightweight network model; S403. Based on the trained lightweight network model, the composite feature vector is input into the lightweight network model, the CNN structure in the lightweight network model is used to extract the local correlation of geometric features, the Transformer structure is used to extract the global dependency of semantic features and assembly relationship features, and the key feature weight values ​​of a single component are output. S404. Compare the key feature weight values ​​with preset weight thresholds. Determine the 3D models with weight values ​​greater than or equal to the preset weight thresholds as key level models. Determine the 3D models with weight values ​​less than the preset weight thresholds as non-key level models and delete them. After deletion, automatically traverse the assembly information of each 3D model in the second STP dataset and update the association relationship and geometric topology connection information of parent and child components to obtain the third STP dataset.

6. The lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features according to claim 1, characterized in that, Step S5 includes the following sub-steps: S501. Based on the third STP dataset, delete the STP data marked with internal hidden redundant geometric elements, perform an edge folding algorithm based on quadratic error metric on non-critical features, and perform order reduction optimization on NURBS surfaces to reduce the number of control points. S502. Inherit the color, material, and function label metadata from the first STP data by defining the attribute fields in the STEP file format; S503. The boundary representation method is used to simplify the geometric description statement, and LZ77 encoding is used to compress continuous geometric coordinate data. Huffman encoding is used to compress high-frequency semantic fields to generate the final lightweight STP model.

7. The lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features according to claim 1, characterized in that, It also includes building intelligent filters based on regular expressions and random forest classification models to identify and remove duplicate components and temporary labeled data in the first STP data.

8. A lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features according to claim 1, characterized in that, When there are multiple STP model lightweighting tasks, the priority of each STP model lightweighting task is configured through the batch task manager based on the actual working conditions. The parallel lightweighting computing resources are allocated to multiple STP model lightweighting tasks according to the task priority, and the task progress is recorded in real time through the local database.

9. A lightweight system for constructing a multi-source heterogeneous STP model based on photon collision and recombination features, the system being implemented based on the lightweight method for constructing a multi-source heterogeneous STP model based on photon collision and recombination features as described in any one of claims 1-8, characterized in that, include: The interface adaptation and export module is used to obtain the native STP output interface information of each 3D industrial design software, and to call the native STP output interface of each 3D industrial design software in batches through the interface adaptation plugin, export the first STP dataset in STEP file format and store it in a distributed cache pool. The model preprocessing module is used to parse the first STP dataset, determine whether each 3D model in the initial STP dataset is a solid model or a surface model, and perform geometric feature enhancement extraction processing on the solid model and surface smoothness regularization processing on the surface model to obtain the second STP dataset. The external contour capture module is used to construct a photon collision detection matrix by simulating the collision process between multi-directional photon beams and the surfaces of various three-dimensional models in the second STP dataset, obtain the external contour data of the three-dimensional model, and mark the external contour data of the three-dimensional model as visible geometric element data and internal hidden redundant geometric element data based on a preset collision intensity threshold. The feature network optimization module is used to extract the geometric features, semantic features, and assembly relationship features of each 3D model in the second STP dataset and construct a composite feature vector. This vector is then input into a pre-trained lightweight network model for inference to obtain key feature weights. Based on these key feature weights, non-key level 3D models are deleted, and the topological hierarchy of the assembly tree is updated synchronously to obtain the third STP dataset. The optimization output and scheduling module is used to perform centralized geometric simplification on STP data marked with redundant geometric elements based on the third STP dataset. It inherits STP metadata through STP attribute definition fields and generates a lightweight STP model through boundary representation optimization algorithm and hybrid compression coding.

10. A lightweight system for a multi-source heterogeneous STP model constructed based on photon collision and recombination features according to claim 9, characterized in that, It also includes an intelligent filtering and batch processing module, comprising an intelligent filtering unit and a batch processing resource allocation unit. The intelligent filtering unit is used to construct an intelligent filter based on regular expressions and a random forest classification model to identify and remove duplicate components and temporary labeled data in the first STP data. The batch processing resource allocation unit is used to configure the priority of multiple STP model lightweight tasks according to actual working conditions through a batch processing task manager, allocate parallel lightweight computing resources for multiple STP model lightweight tasks according to task priority, and record task progress in real time through a local database.

Images