Model feature coding and retrieval management method and system for three-dimensional design engine

Abstract

The application provides a model feature coding and retrieval management method and system for a three-dimensional design engine, relates to the technical field of model feature data processing, obtains an assembly structure tree by analyzing an assembly model, separates the field information of each node in the assembly structure tree, retains preset stable fields, removes preset unstable fields, forms repeatable structure input, generates node stable features based on the stable fields, performs sorting on the child node set corresponding to each parent node in the assembly structure tree based on the node stable features, performs collision detection and context refinement processing when the same sorting basis child nodes appear in the child node set, and finally integrates to obtain the final child node sequence corresponding to each parent node, so that the structural feature jump caused by the number of fasteners or the arrangement density is inhibited, and the coding consistency and retrieval accuracy under the conditions of cross-version and cross-import are further improved.

Description

Technical Field

[0001] This invention relates to the field of model feature data processing technology, and in particular to a model feature encoding and retrieval management method and system for 3D design engines. Background Technology

[0002] Currently, in the feature encoding stage, 3D model management and collaboration platforms need to first perform unified processing on computer-aided design and 3D files from diverse sources, converting the original design files into derived formats that are easy for browsers to load and view in a lightweight manner, simultaneously extracting the object hierarchy tree and various attribute fields of the model, and generating thumbnails of various sizes to enable rapid preview of the model.

[0003] After standardization, the platform further processes the searchable information into structured data or vector form. On one hand, it supplements high-quality auxiliary field information according to a preset classification system and attribute set to support accurate attribute-based filtering; on the other hand, it generates shape features for parts and mines reusable parts based on 3D shape similarity. The retrieval stage adopts a multi-channel parallel mode, providing both keyword and attribute retrieval, supporting search and filtering by name, number, and auxiliary field information, and shape similarity retrieval function, supporting matching similar candidates with sample parts and performing 3D comparison. At the same time, it sets up sketch, outline, and image retrieval entry points and a unified search box to achieve relevance matching of all attribute fields.

[0004] For example, Chinese invention patent CN114461856B discloses a method for encoding and retrieving part features for enterprise design resources, which includes the following steps: (1) Structured representation of part feature information; (2) Constructing part feature encoding rules: adopting a hybrid encoding rule, with both tree-based and chain-based encoding rules, to characterize the management features, shape features, and material features of the parts, and setting additional code positions to more accurately describe the part features; (3) Inputting part feature information, for unstructured data including process documents, using natural language processing to convert unstructured data into structured data; (4) Outputting XML files to part feature encoding; (5) Generating part families with similarity; (6) Retrieving part features.

[0005] For example, Chinese invention patent CN119336938B discloses a method, system, and medium for constructing and retrieving a scene library based on multi-dimensional feature coding, which relates to the field of autonomous driving. The method includes: constructing a hierarchical model; quantifying the feature information of different scene elements in each information layer to obtain the feature codes of scene elements in each information layer; combining the feature codes of scene elements in all information layers in multiple dimensions to obtain the standard codes corresponding to vehicle driving scenarios; constructing a scene library based on the original data, feature information, feature codes, and standard codes of different vehicle driving scenarios; retrieving the standard codes stored in the scene library based on the standard codes of the target driving scenario and / or the feature codes of the target scene elements, and obtaining the original data of the corresponding vehicle driving scenario from the scene library.

[0006] The existing technology has the following technical problems: In cloud-based model management and retrieval scenarios for 3D design engines, assembly models typically organize component relationships in the form of a structure tree. To support requirements such as structural similarity retrieval and solution reuse location, the system often needs to convert this structure tree into indexable structural feature codes (e.g., structural summaries, structural vectors, or structural fingerprints), and calculate similarity and ranking based on these codes during the retrieval phase. This allows for the rapid retrieval of model versions or historical solutions that are similar to the target assembly structure from a vast amount of assets.

[0007] However, existing structure tree coding is prone to relying on factors that are not essential to the structure but are highly variable during actual import and iteration when implemented in engineering projects. This leads to the coding results being overly sensitive to these factors. First, different 3D formats, different parsers, or different import paths can cause changes in the order of sibling nodes, such as inconsistent sorting methods by import order, file name, creation time, or internal ID. At the same time, the number of instances of the same part in the assembly can fluctuate due to adjustments in the number of fasteners, fine-tuning of the layout density, and replacement of alternative parts. If the coding does not distinguish between part definition (type) and assembly instances (quantity) and uses the number of instances as the dominant statistic, then fluctuations in quantity will be mistakenly taken as changes in the essential structure.

[0008] The aforementioned factors are prevalent because 3D model asset management emphasizes cross-format access and multi-role collaboration, resulting in diverse and continuously evolving tree structures. Furthermore, to pursue retrieval speed and index simplification, feature encoding typically tends to compress the tree structure into a fixed dimension or a single fingerprint representation. In this compression process, without unified structural normalization rules and robustness constraints, the encoding results will inevitably carry non-essential noise such as import order, naming conventions, and instance fluctuations, leading to incomparable structural representations across different versions, sources, or parsing conditions. Summary of the Invention

[0009] This invention provides a model feature encoding and retrieval management method and system for 3D design engines, which can generate comparable structural representations under different versions, sources, or parsing conditions. The technical solution provided by this application is as follows: Firstly, a method for model feature encoding and retrieval management for 3D design engines is provided, and the specific implementation of this method is as follows: Step 1: Parse the assembly model to obtain the assembly structure tree, and separate the auxiliary field information of each node in the assembly structure tree. Pre-defined stable fields are retained, and pre-defined unstable fields are removed to form a repeatable structure input. Step 2: Generate node stability features based on the stable fields. Based on the node stability features, sort the child node sets corresponding to each parent node in the assembly structure tree. When child nodes with the same sorting criteria appear in the child node sets, collision detection and context refinement are performed. Finally, integrate to obtain the final child node sequence corresponding to each parent node. Step 3: Aggregate similar child nodes in the final child node sequence and generate a quantitative representation. Based on the quantitative representation, robustly implement the final child node sequence. The process involves several steps: First, updating the final child node sequence corresponding to each parent node to suppress interference from quantity fine-tuning on feature encoding. Second, performing feature encoding based on the updated final child node sequence corresponding to each parent node, and performing stability self-checks when a version chain exists. If an encoding anomaly is detected, parameter loop adjustment is triggered to stabilize the feature encoding. Feature encoding refers to generating structural fingerprints and structural vectors. Third, when the 3D design engine receives a query command, it generates the structural fingerprint and structural vector corresponding to the query command. Based on the structural fingerprint, it recalls a set of candidate model recommendations, then filters the set of candidate model recommendations based on the structural vector, marks it as a model recommendation set, and outputs the model recommendation set.

[0010] Secondly, a model feature encoding and retrieval management system for 3D design engines is provided. This system includes: a field separation module, used to parse the assembly model to obtain the assembly structure tree, and to separate the auxiliary field information of each node in the assembly structure tree, retaining preset stable fields and removing preset unstable fields to form a repeatable structure input; a sequence integration module, used to generate node stable features based on stable fields, and to sort the child node sets corresponding to each parent node in the assembly structure tree based on the node stable features. When child nodes with the same sorting criteria appear in the child node set, collision detection and context refinement processing are performed, and finally, the final child node sequence corresponding to each parent node is obtained; and a robust processing module, used to aggregate similar child nodes in the final child node sequence and generate a quantitative representation. The system robustly processes the final child node sequence based on quantitative representation, thereby updating the final child node sequence corresponding to each parent node to suppress the interference of quantitative fine-tuning on feature encoding. The feature encoding module performs feature encoding based on the updated final child node sequence corresponding to each parent node, performs stability self-check when a version chain exists, and triggers parameter loop adjustment to stabilize feature encoding when an encoding anomaly is detected. Feature encoding refers to generating structural fingerprints and structural vectors. The set output module generates structural fingerprints and structural vectors corresponding to the query command when the 3D design engine receives a query command. Based on the structural fingerprint corresponding to the query command, it recalls a set of candidate model recommendations, then filters the set of candidate model recommendations based on the structural vector corresponding to the query command, marks it as a set of model recommendations, and outputs the set of model recommendations.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: (1) When the 3D design engine performs cloud-based management and similarity retrieval of assembly models, it first parses the model into an assembly structure tree and separates the auxiliary field information of each node. It retains only stable fields that can be repeatedly calculated across import formats, parsers, and import paths, and removes unstable fields that drift with changes in import order, file rewriting, or display strategies. This prevents non-essential information from entering the structural expression and causing coding inconsistencies. Then, it generates stable node features based on stable fields and performs deterministic sorting on the child node set of each parent node. When the same sorting criteria cause the order of sibling nodes to be uncertain, it introduces substructure information through collision detection and context refinement to disambiguate and form a reproducible final child node sequence, solving the problem of misjudgment in similarity retrieval caused by order drift in symmetrical and repetitive parts scenarios. On this basis, it aggregates similar child nodes and generates a quantity expression, and performs robust processing on the final child node sequence accordingly to suppress structural feature jumps caused by fine-tuning of fastener quantity or arrangement density. Furthermore, structural fingerprints and structural vectors are generated based on the updated final child node sequence, and a stability self-check is performed when a version chain exists. When an anomaly is identified where the geometric changes are not significant but the encoding differences are too large, parameter loop adjustment is triggered to stabilize feature encoding, improving encoding consistency and retrieval accuracy across versions and import conditions. In the retrieval phase, corresponding structural fingerprints and structural vectors are generated for each query command. First, a candidate set is recalled based on the structural fingerprint, and then the candidate set is filtered and sorted based on the structural vector to obtain the model recommendation set and output it.

[0012] (2) In the assembly structure tree, this invention statistically analyzes the proportion of child nodes with the same sorting criteria to the total number of child nodes of the corresponding parent node in the child node sequence, and marks this proportion as the sibling node collision probability estimate, which is used to characterize the degree to which child nodes at the same level are difficult to distinguish due to consistent geometric summaries or similar types. The sibling node collision probability estimate is compared with a preset collision threshold. When it is not higher than the collision threshold, the child node sequence of the parent node is confirmed as the final child node sequence and used for feature encoding, thereby ensuring that a reproducible structural expression is obtained under different parsers and different import paths. When the sibling node collision probability estimate is higher than the collision threshold, context refinement processing is triggered, the child nodes that have collided are marked as child nodes to be analyzed and their context summaries are generated. Based on the context summaries and the original sorting criteria, a secondary sorting criteria are constructed, the child nodes to be analyzed are reordered to update the child node sequence, and the sibling node collision probability estimate is recalculated on the updated result. If it is still higher than the collision threshold, the context refinement range is extended to multiple substructures and deepened layer by layer until the collision probability estimate is not higher than the collision threshold. By introducing substructure context as supplementary discriminant information, the order uncertainty in symmetrical and repetitive component scenarios can be effectively suppressed, and the import order difference can be prevented from being amplified into encoding difference, thereby improving the stability of structural feature encoding and the accuracy of retrieval matching.

[0013] (3) When performing feature encoding based on the updated sequence of final child nodes corresponding to each parent node, this invention generates a structural fingerprint and structural vector of the assembly structure tree. If a version chain exists, it obtains the structural fingerprint of the previous version of the same assembly structure tree. By calculating the structural fingerprint similarity between the current structural fingerprint and the previous version's structural fingerprint and simultaneously evaluating the magnitude of geometric changes, it achieves self-checking of the stability of the structural encoding. This processing addresses the common problem of encoding drift caused by non-essential changes in engineering practice. For example, differences in import order, naming adjustments, or minor adjustments to the number of duplicate parts can lead to the same assembly structure being misjudged as a different structure, resulting in inaccurate version tracing and a decrease in the accuracy of similarity retrieval. When the structural fingerprint similarity is not lower than the preset structural fingerprint similarity threshold, or the magnitude of geometric changes is higher than the preset geometric change threshold, the current structural fingerprint is confirmed for database entry and retrieval, ensuring that structural differences are effectively reflected. When the structural fingerprint similarity is lower than the threshold and the magnitude of geometric changes is not higher than the threshold, it is determined that the encoding is too sensitive to non-essential changes and triggers parameter loop adjustment. By adjusting the parameters related to geometric discretization, quantity expression compression, and context disambiguation, the structural fingerprint is regenerated, thereby improving the consistency and reproducibility of structural expression across versions and import conditions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a model feature encoding and retrieval management method for a 3D design engine provided in an embodiment of the present invention; Figure 2 This is a flowchart of the assembly model structure tree sorting and collision handling provided in an embodiment of the present invention; Figure 3 This is a flowchart of the aggregation and quantity robustness processing of similar child nodes provided in the embodiments of the present invention; Figure 4 This is a flowchart of the model feature encoding verification and retrieval management provided in the embodiments of the present invention; Figure 5 This is a structural diagram of the model feature encoding and retrieval management system for 3D design engines provided in an embodiment of the present invention; Figure 6 This is an example diagram of the servo motor planetary reduction module assembly structure tree provided in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0017] Before providing a detailed explanation of the embodiments of this application, the application scenarios of these embodiments will be described first.

[0018] In enterprise-level R&D and collaborative innovation, a large number of 3D assembly models are centrally stored and managed in the cloud. These models come from different design software and file formats, and continuously evolve through design reviews, supply chain collaboration, and multiple version iterations. This objectively leads to problems such as fragmented design data, low reusability, inconsistent drawing and model versions, and low efficiency in cross-team communication. To improve the efficiency of solution reuse and review traceability, platforms typically need to manage and retrieve the assembly model's structure tree, enabling users to quickly locate historical solutions and similar structural models through keyword tags or intelligent searches, and to select reusable representative versions from multiple assembly versions. Therefore, this application addresses the model asset management and retrieval scenarios of the aforementioned cloud-based 3D design engines. Specifically, it proposes a corresponding feature encoding and retrieval management mechanism to address the problem of inconsistent structural representations affecting retrieval accuracy under conditions of cross-format import, parsing differences, and version evolution of the assembly structure tree.

[0019] Among them, the 3D design engine refers to a software platform that carries out 3D model data import and parsing, geometric rendering and interaction, structure tree organization and display, as well as collaborative review and asset management. It is usually cloud-based and provides online viewing and editing capabilities for 3D data in product design and creative design scenarios. It supports multi-terminal access and multi-format compatible import, allowing users to perform interactive operations such as loading, rotating, sectioning, and exploding on complex assembly models without local professional software. Users can also view the complete composition structure of the model based on the model structure tree. At the same time, it combines online collaborative review, annotation and access control mechanisms to realize cross-departmental and cross-regional communication and decision-making, and further provides design asset management, versioning and accumulation, as well as solution location and reuse capabilities based on tags or intelligent search.

[0020] Example 1: This embodiment of the invention provides a model feature encoding and retrieval management method for 3D design engines.

[0021] Figure 1 is a flowchart of the model feature encoding and retrieval management method for 3D design engines provided by the present invention. The processing flow of this method may include the following steps: Taking a servo motor planetary reducer module assembly as an example, after being imported into a 3D design engine, the assembly is parsed into an assembly structure tree to represent its hierarchical composition and assembly hierarchy. The top-level node of this assembly structure tree corresponds to the servo motor planetary reducer module assembly body, with sub-assembly nodes including motor body components, planetary reducer components, output end support components, sensor and wiring harness components, and fastener packages. The fastener package node further includes component nodes such as hexagonal socket head cap screws, flat washers, and spring washers, while the planetary carrier component node includes repeating sub-assembly nodes such as planetary gear components. Through parsing this assembly, the parent node and its corresponding set of child nodes in the assembly structure tree are clearly defined, forming a parent-child relationship structure that can be used for subsequent coding processing.

[0022] Each node in the assembly structure tree of the servo motor planetary reducer module assembly carries a set of auxiliary field information to describe the node's category, identifier, specifications, geometric summary, and import parsing related information. To ensure consistent and repeatable structural input for the assembly under different import formats, parsers, or import paths, relevant technical personnel pre-define field separation rules. These rules divide the node's auxiliary field information set into stable and unstable fields and perform filtering and updates accordingly. Stable fields are pre-defined by technical personnel as field information that stably reflects assembly relationships or can be repeatedly calculated from geometric data, such as part category or type identifiers, standard part specification parameters, bounding box magnitudes, and volume or area magnitudes. Unstable fields are pre-defined by technical personnel as field information that is prone to drifting with changes in the import process, file rewriting, parsing implementation, or display strategy, such as import sequence numbers, parser-generated internal object numbers, display order fields, and name strings carrying migration suffixes or temporary numbers. According to the field separation rules, unstable fields are removed from the auxiliary field information of each node of the servo motor planetary reducer module assembly. Only stable fields are retained and written back to update the corresponding nodes in the assembly structure tree. At the same time, a relationship table between parent nodes and child node sets is established to solidify the hierarchical organization of the assembly and form a repeatable structural input, providing a consistent data foundation for subsequent normalized sorting and feature encoding of child nodes.

[0023] After the servo motor planetary reducer module assembly completes structure tree parsing and field purification, each node in the assembly structure tree retains a set of field information that stably reflects the essence of the structure, such as part category or type identifier, standard part specification parameters, and bounding box magnitude and volume or area magnitude calculated from geometric data. Based on these stable fields, each node of the servo motor planetary reducer module assembly is further generated with node stability features for encoding. The node stability features are structural description information that can be repeatedly compared for that node, used to maintain consistent sorting and alignment under different import conditions.

[0024] Figure 6 This is an example diagram of the servo motor planetary reducer module assembly structure tree provided in this embodiment of the invention. This tree diagram is a schematic diagram of the assembly structure tree of the servo motor planetary reducer module assembly, used to intuitively express which sub-assemblies and parts make up the whole machine level by level and the parent-child assembly relationship between them: The servo motor planetary reducer module assembly at the center of the diagram is the root node, and branches outward to connect the motor body assembly, planetary reducer assembly, output end support assembly, sensor and wiring harness assembly, fastener package and other first-level sub-assembly nodes. Each sub-assembly node is further decomposed into more detailed sub-components and specific part nodes, until the outermost leaf node is a standard part or single part that cannot be further divided; In addition to the name, the node box also gives specification or model information. For example, M6×20 internal hex bolt indicates a fastener specification with a nominal thread diameter of 6 and a length of 20. Bearing 6202 or bearing 6001 indicates the bearing model. Flat washer M6 and spring washer M6 indicate the washer type that is compatible with M6 bolts. The multiplication sign and number following the node name indicate the number of assembly instances of that node at its parent node's corresponding assembly level. That is, the number of similar parts or isomorphic subassemblies that appear repeatedly at that location. For example, "planetary gear assembly ×3" means that the subassembly is repeatedly assembled 3 times in the planet carrier; "permanent magnet ×8" means that 8 permanent magnets are assembled in the rotor assembly; and "M6×20×8 socket head cap screws, M6×8 flat washers, and M6×8 spring washers" means that 8 standard parts of the same specification need to be configured in the fastener package or corresponding connection part. Therefore, by reading from the inside to the outside along any branch, one can clearly understand the composition path of a certain functional module, the types and specifications of the parts it contains, and the quantity configuration of each type of part at that assembly location, thus allowing the reader to accurately understand the hierarchical structure and distribution of repeated parts of the assembly.

[0025] In the assembly structure tree, the fastener package is a parent node with several child nodes. Each child node corresponds to a fastener part category, including sub-nodes such as M6×20 socket head cap screws, M5×16 socket head cap screws, M6 flat washers, and M6 spring washers. The socket head cap screws, flat washers, and spring washers characterize the fastener category. The subsequent M6, M5, M6×20, and M5×16 are specification identifiers, used to characterize dimensional parameters such as nominal thread diameter and length. For example, M6×20 indicates a bolt with a nominal thread diameter of 6mm and a length of 20mm, while M5×16 indicates a bolt with a nominal thread diameter of 5mm and a length of 16mm. For washers, M6 indicates that it is compatible with a thread specification of 6mm. These specifications distinguish fastener types with different sizes and compatibility relationships. The frequency of each child node in the assembly structure tree reflects the actual quantity of that specification of fastener used in the assembly. This quantity information is usually statistically processed in subsequent aggregation and quantity representation. To eliminate differences in the arrangement of sibling nodes caused by different import formats, different parsers, or different import paths, the system constructs a sorting criterion for each child node.

[0026] The sorting criterion can be understood as a comparable representation of the sub-node, formed by a combination of fields that stably reflect the essence of the part. It includes at least the part type identifier and geometric summary information calculated and discretized from geometric data, ensuring that the same sub-node can generate a consistent comparison key under different parsing conditions. It typically consists of two parts: the first part describes the part type of the sub-node, such as bolt, flat washer, or spring washer; the second part describes the geometric specifications of the part, such as the discretized results of diameter, length, or thickness. Each sub-node corresponds to a sorting criterion with a fixed format. For example, the sorting criterion for a spring washer M6 can be understood as the spring washer category plus the geometric magnitude of the smaller thickness within the diameter; the sorting criterion for an internal hex bolt M6×20 can be understood as the bolt category plus the geometric magnitude of the smaller length within the diameter. As long as the model itself remains unchanged, this comparison information can be repeatedly obtained under different import methods.

[0027] A child node sequence refers to a list of all child nodes under the same parent node arranged in a specific order. Since the original order of child nodes may differ across import paths, the system does not use the original order. Instead, it extracts the sorting criteria for each child node and performs a unified sorting to obtain a fixed arrangement. This fixed arrangement is the child node sequence, which is essentially a list specifying the order in which each child node under the fastener package should appear. For example, it might list spring washers M6 first, then flat washers M6, then socket head cap screws M5×16, and finally socket head cap screws M6×20. This process ensures the reproducibility of the child node sequence; that is, with the model structure unchanged, regardless of changes in import and parsing conditions, the child nodes corresponding to the fastener package can be organized into an ordered list in the same order, thus providing a consistent structural input for subsequent feature encoding.

[0028] After completing the normalized sorting, the servo motor planetary reducer module assembly further checks whether there are child nodes with the same sorting basis in the normalized child node sequence. This means that the similarity of the sorting basis of any two child nodes under the same parent node is calculated. When the calculated similarity of the sorting basis is not less than the similarity threshold preset by the relevant technical personnel, it is determined that the sorting basis of the two is consistent. The sorting basis similarity can be calculated using methods such as cosine similarity.

[0029] Figure 2 This is a flowchart of the assembly model structure tree sorting and collision handling provided in this embodiment of the invention. First, the assembly model is parsed to obtain the corresponding assembly structure tree, and a correspondence table between parent nodes and child nodes is established. Then, a unified sorting criterion is constructed for the child nodes under each parent node. After the child nodes are sorted lexicographically, it is determined whether there are child nodes with the same sorting criterion in the child node sequence. If there are no child nodes with the same sorting criterion, the currently sorted child node sequence is directly determined as the final child node sequence of the parent node. If there are child nodes with the same sorting criterion, the parent node and its corresponding child nodes are marked as a collision candidate set, and the collision probability estimate of sibling nodes is calculated simultaneously. Then, the collision probability estimate is compared with a preset collision threshold. If the collision probability estimate is not higher than the preset collision threshold, the current child node sequence is directly confirmed as the final child node sequence. If the collision probability estimate is higher than the preset collision threshold, a context summary is generated for the colliding child nodes, and a secondary sorting criterion is constructed for re-sorting. At the same time, the context refinement range is expanded layer by layer until the collision probability estimate drops to the preset collision threshold or below. Finally, the compliant final child node sequence of the parent node is obtained, completing the structure tree sorting and conflict correction.

[0030] Taking the planetary carrier assembly in a planetary reducer as an example, the planetary carrier assembly contains multiple planetary gear assemblies, and these planetary gear assemblies may be highly consistent in type and geometric summary magnitude, making them difficult to distinguish during sorting. For example, suppose there are two child nodes under the fastener package: Child node 1: M6×20 hex bolt, Child node 2: M6×20 hex bolt (from another import source, with a different name suffix, but the same type and geometric summary). The system encodes the sorting criteria of each child node into a vector, for example, by concatenating the part type identifier, specification discrete code, and geometric summary discrete code, and calculates the similarity between the two using cosine similarity. The relevant technical personnel pre-set the similarity threshold to 0.95. If the calculated similarity between the two is 0.98, then 0.98 is not less than 0.95, and the sorting criteria are determined to be consistent, thereby triggering the subsequent collision candidate set and context refinement mechanism.

[0031] At this point, the system does not use the original import order or the interface display order to forcibly determine the order, but instead includes the child nodes under the parent node that have the same sorting criteria into the collision candidate set and performs a quantitative evaluation of the degree of collision.

[0032] The statistical sorting is based on the proportion of identical child nodes to the total number of child nodes of the corresponding parent node. This proportion is marked as the estimated probability of sibling node collisions and is used to characterize the severity of indistinguishable sibling nodes under that parent node. The higher the proportion, the more likely there is to be unstable sorting or inconsistent cross-parsing conditions at that level.

[0033] The estimated collision probability of sibling nodes is compared with the preset collision threshold in the database; where the collision threshold represents the maximum value allowed for the estimated collision probability of sibling nodes.

[0034] When the estimated collision probability of sibling nodes is not higher than the collision threshold, it indicates that the proportion of cases with the same sorting basis is acceptable, and the order of child nodes under the current parent node has sufficient distinguishability and reproducibility. Therefore, the child node sequence of the parent node is directly identified as the final child node sequence and used for subsequent feature encoding.

[0035] When the estimated collision probability of a sibling node is higher than the collision threshold, it indicates that there are many indistinguishable sibling nodes under that parent node. Simply relying on the current sorting criteria is insufficient to form a stable order, triggering context refinement processing.

[0036] Context refinement is performed on sets of indistinguishable child nodes. In this embodiment, indistinguishable child nodes are marked as nodes to be analyzed. These nodes refer to a set whose order cannot be stably distinguished under the current sorting criteria. The system generates a context summary for each node to be analyzed. This summary describes the underlying assembly characteristics of the node and can be obtained by combining the stable characteristics of its direct child nodes. For example, for the planetary gear assembly, its direct child nodes typically include planetary gears, needle roller bearings, and pins. The system organizes and summarizes the stable characteristics of these direct child nodes, such as their types and geometric outlines, according to unified rules to form a context summary, ensuring that the context summary reflects the differences in the underlying structure of the node to be analyzed.

[0037] After generating the context summary, the system combines the context summary with the original sorting criteria to form a secondary sorting criterion. The secondary sorting criterion can be understood as adding a layer of substructure information on top of the original comparison keys to improve distinguishability. Subsequently, the system reorders the child nodes to be analyzed based on the secondary sorting criterion and merges the reordered results back into the child node sequence of the parent node, thus obtaining an updated child node sequence. After the update, the system again calculates the proportion of child nodes with the same sorting criteria and recalculates the sibling node collision probability estimate for the updated child node sequence to assess whether context refinement has reduced indistinguishable cases to an acceptable range.

[0038] If the recalculated sibling node collision probability estimate is still higher than the collision threshold, it indicates that introducing only direct child node information is insufficient for differentiation. In this case, the system expands the context refinement scope from direct child nodes to multiple substructures, further introducing deeper-level assembly relationships and stable feature information. It then progressively deepens the context summary, repeatedly performing collision statistics, secondary sorting criteria construction, and reordering until the sibling node collision probability estimate is no higher than the collision threshold. If the expansion reaches the upper limit of context deepening, the generation of deeper-level assembly relationships and stable feature information must be stopped. Through this method, even in scenarios with repetitive subassemblies or symmetrical structures, the servo motor planetary reducer module assembly can still obtain a stable and reproducible sequence of parent node final child nodes, thus providing consistent structural input for subsequent feature encoding and reducing the impact of inconsistent structural expressions under different import and parsing conditions on retrieval accuracy.

[0039] It should be explained that the context deepening upper limit refers to the maximum extension depth threshold of context refinement pre-set by relevant technical personnel. It is used to limit the upper limit of the number of levels of substructure information introduced downwards during the collision disambiguation process. Its meaning is that when refining the context of a candidate child node of a parent node, the system allows the system to trace down from that child node and include lower-level assembly relationships and stable feature information not exceeding the upper limit level to generate a context summary. If the level is exceeded, the deepening will stop. This is to avoid introducing too much lower-level information in highly repetitive or deep assembly structure scenarios, which would lead to excessive computational overhead or summary expansion. This maintains a controllable balance between distinguishability and processing cost.

[0040] In this embodiment, after normalized sorting and context refinement, each parent node of the servo motor planetary reducer module assembly yields a reproducible final child node sequence. Since the assembly commonly contains repetitive parts such as fasteners, gaskets, and clips, directly expanding each instance for feature encoding would amplify minor adjustments like adding one or two bolts or fine-tuning the number of clips into structural encoding differences, thus reducing the stability of cross-version retrieval and structural similarity comparison. Therefore, this embodiment further performs similar child node aggregation and quantity robustness processing on the final child node sequence.

[0041] Figure 3 This is a flowchart of the aggregation and robust quantity processing of similar child nodes provided in this embodiment of the invention. Based on the final child node sequence of each parent node, child nodes with completely consistent sorting criteria and secondary sorting criteria within the sequence are categorized and aggregated to form similar child node groups, and the actual occurrence frequency of each group of child nodes is accurately counted. Then, the occurrence frequency is compared with a preset quantity retention threshold. If the occurrence frequency of the child node group is not higher than the preset quantity retention threshold, the actual occurrence frequency is directly retained as the quantity expression, and this value is directly used for subsequent model feature encoding. If the occurrence frequency of the child node group is higher than the preset quantity retention threshold, logarithmic compression processing is performed on the occurrence frequency to generate a compressed quantity expression. This reduces the invalid interference to model feature encoding caused by non-core changes such as fastener quantity fine-tuning and part layout density fine-tuning, achieving robust optimization of the quantity dimension and completing the final update of the child node sequence.

[0042] Taking the parent node of the fastener package in the servo motor planetary reducer module assembly as an example, its child nodes include M6×20 socket head cap screws, M5×16 socket head cap screws, M6 flat washers, M6 spring washers, etc., and each type of child node may correspond to multiple instances in the assembly. In the final child node sequence of the fastener package, the system, based on the comparison formed during the aforementioned sorting process, groups child nodes with consistent sorting criteria and, when necessary, consistent secondary sorting criteria into similar child node groups. Similar child node groups can be understood as collections of parts of the same type at the structural and geometric level; for example, M6×20 socket head cap screws are grouped into one group, and M6 flat washers into another. Subsequently, the occurrence frequency of each group is counted. The occurrence frequency reflects the actual number of instances of the parts in that group used in the assembly, and this occurrence frequency is used as the original value for quantity expression.

[0043] To suppress the interference of quantity fine-tuning on feature encoding, the system compares the occurrence frequency with a quantity retention threshold pre-set by relevant technicians. The quantity retention threshold is used to distinguish between scenarios with a small number of repetitions and a large number of repetitions: when the occurrence frequency is not higher than the quantity retention threshold, it is considered that the quantity change has a certain indicative significance to the structural essence, so the occurrence frequency is directly retained as the quantity expression and participates in subsequent feature encoding; when the occurrence frequency is higher than the quantity retention threshold, it is considered that the group of parts usually belongs to high-frequency repetitive parts, and the quantity is more likely to fluctuate during engineering adjustments. At this time, the occurrence frequency is compressed to map the larger quantity to a smoother quantity expression before participating in feature encoding.

[0044] Among them, quantity compression expression refers to the method of not directly using the occurrence count as the original integer form for feature encoding when the occurrence count of the same type of child node group is large. Instead, according to the compression rules set in advance by relevant technical personnel, the occurrence count is mapped to a quantity representation with a smoother change. It can be mapped to an interval expression. For example, the occurrence count of M6×20 internal hex bolt may change from 8 to the interval [5, 10].

[0045] It needs to be explained that feature encoding refers to using the updated final child node sequence as standardized input to the assembly structure, and comprehensively utilizing the sorting criteria carried by the sequence for encoding processing. The encoding includes at least the hierarchical position and parent-child topological relationship of each child node in the sequence, node type identifier (assembly / sub-assembly / part), part specification semantics (such as bearing model, thread specification, etc.), quantity expression (including quantity code with quantity retention or compression), and aggregation tags and order consistency constraints for repeated sub-assemblies. During encoding, the system first normalizes the above information into stable node tags and relationship tags according to preset rules, and then combines, hashes, and vectorizes the node tags and relationship tags: on the one hand, the key nodes and key relationships of the overall structure are compressed and summarized to form a structural fingerprint, which is used to provide a fast index key; on the other hand, the co-occurrence relationship, hierarchical adjacency relationship, and quantity / specification attributes of the node tags in the sequence are mapped into measurable numerical features to form a structural vector, which is used for subsequent distance or similarity calculations. This allows the generated structural fingerprint and structural vector to reflect the essence of the assembly structure as much as possible and reduce the impact of imported difference noise on the retrieval results.

[0046] It's important to clarify that in this scheme, the purpose of collision detection and context refinement is not to eliminate all identical cases, but rather to prevent instability in the arrangement of sibling nodes due to differences in import or display order, thus making the structural input unreproducible under different parsing conditions. After context refinement, distinguishable structural differences are further amplified into comparable information to improve the determinism of the sorting. However, for duplicate parts or isomorphic subassemblies that objectively exist in the assembly, their type, geometric outline, and even the underlying structure itself are indeed consistent. Even with the introduction of context information, they may still appear to have consistent sorting criteria or secondary sorting criteria. This is a genuine characteristic of the assembly structure, not an anomaly. Meanwhile, the collision threshold measures whether the proportion of identical nodes under the same parent node reaches a level that affects the stability of the sorting. A threshold below the threshold only indicates that the overall sorting is reproducible, not that a small number of identical nodes do not exist. Based on this, sub-nodes with consistent sorting criteria and consistent secondary sorting criteria when needed are grouped into similar sub-node groups and their occurrence counts are counted. This is a reasonable merging expression of duplicates. Replacing the sorting of instances one by one with quantitative expression can avoid introducing unstable fields to distinguish the same instances, thereby suppressing the interference of non-essential changes such as fine-tuning of the number of fasteners on feature encoding and retrieval results.

[0047] Based on the updated final child node sequence, the servo motor planetary reducer module assembly undergoes feature encoding processing, generating two types of structural feature results for database entry and retrieval. The structural fingerprint serves as a fast index key for structural retrieval, representing a compressed summary identifier of the overall assembly structure. The structural vector serves as a feature representation for similarity calculation, representing a numerical structural feature expression capable of participating in distance or similarity calculations. Both the structural fingerprint and structural vector are derived from the updated final child node sequence to reflect the essence of the assembly structure as much as possible while minimizing the impact of imported difference noise.

[0048] Figure 4 This is a flowchart of model feature encoding verification and retrieval management provided in this embodiment of the invention. Based on the updated final child node sequence, corresponding structural fingerprints and structural vectors are generated for the assembly structure tree. Then, the structural fingerprint of the previous version of the assembly structure tree is retrieved, and the similarity and geometric change magnitude of the structural fingerprints of the current version and the previous version are calculated respectively. Next, a dual judgment is performed: if the structural fingerprint similarity is not lower than a preset similarity threshold, or the geometric change magnitude is higher than a preset geometric change threshold, it is judged as compliant, and the currently generated structural fingerprint is directly confirmed as legal and valid, and can be used for model entry into the database and subsequent retrieval; if the structural fingerprint similarity is lower than the preset similarity threshold, and the geometric change magnitude is not higher than the preset geometric change threshold, it is judged as encoding allergy, and parameter loop adjustment is triggered, and the structural fingerprint and structural vector are regenerated. After adjustment, it is judged whether the preset maximum number of loops has been reached. If the maximum number of loops has been reached, the current parameters are fixed and an anomaly marker is recorded before encoding into the database is completed; if the maximum number of loops has not been reached, the calculation of structural fingerprint similarity and geometric change magnitude is returned until compliance is determined. Once the coding compliance is completed, the 3D design engine obtains the query command, generates the corresponding structural fingerprint and vector based on the query command, completes the screening of candidate models, and finally outputs a set of compliant model recommendations, thus completing the entire process of retrieval management.

[0049] Taking a servo motor planetary reducer module assembly as an example, after completing normalized sorting, context refinement, and aggregation and quantity representation of similar child nodes, a structural fingerprint and structural vector can be generated based on the updated final child node sequence. The structural fingerprint is used to summarize the assembly structure; it can be input with a normalized structural sequence summary to perform a fixed-summary operation to obtain a short identifier, such as a 64-bit hexadecimal structural fingerprint (8F73A19264A1632B), used for database indexing and rapid retrieval. The structural vector is used for similarity calculation; it can map statistical features such as the size, hierarchy, and distribution of duplicate parts of the assembly structure into a fixed-length numerical vector, for example, constructing a 12-dimensional structural vector. V = [16, 41, 4, 5, 27, 6, 2, 6, 1, 4, 3, 2], where the first dimension is the number of assembly nodes (16), the second dimension is the number of part nodes (41), the third dimension is the maximum level depth (4), the fourth dimension is the number of top-level sub-assemblies (5), the fifth to eighth dimensions are the number of similar sub-node groups with quantities of 1, 2 to 3, 4 to 7, 8 and above (27, 6, 2, 6 respectively), the ninth dimension is the number of repeating sub-assembly groups (1), and the tenth to twelfth dimensions are the number of similar groups of bearings, gears, and sensors (4, 3, 2 respectively). This allows the same assembly to obtain reproducible structural summary identifiers and computable numerical structural representations under different import and parsing conditions, facilitating subsequent retrieval and sorting.

[0050] When a servo motor planetary gear reducer module assembly has version evolution records in the database, a version chain is formed, meaning that multiple versions of the same assembly entered into the database at different times are associated chronologically. In this embodiment, the structural fingerprint of the previous version of the assembly is further obtained, and the structural fingerprint similarity between the current version and the previous version is calculated. The structural fingerprint similarity can be obtained using a cosine similarity algorithm to measure the degree of structural closeness between the two versions. Simultaneously, the system calculates the geometric change magnitude between the current and previous versions to measure the degree of change in external dimensions or major geometric orders.

[0051] The geometric change magnitude is used to quantify the degree of difference between the current version and the previous version at the geometric level. It is obtained by extracting the geometric summary and calculating the difference between the two versions of the assembly: First, the overall bounding box size parameters of the two versions are calculated under a unified coordinate system and unit system, including length, width and height, and the bounding box volume is obtained accordingly; At the same time, the geometric magnitude indicators of the assembly or key part set are calculated, such as the volume item in the total surface area, total volume or mass attribute, and the above geometric indicators are normalized; Then, the relative change rate of each geometric indicator is calculated, and the change rates of multiple indicators are summed to obtain the geometric change magnitude. The relative change rate of a single item can be obtained by dividing the absolute value of the difference between the two versions by the corresponding indicator of the previous version.

[0052] When the structural fingerprint similarity is not lower than the preset structural fingerprint similarity threshold in the database, or the geometric change magnitude is higher than the preset geometric change threshold in the database, it indicates that the two versions of the structure have good consistency or that the assembly has indeed undergone significant geometric changes. In this case, the structural fingerprint generated by the current version can be used for database entry and retrieval to ensure that the real structural changes can be effectively reflected. The structural fingerprint similarity threshold is used to characterize the minimum similarity requirement that adjacent version assemblies should still be judged as the same or highly similar structures at the structural level, and is used to distinguish the boundary between structural consistency and significant structural changes. The geometric change threshold is used to characterize the minimum change magnitude requirement that adjacent version assemblies are judged to have undergone significant geometric changes at the geometric magnitude or external dimensions level, and is used to distinguish the boundary between basic geometric stability and significant geometric changes.

[0053] When the structural similarity is lower than the structural similarity threshold and the geometric change does not exceed the geometric change threshold, it indicates that the geometry of the assembly has not changed significantly, but the structural fingerprint has changed significantly. In this embodiment, this situation is regarded as an encoding anomaly, which means that the feature encoding is too sensitive to non-essential changes, such as structural summary drift caused by insufficient differentiation of sibling nodes under symmetrical structure, insufficient context refinement depth, or changes in the number of duplicate parts.

[0054] To address this encoding anomaly, this embodiment triggers parameter loop adjustment. Parameter loop adjustment refers to adaptively updating and re-encoding the key parameters upon which the generated structural fingerprint and structural vector depend, so that the structural representation returns to stability. Specifically, the system relaxes the context deepening upper limit, extracts the context deepening upper limit increment from the database, and adds the context deepening upper limit increment to the current relaxed context deepening upper limit. This allows for the introduction of deeper substructure information to generate context summaries when indistinguishable sets of child nodes occur, thereby improving discriminative ability and reducing the risk of ranking dependence on random factors. Subsequently, based on the updated context refinement strategy, the structural fingerprint and structural vector of the servo motor planetary reduction module assembly are regenerated, and the structural similarity and geometric change magnitude are recalculated to verify whether the stability has improved.

[0055] To avoid infinite loops, this embodiment sets a maximum number of loops. If the structural similarity still cannot reach the structural similarity threshold after reaching the maximum number of loops, and the geometric change amplitude still does not exceed the geometric change threshold, then the current parameter value is fixed and an anomaly flag is recorded. The anomaly flag is used to indicate that there is an uncertainty risk in the structural coding of this version, facilitating subsequent review and offline correction. For example, further analysis can be performed manually or offline batch processing to determine whether more granular distinction rules are needed for repeating parts and symmetrical structures at specific levels in this assembly. Through the above-mentioned stability self-check and parameter loop adjustment, the servo motor planetary reducer module assembly can suppress structural coding drift caused by non-essential factors under the condition of version chain existence, improving the reliability of cross-version retrieval and structural similarity comparison.

[0056] The maximum number of loop iterations refers to the upper limit of the parameter loop adjustment iterations preset by relevant technical personnel. It is used to limit the maximum number of attempts to repeatedly adjust the structural coding parameters and regenerate the structural fingerprint and structural vector after the stability self-check is triggered. Its meaning is that when the structural similarity is detected to be lower than the structural fingerprint similarity threshold and the geometric change amplitude does not exceed the geometric change threshold, the system allows multiple parameter adjustments and recodings within the range not exceeding the upper limit in order to eliminate coding anomalies caused by non-essential factors. If the stability judgment condition is not met after reaching this number of iterations, the iteration is stopped, the current parameters are fixed, and the anomaly identifier is recorded to avoid infinite loops and ensure that the data entry process is controllable.

[0057] Several assembly models are pre-stored in the database, and for each model, a structural fingerprint, structural vector, asset identifier, and version number are stored. The asset identifier represents the unified identity of the same assembly in the database, and the version number distinguishes different iterations under the same asset identifier. Example stored records include: Asset identifier A1001, version number 1, corresponding to structural fingerprint 8F73A19264A1632B, structural vector [16, 41, 4, 5, 27, 6, 2, 6, 1, 4, 3, 2], and bounding box size [220, 160, 140]; Asset identifier A1001, version number 2 ... and asset identifier A1001, version number 2, corresponding to structural fingerprint 8F73A19264A1632B, structural vector [16, 41, 4, 5, 27, 6, 2, 6, 1, 4, 3, 2], and bounding box size [220, 160, 140]. [27, 6, 2, 6, 1, 4, 3, 2], bounding box dimensions [221, 160, 140]; the structural fingerprint corresponding to version number 3 of asset identifier A1001 is 9A73A19264A1632B; the structural fingerprint corresponding to version number 1 of asset identifier B2007 is 5C11D0E14B2F9A80; the structural fingerprint corresponding to version number 1 of asset identifier C3012 is 33AB10C9F0D8EE71.

[0058] When the 3D design engine receives a query command, it generates a structural fingerprint and a structural vector for the query command, using the servo motor planetary reducer module assembly as the query object. For example, the structural fingerprint of the query command is 8F73A19264A1632B, and the structural vector is [16, 41, 4, 5, 27, 6, 2, 6, 1, 4, 3, 2]. The bounding box size is [220, 160, 140]. The system sequentially performs consistent matching on the database using the structural fingerprint of the query command. Consistent matching refers to comparing the structural fingerprint of the query command with the structural fingerprints of each assembly model in the database bit by bit, and using Hamming distance to characterize the degree of difference between the two fingerprints. The Hamming distance is the number of differing bits after a bitwise XOR operation between two 64-bit structural fingerprints. The structural fingerprint similarity is defined as 1 − (Hamming distance / 64), representing the consistency ratio of the two fingerprints across 64 bits. A candidate model recommendation set is formed from several assembly models whose structural fingerprint similarity is greater than the preset minimum allowable value for structural fingerprint similarity in the database.

[0059] On the candidate model recommendation set, the system further performs similarity filtering based on structural vectors. The specific filtering process is as follows: the structural vector of the query command is analyzed sequentially with the structural vector of several assembly models in the candidate model recommendation set through similarity algorithms (such as cosine similarity algorithm). Several assembly models whose structural vector similarity is not greater than the maximum allowed structural vector similarity in the database are removed from the candidate model recommendation set, thus forming the model recommendation set. Several assembly models in the model recommendation set are marked as recommended assembly models.

[0060] Subsequently, the system calculates a fusion score for each recommended assembly model in the model recommendation set. The fusion score is used to uniformly measure the matching degree between the structural and geometric levels, which refers to the weighted aggregation of structural fingerprint similarity and structural vector similarity. In this example, weights of 0.7 and 0.3 are used for weighted aggregation. For version A1001, with a structural fingerprint similarity of 1.000 and a structural vector similarity of 1.000, the fusion score is 0.7 × 1.000 + 0.3 × 1.000 = 1. For version A1001, with a structural fingerprint similarity of 0.999 and a structural vector similarity of 0.996, the fusion score is 0.7 × 0.999 + 0.3 × 0.996 = 0.9981.

[0061] When managing searches based on fusion scores, the system uses asset identifiers for grouping and deduplication. Multiple versions belonging to the same asset identifier A1001 are grouped together, and only version number 1 with the highest fusion score is retained as the representative version to avoid multiple versions of the same asset occupying the same space. If the system's preset output number of entries is 3, and the number of entries after deduplication is insufficient, a supplementary update is performed. That is, entries are added from the candidate pool according to the fusion score from high to low until the preset size is restored. In this example, during the supplementary update, version number 1 of asset identifier B2007 with a fusion score of 0.948 and version number 1 of asset identifier C3012 with a fusion score of 0.915 are introduced. Thus, the final output model recommendation set has 3 entries: version number 1 of A1001 with a fusion score of 1.000, version number 1 of B2007 with a fusion score of 0.948, and version number 1 of C3012 with a fusion score of 0.915.

[0062] Secondly, it provides a model feature encoding and retrieval management system for 3D design engines. Figure 5 This is a structural diagram of a model feature encoding and retrieval management system for 3D design engines provided in an embodiment of the present invention, including a field separation module, a sequence integration module, a robust processing module, a feature encoding module, a set output module, and a database.

[0063] The field separation module is connected to the sequence integration module, the sequence integration module is connected to the robust processing module, the robust processing module is connected to the feature encoding module, the feature encoding module is connected to the set output module, and the field separation module, sequence integration module, robust processing module, feature encoding module, and set output module are all connected to the database.

[0064] The database is used to store the parameters involved in the model feature encoding and retrieval management system for 3D design engines. The parameters are regularly updated and calibrated by technical personnel based on actual business needs and system operation status. The parameter data stored in the database is authentic, valid, accurate and complete, providing reliable data support for the stable operation of the entire process of feature encoding, collision detection, similarity screening and retrieval management of the system.

[0065] The field separation module is used to parse the assembly model to obtain the assembly structure tree, and to separate the auxiliary field information of each node in the assembly structure tree, retaining the preset stable fields and removing the preset unstable fields to form a repeatable structure input.

[0066] The sequence integration module is used to generate stable node features based on stable fields. Based on the stable node features, the set of child nodes corresponding to each parent node in the assembly structure tree is sorted. When child nodes with the same sorting criteria appear in the set of child nodes, collision detection and context refinement are performed. Finally, the sequence of child nodes corresponding to each parent node is integrated to obtain the final sequence of child nodes.

[0067] The robust processing module is used to aggregate similar child nodes in the final child node sequence and generate a quantitative representation. Based on the quantitative representation, the final child node sequence is robustly processed to update the final child node sequence corresponding to each parent node, so as to suppress the interference of quantitative fine-tuning on feature encoding.

[0068] The feature encoding module is used to perform feature encoding based on the updated sequence of final child nodes corresponding to each parent node, and performs stability self-check when the version chain exists. When an encoding anomaly is detected, parameter loop adjustment is triggered to stabilize the feature encoding. Feature encoding refers to generating structural fingerprints and structural vectors.

[0069] The set output module is used by the 3D design engine to generate the structural fingerprint and structural vector corresponding to the query command when the query command is received. Based on the structural fingerprint corresponding to the query command, the module recalls a set of candidate model recommendations, then filters the set of candidate model recommendations based on the structural vector corresponding to the query command, marks it as a set of model recommendations, and outputs the set of model recommendations.

[0070] Example 2: Under the condition that other conditions remain unchanged in Example 1, Example 2 further introduces a quantity compression expression mechanism for high-frequency repetitive parts such as fasteners, in order to reduce the impact of fine-tuning of the number of fasteners or fine-tuning of their arrangement density on the feature encoding results.

[0071] Specifically, after completing the tree-structured parsing, the system counts the number of times each part node appears in the parent assembly (i.e., n in quantity × n) and sets a quantity retention threshold N_keep, for example, 4. When the number of occurrences of a part node does not exceed this threshold, the encoding stage retains the accurate quantity representation; when the number of occurrences exceeds this threshold, quantity compression is triggered to avoid unnecessary drastic changes in the structural fingerprint or structural vector due to small fluctuations in the number of fasteners.

[0072] Specifically, for part nodes that appear more than N_keep, the frequency of occurrence is further mapped to a quantity level code. The specific mapping rules can be found in the frequency-quantity level code mapping table stored in the database. For example, when n is 10, it is defined as "medium quantity" and denoted as Q_M@10 (to avoid confusion with M in thread specifications, Q_M here represents the quantity level code Quantity-Medium). Thus, the compressed quantity expression retains the quantity level information and can uniformly absorb fine-tuning such as changing 8 to 9 or 9 to 10 into the same anchor point representation, thereby improving coding stability.

[0073] During feature encoding, the system uses the compressed quantity expression as one of the node encoding fields, which participates in the encoding calculation along with part type, specification parameters, and parent assembly identifier. For example, for the M6×20×8 socket head cap screws in fastener package A5, which appear 8 times and exceed the threshold of 4, the system maps its quantity and generates a quantity code Q_M@10; for the M5×16×10 socket head cap screws, their quantity is also mapped to Q_M@10. This ensures that when the quantity of the above fasteners is slightly adjusted from 8 to 9 due to the adjustment of the arrangement density, the quantity code remains unchanged, thereby reducing the sensitivity of structural feature encoding to such fine-tuning and improving the robustness and comparability of structural fingerprints / structural vectors during version iteration.

[0074] The above-disclosed embodiments are merely some examples of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A model feature encoding and retrieval management method for 3D design engines, characterized in that, The method includes: Step 1: Parse the assembly model to obtain the assembly structure tree, and separate the auxiliary field information of each node in the assembly structure tree. Retain the preset stable fields and remove the preset unstable fields to form a repeatable structure input. Step 2: Generate stable node features based on stable fields. Based on the stable node features, sort the child node sets corresponding to each parent node in the assembly structure tree. When child nodes with the same sorting criteria appear in the child node set, perform collision detection and context refinement. Finally, integrate to obtain the final child node sequence corresponding to each parent node. Step 3: Aggregate similar child nodes in the final child node sequence and generate a quantitative representation. Then, perform robust processing on the final child node sequence based on the quantitative representation to update the final child node sequence corresponding to each parent node, so as to suppress the interference of quantitative fine-tuning on feature encoding. Step 4: Perform feature encoding based on the updated sequence of final child nodes corresponding to each parent node, and perform a stability self-check when a version chain exists. When an encoding anomaly is detected, trigger parameter loop adjustment to stabilize the feature encoding. The feature encoding refers to generating structural fingerprints and structural vectors. Step 5: When the 3D design engine receives a query command, it generates a structural fingerprint and structural vector corresponding to the query command. Based on the structural fingerprint corresponding to the query command, it recalls a set of candidate model recommendations. Then, based on the structural vector corresponding to the query command, it filters the set of candidate model recommendations, marks them as a set of model recommendations, and outputs the set of model recommendations.

2. The model feature encoding and retrieval management method for a 3D design engine as described in claim 1, characterized in that, The process of separating the auxiliary field information of each node in the assembly structure tree is as follows: Obtain the assembly model to be imported into the warehouse, parse it to obtain the assembly structure tree, and clarify the parent-child relationship of each node in the assembly structure tree; Separate stable and unstable fields from the auxiliary field information of each node in the assembly structure tree; Remove unstable fields, keep only stable fields, and establish a relationship table for the set of child nodes corresponding to the parent node. Update the auxiliary field information of each node in the assembly structure tree. The stable field refers to field information that remains consistent or can be repeatedly calculated across different import formats, different parsers, or different import paths; The unstable fields refer to field information that may drift due to changes in the import process, file rewriting, parsing implementation, or display strategy.

3. The model feature encoding and retrieval management method for a 3D design engine as described in claim 1, characterized in that, The specific sorting process for the set of child nodes corresponding to each parent node in the assembly structure tree is as follows: Based on the stable fields of each node in the assembly structure tree, stable node features for encoding are generated for each node. Read the set of child nodes corresponding to each parent node in the assembly structure tree, and construct a sorting criterion for each child node in the set of child nodes based on the characteristics of stable nodes; All child nodes under the same parent node are sorted lexicographically according to the sorting criteria to obtain the sequence of child nodes of that parent node; Determine if there are child nodes in the child node sequence that are sorted by the same criteria; When there are no child nodes with the same sorting criteria, the child node sequence is determined as the final child node sequence of the parent node and used for feature encoding; When there are child nodes with the same sorting criteria, the original import order or display order is not used as the basis for handling a tie. Instead, the parent node and its corresponding child nodes with the same sorting criteria are marked as a collision candidate set, and collision detection and context refinement are performed.

4. The model feature encoding and retrieval management method for a 3D design engine as described in claim 3, characterized in that, The collision detection and context refinement process is as follows: The statistical sorting is based on the proportion of identical child nodes to the total number of child nodes of the corresponding parent node, and is marked as the estimated collision probability value of sibling nodes. The estimated collision probability of sibling nodes is compared with a preset collision threshold. When the estimated collision probability of a sibling node is not higher than the collision threshold, the child node sequence of the parent node is confirmed as the final child node sequence and used for feature encoding. When the estimated collision probability of a sibling node is higher than the collision threshold, context refinement is triggered. The specific process is as follows: mark the child nodes whose estimated collision probability of a sibling node is higher than the collision threshold as child nodes to be analyzed, and generate a context summary for each child node to be analyzed. Based on the context summary of each child node to be analyzed and the original sorting criteria, a corresponding secondary sorting criteria are constructed, and each child node to be analyzed is re-sorted based on the secondary sorting criteria to obtain the updated sequence of child nodes of the parent node. Recalculate the sibling node collision probability estimate for the updated child node sequence; If the estimated collision probability of the sibling node is still higher than the collision threshold, the context refinement range is expanded from direct child nodes to multiple substructures and deepened layer by layer. Collision detection and context refinement are repeated until the estimated collision probability of the sibling node is no higher than the collision threshold, thereby obtaining the final child node sequence of the parent node and using it for feature encoding.

5. The model feature encoding and retrieval management method for a 3D design engine as described in claim 1, characterized in that, The robustness processing of the final child node sequence based on quantitative representation is performed as follows: In the final child node sequence of each parent node, child nodes with the same sorting criteria and secondary sorting criteria are grouped into similar child node groups, and the occurrence frequency of each similar child node group is counted. Compare the number of occurrences with a preset quantity retention threshold; When the number of occurrences is not higher than the quantity retention threshold, the number of occurrences is retained as the quantity representation and used for subsequent feature encoding; When the occurrence count exceeds the quantity retention threshold, the occurrence count is compressed to obtain a compressed quantity expression, which is then used for feature encoding to reduce the impact of fastener quantity fine-tuning or arrangement density fine-tuning on feature encoding.

6. The model feature encoding and retrieval management method for a 3D design engine as described in claim 1, characterized in that, The feature encoding based on the updated sequence of final child nodes corresponding to each parent node is performed as follows: The structural fingerprint and structural vector of the assembly structure tree are generated based on the final child node sequence of each parent node, and the structural fingerprint of the previous version of the same assembly structure tree is obtained. Calculate the structural fingerprint similarity between the current assembly structure tree and the previous version, and simultaneously calculate the magnitude of geometric changes; When the structural fingerprint similarity is not lower than the preset structural fingerprint similarity threshold, or the geometric change amplitude is higher than the preset geometric change threshold, the structural fingerprint of the current assembly structure tree is confirmed to be used for database entry and retrieval. When the structural fingerprint similarity is lower than the preset structural fingerprint similarity threshold and the geometric change amplitude is not higher than the preset geometric change threshold, it is determined that the current encoding is too sensitive to non-essential changes, and parameter loop adjustment is triggered.

7. The model feature encoding and retrieval management method for a 3D design engine as described in claim 6, characterized in that, The adjustment process for the trigger parameter loop is as follows: Relax the context to deepen the upper bound, and regenerate the structural fingerprint and structural vector of the assembly structure tree; Set a maximum number of loops. If the structural fingerprint similarity is not lower than the preset structural fingerprint similarity threshold after reaching the maximum number of loops, or if the geometric change amplitude is higher than the preset geometric change threshold, fix the current parameters and record the anomaly identifier for subsequent review and offline correction.

8. The model feature encoding and retrieval management method for a 3D design engine as described in claim 1, characterized in that, The specific recall process for the candidate model recommendation set based on the structural fingerprint corresponding to the query command is as follows: Obtain the query command and generate the structural fingerprint and structural vector of the query command; Consistent matching is performed based on the structural fingerprint of the query command to recall the candidate model recommendation set; Based on the structural vector corresponding to the query command, the candidate model recommendation set is filtered by similarity and marked as the model recommendation set; Calculate the fusion score for each recommended assembly model in the model recommendation set; Search management is based on fusion scores.

9. The model feature encoding and retrieval management method for a 3D design engine as described in claim 8, characterized in that, The retrieval management based on fusion scores is specifically managed as follows: Group multiple recommended assembly models with the same identifier in the model recommendation set into one group, and retain only the recommended assembly model with the highest fusion score in the group; At the same time, the model recommendation set is supplemented and updated to ensure that the number of recommended items in the updated model recommendation set is consistent with that before the update.

10. A model feature encoding and retrieval management system for 3D design engines, characterized in that, include: The field separation module is used to parse the assembly model to obtain the assembly structure tree, and to separate the auxiliary field information of each node in the assembly structure tree, retaining the preset stable fields and removing the preset unstable fields to form a repeatable structure input. The sequence integration module is used to generate stable node features based on stable fields. Based on the stable node features, the set of child nodes corresponding to each parent node in the assembly structure tree is sorted. When child nodes with the same sorting criteria appear in the set of child nodes, collision detection and context refinement are performed. Finally, the sequence of child nodes corresponding to each parent node is integrated. The robust processing module is used to aggregate similar child nodes in the final child node sequence and generate a quantitative representation, and to perform robust processing on the final child node sequence based on the quantitative representation, thereby updating the final child node sequence corresponding to each parent node to suppress the interference of quantitative fine-tuning on feature encoding. The feature encoding module is used to perform feature encoding based on the updated sequence of final child nodes corresponding to each parent node, and to perform stability self-check when the version chain exists. When an encoding anomaly is detected, parameter loop adjustment is triggered to stabilize the feature encoding. The feature encoding refers to generating structural fingerprints and structural vectors. The set output module is used by the 3D design engine to generate the structural fingerprint and structural vector corresponding to the query command when the query command is received. Based on the structural fingerprint corresponding to the query command, the module recalls a set of candidate model recommendations, then filters the set of candidate model recommendations based on the structural vector corresponding to the query command, marks them as the model recommendation set, and outputs the model recommendation set.

Images