A manufacturing process feature intelligent identification method and system based on 3D space reasoning

Abstract

The application discloses a kind of manufacturing process feature intelligent identification method and system based on 3D space reasoning, belong to intelligent manufacturing technical field, method includes: S1. associated world model pre-training stage: from mass {B-Rep, PMI} pairing data, the stable correlation law between PMI semantics and B-Rep geometric structure is learned unsupervisedly;S2. space reasoning and feature identification stage based on memory: the knowledge assets of step S1 are used to identify new CAD model containing only B-Rep geometric information, and complete manufacturing process feature is obtained;S3. structured feature output stage: based on the manufacturing process feature, the structured feature semantic tree that can directly drive downstream CAPP / CAM system is output.The application solves the problem that geometry and semantics are separated in the prior art, lack reasoning ability and have strong data dependency, realizes high-precision, high-robustness, strong generalization ability and feature recognition rich in engineering semantics, and can effectively drive automatic process planning and intelligent numerical control programming.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and computer-aided design and manufacturing (CAD / CAM), and in particular to an intelligent identification method and system for manufacturing process features based on 3D spatial reasoning. Background Technology

[0002] In digital manufacturing processes, the conversion from design models (CAD) to manufacturing instructions (CAM / G code) is a core bottleneck. Among these, the identification of manufacturing process features (such as holes, slots, cavities, chamfers, etc.) is a crucial step, directly determining the level of automation in Computer-Aided Process Planning (CAPP) and CNC programming. Currently, the industry's demand for manufacturing process feature identification is increasingly urgent, mainly reflected in: 1. Automated process planning requirements: Traditional process planning relies on process engineers manually interpreting drawings, which is inefficient and inconsistent.

[0003] 2. Requirements for intelligent CNC programming: Modern intelligent CNC programming systems require structured feature information to achieve automated toolpath generation.

[0004] 3. AI Factory Integration Needs: Feature information is key data that connects the entire process of design, process, manufacturing, and testing.

[0005] Currently, mainstream automated manufacturing process feature recognition methods in industry can be divided into three categories: rule-based methods, graph matching-based methods, and machine learning-based methods. Rule-based methods rely on expert-defined rule bases, which are effective for simple process features but struggle with complex, intersecting features, and rule maintenance is costly. Graph matching-based methods represent the Boundary Representation Structure (B-Rep) model as an Attribute Adjacency Graph (AAG), matching predefined feature templates through subgraph isomorphism. However, they suffer from high computational complexity and are sensitive to noise and model changes. In recent years, Graph Neural Network (GNN)-based methods have shown potential, using the surfaces of the B-Rep model as nodes and learning node features for feature classification. However, they still have the following significant drawbacks and limitations: 1. Geometric and semantic separation: This includes GNN-based methods that treat Product Manufacturing Information (PMI) as an additional label, failing to model the deep semantic association and spatial constraint relationship between PMI and B-Rep geometric entities.

[0006] 2. Lack of causal reasoning: Existing models are mostly end-to-end black-box classifiers, whose decision-making processes are difficult to interpret. They cannot simulate the cognitive process by which process engineers reason based on the causal relationship between design intent (such as the functional requirements implied by PMI) and geometric implementation, resulting in a significant drop in recognition accuracy and robustness when faced with complex interactive features such as intersecting holes and nested slots.

[0007] 3. Strong Data Dependence: Supervised learning-based models heavily rely on a large amount of precisely labeled {B-Rep, PMI} pairing data for training. In industrial practice, such data is scarce, and obtaining such high-quality labeled data is extremely costly and difficult. This limits the model's generalization ability, making it difficult to adapt to new parts or unseen feature combinations of different industries and complexities.

[0008] In summary, existing technologies cannot achieve intelligent manufacturing process feature recognition with engineering semantic understanding, strong generalization ability, high reliability, and easy engineering implementation while maintaining geometric accuracy. Summary of the Invention

[0009] The purpose of this invention is to overcome the problems existing in the prior art and provide a method and system for intelligent recognition of manufacturing process features based on 3D spatial reasoning, proposing the "Associative Embedding Spatial Reasoning" (AESR) framework. Its core innovation lies in adopting a two-stage paradigm: first, a world model of the association between "PMI semantic-geometric space" is constructed through pre-training; then, this model is used as a prior knowledge base to guide a novel B-Rep-PMI-GAT network for interpretable spatial reasoning and feature recognition.

[0010] The objective of this invention is achieved through the following technical solution: A first aspect of the present invention provides a method for intelligent recognition of manufacturing process features based on 3D spatial reasoning, comprising the following steps: S1. Pre-training stage of the associated world model: This stage is the offline knowledge learning stage. The goal is to learn, in unsupervised, stable association rules between PMI semantics and B-Rep geometric structure from massive {B-Rep, PMI} paired data. B-Rep represents the boundary representation structure, and PMI represents product manufacturing information. S2. Memory-based spatial reasoning and feature recognition stage: Using the knowledge assets constructed in step S1, the new CAD model containing only B-Rep geometric information is identified to obtain complete manufacturing process features; S3. Structured Feature Output Stage: Based on the manufacturing process features, output a structured feature semantic tree that can directly drive the downstream CAPP / CAM system.

[0011] In some embodiments, step S1 specifically includes: S11. Input a CAD model file in ISO 10303 STEP AP242 format; S12. Extract the geometric structure code and PMI semantic code of the model by means of a dual-stream associative encoder; the dual-stream associative encoder includes a geometric encoder and a PMI encoder, wherein the geometric encoder contains a spatial partition head, which forces it to predict the potential spatial layout that can reflect the macroscopic functional area of ​​the part. S13. By training the dual-stream association encoder through the contrastive learning objective loss function, the mutual information between the geometric structure code and the PMI code from the same part is maximized, capturing the most essential association between the geometric structure code and the PMI code, and obtaining the joint association embedding vector. S14. After training, the joint association embedding vectors of all training data are summarized into K prototype vectors through a clustering algorithm to construct a differentiable association memory M, where each prototype vector represents a "design pattern prototype"; at the same time, a PMI semantic decoder is trained for subsequent prediction of PMI information from geometric features; the pre-trained two-stream association encoder parameters and PMI semantic decoder are saved.

[0012] In some embodiments, step S1 further includes preprocessing the CAD model file: B-Rep parsing: Extract faces, edges, vertices and their topological relationships, calculate the geometric descriptor of each face, and construct a facet graph; PMI parsing and graph construction: Dimensions, tolerances, datums, and annotations are parsed into structured entities to construct a PMI semantic graph.

[0013] In some embodiments, the contrastive learning objective loss function is the InfoNCE loss.

[0014] In some embodiments, step S2 specifically includes: S21. Input the B-Rep structure of the new CAD model to be recognized, and load the pre-trained geometric encoder and associated memory library M; S22. Construct a B-Rep-PMI graph attention network to represent the B-Rep structure of the new CAD model as a surface node graph; S23. In the B-Rep-PMI graph attention network, a query-memory space reasoning module is introduced as the core reasoning unit. The query-memory space reasoning module performs a query operation for each node (geometric patch) to generate semantically enhanced node geometric features. S24. Through multi-layer information transmission and aggregation of graph attention network, and combined with iterative controller (which guides the network to recall more specific memory prototypes for multiple rounds of refinement when feature conflict is detected), the geometric features of nodes are clustered and decoded into complete manufacturing process features.

[0015] In some embodiments, the query operation specifically includes: Query: Convert the geometric features of nodes into query vectors; Addressing and reading: Calculate the similarity (attention weight) between the query vector and all prototype vectors in the associated memory M, and sum them by weight to read out the most relevant semantic memory content; Fusion: The read memory content (containing PMI semantic priors) is fused with the original geometric features to generate semantically enhanced node geometric features.

[0016] In some embodiments, the implementation of the iteration controller includes: The classification confidence of the semantically enhanced node geometric features is monitored after each round of inference. When a high-confidence conflict is detected in the node attribution, an intervention mechanism is triggered to adjust the query parameters or apply symbolic constraints and guide the network to the next round of refined inference.

[0017] In some embodiments, the structured feature semantic tree includes feature type, geometric parameters, topological references, predicted PMI information, and interpretable reasoning basis (such as activated memory prototype IDs), and can be visualized and highlighted on the original CAD model.

[0018] In some embodiments, adjusting the query parameters includes: Increase the query temperature parameter for conflict nodes; The application of symbolic constraints includes: It invokes built-in rules to directly correct the probability distribution of conflicting nodes.

[0019] A second aspect of the present invention provides a manufacturing process feature intelligent recognition system based on 3D spatial reasoning, comprising: The associated world model pre-training module is used to learn, in unsupervised, stable association rules between PMI semantics and B-Rep geometric structure from massive {B-Rep, PMI} paired data, where B-Rep represents boundary representation structure and PMI represents product manufacturing information. The spatial reasoning and feature recognition module is used to identify new CAD models containing only B-Rep geometric information using knowledge assets built by the associated world model pre-training module, and obtain complete manufacturing process features; The structured feature output module is used to output a structured feature semantic tree that can directly drive the downstream CAPP / CAM system based on the manufacturing process features.

[0020] It should be further noted that the technical features corresponding to the above-mentioned options and embodiments can be combined or substituted with each other to form new technical solutions without conflict.

[0021] Compared with the prior art, the beneficial effects of the present invention are: 1. Originality and innovation: It proposes a brand-new paradigm of "learning the associated world model first and then performing memory-driven reasoning", which upgrades the recognition of manufacturing features from passive classification to active, experience-based analogical reasoning.

[0022] 2. Strong generalization and robustness: By separating "knowledge learning" from "knowledge application," the model exhibits stronger adaptability to feature combinations not seen during training, incomplete geometry, or missing PMIs. The iterative inference mechanism is specifically designed to handle complex feature intersections.

[0023] 3. The output is rich in semantics and interpretability: It not only outputs geometric features, but also predicts their associated PMI design intent, and provides a traceable explanation based on similar design patterns for each identification decision through "activated memory prototype ID", which greatly improves the engineering credibility.

[0024] 4. Higher data efficiency: The first stage can be pre-trained using a large number of inaccurately labeled CAD models (only requiring a weak correlation between the model's own B-Rep and PMI). The second stage only requires a small amount of finely labeled data to fine-tune the inference network to generate a hierarchical manufacturing feature tree, which includes feature types, parameters, topological references, and manufacturing semantics, and can directly drive downstream CAPP / CAM systems.

[0025] 5. A "differentiable relaxed Voronoi segmentation algorithm" is proposed. This algorithm does not output hard boundaries, but instead outputs a "probability distribution vector belonging to K functional prototype regions" for each B-Rep patch. This is not only a loss function, but also a novel algorithm module that can be embedded in neural network training.

[0026] 6. An "online memory enhancement mechanism" is proposed. During the second-stage inference, if a new feature pattern with high confidence but insufficient similarity to all prototype vectors in M ​​is encountered, the system can generate a new, temporary memory prototype and store it in a "short-term memory bank." When such patterns accumulate to a certain number, they can trigger incremental learning of the model and update the main memory bank M. This enables the system to have the evolutionary ability to "learn from practice."

[0027] 7. A "neural-symbolic hybrid inference controller" is proposed. Internally, it encapsulates a set of lightweight, domain-knowledge-driven symbolic rules (e.g., "a cylindrical surface cannot simultaneously be the side of a through hole and the bottom of a blind hole"). When conflicts arise in neural network inference, the controller not only guides "recall" but also directly invokes these symbolic rules to impose hard constraints on node labels, forcibly resolving conflicts and achieving a closed loop of "data-driven and knowledge-driven" inference.

[0028] 8. Propose "Feature Context-Based PMI Chain Decoding". For example, after identifying "datum feature A", when identifying "feature B", the "datum system" prediction module for its position tolerance will actively query the geometry in the drawing that has been identified as "datum A", thereby generating a more structurally accurate and logically consistent PMI. This simulates the sequential dependency when engineers read drawings. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating an intelligent identification method for manufacturing process features based on 3D spatial reasoning, as shown in an embodiment of the present invention. Detailed Implementation

[0030] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] It should be noted that the defects in the solutions in the prior art are all the results of the inventors' practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of this application in the following text should be the inventors' contributions to this application in the process of invention and creation, and should not be understood as technical content known to those skilled in the art.

[0032] In view of the technical problems pointed out in the background art, the present invention provides the following embodiments: In one exemplary embodiment, a method for intelligent recognition of manufacturing process features based on 3D spatial reasoning is provided, such as... Figure 1 As shown, it includes the following steps: S1. Pre-training stage of the associated world model: From massive {B-Rep, PMI} paired data, learn the stable association rules between PMI semantics and B-Rep geometric structure in an unsupervised manner, where B-Rep represents the boundary representation structure and PMI represents product manufacturing information; S2. Memory-based spatial reasoning and feature recognition stage: Using the knowledge assets constructed in step S1, the new CAD model containing only B-Rep geometric information is identified to obtain complete manufacturing process features; S3. Structured Feature Output Stage: Based on the manufacturing process features, output a structured feature semantic tree that can directly drive the downstream CAPP / CAM system.

[0033] Specifically, the principle of this method is as follows: In step S1, the concept of a "world model" is adopted: drawing on the core idea of ​​"world model" in reinforcement learning to understand and predict the environment by constructing an internal model. This invention proposes to construct a "CAD design world model", the core objective of which is to learn the causal mapping relationship between design semantics (PMI) and geometric reality (B-Rep), that is, "a certain design intention is usually realized as a certain spatial geometric composition".

[0034] In step S2, the "structured representation learning" is modified: referencing cutting-edge ideas in structured representation learning such as "Split-and-Fit" in computer vision, but shifting its goal from "reconstructing geometry" to "deconstructing design intent." Instead of requiring the model to accurately reconstruct every facet of the B-Rep, it is required to learn to predict a potential spatial partition that reflects the "sphere of influence" of different manufacturing features in space, thereby gaining a high-level understanding of the macroscopic functional layout of the part.

[0035] Furthermore, the invention incorporates a "query-memory space reasoning" mechanism: this is the core original innovation of the invention. It creatively stores the "world knowledge" learned in the first stage in a differentiable associative memory. In the second stage, the recognition task is reconstructed as a memory-based reasoning process: faced with new B-Rep geometry, the model "recalls" the most relevant design patterns in the memory through a "query" operation, and actively guides the aggregation and interpretation of geometric features using the "recalled" knowledge. This mechanism simulates the cognitive process of experts "making analogical reasoning based on experience."

[0036] The following is a specific implementation method based on the above method. The overall process of the method of the present invention is divided into two stages: offline pre-training and online inference.

[0037] Step S1 (Offline Pre-training): The goal of this stage is to learn stable association patterns between PMI and B-Rep from massive paired data in an unsupervised manner, constructing a "world model". The input is a massive amount of STEP AP242 files. PMI annotations are pre-parsed into structured semantic graphs by the PMI parser. In the dual-stream encoder, the geometric encoder (such as a patch-based GNN) extracts geometric features and predicts the latent spatial layout through a "spatial partitioning head"; the PMI encoder (such as a graph Transformer-based network) extracts semantic features. The two are aligned by contrastive learning objectives. After training, an association memory M and a PMI decoder are constructed.

[0038] S10. Data Preprocessing Input the original STEP AP242 file set.

[0039] Processing procedure: 1. B-Rep parsing: Extract faces, edges, vertices and their topological relationships, calculate the geometric descriptor (normal, curvature, area, etc.) of each face, and construct a patch graph G_brep = (V, E).

[0040] 2. PMI parsing and graph construction: Dimensions, tolerances, datums, annotations, etc. are parsed into structured entities, and a PMI semantic graph G_pmi = (P, R) is constructed, where node P is a PMI entity and edge R represents its association relationship (such as "constraint" or "reference").

[0041] Output: 1. Normalized B-Rep patch feature matrix X_geo and adjacency matrix A. 2. PMI semantic feature matrix X_pmi and PMI-geometric association matrix R_link (indicating which PMI is associated with which face IDs).

[0042] S11. Dual-stream coding Input X_geo, A; X_pmi, R_link.

[0043] Processing procedure: 1. Geometric Encoder (E_geo): A GNN. Its innovation lies in its output layer having two heads connected in parallel: - Feature Header: Outputs surface-level geometric code H_geo.

[0044] - Spatial partition header: Outputs the probability of space affiliation at the surface level, S_partition.

[0045] 2. PMI Encoder (E_pmi): A Graph Transformer that processes G_pmi and outputs the global semantic code z_pmi for PMI.

[0046] Output: 1. Part-level geometric code g_brep = READOUT(H_geo). 2. Part-level PMI semantic code s_pmi = z_pmi. 3. (Optional) Spatial partition map S_partition.

[0047] S12. Contrastive Association Learning Input a batch of N pairs of parts (g_brep^i, s_pmi^i).

[0048] Processing procedure: 1. Calculate the association embedding: e_assoc^i = [g_brep^i; s_pmi^i].

[0049] 2. Apply InfoNCE loss: Treat (g_brep^i, s_pmi^i) of the same part as positive sample pairs, and treat all other combinations in the batch as negative sample pairs, and train two encoders.

[0050] Output: The trained encoder parameters θ_geo and θ_pmi, which can map semantically related geometry and PMI to similar positions in the embedding space.

[0051] S13. Memory Construction Input e_assoc for all training samples.

[0052] Processing procedure: 1. Use a clustering algorithm (such as K-Means) to cluster e_assoc to obtain K cluster centers {c_1,c_2, ..., c_K}.

[0053] 2. Initialize each cluster center as a differentiable memory vector m_k, forming an association memory matrix M (K x D).

[0054] 3. Train a lightweight PMI decoder D_pmi, with g_brep as input and the predicted PMI semantic vector s'_pmi as output.

[0055] Output: 1. Core assets: Associative memory M. 2. Frozen pre-trained encoders E_geo and E_pmi. 3. PMI semantic decoder D_pmi.

[0056] Furthermore, during the pre-training phase, the geometric encoder E_geo not only needs to extract surface-level geometric features but also needs to understand the functional partitioning of the part in macroscopic space. Traditional hard clustering (such as K-Means) or semantic segmentation (assigning a discrete region label to each face) is non-differentiable and cannot express the ambiguity that a face may be associated with multiple functional regions simultaneously (e.g., a face may be part of a "hole" and also constitute the sidewall of a "cavity"). To address this issue, this invention proposes a differentiable relaxed Voronoi segmentation algorithm. This algorithm does not output a hard decision on which region each face belongs to, but instead outputs a "probability distribution belonging to K potential functional prototype regions" for each face. This is equivalent to performing a "soft" spatial partitioning in the feature space, allowing the entire process to be optimized end-to-end through gradient descent. The specific implementation of the algorithm is as follows: Algorithm input and initialization: Input features: Surface-level geometric feature vectors H∈R from the output of the geometric encoder backbone network. N×D , where N is the number of patches in the B-Rep model and D is the feature dimension.

[0057] Learnable parameters: Initialize K learnable "region prototype" vectors P = {p1, p2, ..., p...} K}, where p k ∈R D These K prototype vectors represent the "centers" of the potential functional regions. (Note: K here can be the same size as the final constructed memory M, or it can be an independent hyperparameter used to capture more fine-grained spatial partitioning.)

[0058] The core of this algorithm is to relax the hard boundaries of the Voronoi partition into a differentiable soft allocation process. The specific process is as follows: Step 1: Calculate membership degree (soft assignment) For each feature vector h of facet i i and each region prototype p k The "negative distance" is calculated as a similarity score, and then transformed into a probability distribution using either Gumbel-Softmax or a regular Softmax function. This simulates the "nearest neighbor" decision rule in a Voronoi diagram, but makes it differentiable. The formula is as follows: score i,k =−‖h i -p k || 2 s i,k = Among them, s i,k Let be the probability that patch i belongs to the prototype of the k-th region. 2 It is the square of the Euclidean distance, conforming to the basic principle of Voronoi diagrams based on distance metrics. τ is a temperature parameter. When τ→0, the distribution approaches a one-hot hard Voronoi partition; when τ>0, it is a relaxed, differentiable soft partition. In the early stages of training, a higher temperature τ can be used to encourage exploration, and later annealing can be gradually applied to a lower temperature to approximate the true Voronoi boundary.

[0059] Step 2: Generate a spatial partition probability map Through the above calculations, a spatial partitioning probability matrix S∈R is obtained. N×K Each row S in the matrix i This refers to the "surface-level spatial attribution probability Spartition" mentioned in step S12.

[0060] Step 3: Construct Partition Consistency Loss Simply finding the nearest prototype for each face is insufficient. To ensure these prototypes truly reflect coherent "functional regions," constraints need to be imposed such that faces belonging to the same prototype should also be close to each other in the geometric feature space. To this end, this invention designs a partition consistency loss function. This loss function aims to maximize the similarity of features among all faces within the same region, while minimizing the similarity between prototypes in different regions. The loss function is as follows: The attraction term is a weighted sum, with the weights being the soft-assignment probability s. i,k It encourages the use of the h feature of each facet. i To its most likely prototype p k The repulsion term is a hinge loss with a spacing m. It prevents different regional prototypes pk and pl from getting too close to each other, ensuring that the spatial regions they represent are separated from each other in the feature space. λ is a weighting coefficient that balances the importance of the two terms.

[0061] m is a hyperparameter that defines the minimum distance interval required between different prototypes.

[0062] Furthermore, the spatial partitioning head, as an auxiliary output branch of the geometric encoder E_geo, is jointly trained with the main feature extraction network. The final pre-trained loss function combines contrastive learning loss and partition consistency loss: L total =L InfoNCE +αL partition Where α is the equilibrium hyperparameter. In this way, the gradient not only aligns the geometry with the PMI through contrastive learning, but also through L... partition Backpropagation continuously optimizes the patch features H and region prototypes P. After training, these optimized region prototypes P can be further aggregated or directly used to initialize the subsequent association memory M, providing the memory with initial "anchor points" containing spatial semantic information.

[0063] Step S2 (Online Inference): This stage is the online inference process. A novel B-Rep-PMI-GAT network is constructed, the core of which utilizes the "world knowledge" (memory M) learned in the first stage to complete recognition through active querying and iterative inference. The new B-Rep model is input to construct a surface-node graph. The key innovation lies in the query-memory space inference module. This module works before propagation at each graph layer: it uses node features as queries, performs attention addressing in the associative memory M, reads the weighted averaged memory vector, and fuses it with the original node features, thereby enhancing the geometric representation with prior semantic knowledge. The iterative controller monitors node classification conflicts, triggering a new round of "recall-fusion" process to achieve refined inference.

[0064] S20. Inference Initialization Input: 1. New part STEP file (B-Rep only). 2. Pre-trained model (E_geo, D_pmi) and memory bank M.

[0065] Processing procedure: 1. Use a clustering algorithm (such as K-Means) to cluster e_assoc to obtain K cluster centers {c_1,c_2, ..., c_K}.

[0066] 2. Use the frozen E_geo to extract the initial surface-level geometric features H_init.

[0067] Output: Initialized graph node features H^0 = H_init.

[0068] S21. Constructing the Reasoning Network Input: Graph G_new and node features H^l (layer l).

[0069] Processing procedure: Construct an L-layer B-Rep-PMI-GAT network. Insert a query-memory space inference module before the standard GAT operation in each layer.

[0070] Output: A neural network structure waiting to perform inference.

[0071] S22. Iterative Reasoning Process Input: The node features h_i^l of the l-th layer.

[0072] Processing procedure: For each node i, execute at each level: 1. Query: q_i^l = W_q * h_i^l.

[0073] 2. Addressing: Calculate the similarity between q_i^l and all m_k in M, a_{i,k}^l = softmax(λ * q_i^l * m_k^T).

[0074] 3. Read: m_i^l = Σ_k (a_{i,k}^l * m_k).

[0075] 4. Fusion: h_i^{l'} = FFN([h_i^l; m_i^l]), generating semantically enhanced features.

[0076] 5. GAT propagation: Perform standard graph attention aggregation on h_i^{l'} to obtain h_i^{l+1}.

[0077] Output: Updated node features H^{l+1}, which incorporates semantic priors from the memory.

[0078] S23. Conflict Detection and Refinement (Iterative Controller) Input: The probability distribution of the predicted feature type of the node after each round of inference.

[0079] Processing procedure: For each node i, execute at each level: 1. Monitoring: Check if there is a facet being contested by two conflicting features with a high probability (e.g., facet j simultaneously belongs to "hole side" and "groove bottom" with a probability >0.8).

[0080] 2. Intervention: - Neural guidance: Increase the query temperature parameter λ of the conflict node, forcing it to "extensively recall" more memory prototypes.

[0081] - Symbolic constraints: invoke built-in rules to directly correct the probability distribution of conflicting nodes (e.g., the rule determines that it should be "hole side").

[0082] 3. Feedback: Use the corrected information as additional input for the next round of reasoning.

[0083] Output: Trigger a new round of inference (layer l+1), or output the final stable node feature H_final.

[0084] S24. Feature Decoding and Structuring Input: Final node feature H_final.

[0085] Processing procedure: For each node i, execute at each level: 1. Node clustering: Clustering is performed based on H_final, with each cluster corresponding to a manufacturing feature.

[0086] 2. Parameter calculation: For each cluster, feature parameters (such as aperture, depth, and axis) are fitted based on the geometric information of the patches it contains.

[0087] 3. PMI Prediction: For each feature, use D_pmi to decode its predicted PMI information.

[0088] 4. Generate a feature tree: Assemble all the information to form a hierarchical feature structure.

[0089] Output: Structured feature semantic tree (JSON / XML format), containing: feature ID, type, confidence, list of reference surface IDs, geometric parameters, predicted PMI, and list of activated memory prototype IDs.

[0090] Step S3 (output) is implemented as follows: S31. Format Conversion Input: Structured feature semantic tree.

[0091] Processing procedure: Based on the API requirements of the target downstream systems (such as NX, CATIA, and self-developed CAPP), the semantic tree is converted into a specific format; Outputs: 1. General exchange file (JSON). 2. CAM system specific feature file (e.g., .feature). 3. Process specification template population data.

[0092] S32. Visual Feedback Input: Original CAD model + Feature recognition results.

[0093] Processing procedure: In the CAD GUI, different identified features are highlighted with different colors, and their types and key parameters are labeled. Output: Enhanced visual CAD model for engineers to review and confirm.

[0094] The system's final output includes: Structured feature file (XML / JSON format) CAM system interface files (such as NX's .feature file) Visualized highlighting and annotation on the original CAD model, etc.

[0095] The node clusters and feature labels output by the network are converted into a structured semantic tree, output in JSON / XML format. An example is shown below: { "part_id": "engine_connecting_rod_001", "features": [ { "feature_id": "F001", "feature_type": "THROUGH_HOLE", "confidence": 0.98, "geometry": {"referenced_face_ids": [101, 102, 201], "parameters": {"diameter": 12.5}}, "predicted_pmi": {"geometric_tolerance": [{"type": "location degree", "value":0.1}]}, "reasoning_basis": {"activated_memory_prototypes": [42, 87]} } ] } The system can ultimately output CAM interface files and visualization results.

[0096] Model training and deployment scheme: Pre-training: The encoder is trained on a million-level unlabeled STEP model through self-supervised tasks such as mask surface reconstruction and contrastive learning.

[0097] Fine-tuning: On tens of thousands of labeled data, most of the encoder parameters are frozen, and the decoder and rule adaptation module are trained.

[0098] Deployment: The model is converted to ONNX format, encapsulated as a C++ service, and a CAD plugin interface is provided.

[0099] In another exemplary embodiment, based on the same inventive concept as the method embodiment, a manufacturing process feature intelligent recognition system based on 3D spatial reasoning is provided, comprising: The associated world model pre-training module is used to learn, in unsupervised, stable association rules between PMI semantics and B-Rep geometric structure from massive {B-Rep, PMI} paired data, where B-Rep represents boundary representation structure and PMI represents product manufacturing information. The spatial reasoning and feature recognition module is used to identify new CAD models containing only B-Rep geometric information using knowledge assets built by the associated world model pre-training module, and obtain complete manufacturing process features; The structured feature output module is used to output a structured feature semantic tree that can directly drive the downstream CAPP / CAM system based on the manufacturing process features.

[0100] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.

Claims

1. A method for intelligent recognition of manufacturing process features based on 3D spatial reasoning, characterized in that, Includes the following steps: S1. Pre-training stage of the associated world model: From massive {B-Rep, PMI} paired data, learn the stable association rules between PMI semantics and B-Rep geometric structure in an unsupervised manner, where B-Rep represents the boundary representation structure and PMI represents product manufacturing information; S2. Memory-based spatial reasoning and feature recognition stage: Using the knowledge assets constructed in step S1, the new CAD model containing only B-Rep geometric information is identified to obtain complete manufacturing process features; S3. Structured Feature Output Stage: Based on the manufacturing process features, output a structured feature semantic tree that can directly drive the downstream CAPP / CAM system.

2. The intelligent recognition method for manufacturing process features based on 3D spatial reasoning according to claim 1, characterized in that, Step S1 specifically includes: S11. Input a CAD model file in ISO 10303 STEP AP242 format; S12. Extract the geometric structure code and PMI semantic code of the model by means of a dual-stream associative encoder; the dual-stream associative encoder includes a geometric encoder and a PMI encoder, wherein the geometric encoder contains a spatial partition head, which forces it to predict the potential spatial layout that can reflect the macroscopic functional area of ​​the part. S13. By training the dual-stream association encoder through the contrastive learning objective loss function, the mutual information between the geometric structure code and the PMI code from the same part is maximized, capturing the most essential association between the geometric structure code and the PMI code, and obtaining the joint association embedding vector. S14. After training, the joint association embedding vectors of all training data are summarized into K prototype vectors through a clustering algorithm to construct a differentiable association memory M; at the same time, a PMI semantic decoder is trained for subsequent prediction of PMI information from geometric features; the pre-trained two-stream association encoder parameters and PMI semantic decoder are saved.

3. The intelligent recognition method for manufacturing process features based on 3D spatial reasoning according to claim 2, characterized in that, Step S1 also includes preprocessing the CAD model file: B-Rep parsing: Extract faces, edges, vertices and their topological relationships, calculate the geometric descriptor of each face, and construct a facet graph; PMI parsing and graph construction: Dimensions, tolerances, datums, and annotations are parsed into structured entities to construct a PMI semantic graph.

4. The intelligent recognition method for manufacturing process features based on 3D spatial reasoning according to claim 2, characterized in that, The objective loss function for the contrastive learning is the InfoNCE loss.

5. The intelligent recognition method for manufacturing process features based on 3D spatial reasoning according to claim 2, characterized in that, Step S2 specifically includes: S21. Input the B-Rep structure of the new CAD model and load the pre-trained geometric encoder and associated memory library M; S22. Construct a B-Rep-PMI graph attention network to represent the B-Rep structure of the new CAD model as a surface node graph; S23. In the B-Rep-PMI graph attention network, a query-memory space reasoning module is introduced as the core reasoning unit. The query-memory space reasoning module performs a query operation for each node and generates semantically enhanced node geometric features. S24. Through multi-layer information transmission and aggregation of graph attention networks, and combined with an iterative controller, the geometric features of nodes are clustered and decoded into complete manufacturing process features.

6. The intelligent recognition method for manufacturing process features based on 3D spatial reasoning according to claim 5, characterized in that, The query operation specifically includes: Query: Convert the geometric features of nodes into query vectors; Addressing and reading: Calculate the similarity between the query vector and all prototype vectors in the associated memory M, and sum them by weight to read out the most relevant semantic memory content; Fusion: The read memory content is fused with the original geometric features to generate semantically enhanced node geometric features.

7. The intelligent recognition method for manufacturing process features based on 3D spatial reasoning according to claim 5, characterized in that, The implementation of the iterative controller includes: The classification confidence of the semantically enhanced node geometric features is monitored after each round of inference. When a high-confidence conflict is detected in the node attribution, an intervention mechanism is triggered to adjust the query parameters or apply symbolic constraints and guide the network to the next round of refined inference.

8. The intelligent recognition method for manufacturing process features based on 3D spatial reasoning according to claim 1, characterized in that, The structured feature semantic tree includes feature type, geometric parameters, topological references, predicted PMI information, and interpretable reasoning basis.

9. The intelligent recognition method for manufacturing process features based on 3D spatial reasoning according to claim 7, characterized in that, The adjustment of query parameters includes: Increase the query temperature parameter for conflict nodes; The application of symbolic constraints includes: It invokes built-in rules to directly correct the probability distribution of conflicting nodes.

10. A manufacturing process feature intelligent recognition system based on 3D spatial reasoning, characterized in that, include: The associated world model pre-training module is used to learn, in unsupervised, stable association rules between PMI semantics and B-Rep geometric structure from massive {B-Rep, PMI} paired data, where B-Rep represents boundary representation structure and PMI represents product manufacturing information. The spatial reasoning and feature recognition module is used to identify new CAD models containing only B-Rep geometric information using knowledge assets built by the associated world model pre-training module, and obtain complete manufacturing process features; The structured feature output module is used to output a structured feature semantic tree that can directly drive the downstream CAPP / CAM system based on the manufacturing process features.

Images