OCCT intelligent fuzzy reasoning code generation method based on physical constraint VLA

Abstract

The application discloses an OCCT intelligent fuzzy reasoning code generation method based on a physical constraint VLA, and belongs to the technical field of industrial intelligent manufacturing and computer vision. Firstly, a geometry self-attention mechanism with physical rationality constraints and a multi-view consensus alignment mechanism are fused to realize millimeter-level high-fidelity three-dimensional reconstruction and point cloud registration. Secondly, component perception graph attention aggregation, industrial semantic interaction matrix and connectivity-bound consistency joint constraint training are combined to complete accurate segmentation and engineering understanding from geometry and semantics. Then, natural language fuzzy intent and accurate geometric parameters are mapped into executable parameterized modeling codes conforming to topological-syntactic constraints. Driven by fuzzy comprehensive rewards, strategy gradient optimization and causal root cause analysis are combined to realize continuous optimization of the model. Finally, through elastic weight consolidation and geometry perception knowledge distillation, the system is ensured not to forget the key features, and high-fidelity CAD models directly used for simulation and quality inspection are autonomously and intelligently generated from multi-modal raw data.

Description

A method for generating intelligent fuzzy inference code for OCCT based on physical constraint VLA Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of industrial intelligent manufacturing and computer vision, specifically involving an intelligent fuzzy inference code generation method for OCCT (Open CASCADE Technology, an open-source 3D geometric modeling kernel) based on physical constraint VLA (Vision-Language-Action Model). Background Technology

[0002] The development of industrial software is rooted in rigorous mathematical and physical models, consistently pursuing near-obsessive precision and reliability when describing the geometric forms and physical changes of the real world. Traditional 3D modeling and quality inspection technologies heavily rely on comparison with computer-aided design (CAD) models based on preset rules, or on 2D visual inspection based on traditional image processing. However, when faced with the increasingly complex components, composite material structures, and stringent online inspection requirements of modern intelligent manufacturing, existing technologies reveal fundamental bottlenecks: for example, in industrial processes such as parts cleaning, traditional methods struggle to adapt to conceptual drift caused by equipment aging and changes in cleaning cycles, leading to a decline in model prediction accuracy over time and an inability to maintain long-term reliable performance.

[0003] Traditional geometric algorithms and multi-view stereo vision (MVS) modeling methods struggle to achieve complete, high-fidelity 3D reconstruction for highly reflective surfaces, transparent materials, areas with weak textures, and complex assemblies with severe visual occlusion. Models commonly suffer from accuracy loss at edges, holes, and minute features (such as micron-level scratches and assembly burrs), requiring significant manual intervention for subsequent repairs, thus failing to meet the demands of fully automated closed-loop quality inspection.

[0004] Most existing methods focus on geometric shape reconstruction, lacking a deep understanding of the functional semantics of components (such as assembly surfaces, sealing grooves, and conductive contacts) and physical constraints (such as tolerance zones and deformation patterns). This leads to a lack of generalization understanding of physical semantics. When the object under test exhibits novel defects inconsistent with the distribution of training data or undergoes morphological changes due to changes in operating conditions (such as thermal deformation), the model is prone to misjudgment and exhibits poor generalization ability.

[0005] While purely data-driven deep learning models (such as voxel networks, PointNet++, and NeRF) have made progress in visual reconstruction, they are "black boxes" with a lack of interpretability in their decision-making process. They are also difficult to integrate known physical laws and industrial knowledge (such as material mechanical properties and assembly process specifications), resulting in a disconnect between data-driven approaches and physical priors. Consequently, their results are often deemed "unreliable" and difficult to adopt in harsh industrial environments.

[0006] It is worth emphasizing that the rigorous numerical computation approach and the ability to uncover complex nonlinear mapping relationships from massive amounts of data inherent in the core of artificial intelligence are highly compatible with the inherent requirements of high precision, quantifiability, and verifiability pursued in the development of industrial software. Specifically, the outstanding performance of continuous learning methods in handling the concept drift problem makes it possible to solve the problem of model failure over time in industrial 3D modeling; while the elastic weight consolidation mechanism provides a theoretical basis for retaining historical design knowledge and adapting to new changes.

[0007] However, directly applying existing general-purpose AI vision algorithms (such as standard convolutional networks and the original Transformer architecture) to industrial quality inspection scenarios often yields unsatisfactory results. This is because general-purpose algorithms lack specific designs tailored to the unique and extreme requirements of industrial vision. For example, while standard continuous learning methods can adapt to gradual drift when handling short-term load prediction, they cannot effectively cope with systematic changes in geometric features when applied to 3D modeling. The computational overhead of global concept drift detection makes it difficult to handle high-resolution industrial point clouds, and its homogenized adaptation method may not be able to effectively focus on minute defect areas crucial for quality inspection.

[0008] For example, the large scale variations, complex lighting, and precise localization requirements commonly found in industrial point clouds pose challenges to the model's adaptive field and knowledge preservation capabilities that far exceed those of conventional applications.

[0009] Therefore, the industrial sector urgently needs an intelligent modeling algorithm specifically designed for harsh industrial environments. It must not only inherit the data-driven advantages of artificial intelligence but also deeply integrate continuous learning mechanisms with the constraints of industrial knowledge, achieving industrial-grade standards in accuracy, robustness, and efficiency. This will truly empower core aspects such as industrial design, production quality inspection, and operation and maintenance diagnosis, enabling a leap from perceptual intelligence to decision-making intelligence. Summary of the Invention

[0010] This invention proposes an intelligent fuzzy reasoning code generation method for OCCT based on Physically Constrained Virtual Aspects (VLA). The method first identifies the geometric features of industrial components in the 3D world through multi-view heterogeneous views. Then, it develops a VLA mechanism based on physically constrained rationality to achieve semantic understanding and causal reasoning of input images and instructions. Furthermore, an intent-driven active perception module is designed to understand and correct potentially fuzzy semantic instructions input by the user.

[0011] Building upon this foundation, the system design employs a reinforcement learning strategy to achieve self-correction and refinement of the 3D model. Furthermore, the system constructs an autonomous optimization mechanism driven by physical rationality, utilizing a multi-objective trade-off optimization algorithm with physical rule compliance as the core reward function to drive iterative model optimization. This enables intention-driven proactive perception and model self-correction; ultimately generating high-fidelity OCCT code sequences that can be directly used for industrial simulation and quality inspection. The specific technical solution is as follows:

[0012] A method for generating intelligent fuzzy inference code for OCCT based on physical constraint VLA includes the following steps:

[0013] S1: Alignment preprocessing is performed on multi-view images of industrial parts. Then, a multi-view shared feature containing global geometric context information is generated based on a geometric self-attention mechanism that incorporates physical rationality constraints. The multi-view shared feature is decoupled based on appearance information and spatial information. Then, the features of appearance information and spatial information are updated using industrial structure priors and saliency weights that reflect the importance of part regions. The updated features are converted into a dense depth map. Finally, after optimization using a nonlinear filtering module and an edge-aware repair algorithm, a 3D point cloud model is obtained by back-projection.

[0014] S2: Constructing a point cloud segmentation network. First, a multi-scale industrial semantic graph structure is constructed based on a 3D point cloud model, including a geometric proximity graph based on geometric spatial information, a functional semantic graph based on functional regions, and a manufacturing constraint graph based on manufacturing process constraints. Then, the geometric proximity graph, functional semantic graph, and manufacturing constraint graph are dynamically fused to obtain a dynamic fused graph. The nodes in the dynamic fused graph correspond to a point or a feature aggregation unit in the 3D point cloud model. After calculating the association weights of the nodes in the dynamic fused graph based on local-global interactive attention, the neighboring nodes are weighted based on the association weights to obtain the enhanced features of the nodes. Then, the enhanced features are updated based on a learnable industrial semantic interaction matrix. Finally, the component semantic descriptors of the nodes are obtained through a classifier, completing the semantic reorganization processing of all points in the point cloud and obtaining a semantic 3D model rich in component semantic descriptors.

[0015] S3: Based on the semantic 3D model feature extraction, rich semantic feature vectors are extracted. At the same time, the user's natural language input commands are mapped into embedding vectors and combined with the rich semantic feature vectors to obtain a set of action candidates. The candidate actions are subjected to parameter-level reasoning, screening and correction through a preset fuzzy rule base to determine the modeling action and its specific parameter configuration that best matches the current geometric state and industrial design logic. Finally, OCCT code is automatically generated to obtain the reconstructed CAD model.

[0016] S4: Based on 3D geometric concept drift detection and continuous learning; extract multi-dimensional geometric and semantic statistical features from the reconstructed CAD model, and perform concept drift detection by calculating the probability distribution of the new and old tasks; when the system detects concept drift, it enters the continuous learning stage, and monitors the balance between the new and old knowledge by calculating the change rate of the loss function; when training a new task, the soft target distillation and geometric feature alignment of the new and old tasks ensure a smooth transition in the semantic understanding and geometric feature extraction capabilities of the new and old models in 3D shape.

[0017] S1 is specifically as follows:

[0018] Acquire multi-view images of industrial parts and perform alignment preprocessing;

[0019] The aligned multi-view images are first fed into a geometric self-attention mechanism that incorporates physical rationality constraints to generate multi-view shared features containing global geometric context information;

[0020] In the multi-view shared feature generation stage, based on the domain-specific projection and constraint mechanism, the multi-view shared features are decomposed into appearance domain features that encode the texture, brightness and local appearance change information of the image, and geometric domain features that encode the depth cues, disparity relationships and spatial structure information.

[0021] Based on the pre-defined industrial structure priors, the component-level importance of multi-view shared features is evaluated, and the significant weights reflecting the geometric contribution of different component regions in industrial parts are output.

[0022] Based on the multi-view consistency constraint, the shared features from different viewpoints are aligned and fused to obtain the multi-view matching consistency measure.

[0023] Based on saliency weights and multi-view matching consistency measures, appearance domain features and corresponding geometric domain features under different perspectives are co-encoded to obtain the final multi-view features.

[0024] Based on the original image patch and multi-view features, appearance perception feature extraction and geometric structure feature extraction are performed. The two feature vectors obtained are used to update the appearance domain features and geometric domain features, and the optimization is continuously iterated until it stops.

[0025] The final multi-view features are used to generate a dense depth map with millimeter-level precision. The dense depth map and multi-view features are then jointly encoded to construct a unified and reversible 3D descriptor.

[0026] Dense depth maps and unified reversible 3D descriptors are fused in parallel with deformable 3D rolls and graph attention mechanisms to obtain fused feature representations. Then, nonlinear filtering modules and edge-aware repair algorithms are used to optimize the fused feature representations to obtain ultra-high precision depth maps.

[0027] Finally, point cloud back projection calculation is performed to transform the two-dimensional depth information of the ultra-high precision depth map into a three-dimensional point cloud model.

[0028] S2 is specifically as follows:

[0029] S201: Construct a multi-scale industrial semantic graph structure based on a 3D point cloud model; the geometric proximity graph is constructed based on the k-nearest neighbors of the point cloud spatial coordinates; the functional semantic graph is based on the functional area division extracted from industrial design knowledge, and each point is initially labeled with a lightweight functional classifier to connect points of the same functional category; the manufacturing constraint graph connects point pairs that satisfy specific manufacturing relationships based on manufacturing process constraints.

[0030] S202: Based on the adjacency matrix of the graph structure and their respective learnable weight matrices, the three graph structures are dynamically fused to obtain a learnable dynamic fused graph; the attention weights between nodes in the dynamic fused graph are calculated through local-global interactive attention, as shown in the following formula:

[0031] ;

[0032] in, Represents nodes in the dynamic fusion graph , Attention weights between nodes; nodes refer to vertices of the dynamic fusion graph, with each node corresponding to a point or a feature aggregation unit in the point cloud; For activation functions; and Representing nodes respectively With nodes eigenvectors; These are learnable fusion weight parameters; For local feature interaction functions; For mapping functions; These are prior features of industrial semantics; Manufacturing functions constrained by process technology. Encoding the process constraint relationships between nodes. This indicates a gating feature fusion operation;

[0033] Feature points to be processed in a point cloud model As the core point, for each neighboring point Perform a sequential traversal, based on attention weights The input features of neighboring points are weighted and modulated as follows:

[0034] ;

[0035] in, For point Enhanced features, Represents the set of neighborhood points. This indicates element-wise gated modulation operation. As the modulation vector, through a learnable industrial semantic interaction matrix The calculation yielded:

[0036] ;

[0037] in, and Representing points respectively The predicted functional category probability distribution of j;

[0038] The enhanced features are input into the industrial semantic interaction query module and combined with the learnable industrial semantic interaction matrix. Perform semantic consistency matching; if the matching result does not meet the preset semantic confidence threshold, then update the parameters of the industrial semantic interaction matrix in reverse and return to the aforementioned weighted modulation step to re-execute;

[0039] If the matching result meets the semantic confidence requirement, the enhanced features of each node are input into a lightweight point-independent MLP classifier, and the component semantic descriptor is output.

[0040] S203: During the training phase, joint optimization is performed based on the cross-entropy loss function, component connectivity constraints, and boundary consistency constraints; the component connectivity constraints identify connected regions in the point cloud through the flood fill algorithm, and the boundary consistency constraints measure boundary alignment through the similarity of point cloud normal vectors.

[0041] S204: A closed-loop semantic correction unit is introduced during the inference phase. It receives semantic feedback from S3. Specifically, the semantic feedback is the semantic representation of modeling actions and their specific parameter configurations that cannot be matched with the current geometric state and industrial design logic. The semantic feedback is input into the embedded encoder to generate a semantic correction signal. Based on the feedback, the corresponding point cloud problem subset is identified. The cross-entropy loss and boundary consistency constraints are calculated based on the point cloud problem subset. The correction gradient is calculated by fusing the semantic correction signal. The learnable weight matrix of the last layer of the point cloud segmentation network is updated based on the correction gradient.

[0042] S3 is specifically as follows:

[0043] S301: Extract semantically rich feature vectors based on semantic 3D models, including point function labels, geometric parameters, and spatial relationship information;

[0044] S302: Use a language encoder to map user natural language input commands into embedding vectors. Meanwhile, the semantically rich feature vector corresponding to each component is mapped to a component embedding representation. Calculate the strength of causal relationship:

[0045] ;

[0046] in This is a priori confidence bias. The term represents the strength of the causal relationship between instructions and components, where d is the feature dimension; the term with the strongest causal relationship is selected. The corresponding component is used as the target operation object, and the final output is a structured set of action candidates. :

[0047] ;

[0048] in Indicates the target object. As candidate actions, Action confidence;

[0049] S303: Design a hierarchical fuzzy inference engine that receives geometric parameters from a candidate action set and semantically rich feature vectors as input. It performs parameter-level inference, filtering, and correction on candidate actions using a pre-defined fuzzy rule base to determine the modeling action and its specific parameter configuration that best match the current geometric state and industrial design logic. Specifically, for each geometric parameter, a fuzzy membership function is defined, and an expert fuzzy rule base initialized based on industrial experience is designed to map natural language instructions and geometric parameters to specific modeling operations. After defuzzification, the output is a cleared result of the fuzzy set.

[0050] S304: Based on the clarification results, OCCT code is automatically generated. The reconstructed CAD model is compared with the original semantic 3D model in multiple dimensions to obtain geometric error, semantic consistency error and physical rationality index. Based on these, a fuzzy comprehensive reward function is defined. If the fuzzy comprehensive reward function is lower than the threshold, the corresponding fuzzy rule weights are updated, the action confidence is adjusted, and finally the OCC code is regenerated and re-simulated.

[0051] S4 is specifically as follows:

[0052] S401: Extract multidimensional geometric and semantic statistical features from the reconstructed CAD model. Then, within a continuous time period, calculate the probability distribution of the multidimensional geometric and semantic statistical features at the current time and the probability distribution at the previous time. Use KL divergence to quantify the difference between the two. When the KL divergence is greater than the drift threshold, the system determines that there is significant concept drift and triggers the continuous learning mechanism; or it judges concept drift by calculating the distance between the probability distributions of the new and old geometric features.

[0053] S402: Entering the continuous learning phase; let the current task number be... , Indicates the current task The following network parameter set, For network parameter indexing, This represents the total number of network parameters, including all trainable parameters that participate in gradient updates during training.

[0054] Dynamic adaptive geometric regularization loss function for current task K Defined as:

[0055] ;

[0056] in The standard loss function for the current task measures the model's predicted output. With real labels The error between; Representing historical tasks Network parameters The optimal value, The diagonal element of the Fisher information matrix corresponding to this historical task. The geometric regularization intensity hyperparameter;

[0057] The update formula for the Fisher information matrix is ​​as follows:

[0058] ;

[0059] in, This represents the accumulated Fisher information value for the current task. This represents the cumulative value from the previous task; This provides real-time Fisher information for this parameter in the current task. Historical information attenuation coefficient; initial conditions This indicates that the cumulative matrix of the first task is its Fisher information matrix;

[0060] S403: Through the dual constraints of soft target distillation and geometric feature alignment, a smooth transition is ensured between the old and new models in terms of semantic understanding of 3D shapes and geometric feature extraction capabilities. A multi-objective optimized geometry-aware distillation loss function is used. as follows:

[0061] ;

[0062] in, Indicates the geometric distillation weighting coefficient; Indicates the current task Task loss; This represents the soft-target distillation loss and constrains the student network output for the current task. Teacher network output for historical tasks Consistency in the final category probability distribution; ξ represents the feature alignment weight coefficient; This represents the geometric feature alignment loss, which constrains the current student network in the intermediate feature layers. Feature layer corresponding to the historical task teacher network Consistency of output.

[0063] The beneficial effects of this invention are as follows:

[0064] (1) Compared with traditional 3D modeling methods that rely on manual interaction and rule definition, this invention can realize end-to-end intelligent generation from multimodal data to parametric CAD models. Through industrial-grade multi-scale feature enhancement architecture and progressive depth inference engine, it significantly improves the integrity and millimeter-level accuracy of 3D reconstruction under weak texture, high reflectivity and complex occlusion conditions.

[0065] (2) Compared with the traditional point cloud processing method's reliance on data sorting and regularization, this invention proposes a three-dimensional shape direct recognition and segmentation technology based on point independent perception and symmetry function. Through the permutation invariant network structure and local-global feature fusion mechanism, it achieves high-fidelity semantic segmentation and geometric understanding of the original disordered point cloud.

[0066] (3) Compared with traditional methods that only restore geometry but lack modeling semantics, this invention constructs an OCCT code automatic generation pipeline that integrates fuzzy reasoning. Through multimodal uncertainty calibration and hierarchical fuzzy rule base, low-level geometric features are transformed into high-level modeling intentions, driving the generation of parametric modeling code sequences that conform to industry standards.

[0067] (4) Compared with traditional static modeling systems, this invention proposes a learning mechanism of autonomous evolution and closed-loop optimization. By integrating a reinforcement learning framework with fuzzy rewards and a causal diagnostic strategy, it realizes the continuous self-optimization of modeling strategies and the dynamic evolution of design specifications, enabling the system to evolve from an automated tool to an autonomous intelligent agent.

[0068] (5) Compared with the traditional single visual perception modeling process, the present invention realizes the deep integration of multimodal perception and physical semantics. The image embedding and point cloud geometric features are uniformly represented by the graph attention fusion module, and combined with industrial prior knowledge, the semantic understanding of component functions and assembly relationships is enhanced.

[0069] (6) Compared with visual algorithms with fixed parameters, this invention establishes an adaptive optimization system for industrial scenarios. Its key parameters such as matching cost weight, pooling strategy, and fuzzy rule confidence can be dynamically adjusted according to sensor type, part complexity and task objectives, which significantly improves the generalization and practicality of the system in different industrial environments.

[0070] (7) Compared with black-box deep learning models, this invention integrates industrial knowledge constraints while maintaining data-driven performance through an interpretable fuzzy reasoning system and orthogonal constraint feature alignment mechanism, making the model's decision-making process both high-performance and interpretable, which is more in line with the requirements of industrial quality inspection for reliability and credibility.

[0071] (8) Compared with isolated modeling and quality inspection processes, this invention constructs an integrated closed-loop system of “modeling-simulation-diagnosis-optimization”, which can embed manufacturability analysis and performance simulation requirements in the modeling stage, realize the advanced concept of Design-Right-First, and significantly shorten the product development cycle. Attached Figure Description

[0072] Figure 1 is a flowchart of the specific process of the industrial prior-guided progressive deep inference engine.

[0073] Figure 2 shows the iterative network diagram for deconstructing multi-dimensional feature extraction.

[0074] Figure 3 is the execution flowchart of the adaptive point cloud understanding architecture module.

[0075] Figure 4 shows a scanning example of a multi-scale industrial semantic graph.

[0076] Figure 5 shows the feature reorganization diagram guided by industrial semantics.

[0077] Figure 6 is a flowchart of the OCCT code generation and closed-loop correction unit.

[0078] Figure 7 is a framework diagram of the knowledge distillation model for continuous learning.

[0079] Figure 8 shows a comparison between the adaptive modeling framework based on concept drift detection and continuous learning and the traditional model.

[0080] Figure 9 shows the performance evolution of the geometry-sensing distillation framework. Detailed Implementation

[0081] This invention provides an intelligent fuzzy inference code generation method for OCCT based on physical constraint VLA. As an intelligent modeling and refined reconstruction system for industrial parts, it is driven by industrial quality requirements. By constructing a fully closed-loop intelligent processing architecture from multi-view 3D perception to parametric CAD code generation and adaptive evolution, it achieves end-to-end automated processing from raw scan data to high-precision part-level semantic models and their executable design code. The specific steps of this method are as follows:

[0082] S1: Based on an industrial prior-guided progressive depth inference engine, multi-view images of industrial parts are processed in an integrated manner. Through a collaborative multi-view representation framework with physical rationality constraints and a geometric topology inferencer, a 3D point cloud model with millimeter-level accuracy is generated. As shown in Figure 1, the specific steps include:

[0083] S101: To address the visual challenges in industrial part modeling, such as weak textures, metallic reflections, and strong geometric repetition, a robust 3D shape representation is constructed by rendering 3D models from different perspectives, generating a view descriptor for each view image, and then integrating the information from these multiple perspectives through a specific strategy, as detailed below.

[0084] Acquire multi-view images of industrial parts and perform alignment preprocessing, specifically using one reference image and its corresponding N images (suggested). The source image is used as input, and all images are uniformly scaled to [size missing]. Resolution (of which) Indicates the width and height of the image, such as The pixel values ​​are normalized. Features are extracted through shared convolution or Transformer encoders and multi-scale deep feature encoding is performed at different spatial resolutions to form a preliminary multi-scale deep feature representation. Subsequently, multi-scale deep features from different perspectives are spatially aligned and projected to generate multi-view context information.

[0085] The spatial alignment and projection, through camera calibration parameters, align each source image with each candidate depth plane of the reference image via a differentiable homography transformation.

[0086] S102: To learn and extract industrially robust feature representations from aligned multi-view images, this invention designs a collaborative multi-view representation framework based on a three-stage reasoning paradigm of decoupling-analysis-fusion. This framework, based on unified physical rationality constraints, introduces component-level structural priors to achieve collaborative fusion of multi-view information. As shown in Figure 2, the collaborative multi-view representation framework includes at least: a geometric self-attention mechanism incorporating physical rationality constraints, a geometric saliency inference submodule driven by industrial structural priors, and a multi-view consensus alignment mechanism. The core of this framework is an iterative network used to deconstruct multi-dimensional features and progressively optimize multi-view information. By constructing an iterative network for deconstructing multi-dimensional feature extraction and using a geometric self-attention mechanism as the core to perform multi-level feature interaction and updates, decoupled modeling and joint reasoning of the appearance and geometric domains for weak textures and occluded regions are achieved, thereby coordinating the consistent expression of multi-view features at a global scale.

[0087] S1021: Aligned multi-view images are first input into a geometric self-attention mechanism that incorporates physical rationality constraints to generate multi-view shared features containing global geometric context information;

[0088] Furthermore, the geometric self-attention mechanism explicitly models the overall geometric continuity of the scene across multiple view domains. During the attention weight calculation process, it introduces a 3D spatial relative position encoding constructed based on camera geometric parameters and depth assumptions obtained in the preprocessing stage, injecting real 3D spatial relationships into feature interactions, thereby obtaining a globally geometric context feature representation with higher physical consistency. The attention calculation form is as follows:

[0089]

[0090] in, For the calculation of attention mechanism, for function, These are query, key, and value matrices, respectively. For feature dimension, Relative position encoding represents the relative positional relationship of features from different perspectives in three-dimensional space. By introducing this relative position encoding as an additive bias term into the attention weight calculation process, the geometric continuity and physical rationality constraints in three-dimensional space can be prioritized during feature aggregation. This allows it to transcend simple image appearance similarity and fundamentally understand and reconstruct a continuous, smooth, and physically reasonable global morphology in three-dimensional space. The global contextual information it captures directly serves to construct a geometrically consistent overall three-dimensional structure, rather than scattered local features.

[0091] In the multi-view shared feature generation stage, a feature deconstruction module is introduced. Based on domain-specific projection and constraint mechanisms, the multi-view shared features are structurally decomposed into two sub-representations: appearance domain features and geometric domain features. The appearance domain feature branch focuses on encoding the image's texture, brightness, and local appearance variation information; the geometric domain feature branch focuses on encoding depth cues, disparity relationships, and spatial structure information.

[0092] S1022: The geometric saliency inference submodule, based on a preset industrial structure prior, performs component-level importance assessment on multi-view shared features and outputs saliency weights reflecting the geometric contribution of different component regions in an industrial part. These saliency weights characterize the importance of different component regions. The preset industrial structure prior is used to characterize the importance of functional and structural key regions in an industrial object.

[0093] The multi-view consensus alignment mechanism first calculates a multi-view matching consistency metric based on the correspondence between appearance domain features and geometric domain features under different perspectives. This metric is used to quantify the stability and reliability of the same spatial hypothesis or feature response under multi-view conditions. Specifically, for any candidate spatial hypothesis... Let it be in the first The multi-view shared feature response from each perspective is By calculating the consistency of multi-view feature similarity and introducing 3D geometric projection consistency... As a weighted constraint, the multi-view matching consistency metric score for:

[0094]

[0095]

[0096] in, The number of effective perspectives participating in the integration; For reference perspective; Indicates cross-view normalized similarity; Indicates the candidate hypothesis in perspective The three-dimensional coordinates obtained by downward back projection. Its multi-view mean, Represents the L2 norm; Represents the variance operator. To prevent numerically unstable minimal constants; Represents the normalization function. This represents the weighted norm.

[0097] S1023: Construct a cross-domain matching cost fusion module; this fusion module performs joint modeling of the appearance domain and geometric domain and adaptive modulation.

[0098] Based on industrial structure priors, geometric domain features, and multi-view matching consistency measures, joint features are obtained by weighting them. Specifically, for each feature channel and spatial region, channel-by-channel weighting is performed by multiplication or by using a gate function as a modulation factor. This allows the joint features to focus on functional and structural key regions during the update process, enhancing feature responses in regions with significant structures and high multi-view consistency, and suppressing unreliable matching in regions with weak textures, occlusion, or low consistency.

[0099] Finally, based on this joint feature, a matching cost fusion strategy is executed to integrate appearance similarity and geometric consistency information to generate multi-view features with higher physical plausibility and less ambiguity, providing optimized input for the subsequent industrial physical guidance patch matching module. Here, appearance similarity information refers to the matching similarity or cost between appearance domain features from different viewpoints, reflecting the consistency of texture, brightness, and local appearance changes; geometric consistency information refers to the consistency of 3D geometric projection, reflecting the 3D spatial plausibility of the matching relationship. By fusing these two types of information, multi-view features with higher physical plausibility and less matching ambiguity are generated.

[0100] S1024: Building upon this, a learnable industrial physics-guided patch matching module is constructed that jointly optimizes color similarity and spatial distance. This patch matching module receives multi-view features and original image patch pairs as input. An appearance-aware feature extraction mechanism is designed to calculate the statistical distribution similarity of patch pairs in different color spaces and extract texture descriptors robust to illumination changes, generating appearance-aware feature vectors. Simultaneously, a geometric structure feature extraction mechanism is designed to construct a local spatial relationship graph based on patch coordinates and initial depth assumptions in the camera coordinate system, calculating geometric structure feature vectors representing surface continuity and orientation consistency. These two feature vectors are fed back into the iterative network for updating appearance and geometric features in the next iteration, achieving continuous optimization.

[0101] In its first iteration, the iterative network decouples multi-view shared features into appearance domain features and geometric domain features. In subsequent iterations, it directly utilizes the appearance-aware feature vectors and geometric structure feature vectors output by the patch matching module to specifically update and optimize the current appearance domain features and geometric domain features. During the update process, the iterative network adaptively adjusts the weights for each feature channel and spatial location based on industrial structure priors, multi-view matching consistency metrics, geometric domain features, and saliency weights. This enhances the response to key geometric structures and functional regions while suppressing matching ambiguities caused by weak textures, occlusion, or geometrically repetitive structures. In the fusion phase, a joint weighting strategy of saliency and consistency is used to fuse appearance similarity information and geometric consistency information at matching costs, generating multi-view features with less ambiguity and higher physical consistency.

[0102] The optimal depth hypothesis is selected through global or local optimization methods. After reaching the set iteration stopping condition, the final multi-view features are used to generate a dense depth map with millimeter-level precision. The dense depth map and the corresponding multi-view features are jointly encoded to construct a unified reversible 3D descriptor. In this process, each 3D point or voxel not only records spatial coordinate information but also integrates local texture, brightness, normal vector, and geometric structure features, so that it simultaneously reflects the object's geometric shape, topological relationships, and industrial functional information. The unified reversible 3D descriptor adopts a structured or reversible encoding method, enabling it to reconstruct high-density point clouds or depth maps from the descriptor, achieving reversibility, and providing optimized input for subsequent fine registration, parameter inversion, and functional analysis of industrial parts. Through the above methods, the finally obtained multi-view features and unified reversible 3D descriptor have high consistency, low matching ambiguity, and physical rationality, providing reliable support for end-to-end automated modeling of industrial parts.

[0103] S103: Feed the high-quality cost volume (high-fidelity dense depth map and unified reversible 3D descriptor) into the geometric topology inferencer.

[0104] First, a deformable 3D convolution and a graph attention mechanism are fused in parallel to achieve collaborative modeling of local geometry and global topological relationships. The deformable 3D convolution captures local surface curvature and detailed features, while the graph attention mechanism, based on the graph topology constructed from the cost volume, strengthens the global dependencies between different spatial nodes. The two are then fused using weighted fusion to obtain the final feature representation.

[0105]

[0106] in, This represents the fusion feature representation. For deformable 3D convolution output. To achieve attention output, To integrate weights and ensure that local and global information are complementary rather than unidirectionally linked, the local geometry (such as surface curvature) and global topological connectivity of the scene are explicitly modeled while regularizing the cost volume.

[0107] The fused feature representation is optimized using a nonlinear filtering module and an edge-aware repair algorithm. The nonlinear filtering module, based on anisotropic diffusion smoothing constraints, effectively suppresses noise and outliers in the dense depth map while preserving important geometric edges. Then, the edge-aware repair algorithm, by jointly optimizing gradients and region consistency at depth discontinuities, accurately repairs holes and artifacts caused by occlusion or weak textures. The final output is an industrial-grade ultra-high-precision depth map that combines geometric smoothness, structural integrity, and sharp boundaries.

[0108] Furthermore, the execution process of the edge-aware repair algorithm includes at least the following steps:

[0109] First, based on the gradient distribution information of the dense depth map in the spatial domain, the geometric edge response intensity at each pixel location is calculated. By jointly analyzing the first or second-order differences of the depth values ​​in the horizontal and vertical directions, an edge response map reflecting the degree of depth discontinuity and structural abrupt changes is constructed, used to characterize key geometric locations such as contour boundaries, edges, and hole edges in industrial parts. Edge detection based on the edge response map calculated from the depth gradient is as follows:

[0110] .

[0111] in, The edge response intensity is used to reflect the degree of depth discontinuity or structural abrupt change. This is a dense depth map; This represents the gradient of the dense depth map in the horizontal direction. This represents the gradient of the dense depth map in the vertical direction.

[0112] Guided by the edge response map, regions with missing depth, abnormal fluctuations, or significant inconsistencies in the dense depth map are automatically marked, forming a set of regions to be repaired. Among them, regions with missing depth mainly correspond to areas that are not visible from multiple views or have weak texture, while regions with abnormal fluctuations correspond to outliers introduced by noise interference or mismatches.

[0113] Specifically, by constructing an energy minimization model that incorporates edge-aware weights, the depth distribution within the repair area is optimized as a whole, and its objective function is expressed as:

[0114]

[0115] in, This is the repaired ultra-high precision depth map. This is the initial depth estimation result. Indicates the area to be repaired. For pixel position, For depth gradient operators; The edge-aware weighting function is adaptively calculated based on the edge response intensity of the dense depth map.

[0116]

[0117] in, Indicates the geometric edge response intensity at the pixel location. These are edge-sensitive control parameters.

[0118] Through the above weight modulation mechanism, smoothness consistency constraints are enhanced in non-edge regions, ensuring continuous depth filling and maintaining local smoothness. Meanwhile, smoothing weights are automatically reduced near geometric edges, explicitly suppressing excessive smoothing across edges and thus preventing depth penetration between different geometric components or surfaces. This joint optimization process effectively preserves the original or inferred geometric discontinuities while ensuring depth continuity in the repaired area.

[0119] S104: Perform point cloud back projection calculations to convert the two-dimensional depth information of the ultra-high precision depth map into three-dimensional geometric data.

[0120] Specifically, let the pixel coordinates in the ultra-high precision depth map be... Its corresponding depth value The camera intrinsic parameter matrix is It includes the focal length in the horizontal direction. Vertical focal length and principal point coordinates The pixels are mapped to three-dimensional points in the camera coordinate system using the following back projection relationship. :

[0121]

[0122] Where X, Y, and Z are three-dimensional points. The three-dimensional coordinates are obtained. By traversing all valid pixels in the ultra-high precision depth map and repeating the above calculation process, a complete 3D point cloud model with millimeter-level precision can be generated. The 3D point cloud model fully preserves the geometric details recovered in the reconstruction stage and serves as direct input for subsequent point cloud semantic understanding and component-level segmentation modules, providing a reliable geometric foundation for subsequent component recognition and semantic analysis.

[0123] S2: Construct a point cloud segmentation network, and perform component-level semantic segmentation and understanding of point clouds through multi-scale industrial semantic graphs and local-global interactive attention mechanisms, and output a structured semantic model;

[0124] Traditional point cloud processing methods are mainly based on pure geometric reconstruction, which struggles to capture the functional and manufacturing relationships of industrial components, resulting in segmentation results lacking engineering value. This invention constructs a unified semantic segmentation framework based on industrial priors constrained by physical rationality (such as functional area division and manufacturing constraints). This framework overcomes the limitations of traditional point cloud processing, achieving a deep understanding from 3D shape to component semantics. It effectively handles the complexity and diversity of industrial point cloud data, ensuring that the output possesses physical rationality and engineering applicability. The process is shown in Figure 3, and the specific steps are as follows:

[0125] S201: Feature encoding of the 3D point cloud model. To address the heterogeneity and multi-scale characteristics of industrial point clouds, a multi-scale industrial semantic graph structure is first constructed. Instead of using a single k-nearest neighbor graph, a three-level graph structure is constructed, as shown in Figure 4:

[0126] Geometric proximity graph 𝐺_𝑔𝑒𝑜: Represented by solid blue lines, it connects k-nearest neighbors (k=16) based on the spatial coordinates of the point cloud. This graph structure primarily captures the local geometric features and spatial distribution information of the point cloud, providing basic geometric support for subsequent semantic understanding. The blue lines in the graph connect spatially close points, forming a grid-like structure.

[0127] Functional semantic graph 𝐺_𝑓𝑢𝑛𝑐: Represented by solid green lines, it is based on functional area divisions extracted from industrial design knowledge. A lightweight functional classifier initially labels each point with its function, connecting points of the same functional category. In the engine block, this includes functional areas around cylinder bores, bolt hole areas, etc. This graph structure incorporates functional area divisions from industrial design knowledge, aiding in understanding the working principles of parts.

[0128] Manufacturing constraint diagram 𝐺_𝑚𝑎𝑛𝑢: Based on manufacturing process constraints (such as minimum wall thickness, assembly clearance, and other industrial a priori requirements), it connects pairs of points that satisfy specific manufacturing relationships, represented by red dashed lines. This diagram structure explicitly encodes industrial manufacturing knowledge, helping to ensure the manufacturability and assemblability of the design.

[0129] S202: By designing a learnable dynamic graph fusion network, it can adaptively fuse three graph structures to generate a unified industrial semantic graph that reflects both geometric layout and functional and manufacturing relationships. The learnable dynamic graph fusion method is as follows:

[0130]

[0131] in, Represents a learnable, dynamically fused graph. The adjacency matrix, representing the geometric proximity graph, is constructed from the k-nearest neighbor relationships of spatial coordinates and is used to characterize the local geometric adjacency relationships between points. The adjacency matrix represents the functional semantic graph, which connects points with the same or related functional attributes based on the initial functional region labeling results. The adjacency matrix represents the manufacturing constraint graph, which connects pairs of points that satisfy specific manufacturing constraints based on manufacturing process priors (such as wall thickness and assembly relationships). It represents the three graph structures—geometric proximity graph, functional semantic graph, and manufacturing constraint graph—that are corresponding to each other, independent, and learnable fusion weight matrices, used to adaptively adjust the contribution ratio of different types of relations in the final semantic graph; This represents a nonlinear normalization function used to limit the numerical range of the adjacency weights after fusion and enhance numerical stability.

[0132] Building upon this foundation, and abandoning traditional k-nearest neighbor graph construction and standard graph attention mechanisms, this invention proposes a component-aware graph attention aggregation mechanism driven by physical rationality. On a learnable, dynamically fused graph structure, a dual-path interactive attention mechanism is designed to calculate the association weights between nodes. This is a novel attention computation paradigm: local-global interactive attention.

[0133]

[0134] in, Represents nodes in the dynamic fusion graph , Attention weights between nodes. A node here refers to a vertex in the dynamic fusion graph, with each node corresponding to a point or a feature aggregation unit in the point cloud. As an activation function; and Representing nodes respectively With nodes The feature vectors, as local features, specifically include geometric information (such as 3D coordinates, normal vectors, and local curvature) and semantic information (such as functional area categories or component labels). This is a local feature interaction function used to model fine-grained geometric relationships between nodes; These are prior features for industrial semantics, serving as global features and including the functional area division of the entire part, assembly relationships, manufacturing constraints, and global geometric statistics. In the attention mechanism, As a mapping function, it maps global features to modulation terms, guiding the aggregation of information between nodes to follow the overall industrial design logic; The encoding of the process constraint relationships between nodes comes from the manufacturing constraint diagram. Manufacturing functions constrained by process Used to depict engineering constraints such as manufacturing and assembly; This indicates a gating feature fusion operation; These are learnable fusion weight parameters. This mechanism simultaneously considers three factors: the detailed geometric feature relationships and local feature interactions between points, global semantic modulation, and industrially constrained coding.

[0135] S203: In the feature aggregation stage, features are modulated based on the predicted functional category of points through an industrial semantic-guided reorganization mechanism. A gated fusion unit is designed to integrate these three information streams and calculate attention weights that better reflect the physical rationality logic. As shown in Figure 5, this reorganization mechanism aims to overcome the limitations of traditional geometric feature aggregation by injecting functional correlations between components, making feature propagation more consistent with the physical design logic of industrial parts. It includes the following steps:

[0136] First, the feature points to be processed in the point cloud model. As the core point, obtain its corresponding Nearest neighbor set To determine the local semantic interaction range of the feature point;

[0137] Subsequently, for each neighboring point in the neighborhood point set The process is performed iteratively. For each neighboring node in the current iteration, attention weights between nodes are generated using the aforementioned local-global interactive attention calculation mechanism. and according to The input features of neighboring points are weighted and modulated. The calculation process is as follows:

[0138]

[0139] in, For point Enhanced features, Represents the set of its neighborhood points. The attention weights between nodes For neighborhood points The input feature vector, This indicates element-wise gated modulation operation. This is the modulation vector.

[0140] Next, the enhanced features are input into the industrial semantic interaction query module and combined with the learnable industrial semantic interaction matrix. Perform semantic consistency matching; if the matching result does not meet the preset semantic confidence threshold, then update the parameters of the industrial semantic interaction matrix in reverse and return to the aforementioned weighted modulation step to re-execute;

[0141] Industrial semantic interaction matrix here It is a miniature knowledge base whose parameters record typical feature interaction patterns between different functional components (such as assembly surfaces and bolt holes), used to generate corresponding feature modulation weights based on the functional semantic relationships of point pairs. The industrial semantic interaction matrix is ​​based on points... With point The predicted semantic category is dynamically generated to modulate the transmitted features. Perform refined weighting:

[0142]

[0143] in, For industrial semantic interaction mapping functions, and They represent based on feature points The function category probability distribution predicted by j. This modulation step means that feature propagation depends not only on geometric proximity, but also on the functional correlation between components. For example, the mechanism learns that the features of the "sealing groove" are crucial to the optimization of the "mating surface," and that the feature correlation between them can be strengthened even if they are at a certain distance.

[0144] If the matching result meets the semantic confidence requirement, the enhanced features of each point are input into a lightweight point-independent MLP classifier, which outputs the component semantic descriptor corresponding to that point. The above process is repeated until the semantic reorganization of all feature points in the point cloud is completed. Thus, a disordered point cloud with only geometric information is transformed into a structured semantic 3D model rich in component-level functional semantics.

[0145] S203: During the training phase, to fundamentally ensure that the segmentation results conform to the physical manufacturing logic of industrial parts and to achieve system self-optimization, a training paradigm with physical consistency constraints was designed for the generated semantic 3D model. Based on the standard cross-entropy loss function, component connectivity constraints and boundary consistency constraints were added, thus forming a joint loss function with industrial logic adaptive capabilities. Specifically, the cross-entropy loss function... This is used to measure the difference between the point-level predicted semantic distribution and the actual component labels. By minimizing the cross-entropy loss, the model can learn accurate component semantic discrimination ability at the point level.

[0146] To prevent discrete noise points from appearing within the same component, a component connectivity constraint loss is constructed. This loss is achieved by identifying connected regions in the point cloud using a flood-fill algorithm. When predicting components... Number of connected components When greater than 1, the definition is:

[0147]

[0148] in This is the weighting factor (typical value 0.1). For components The number of points in the data is used to normalize the noise effect.

[0149] On the other hand, to ensure that the boundaries of the predicted components are aligned with the actual geometric boundaries, boundary consistency constraints are introduced. Boundary alignment is measured by the similarity of point cloud normal vectors and is defined as follows:

[0150]

[0151] in and These are the predicted boundary point set and the actual geometric edge point set, respectively. To predict boundary points The normalized normal vector, For its real edge point set The nearest neighbor normal vector is used, and nearest neighbor matching is accelerated by KD-Tree (neighborhood k=5). Therefore, the joint loss function is defined as follows:

[0152]

[0153] in and To balance the hyperparameters (typical values) of the corresponding loss function , The training paradigm can be determined through grid search optimization. This training paradigm is implemented in the PyTorch framework, using the Adam optimizer (learning rate 1×10⁻³, weight decay 1×10⁻³). 5 The batch size is 32, and the total number of iterations is 1000 rounds. The mIoU metric is evaluated on the validation set every 50 rounds. If mIoU < 0.85, the batch size is dynamically increased. and The weights are used to strengthen the constraints, thereby achieving co-optimization of semantics and geometry during the training process.

[0154] S204: During the inference phase, to further enhance the model's adaptability and industrial robustness, this invention designs a closed-loop semantic correction unit. The closed-loop semantic correction unit receives semantic feedback from the subsequent VLA inference engine. This feedback specifically represents the semantic representation of modeling actions and their specific parameter configurations that cannot be matched with the current geometric state and industrial design logic, such as "this area is identified as a mounting hole, but the shape irregularity exceeds the threshold." This feedback is then input into an embedded encoder (such as a BERT-like model) to generate a 512-dimensional semantic correction signal. The system first identifies the corresponding subset of point cloud problems based on the feedback. (Typically 100 to 500 points), and calculate the local loss only in this local area:

[0155]

[0156] The correction gradient of the fused semantic signal is then calculated. :

[0157]

[0158] in The learnable weight matrix for the last layer of the point cloud segmentation network has an output dimension equal to the number of semantic categories. For the corresponding feature representation, Indicates local loss For the last layer of weights The gradient, i.e., when The rate of change of the loss function when small changes occur. Represents the feature vector of the last layer Weights The gradient is used to transform the semantic correction signal. Inject weight update. Here are the semantic fusion coefficients (typically 0.2). The learnable weight matrix update rule is defined as follows:

[0159]

[0160] Learning rate Typical initial learning rate Learning rate decay coefficient , This is an iterative step. The update process loops 1 to 5 times until the shape irregularity is below a threshold;

[0161] Irregularity is defined as curvature variance ,in, The number of points in a subset of the point cloud problem is used for normalization and variance calculation. To traverse each point in the subset; For point The local curvature was estimated using principal component analysis (PCA); This represents the mean local curvature.

[0162] If the accuracy requirements are still not met after multiple iterations and the confidence level is below 0.7, manual intervention is triggered. Through this closed-loop mechanism, the system can achieve semantic-aware feedback-driven self-correction and dynamic learning during the inference process, continuously improving the geometric accuracy and semantic consistency of the segmentation results, significantly reducing reliance on manual correction, and providing a high-precision, interpretable, and evolvable 3D basic model for subsequent intelligent modeling and parameterized OCC code generation.

[0163] S3: A semantic instruction parsing method based on physical constraints and causal alignment intervention. It receives natural language instructions from the user and extracts a rich semantic feature vector from a semantic 3D model with component semantic labels from a point cloud understanding architecture. The VLA semantic instruction understanding module then predicts and analyzes the user's actual natural language requirements.

[0164] This step automatically converts the semantic 3D model obtained through the preceding steps into executable parametric CAD modeling code. The method of this invention introduces a visual language causal reasoning mechanism and a hierarchical fuzzy reasoning engine onto the traditional geometry-code mapping, achieving adaptive intelligent generation from visual semantics to parametric code. It can also perform online calibration and iterative optimization of the modeling results based on uncertainty feedback, thus forming a closed loop of code generation with physical rationality and industrial interpretability. Its main process is shown in Figure 6, and the specific steps are as follows.

[0165] S301: Based on the generated semantic 3D model, a semantically rich feature vector is extracted and generated. This vector is a concatenation of functional labels, geometric parameters, and spatial relationship information. Functional labels are predicted labels directly from each point in the point cloud. For example, the system knows which points belong to "mounting surfaces" and which belong to "threaded holes." Geometric parameters are calculated from the labeled point set. Spatial relationships are obtained by comparing the geometric parameters of different components. The feature vector for each component is defined as follows:

[0166]

[0167] in, Let gi represent the semantically rich feature vector of the i-th component, and let gi represent the geometric dimension feature vector. Represents surface curvature and topological complexity; This is the embedding vector of the semantic label of the component; It represents the rotation angle of the component around the global X-axis, used to characterize the component's attitude change in the pitch direction; It represents the rotation angle of the component around the global Y-axis, used to characterize the attitude change of the component in the yaw direction; It represents the rotation angle of the component around the global Z-axis, used to characterize the attitude change of the component in the roll direction; Confidence level of component functional attributes.

[0168] S302: Design a VLA semantic instruction understanding module responsible for parsing user natural language input instructions (e.g., "thicken the shell" or "add a keyway to the shaft end") into actionable semantic units (ActionUnits). This module is based on a visual semantic alignment network structure, the core of which is a reasoning mechanism that jointly embeds physical common sense and design rules. The VLA semantic instruction understanding module first parses the operation intention from the user's natural language to describe the design goal or functional adjustment direction that the user expects to apply. Based on this, and combining the semantic features and geometric context information of each component, the operation intention is further instantiated into a set of executable modeling action candidate sets.

[0169] First, a language encoder is used to map user input commands (such as increasing the stiffness of the base plate) into embedding vectors. At the same time, the semantically rich feature vector corresponding to each component is mapped to a component embedding representation. Calculate the strength of causal relationship:

[0170]

[0171] in This serves as the prior confidence bias, reflecting the operational frequency of each component in historical data. The resulting... This indicates the strength of the causal relationship between the instruction and the component. The system selects the association with the strongest causal relationship. The corresponding component is used as the target operation object. Ultimately, VLA outputs a structured set of action candidates. :

[0172]

[0173] in Indicates the target object. Candidate actions (such as stretching, cutting, chamfering, etc.) For action confidence.

[0174] S303: Design a hierarchical fuzzy inference engine, which receives an action candidate set. The modeling action, along with the geometric parameters in the semantically rich feature vector, is used as input. A pre-defined fuzzy rule base is used to perform parameter-level reasoning, filtering, and correction on candidate actions, thereby determining the modeling action and its specific parameter configuration that best match the current geometric state and industrial design logic. For each geometric parameter (such as thickness, angle, curvature), a fuzzy membership function is defined. Taking the thinner and thicker parts of the fuzzy semantic set as examples:

[0175]

[0176]

[0177] in, The values ​​of the geometric parameters to be evaluated are preferred to be the actual physical dimensions of the component, such as local thickness values, wall thickness values, or structural height values. Representing geometric parameters The membership degree of "thin" in the fuzzy semantic set has a range of values. The larger the value, the more likely it is to be true. The more it fits the semantic concept of "thin"; Representing geometric parameters The membership degree of "thick" in the fuzzy semantic set also has a range of values. The larger the value, the more likely it is to be true. The more it fits the semantic concept of "thicker"; The central parameter value represents the semantic meaning of "thin", used to characterize the "thin" structural dimensions that are considered most representative in industrial experience or design specifications; This represents the center parameter value corresponding to the "thicker" semantic. The scale control parameter, representing the semantics of "thin", is used to adjust the width of the fuzzy membership function; : A scale control parameter representing the semantics of "thickness", its function is related to Similarly, it is used to control the fuzzy extension range of the "thicker" semantics; The natural exponential function is used, and the membership function above adopts a Gaussian form to ensure that the membership degree changes continuously near the central value and decays smoothly when it is far from the central value. For example, when the thickness is 2.3mm, the system can determine that its membership degree is too thin (0.8) and normal (0.4).

[0178] Preferably, an expert fuzzy rule base is designed based on industrial experience initialization. This is a decision core that encapsulates domain expertise and physical common sense. It maps natural language instructions and precise geometric parameters into specific modeling operations through M "IF-THEN" fuzzy rules. This rule base deeply integrates classic design specifications, CAE simulation experience, and manufacturing process constraints (such as stiffener layout criteria, stress concentration handling principles, and draft angle requirements). Its rules handle continuous concepts such as "too thin" and "too tall" through fuzzy membership functions, and assign dynamic weights to each rule to reflect its historical reliability. When receiving an intention such as "reinforce the base plate" and current thickness and stress data, the rule base can simultaneously perform parallel reasoning based on physical causal logic and industrial best practices to calculate an optimal solution that is both reasonable and manufacturable, such as "increase the thickness by 1.5mm and add a grid-shaped stiffener." Finally, it outputs the precise parameters that drive the CAD software execution through defuzzification. Each rule... The format is as follows:

[0179]

[0180]

[0181] The meanings of each component are as follows:

[0182] This indicates that the target object of the current rule is the first... One industrial component, among which These are derived from component instances already parsed in the aforementioned semantic 3D model;

[0183] This indicates that the thickness parameter of the component has a high degree of membership in the fuzzy semantic set "Thin". This judgment is based on the fuzzy membership function corresponding to the thickness. Give;

[0184] This indicates that, under the current operating or analytical conditions, the stress level of the component has a relatively high degree of membership in the fuzzy semantic set "High". This judgment is based on the fuzzy membership function corresponding to the stress parameter. Give;

[0185] This indicates that when the rule prerequisites are activated, the system suggests performing the specific modeling operation of "adding or reinforcing ribs".

[0186] Indicates the first The rule weight of a fuzzy rule is used to characterize the reliability and priority of the rule in industrial experience or historical tasks. Its value can be set by expert experience or adaptively learned and updated through historical execution results during system operation.

[0187] Calculate the activation intensity of the rule, taking thickness and stress parameters as examples:

[0188]

[0189] in: Indicates the first The activation strength of the fuzzy rule, with a value range of . The larger the value, the more the rule conforms to the current geometric and physical state of the component; This indicates the current thickness parameter value of the component; This indicates the current stress parameter value of the component; Represents the membership function of stress parameters;

[0190] The sharpening result of the output fuzzy set is determined by the defuzzification function:

[0191]

[0192] in The output action parameters for the i-th rule are used to obtain the optimal operation parameters. (e.g., thickness increment = 1.2mm).

[0193] S304: A parameterized OCCT code generation method based on hierarchical fuzzy reasoning generates code sequences by receiving action candidate sets and specific geometric parameters extracted from the semantic 3D model.

[0194] Based on the clarified results from the fuzzy inference output, the system enters the automatic OCCT code generation stage. The code generator operates in the form of "topological constraints + syntax templates" to ensure the geometric rationality and syntactic correctness of the generated code. During this process, the system automatically verifies whether the input surface set is closed and whether the topological constraints satisfy the constructibility conditions. Subsequently, code perfection verification inference is performed: after executing the generated OCC code, the system will construct a 3D CAD model in a virtual OCC sandbox.

[0195] The input face set is an intermediate geometric representation explicitly constructed from the aforementioned semantic 3D model and inference results. Specifically, in completing the action candidate set... After the screening and fuzzy inference parameters are determined, the system first determines the target component. In the semantic 3D model, the geometric description is performed on the corresponding point cloud subset or mesh region. The patch-level reconstruction and semantic aggregation are performed to generate a set of candidate geometric surfaces. Each surface is obtained by fitting the point set or mesh fragment under the semantic label of the same component. It can be a plane, cylindrical surface, conical surface or free surface. The input surface set also carries surface normal, boundary contour and adjacent topological relationship information for subsequent topological construction and constructability verification.

[0196] After the generated OCC code is executed in the virtual OCCT sandbox environment, the system will perform a multi-dimensional quantitative comparison between the reconstructed CAD model and the original semantic 3D model to obtain the geometric error. Semantic consistency error and physical rationality indicators Based on the above indicators, a fuzzy comprehensive reward function is defined. :

[0197]

[0198] in These are adjustable weighting coefficients. If the CAD model fails geometric verification after the generated code is executed (e.g., volume discontinuities or normal conflicts), then... Below the threshold The system will automatically activate the calibration process (adaptive feedback mechanism) and update the corresponding fuzzy rule weights. , Update the coefficients for the fuzzy rule weights; then adjust the VLA action confidence. Finally, the OCC code is regenerated and re-simulated.

[0199] Wherein, the geometric error Semantic consistency error and physical rationality indicators All methods are constructed based on existing geometric deviation metrics, semantic consistency evaluation methods, and industrial manufacturing constraint rules in this field, and are used to quantitatively evaluate the reconstructed model from three dimensions: geometric accuracy, semantic preservation, and engineering manufacturability. When the generated model fails the geometric or topological verification due to insufficient construction of the input surface set or unreasonable inference parameters, the value of the fuzzy comprehensive reward function will be significantly reduced, thereby triggering an adaptive feedback mechanism to jointly calibrate the fuzzy rule weights and action confidence.

[0200] S4: Applying machine learning models in industrial environments presents a series of challenges because the underlying production systems and their subsystems are constantly evolving, affecting model accuracy and accountability over time. These factors include environmental factors. Therefore, this study proposes an adaptive 3D modeling framework based on 3D geometric concept drift detection and continuous learning. This framework constructs an autonomous intelligent learning agent capable of continuously learning industrial knowledge and design theories. This agent automatically performs self-learning and optimization processes upon detecting model performance degradation or distribution shifts, achieving long-term stable model updates. As shown in Figure 7, the specific process is as follows.

[0201] S401: First, the concept of drift in 3D modeling is proposed. In 3D modeling of industrial parts, conceptual drift can be defined as the systematic changes in the geometric features, physical properties, or performance requirements of a part during its design, manufacturing, and use. These changes cause the original 3D model to gradually deviate from the actual situation. Analogous to drift in part cleaning load prediction (such as energy mode shift caused by changes in the cleaning cycle), 3D modeling drift often manifests as gradual (slow aging) or sudden (sudden damage), and can be adapted in real time through continuous learning technology.

[0202] To quantitatively detect feature drift, this invention employs a sliding window statistical method to monitor changes in feature distribution. The specific process is as follows: First, multidimensional geometric and semantic statistical features are extracted from the reconstructed CAD model. Then, within a continuous time period, the probability distribution of the current multidimensional geometric and semantic statistical features is... The probability distribution at the previous time step was: The Kullback-Leibler (KL) divergence is used to quantify the difference between the two concepts. When the KL divergence exceeds a drift threshold, the system determines that there is significant concept drift and triggers a continuous learning mechanism. Furthermore, for the geometric features output by the modeling (such as surface normals, curvature distribution, and assembly hole position deviations), the system introduces auxiliary drift detection metrics:

[0203]

[0204] in This represents the probability distribution of the geometric features of the current model. Represents the historical baseline distribution. The L2 norm is used to measure the linear distance between two vectors in a multidimensional feature space; when the calculation result... When the value exceeds a set threshold, a subsequent continuous learning phase is triggered.

[0205] S402: When the system detects concept drift, it enters a continuous learning phase to maintain the model's ability to acquire new knowledge and prevent the forgetting of existing geometric knowledge. To this end, this invention proposes a dynamic adaptive learning strategy based on regularized information coverage. By calculating the rate of change of the loss function, the balance between the agent and old knowledge is monitored, thereby ensuring that the model simultaneously possesses the knowledge acquisition ability required to overcome the confidence index and the stable memory ability required to overcome confidence decay.

[0206] Preferably, the present invention designs a geometric feature importance assessment mechanism based on elastic weight consolidation to protect key geometric parameters. This mechanism quantifies each network parameter by calculating the Fisher information matrix of network parameters against historical task data (especially key geometric features) to identify network parameters that play a crucial role in the function of the part. The importance of learned geometric shapes and functional structures. The diagonal elements of the Fisher information matrix. A higher value indicates that the network parameter is more important for modeling the geometric features of historical tasks, and should be subject to stronger constraints in subsequent learning to prevent the forgetting of important geometric features. Let the current task number be... The network parameters are , indicating the current task The set of network parameters below. The total number of network parameters includes all trainable parameters (such as weight matrices and biases) in the neural network of this method (point cloud understanding module, feature reconstruction module, semantic reasoning module, etc.), which participate in gradient updates during training.

[0207] The dynamic adaptive geometric regularization loss function for the current task K is defined as:

[0208]

[0209] in, This represents the total number of parameters. This is the standard loss function for the current task, used to measure the model's predicted output. With real labels The error between; Representing historical tasks Network parameters The optimal value, The diagonal elements of the Fisher information matrix corresponding to this historical task are used to quantify the network parameters. The value indicates the importance of key geometric features or functional structures; the larger the value, the more critical the parameter. The geometric regularization intensity hyperparameter is used to control the degree of protection for key parameters, thereby limiting the over-updating of key parameters during continuous learning and ensuring that the model retains the learned geometric knowledge and functional structure.

[0210] Preferably, to meet the industrial demand for online learning of continuously arriving 3D scanning data streams (i.e., continuous, real-time sequences of 3D point cloud and mesh information used to capture changes in the geometric and semantic features of parts), this invention further designs an online geometric elastic weight consolidation mechanism for continuous 3D scanning. The core advantage of online learning lies in its dynamic update strategy; the system no longer needs to store the complete Fisher information matrix separately for each historical task. Instead of using a traditional method, it maintains a streamlined integration matrix through weighted accumulation. This mechanism achieves lightweight storage and rapid adaptation by dynamically maintaining a weighted accumulated Fisher information matrix. Its update formula is as follows:

[0211]

[0212] Specifically, This represents the accumulated Fisher information value for the current task, used to quantify the contribution of this parameter to key geometric features and functional structures; This represents the cumulative value from the previous task; This provides real-time Fisher information for the parameter in the current task, reflecting its importance in modeling the current geometric features; This is a historical information attenuation coefficient, used to weight and attenuate long-term task information during the accumulation process, thereby balancing the importance of recent and historical tasks; initial conditions. This indicates that the cumulative matrix of the first task is its Fisher information matrix. Through this dynamic accumulation strategy, this mechanism achieves lightweight storage and rapid adaptation while ensuring the protection of key geometric parameters. This enables the system to efficiently update model parameters online in the continuously arriving 3D scan data stream, while preventing the forgetting of learned knowledge.

[0213] Through this dynamic adjustment mechanism, the system achieves self-balancing learning in a 3D modeling environment. It can quickly adapt to changes in industrial part design rules or geometric distributions while maintaining an accurate understanding of the learned geometric structures, thus ensuring high stability and evolutionary capability of the model during long-term operation. The efficiency of the method in this invention compared to traditional methods is shown in Figure 8. The invention simulates the performance evolution of the model under long-term operation in an industrial environment. The blue curve represents the traditional model (without an adaptive mechanism), whose performance initially improves slightly but then quickly forgets and drops below the initial baseline. The orange curve represents the Elastic Weight Consolidation (EWC) method, which only alleviates the forgetting phenomenon but still shows a slow downward trend. The green curve represents the General Continuous Learning (GEWC) method, which can basically maintain the performance level but without significant improvement. The red curves represent the two implementations of this invention. The thinner red curve represents the method of geometric elastic weight consolidation combined with geometric perception distillation, which has shown a clear trend of continuous performance improvement. The thicker red curve represents the complete framework (concept drift detection + adaptive learning + geometric perception distillation), which has the most significant performance improvement, maintaining a stable increase throughout the entire time step, far exceeding all compared methods.

[0214] S403: The elastic weight consolidation mechanism effectively anchors key geometric features by imposing constraints in the parameter space. To achieve more refined knowledge preservation at the output and feature levels, this invention proposes a geometry-aware distillation framework. Based on knowledge distillation technology, this framework ensures a smooth transition between the old and new models in terms of semantic understanding of 3D shapes and geometric feature extraction capabilities through the dual constraints of soft target distillation and geometric feature alignment.

[0215] Soft-objective distillation: When training a new task, a network model capable of modeling historical tasks is fixed as the teacher network, and the currently learned network model is used as the student network to fit the probability distribution of the teacher network's output (soft objective). Compared to using only the true labels (hard objective), soft objectives contain rich information about the relationships between categories and can more effectively convey historical knowledge.

[0216] Geometric Feature Alignment: To address the characteristics of 3D modeling, an alignment loss for intermediate layer geometric features is further introduced. These intermediate layer geometric features refer to the encoded representations of the 3D model's geometric information by each hidden layer within the neural network. They do not directly output semantic labels but reflect the model's understanding of the point cloud / geometric structure. This loss directly constrains the consistency of the student network and the teacher network's outputs on key geometric feature maps, forcing the new model to inherit the ability to model geometric properties such as point cloud distribution, surface continuity, and topological structure at the internal feature extraction level.

[0217] Therefore, the geometry-aware distillation loss function for multi-objective optimization as follows:

[0218]

[0219] in, This represents the geometric distillation weighting coefficient, used to balance the contribution of the new task loss and the distillation constraint; Indicates the current task Task loss (e.g., cross-entropy loss for point cloud segmentation) ensures the model's basic performance on new tasks; This represents the total number of historical tasks, used to normalize the accumulated distillation constraints. This indicates that all historical tasks are iteratively accumulated to achieve multi-task distillation; This represents the soft-target distillation loss, used to constrain the current student network output. Teacher network output for historical tasks Consistency in the final class probability distribution; ξ represents the feature alignment weight coefficient, used to balance the ratio of soft target distillation and geometric feature alignment loss; This represents the geometric feature alignment loss, which constrains the current student network in the intermediate feature layers. Feature layer corresponding to the historical task teacher network This ensures consistent output, thereby guaranteeing that critical geometry is preserved.

[0220] This loss function optimizes both the performance of new tasks and the preservation of historical geometric knowledge, enabling the student network to smoothly inherit key 3D structural and semantic information when learning new tasks, thus avoiding the forgetting of key geometric features during continuous learning.

[0221] Furthermore, this framework includes two optimization strategies:

[0222] To address the multi-layered feature representations formed in the aforementioned point cloud understanding module, feature reorganization module, and semantic reasoning module, differentiated knowledge transfer intensities are implemented at different network layers. Specifically, the lower layer corresponds to the geometric detail features formed during point cloud voxelization and local geometric encoding; the middle layer corresponds to the local structural features formed during neighborhood feature reorganization and component-level structural modeling; and the higher layer corresponds to the global component functional semantic features output by the semantic reasoning module. By introducing distillation constraints at each of these layers, comprehensive knowledge retention, from minute geometric structures to macroscopic shape and functional semantics, is ensured.

[0223] Dynamic distillation weight adjustment: Dynamically adjust the geometric feature alignment loss based on the importance of geometric features. The weighting of features is adjusted. Stronger alignment constraints are imposed on feature maps involving critical functional areas (such as mounting holes and load-bearing wall panels), while greater adaptability is allowed for secondary features, thereby achieving efficient and accurate knowledge consolidation.

[0224] To verify the effectiveness of the aforementioned geometry-aware distillation framework, comparative experiments were conducted on a typical industrial parts 3D semantic segmentation dataset. The experimental results are shown in Figure 9. The red curve represents the traditional fine-tuning method, which exhibits significant overfitting in the later stages of training, leading to a significant decrease in semantic accuracy. The purple curve represents the standard knowledge distillation (Vanilla KD) method, which, while mitigating overfitting, has a slow convergence speed and limited final accuracy. The blue curve represents the proposed geometry-aware distillation (GMAS) method, which demonstrates the fastest convergence speed and highest final accuracy throughout the entire training process, and stably approaches the theoretical upper limit of knowledge distillation (green dashed line) within 25 training epochs. This fully demonstrates that the dual constraint mechanism of soft-target distillation combined with geometric feature alignment can significantly improve the model's continuous learning ability and geometric knowledge retention capability in complex industrial scenarios.

[0225] Finally, through hyperparameters Together with the contribution of these three losses in ξ precise balancing, they guide the model to solidly maintain its historical geometric cognition while learning new knowledge.

[0226] This invention aims to construct an intelligent fuzzy inference code generation method for OCCT (Open CASCADE Technology) based on physical constraint VLA, realizing the entire process of autonomous intelligent modeling from raw visual data to high-fidelity, directly driveable industrial CAD models.

[0227] The core innovation of this method lies in constructing an end-to-end intelligent pipeline of "perception-understanding-reasoning-generation-evolution". At the perception layer, the system solves industrial challenges such as weak texture and high reflectivity through an industrially prior-guided progressive deep inference engine, generating millimeter-level precision 3D point clouds from multi-view images. Building upon this, a physically rationality-driven adaptive point cloud understanding architecture, by introducing a component-aware graph attention aggregation mechanism and a physical constraint training paradigm, achieves a profound understanding from raw geometry to component-level functional semantics, completing the "semantic" transformation of 3D data.

[0228] At the cognitive and decision-making levels, the system possesses powerful intent understanding and reasoning capabilities. A visually semantically aligned instruction understanding module parses the user's natural language through causal reasoning, transforming it into structured semantic actions. Subsequently, a hierarchical fuzzy reasoning engine utilizes an expert rule base initialized with industrial knowledge and physical common sense to fuse operational intent with precise geometric parameters, simulate expert decision-making, and output optimal modeling parameters.

[0229] At the execution and optimization layer, this invention proposes an automatic OCCT code generation technology that integrates fuzzy reasoning, transforming decisions into geometrically correct code sequences that conform to "topological constraints + syntax templates." More importantly, the system embeds a closed-loop self-evolving engine that integrates "simulation-verification-diagnosis-update." By executing code in the OCC sandbox, reconstructing the model, and performing multi-dimensional quantitative comparisons, it generates a fuzzy reward signal that comprehensively considers geometric fidelity, semantic consistency, and physical rationality. This signal then dynamically feeds back to correct the inference rules and model parameters at the front end, ensuring high reliability and robustness of the output.

[0230] To ensure the long-term viability of the system, this invention proactively introduces a continuous learning framework based on 3D geometric concept drift detection. This framework actively monitors the evolution of industrial design data and the environment, and through mechanisms such as geometric elastic weight consolidation and geometric perception knowledge distillation, enables the core model to continuously absorb new knowledge while firmly maintaining its memory of existing key geometric structures and functional features. This ensures that the method has the ability to self-evolve in dynamic industrial environments.

[0231] In summary, this invention, through the deep integration of five technological pillars—high-precision perception, semantic understanding, intelligent reasoning, automated generation, and closed-loop evolution—creates an autonomous intelligent code generation method capable of understanding images, comprehending instructions, thinking according to physical laws, generating precise code, and continuously learning and improving. It fundamentally changes the traditional CAD system's reliance on manual labor and fragmented processes, providing a complete solution for industrial design towards full-process intelligence and automation.

[0232] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for generating intelligent fuzzy inference code for OCCT based on physical constraint VLA, characterized in that, Includes the following steps: S1: Alignment preprocessing is performed on multi-view images of industrial parts, and then multi-view shared features containing global geometric context information are generated based on a geometric self-attention mechanism that incorporates physical rationality constraints. Based on appearance and spatial information, the shared features of multiple view domains are decoupled. Then, the features of appearance and spatial information are updated using industrial structure priors and saliency weights reflecting the importance of component regions. The updated features are converted into dense depth maps. Finally, after optimization using a nonlinear filtering module and an edge-aware repair algorithm, a 3D point cloud model is obtained by back-projection. S2: Construct a point cloud segmentation network. First, a multi-scale industrial semantic graph structure is constructed based on the 3D point cloud model, including a geometric proximity graph based on geometric spatial information, a functional semantic graph based on functional regions, and a manufacturing constraint graph based on manufacturing process constraints. Then, the geometric proximity graph, functional semantic graph, and manufacturing constraint graph are dynamically fused to obtain a dynamic fused graph. The nodes of the dynamic fused graph correspond to a point or a feature aggregation unit in the 3D point cloud model. After calculating the association weights of nodes in the dynamic fusion graph based on local-global interactive attention, neighboring nodes are weighted based on their association weights to obtain enhanced features of the nodes. Then, the enhanced features are updated based on a learnable industrial semantic interaction matrix. Finally, a classifier is used to obtain the component semantic descriptors of the nodes, completing the semantic reorganization processing of all points in the point cloud and obtaining a semantic 3D model rich in component semantic descriptors; S3: Based on the semantic 3D model features, rich semantic feature vectors are extracted. At the same time, the user's natural language input commands are mapped into embedding vectors and combined with the rich semantic feature vectors to obtain a set of action candidates. The candidate actions are subjected to parameter-level reasoning, screening and correction through a preset fuzzy rule base to determine the modeling action and its specific parameter configuration that best matches the current geometric state and industrial design logic; Finally, OCCT code is automatically generated to obtain the reconstructed CAD model; S4: Based on 3D geometric concept drift detection and continuous learning; Multidimensional geometric and semantic statistical features are extracted from the reconstructed CAD model. Concept drift is detected by calculating the probability distribution of the model under new and old tasks. When the system detects concept drift, it enters the continuous learning stage. The balance between the model and the old knowledge is monitored by calculating the change rate of the loss function. When training a new task, soft target distillation and geometric feature alignment between the new and old tasks are used to ensure a smooth transition in the semantic understanding and geometric feature extraction capabilities of the new and old models in three-dimensional shape.

2. The OCCT intelligent fuzzy inference code generation method based on physical constraint VLA according to claim 1, characterized in that, S1 is specifically as follows: acquire multi-view images of industrial parts and perform alignment preprocessing; the aligned multi-view images are first input into a geometric self-attention mechanism that incorporates physical rationality constraints to generate multi-view shared features containing global geometric context information; in the multi-view shared feature generation stage, based on a domain-specific projection and constraint mechanism, the multi-view shared features are decomposed into appearance domain features that encode the texture, brightness, and local appearance change information of the image, and geometric domain features that encode depth cues, disparity relationships, and spatial structure information; Based on the pre-defined industrial structure priors, the component-level importance of multi-view shared features is evaluated, and the significant weights reflecting the geometric contribution of different component regions in industrial parts are output. Based on the multi-view consistency constraint, the shared features from different viewpoints are aligned and fused to obtain the multi-view matching consistency measure. Based on saliency weights and multi-view matching consistency measures, appearance domain features and corresponding geometric domain features under different perspectives are co-encoded to obtain the final multi-view features. Based on the original image patches and multi-view features, appearance perception feature extraction and geometric structure feature extraction are performed. The two feature vectors obtained are used to update the appearance domain features and geometric domain features, and the optimization is continuously iterated until it stops. The final multi-view features are used to generate a dense depth map with millimeter-level precision. The dense depth map and the multi-view features are jointly encoded to construct a unified reversible 3D descriptor. The dense depth map and the unified reversible 3D descriptor are fused in parallel with a deformable 3D roll and a graph attention machine to obtain a fused feature representation. Then, the fused feature representation is optimized by a nonlinear filtering module and an edge-aware repair algorithm to obtain an ultra-high precision depth map. Finally, point cloud back projection calculation is performed to transform the 2D depth information of the ultra-high precision depth map into a 3D point cloud model.

3. The OCCT intelligent fuzzy inference code generation method based on physical constraint VLA according to claim 2, characterized in that, S2 is as follows: S201: Construct a multi-scale industrial semantic graph structure based on a 3D point cloud model; the geometric proximity graph is constructed based on the k nearest neighbors of the point cloud spatial coordinates; the functional semantic graph is based on the functional area division extracted from industrial design knowledge, and each point is initially labeled with a lightweight functional classifier, and points of the same functional category are connected. The manufacturing constraint graph connects point pairs that satisfy specific manufacturing relationships based on manufacturing process constraints; S202: Based on the adjacency matrix of the graph structure and their respective learnable weight matrices, the three graph structures are dynamically fused to obtain a learnable dynamic fused graph; the attention weights between nodes in the dynamic fused graph are calculated through local-global interactive attention, as shown in the following formula: ;in, Represents nodes in the dynamic fusion graph 、 Attention weights between nodes; nodes refer to vertices of the dynamic fusion graph, with each node corresponding to a point or a feature aggregation unit in the point cloud; For activation functions; and Representing nodes respectively With nodes eigenvectors; These are learnable fusion weight parameters; For local feature interaction functions; For mapping functions; These are prior features of industrial semantics; Manufacturing functions constrained by process technology. Encoding the process constraint relationships between nodes. This indicates a gated feature fusion operation; the feature points to be processed in the point cloud model. As the core point, for each neighboring point Perform a sequential traversal, based on attention weights The input features of neighboring points are weighted and modulated as follows: ;in, For point Enhanced features, Represents the set of neighborhood points. This indicates element-wise gated modulation operation. As the modulation vector, through a learnable industrial semantic interaction matrix The calculation yielded: ;in, and Representing points respectively The predicted functional category probability distribution of j; the enhanced features are input into the industrial semantic interaction query module and combined with the learnable industrial semantic interaction matrix. Perform semantic consistency matching; if the matching result does not meet the preset semantic confidence threshold, then update the parameters of the industrial semantic interaction matrix in reverse and return to the aforementioned weighted modulation step to re-execute; if the matching result meets the semantic confidence requirement, then input the enhanced features of each node into a lightweight point-independent MLP classifier and output the component semantic descriptor; S203: During the training phase, perform joint optimization based on the cross-entropy loss function, component connectivity constraints, and boundary consistency constraints; the component connectivity constraints identify connected regions in the point cloud through the flood filling algorithm, and the boundary consistency constraints identify connected regions in the point cloud through the point cloud. The similarity of normal vectors is measured by boundary alignment; S204: A closed-loop semantic correction unit is introduced in the inference stage to receive semantic feedback from S3. The semantic feedback is specifically the semantic representation of modeling actions and their specific parameter configurations that cannot be matched with the current geometric state and industrial design logic. The semantic feedback is input into the embedded encoder to generate a semantic correction signal. Based on the feedback, the corresponding point cloud problem subset is identified. The cross-entropy loss and boundary consistency constraint are calculated based on the point cloud problem subset. The correction gradient is calculated by fusing the semantic correction signal. The learnable weight matrix of the last layer of the point cloud segmentation network is updated based on the correction gradient.

4. The OCCT intelligent fuzzy inference code generation method based on physical constraint VLA according to claim 3, characterized in that, Specifically, S3 is as follows: S301: Extract semantically rich feature vectors based on the semantic 3D model, including point function labels, geometric parameters, and spatial relationship information; S302: Use a language encoder to map user natural language input commands into embedded vectors. Meanwhile, the semantically rich feature vector corresponding to each component is mapped to a component embedding representation. Calculate the strength of causal relationship: ;in This is a priori confidence bias. The causal relationship between instructions and components is represented by d, where d is the feature dimension. Select the association with the strongest causal relationship. The corresponding component is used as the target operation object, and the final output is a structured set of action candidates. ： ;in Indicates the target object. As candidate actions, S303: Design a hierarchical fuzzy inference engine that receives geometric parameters from the action candidate set and semantically rich feature vectors as input. It performs parameter-level inference, filtering, and correction on candidate actions using a pre-set fuzzy rule base to determine the modeling action and its specific parameter configuration that best matches the current geometric state and industrial design logic. Specifically, for each geometric parameter, a fuzzy membership function is defined, and an expert fuzzy rule base initialized based on industrial experience is designed to map natural language instructions and geometric parameters to specific modeling operations. After defuzzification, the clarified result of the fuzzy set is output. S304: Based on the clarified result, OCCT code is automatically generated. The reconstructed CAD model is compared with the original semantic 3D model in multiple dimensions to obtain geometric error, semantic consistency error, and physical rationality indicators. A fuzzy comprehensive reward function is defined based on these indicators. If the fuzzy comprehensive reward function is below a threshold, the corresponding fuzzy rule weights are updated, the action confidence is adjusted, and finally, the OCC code is regenerated and re-simulated.

5. The OCCT intelligent fuzzy inference code generation method based on physical constraint VLA according to claim 4, characterized in that, S4 is specifically as follows: S401: Extract multidimensional geometric and semantic statistical features from the reconstructed CAD model. Subsequently, within a continuous time period, calculate the probability distribution of the multidimensional geometric and semantic statistical features at the current time and the probability distribution at the previous time. Use KL divergence to quantify the difference between the two. When the KL divergence is greater than the drift threshold, the system determines that there is significant concept drift and triggers the continuous learning mechanism; or it judges concept drift by calculating the distance between the probability distributions of the new and old geometric features; S402: Enter the continuous learning phase; Let the current task number be... ， Indicates the current task The following network parameter set, For network parameter indexing, The total number of network parameters, including all trainable parameters, which participate in gradient updates during training; the dynamic adaptive geometric regularization loss function for the current task K. Defined as: ;in The standard loss function for the current task measures the model's predicted output. With real labels The error between; Representing historical tasks Network parameters The optimal value, The diagonal element of the Fisher information matrix corresponding to this historical task. Let be the geometric regularization intensity hyperparameter; the update formula for the Fisher information matrix is ​​as follows: ;in, This represents the accumulated Fisher information value for the current task. This represents the cumulative value from the previous task; This provides real-time Fisher information for this parameter in the current task. Historical information attenuation coefficient; initial conditions This indicates that the cumulative matrix of the first task is its Fisher information matrix; S403: Through the dual constraints of soft target distillation and geometric feature alignment, a smooth transition is ensured between the old and new models in terms of semantic understanding of 3D shapes and geometric feature extraction capabilities, with a multi-objective optimized geometric-aware distillation loss function. as follows: ;in, Indicates the geometric distillation weighting coefficient; Indicates the current task Task loss; This represents the soft-target distillation loss and constrains the student network output for the current task. Teacher network output for historical tasks Consistency in the final category probability distribution; ξ represents the feature alignment weight coefficient; This represents the geometric feature alignment loss, which constrains the current student network in the intermediate feature layers. Feature layer corresponding to the historical task teacher network Consistency of output.

6. The OCCT intelligent fuzzy inference code generation method based on physical constraint VLA according to claim 5, characterized in that, Based on the camera geometry parameters and depth assumptions during alignment preprocessing, a 3D spatial relative position code is constructed. The geometric self-attention mechanism, which incorporates physical rationality constraints, introduces the relative position code as an additive bias term into the attention weight calculation.

7. The OCCT intelligent fuzzy inference code generation method based on physical constraint VLA according to claim 6, characterized in that, The multi-view feature is obtained through the following steps: the multi-view matching consistency metric calculates the consistency of the similarity of shared features across multiple views and introduces three-dimensional geometric projection consistency. As a weighted constraint; specifically for any candidate space assumption Let it be in the first The joint feature response from each perspective is Multi-view matching consistency metric score The calculation is as follows: ； ;in, The number of effective perspectives participating in the integration; Represents the normalization function. It is a very small constant. Denotes the weighted norm; For reference perspective; Indicates cross-view normalized similarity; This represents the joint characteristic response from the reference perspective; Indicates the candidate hypothesis in perspective The three-dimensional coordinates obtained by downward back projection. Its multi-view mean; Represents the variance operator. The L2 norm is used to represent the joint features obtained by weighting the industrial structure prior, geometric domain features, and multi-view matching consistency measure. Finally, based on the joint features, a matching cost fusion strategy is executed to integrate appearance similarity and geometric consistency information to generate multi-view features. Here, appearance similarity information refers to the matching similarity between appearance domain features under different viewpoints, and geometric consistency information refers to the consistency of three-dimensional geometric projection.

8. The OCCT intelligent fuzzy inference code generation method based on physical constraint VLA according to claim 7, characterized in that, The component connectivity constraints and boundary consistency constraints are calculated as follows: when predicting the component Number of connected components When the value is greater than 1, define the component connectivity constraint loss. ： ;in As a weighting factor, For components Number of points in the data; boundary consistency constraints ： ;in and These are the predicted boundary point set and the actual geometric edge point set, respectively. To predict boundary points The normalized normal vector, For its real edge point set The nearest neighbor normal vector in the equation; during the inference phase, the gradient is corrected. The calculation is as follows: ;in, The total loss is calculated based on a subset of the point cloud problem, including component connectivity constraints and boundary consistency constraints. The learnable weight matrix for the last layer of the point cloud segmentation network. For the corresponding feature representation, For semantic correction signals, The semantic fusion coefficients are used; the learnable weight matrix update rule is defined as follows: ;in For the updated learnable weight matrix, The learning rate is used; the update process loops until the shape irregularity is lower than a set threshold. If the accuracy requirement is still not met after multiple iterations, manual intervention is triggered.

9. The OCCT intelligent fuzzy inference code generation method based on physical constraint VLA according to claim 8, characterized in that, The hierarchical fuzzy inference engine is specifically defined as follows: a fuzzy membership function is defined for each geometric parameter, taking the thinner part of the fuzzy semantic set as an example: ;in, This indicates the values ​​of the geometric parameters to be evaluated. Representing geometric parameters For the thin membership degree in the fuzzy semantic set, This represents the center parameter value corresponding to the thin semantics. Scale control parameters representing the thinness semantics. The natural exponential function is represented, and the membership function adopts a Gaussian function form. An expert fuzzy rule base initialized based on industrial experience is designed. Through M IF-THEN fuzzy rules, natural language instructions and precise geometric parameters are mapped to specific modeling operations. This expert fuzzy rule base integrates classic design specifications, CAE simulation experience, and manufacturing process constraints, and assigns dynamic weights to each rule. The rule activation intensity is calculated, taking thickness and stress parameters as examples. ;in: Indicates the first The activation strength of a fuzzy rule This indicates the current thickness parameter value of the component; This indicates the current stress parameter value of the component; The membership function represents the stress parameter; the sharpening result of the output fuzzy set is determined by the defuzzification function. ;in The output action parameters for the i-th rule are... These are the optimal operating parameters.

10. The OCCT intelligent fuzzy inference code generation method based on physical constraint VLA according to claim 9, characterized in that, The execution process of the edge-aware repair algorithm includes at least the following steps: First, based on the gradient distribution information of the dense depth map in the spatial domain, the geometric edge response intensity at each pixel location is calculated. Then, by jointly analyzing the first or second-order differences of the depth values ​​in the horizontal and vertical directions, an edge response map reflecting the degree of depth discontinuity and structural abrupt changes is constructed. Edge detection based on the edge response map calculated from the depth gradient is as follows: ;in, The edge response intensity is used to reflect the degree of depth discontinuity or structural abrupt change. This is a dense depth map; This represents the gradient of the dense depth map in the horizontal direction. Let be the gradient of the dense depth map in the vertical direction; construct an energy minimization model with edge-aware weights to optimize the overall depth distribution within the repair area, and its objective function is expressed as: ;in, This is the repaired ultra-high precision depth map. This is the initial depth estimation result. Indicates the area to be repaired. For pixel position, For depth gradient operators; The edge-aware weighting function is adaptively calculated based on the edge response intensity of the dense depth map. ;in, Indicates the geometric edge response intensity at the pixel location. These are edge-sensitive control parameters.

Images