Geometric defect identification and processing method based on AI large model

Abstract

The application discloses a kind of geometric defect identification and processing method based on AI big model, it is related to aerospace engineering field.The method aims at solving the problem of low efficiency and poor accuracy of aircraft CAD model geometric defect artificial repair.The steps include: first, CAD assembly body model is intelligently segmented into multiple sub-components;Second, the part geometry is converted into a structured feature representation, and a multi-task hierarchical graph neural network model incorporating domain knowledge is trained using labeled data;Then, the model is used to identify defects in sub-components, outputting defect type and spatial location information;Finally, according to the identification results, the repair topology is connected first, and the geometric anomaly is processed in sequence, and a watertight surface mesh is generated.The application realizes the whole process automation of geometric defect identification and repair, greatly improves the pre-processing efficiency and quality, and reduces the dependence on expert experience.

Description

Technical Field

[0001] This invention relates to the field of aerospace engineering, and more specifically, to a method for identifying and processing geometric defects based on a large AI model. Background Technology

[0002] In the field of aerospace engineering, the quality of computer-aided design (CAD) models directly determines the accuracy and reliability of subsequent digital engineering processes such as computational fluid dynamics (CFD) analysis and structural strength simulation. As a highly complex system integrator, an aircraft's CAD model typically consists of thousands of surfaces, curves, and topological relationships. During the export from design software, transfer between multiple systems, and model simplification processes, CAD models inevitably develop various geometric defects, including but not limited to: tiny gaps between surfaces, surface overlaps, discontinuous short edges, unnecessary small silver faces, unexpected holes, illegal intersections between surfaces, and inconsistent normals.

[0003] The presence of these geometric defects can severely impact the subsequent meshing process. CFD analysis requires high-quality watertight surface meshes as the computational domain boundary, and geometric defects can directly lead to mesh generation failure, low mesh quality, or difficulty in computational convergence.

[0004] In traditional CFD preprocessing workflows, engineers spend a significant amount of time (typically 60%-70% of the entire analysis cycle) on geometry cleaning and repair. This approach relies heavily on human experience and suffers from the following significant technical drawbacks:

[0005] 1. Inefficient and highly experience-dependent: Engineers need to inspect each surface individually, manually identify defect types, and select repair strategies. The geometry cleanup of a complete aircraft model (such as a commercial airliner) typically takes weeks or even months, severely limiting the speed of design iteration.

[0006] 2. Difficulty in guaranteeing identification accuracy: Manual inspection is prone to omission or misjudgment of defects due to visual fatigue and differences in experience, especially for micron-level gaps or complex situations with multiple intertwined defects, where manual identification is almost impossible to achieve full coverage.

[0007] 3. Lack of systematic defect quantitative assessment: Existing technologies are unable to quantify and classify the severity of defects, and cannot automatically determine repair priorities based on their impact on subsequent analysis. This may cause engineers to spend a lot of time dealing with defects that have little impact on mesh generation.

[0008] 4. Limited recognition capabilities of traditional algorithms: Currently, the automatic defect detection functions in commercial CAD / CAE software are mainly based on simple geometric rules and threshold judgments (such as side length thresholds, included angle thresholds, etc.), which have the following shortcomings:

[0009] It can only detect obvious geometric anomalies on the surface, but cannot understand the design intent and functional semantics;

[0010] It has a weak ability to identify hidden defects in complex topological relationships (such as logical gaps caused by inconsistencies in surface normals);

[0011] The parameter settings are sensitive, and the same set of parameters is difficult to adapt to the different requirements of different parts of the aircraft (such as the leading edge of the wing and the fuselage skin);

[0012] The lack of ability to analyze the correlation between defects makes it impossible to identify systemic quality problems formed by multiple small defects.

[0013] 5. Inability to leverage historical experience: The experience of each engineer is difficult to effectively accumulate and reuse, leading to repetitive work on similar problems. Furthermore, with the departure of senior engineers, the company faces the risk of experience gaps.

[0014] 6. Disconnected from automated repair processes: Existing defect identification tools typically only provide defect marking functions and lack intelligent integration with subsequent automated repair algorithms, thus failing to form a closed-loop workflow of identification-evaluation-repair.

[0015] With the rapid development of artificial intelligence technology, especially the breakthroughs in multimodal understanding using Large Language Models (LLM) and Large Visual Models (VLM), a new technological path has been provided for the intelligent identification of complex geometric defects. However, the application of AI technology in the field of 3D geometry processing still faces the following challenges:

[0016] There is a lack of large-scale, high-quality labeled datasets for the aerospace field;

[0017] Existing 3D deep learning models do not adequately support engineering accuracy requirements;

[0018] There is a lack of effective mechanisms for incorporating domain knowledge (such as aerodynamic requirements and manufacturing process constraints);

[0019] The model has poor interpretability and is difficult to gain the trust of engineers. Summary of the Invention

[0020] The purpose of this invention is to solve the above problems by providing a geometric defect identification and processing method that can intelligently, accurately, and efficiently identify and process geometric defects, thereby overcoming the limitations of traditional methods, significantly improving the level of pre-processing automation, and shortening the research and development cycle.

[0021] To achieve the above-mentioned objectives, this invention provides a method for geometric defect identification and processing based on a large AI model, the method comprising:

[0022] Step 1: Obtain the CAD model of the aircraft to be processed, and divide the CAD model of the aircraft into multiple parts;

[0023] Step 2: Obtain multiple labeled sample parts, and convert the geometric data of each sample part into a structured feature representation for processing by the deep learning model to obtain a training set. The structured feature representation includes at least one of the graph structure, point cloud, and feature vector of the sample part; the labeling includes the type and spatial location information of the geometric defects existing in the sample part.

[0024] Step 3: Use the training set to train a multi-task hierarchical graph neural network model to obtain the trained multi-task hierarchical graph neural network model;

[0025] Step 4: For any component obtained from the segmentation in Step 1, convert the geometric data of the component into the same structured feature representation as in Step 2, and input the converted data into the trained multi-task hierarchical graph neural network model; the trained multi-task hierarchical graph neural network model outputs the type and spatial location information of the geometric defects contained in the component.

[0026] Step 5: Based on the geometric defect type and spatial location information output in Step 4, firstly, execute the topology repair algorithm for defects identified as gaps to restore the topological connectivity at that location; then, execute the corresponding geometric repair algorithm for defects identified as geometric anomalies, and perform mesh stitching on the repaired results of all components to generate the final surface mesh.

[0027] While artificial intelligence technologies, particularly large language models (LLM) and visual large models (VLM), have made progress in multimodal understanding, their direct application in 3D geometry processing still faces challenges, including: a lack of domain-specific large-scale labeled datasets; insufficient support for engineering-level accuracy in existing models; difficulty in effectively integrating prior knowledge from fields such as aerodynamics and manufacturing; and poor interpretability of model decisions leading to low engineer trust. This method first intelligently segments complex CAD assembly models into more regular sub-components; then, it trains a multi-task hierarchical graph neural network (MT-HGNN) that incorporates domain knowledge to automatically identify and evaluate geometric defects in each sub-component; finally, based on the identification results, it drives a phased repair algorithm process following the principle of topology first, then geometry, ultimately generating a watertight overall surface mesh. This solves the problems of low efficiency and insufficient accuracy in geometric defect identification caused by the reliance on human experience in traditional methods, achieving automated, intelligent identification and precise location of various defects in complex CAD models. Overcoming the limitations of traditional algorithms that are based on simple rules and lack semantic understanding, this approach enhances the engineering relevance and accuracy of the identification process by encoding aerospace expertise (such as aerodynamic sensitivity and manufacturing constraints) into the AI ​​model. A closed-loop automated process from intelligent defect identification to targeted repair has been established. Through a phased repair strategy that prioritizes topology over geometry, the identification results are efficiently and reliably transformed into high-quality simulation mesh models, freeing engineers from tedious manual labor and improving the overall efficiency and quality of digital simulation preprocessing.

[0028] This method pioneers a closed-loop paradigm of divide-and-conquer, identification, and repair. Its principles are: 1) Divide and Conquer (Step 1): Reducing the dimensionality of complex problems. Through component segmentation, the disordered and diverse defects at the assembly level are transformed into relatively regular and finite defect patterns within the component level, creating conditions for AI identification. 2) Intelligent Identification (Steps 2-4): Datafication of expert experience. By training a dedicated AI model (MT-HGNN), the model learns the mapping from component geometry to defects, mimicking and surpassing the identification and judgment capabilities of the human eye and brain. 3) Tiered Repair (Step 5): Operationalizing the identification results. Based on the structured defect report output by the AI, a repair algorithm is driven to execute in a strict order (topology first, then geometry), transforming intelligent judgment into deterministic geometric operations, ultimately outputting a high-quality mesh that can be directly used for simulation.

[0029] Preferably, step 1 specifically includes:

[0030] Step S1: Obtain the CAD model of the aircraft to be processed, wherein the CAD model of the aircraft has a boundary representation structure;

[0031] Step S2: Based on the engineering requirements of mesh generation, perform engineering semantic annotation on the faces in the boundary representation structure of the aircraft CAD model to obtain component semantic labels; extract the geometric parameters of each face from the boundary representation structure; generate a multi-view 2D rendering image for each annotated face; associate and store the multi-view 2D rendering image, the geometric parameters, and the component semantic labels to form training samples.

[0032] Step S3: Train a multimodal recognition model using the training samples. The multimodal recognition model is configured to fuse input image features and language semantic information to output semantic labels for components on the surface of the aircraft CAD model.

[0033] Step S4: Input the aircraft CAD model to be identified into the trained multimodal recognition model, identify the surfaces in the aircraft CAD model, and obtain the preliminary surface-part semantic label mapping relationship;

[0034] Step S5: Based on the preliminary face-part semantic label mapping relationship, for faces whose recognition confidence output by the multimodal recognition model is lower than a preset threshold or whose geometric area ratio is less than a set threshold, the geometric rule auxiliary module is activated to perform semantic completion to obtain the final part segmentation result.

[0035] Step S6: Based on the final component segmentation result, extract the key geometric feature lines corresponding to each component from the boundary representation structure of the aircraft CAD model;

[0036] Step S7: Convert the final component segmentation results and the extracted key geometric feature lines into a structured engineering semantic format for the target mesh generation software to read and utilize, and then output it.

[0037] The specific steps in step 1 above address the critical upstream issue of how to accurately and semantically conform to component-level segmentation of complex aircraft assemblies. Traditional manual segmentation is inefficient and prone to subjective variations; general segmentation methods cannot understand the engineering meaning. This approach provides high-quality, engineering-meaning input units (components) for defect identification. It ensures that the segmentation results are not only based on geometry but also incorporate the engineering semantics required for mesh generation (such as distinguishing between wings and fuselages), making subsequent defect identification more targeted. First, a multimodal fusion model (vision + geometry + language) is used to mimic the ability of engineers to understand components from shapes and labels, performing large-scale preliminary intelligent identification. Subsequently, for details that the AI ​​model struggles to grasp (small features, low-confidence regions), an auxiliary module based on deterministic geometric rules is activated for completion and correction. This AI + rule-based collaborative mechanism combines the flexibility of data-driven approaches with the precision of knowledge-driven approaches, ensuring the comprehensiveness and engineering accuracy of the segmentation.

[0038] Preferably, in step 2, the structured features are represented as a graph structure; wherein, the faces of the components are used as nodes of the graph, and the shared edges between faces are used as edges of the graph; the node features include the geometric properties of the faces, and the edge features include the geometric properties of the shared edges.

[0039] The above solution addresses how to effectively transform unstructured CAD boundary representation (B-Rep) data into structured input suitable for deep learning models, providing a high-fidelity, information-rich geometric data representation method. It completely maps the faces, edges, relationships, and attributes of a CAD model into an attribute graph, enabling graph neural networks (GNNs) to directly reason on the model's topology and geometry.

[0040] Preferably, in step 3, the multi-task hierarchical graph neural network model includes:

[0041] A multi-scale feature extraction layer is used to encode geometric features at the node, edge, and subgraph levels.

[0042] Hierarchical graph pooling and attention layers are used to perform hierarchical clustering of patches in different regions and assign attention weights;

[0043] The multi-task output head includes a gap recognition head, a surface recognition head, a hole boundary recognition head, and a defect severity assessment head connected in parallel.

[0044] The proposed solution addresses the challenge of existing general-purpose neural network models simultaneously and accurately identifying multiple types of geometric defects and assessing their engineering impact. Single models cannot handle tasks with varying granularities, such as gaps (edge ​​level), fragmented surfaces (face level), and holes (ring level). A multi-functional, high-precision AI recognition core specifically designed for geometric defect analysis was constructed. This model can output the recognition results and severity assessments of multiple defects in parallel, achieving significantly higher recognition accuracy and engineering practicality than algorithms based on simple rules. The system employs a multi-scale feature extraction layer, extracting features from three levels: nodes (faces), edges (connections), and subgraphs (local regions), ensuring the model can perceive geometric information from microscopic to macroscopic scales. Hierarchical graph pooling and attention mechanisms allow the model to adaptively focus on feature regions at different scales, such as simultaneously focusing on micrometer-level gaps and defect clusters in high-curvature regions. A multi-task output head is designed for different defect types, allowing the network to learn optimal discriminative features for different tasks while sharing the underlying feature extractor. This design enables different defect recognition tasks to mutually reinforce each other, sharing learned geometric context knowledge and thus improving overall performance.

[0045] Preferably, in the hierarchical graph pooling and attention layer, an attention weight guidance mechanism based on domain knowledge is introduced; different prior attention weight coefficients are assigned to the facets according to the aerodynamic critical region or manufacturing process region where the facets are located.

[0046] In the computation of the neural network attention mechanism, attention weights are not entirely driven by data. Instead, a prior attention bias is applied to surfaces with different functional attributes based on a predefined domain knowledge rule base (e.g., if a surface belongs to the leading edge of an airfoil, the weight coefficient is increased by 20%). This is equivalent to injecting engineers' experience into the model early in its training, guiding the model to learn patterns in key areas more quickly, thereby achieving higher recognition sensitivity and accuracy in these areas. This solves the problem that AI models treat all defects equally when identifying them, failing to prioritize key aerodynamic or process-constrained areas that have a greater impact on the final simulation results; that is, the model lacks an engineering perspective for prioritization. It endows the AI ​​model with expert vision, enabling it to identify defects in key areas more preferentially and accurately. This optimizes the allocation of computational resources and ensures that defects with the greatest impact on simulation quality are found and addressed first.

[0047] Preferably, when training the multi-task hierarchical graph neural network model, a weighted multi-task loss function is used; in the weighted multi-task loss function, the weight of the loss term for each defect category is inversely proportional to the number of samples of that defect category in the training set.

[0048] In the backpropagation optimization process, by assigning greater weights (inversely proportional to the number of samples) to the loss term of defect categories with fewer samples, the penalty for misclassification of these categories is amplified. This forces the model to devote more attention during training to learning how to correctly identify these rare defect samples, thereby correcting its natural bias towards the majority class and achieving a balanced improvement in the performance of various defect recognition types. This solves the model training bias problem caused by the extreme imbalance in the number of defect samples in the training data (e.g., more normal edges than gap edges), where the model easily ignores minority classes (important defects), resulting in high overall accuracy but serious missed detections. It significantly improves the recall rate of minority class defects (such as gaps and fragmented surfaces), avoiding the model's neglect of these key but rare defect types due to differences in sample numbers, making the model's recognition results more comprehensive and practical.

[0049] Preferably, in step 3, a domain knowledge-enhanced data augmentation strategy is adopted during the training phase of the multi-task hierarchical graph neural network model. The data augmentation strategy includes simulating the defect generation pattern in the aerospace field, specifically: randomly introducing micron-level gaps at the surface splicing points, and / or generating irregular holes in the planar area to simulate process damage.

[0050] Unlike general image rotation and cropping, this method specifically simulates the physical processes of defects that may actually occur in aerospace manufacturing and assembly processes. For example, it introduces tiny gaps at curved surface joints (simulating assembly gaps) and generates irregular holes in planar areas (simulating impacts or corrosion). Adding these simulated, domain-specific defect data to the training set allows the model to learn in virtual engineering practice, enabling it to better understand and identify various real-world defects it may encounter. This solves the problem that models trained on limited historical labeled data lack generalization ability and struggle to handle the complex and varied defect morphologies in real engineering. Purely data-driven approaches may overfit the training set and have weak ability to identify new defect morphologies. This significantly enhances the model's robustness in identifying defects that may occur in real-world engineering scenarios but are not present in the original dataset. By expanding the diversity and realism of the training data, the model's generalization ability and engineering applicability are improved.

[0051] Preferably, step 5 specifically includes:

[0052] Step 51: Based on the spatial location information of the gap, call the gap repair algorithm at the boundary representation level to restore the topological connection of adjacent surfaces and obtain the intermediate model after B-Rep level topological repair;

[0053] Step 52: Discretize the intermediate model into a triangular shape to obtain a surface mesh; based on the spatial location information of holes, overlaps, and fragments, sequentially execute the hole-filling algorithm, the algorithm for deleting or merging overlapping patches, and the algorithm for fusing or removing fragments on the surface mesh;

[0054] Step 53: Perform a mesh stitching operation at adjacent boundaries on the surface meshes corresponding to all components that have completed step 52 to generate the final surface mesh.

[0055] The repair process follows two sequences: 1) Representational hierarchy (B-Rep before discrete mesh): Gaps are repaired first at the precise boundary representation (B-Rep) level to restore the mathematical integrity of the model; then other operations are performed on the discretized surface mesh, utilizing the flexibility of the mesh algorithm. 2) Defect nature sequence (topology before geometry): Topological defects (gaps) that affect the continuity of the model are addressed first, providing a closed and connected geometric framework for subsequent processing; then geometric anomalies (holes, overlaps, etc.) that do not affect the topology but affect shape integrity are addressed. This sequence aligns with the inherent logic of the data structure and geometric problems of CAD models and is a core engineering principle ensuring the success of automated repair. It solves the problems of blind and disordered repair processes, which can easily introduce new defects or lead to repair failure due to improper repair order. For example, filling a fake hole composed of gaps first may fail due to boundary discontinuities. A stable, reliable, and predictable automated repair process is established. This ensures a high success rate and high-quality output for repair work, avoiding the repeated trial and error and secondary damage to the model common in manual repair.

[0056] Preferably, in step 2, when converting the geometric data of the sample component into a structured feature representation, the extracted geometric features include aerodynamic sensitivity features and / or process constraint features.

[0057] In this process, when converting geometric data into model input (feature representation), not only are general area and curvature calculated, but additional domain-specific features are also calculated and added. For example, aerodynamic sensitivity features might include local angle of attack and estimated pressure gradient; process constraint features might include manufacturing tolerances and material type for that area. By using these features as part of the input, the model can establish a correlation between specific geometric shapes, specific engineering attributes, and defect importance during the learning process. This allows it to learn to assess the severity of a defect from an engineering value perspective, much like a domain expert. This solves the problem that AI models, by only learning general geometric features, cannot understand the engineering significance of defects, potentially leading to a defect list that does not match actual engineering priorities. For example, it might identify a process defect with no aerodynamic impact but miss a critical micro-gap that causes airflow separation. This ensures that the AI ​​model's identification results are directly aligned with the final engineering goals (such as aerodynamic performance and manufacturability). The defect severity assessment output by the model has greater engineering reference value and can truly guide the optimal allocation of repair resources.

[0058] Preferably, in step 4, the spatial location information output by the multi-task hierarchical graph neural network model is used to identify edges for gap defects, boundary loops for hole defects, and surfaces for overlapping or fragmented defects.

[0059] Based on the geometric nature of the defects, the output of the identification results must be mapped back to the original data structure of the CAD model: gaps must be associated with one or a set of edges; holes must be defined by a loop formed by their boundary edges; overlaps or fragmented surfaces must be associated with one or a set of faces. This mapping ensures that the spatial location information output by the AI ​​is not a vague coordinate range, but a specific geometric entity ID or index that downstream B-Rep or mesh repair algorithms can directly identify and manipulate, thus achieving precise repair. This solves the problem that the defect information output by the AI ​​model is inconsistent in format and too abstract, making it unable to directly and unambiguously drive downstream repair algorithms. Simply outputting "holes" or "gap" is insufficient to guide specific geometric operations. Seamless instruction transmission from intelligent identification to automated repair is achieved. The AI's identification results are transformed into precise geometric element identifiers that the repair algorithm can directly execute, bridging the final link between intelligent decision-making and deterministic operation.

[0060] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0061] Based on the integrated model segmentation, intelligent recognition, and phased repair technology proposed in this invention, the following significant advantages are achieved compared to the traditional completely manual processing mode:

[0062] (1) Significantly reduce the dependence of geometric repair on the experience of senior experts: Due to the adoption of AI-based automated identification and marking of component-level geometric defects, the system can mimic the perspective of experts and automatically and accurately locate various complex defects. This eliminates the need for engineers to rely on their eyes and experience to conduct time-consuming and easily overlooked global searches, enabling junior engineers to start the repair process efficiently and accurately, and significantly reducing the technical threshold for high-quality mesh preprocessing.

[0063] (2) Achieve an order-of-magnitude improvement in geometric repair efficiency: Due to the adoption of the technical means of component segmentation + AI batch recognition to replace the overall model + manual inspection, the open and serial defect finding process is transformed into a structured and parallel intelligent detection process, which can shorten the model cleaning work that may have taken several days or even weeks to several hours, greatly compressing the overall cycle of simulation analysis.

[0064] (3) Systematically accumulate and pass on repair strategies and experience: By using the technical means of encoding defect identification experience into trainable and iterative AI models, the originally discrete, implicit and highly personalized expert experience (such as the skills to identify specific types of gaps) is transformed into standardized digital assets that can be permanently saved, continuously optimized and shared within the team, ensuring the consistency and traceability of repair quality and realizing the efficient reuse of organizational knowledge.

[0065] (4) Achieving a closed-loop repair process from random patching to reliable and standardized repair: Due to the adoption of AI precise positioning and phased targeted repair algorithms, the system can automatically call the optimal repair strategy (such as topological fusion followed by geometric cladding) according to the identified defect type, avoiding the secondary damage to the model caused by improper methods in traditional manual repair, and finally forming a stable, reliable and predictable automated processing flow, which fundamentally improves the robustness and success rate of the mesh generation stage. Attached Figure Description

[0066] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0067] Figure 1 This is a simplified flowchart of a geometric defect identification and processing method based on a large AI model.

[0068] Figure 2 This is a schematic diagram of the architecture of the technical solution corresponding to this method;

[0069] Figure 3 This is a detailed flowchart illustrating a method for identifying and processing geometric defects based on a large AI model.

[0070] Figure 4 This is a detailed flowchart illustrating the defect identification and repair process;

[0071] Figure 5 This is a flowchart illustrating the intelligent component recognition and segmentation method for aircraft CAD models based on three-modal fusion.

[0072] Figure 6 This is a schematic diagram of the architecture of an intelligent component recognition and segmentation system for aircraft CAD models based on three-modal fusion. Detailed Implementation

[0073] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0074] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0075] Those skilled in the art should understand that, in the disclosure of this invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting this invention.

[0076] It is understood that the term "a" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0077] Example 1;

[0078] Please refer to Figures 1-4 This invention provides a method for geometric defect identification and processing based on a large AI model, the method comprising:

[0079] Step 1 - Model Segmentation: Obtain the CAD model of the aircraft to be processed, and segment the aircraft CAD model into multiple parts;

[0080] Step 2 - Training Data Preparation: Obtain multiple labeled sample parts, and convert the geometric data of each sample part into a structured feature representation for processing by the deep learning model to obtain a training set. The structured feature representation includes at least one of the graph structure, point cloud, and feature vector of the sample part; the annotation includes the type and spatial location information of the geometric defects existing in the sample part.

[0081] Step 3 - AI Model Training: Use the training set to train a multi-task hierarchical graph neural network model to obtain the trained multi-task hierarchical graph neural network model;

[0082] Step 4 - Defect Identification: For any component obtained from the segmentation in Step 1, the geometric data of the component is converted into the same structured feature representation as in Step 2. The converted data is then input into the trained multi-task hierarchical graph neural network model. The trained multi-task hierarchical graph neural network model outputs the type and spatial location information of the geometric defects contained in the component.

[0083] Step 5 - Phased Repair and Mesh Generation: Based on the geometric defect type and spatial location information output in Step 4, firstly, a topology repair algorithm is executed for defects identified as gaps to restore the topological connectivity at that location; then, a corresponding geometric repair algorithm is executed for defects identified as geometric anomalies, and mesh stitching is performed on the repaired results of all components to generate the final surface mesh.

[0084] This invention provides an efficient method for handling geometric defects in complex CAD models based on artificial intelligence. Its core principle lies in addressing the pain points of traditional manual processing—namely, its lengthy process, diverse problems, and lack of unified rules—by proposing a new paradigm of model segmentation, intelligent identification, and tiered repair. First, the complex assembly model is intelligently decomposed into more predictable components based on engineering semantics or geometric features. The purpose of splitting into components is that, because a single model contains many defect categories that are difficult to classify effectively, dividing it into individual components—for example, the defects in an aircraft nose section—are relatively fixed. Therefore, large AI models can be trained for identification, resulting in better accuracy and generalization. Subsequently, a deep learning model is constructed to learn the mapping relationship between component geometric features and defect categories and spatial locations, enabling automated identification and labeling of geometric defects (such as gaps, overlaps, holes, and fragments) within each component. Finally, the AI ​​identification results are seamlessly integrated with a carefully designed phased geometric repair algorithm flow, first repairing the topology (gaps), then processing the geometry (holes, overlaps, etc.), and finally obtaining a watertight overall surface mesh through mesh stitching, thereby freeing engineers from heavy and repetitive manual inspection and repair work.

[0085] (1) Complete method flow:

[0086] This method, from model preprocessing to final mesh generation, includes the following steps in sequence:

[0087] 1. Model preparation and component segmentation: Based on geometric continuity, engineering functional semantics, or user-defined rules, the original CAD assembly model is automatically or semi-automatically segmented into multiple more regular sub-components.

[0088] 2. Data preparation and feature representation: Collect geometric samples of components with defect annotations and convert them into multimodal feature representations suitable for deep learning models (such as figure structure, point cloud, feature vector, etc.) to build a training dataset.

[0089] 3. Training the AI ​​geometric defect recognition model: Using the training dataset constructed in step 2, train a deep learning model (as shown in the figure neural network) to accurately identify and locate various geometric defects and their boundaries in the input component.

[0090] 4. Intelligent identification and localization of defects in new components: For the segmented new components, extract their features and input them into the trained AI model. The model outputs the type, confidence level and precise spatial location marking of all defects on the component.

[0091] 5. Staged geometric repair and global mesh stitching: Based on the defect markers provided by AI, the gap repair algorithm is first called at the B-Rep level to restore the topological connectivity; then the repaired surface is discretized, and discrete repair and surface wrapping are performed for problems such as holes and overlaps to obtain a component-level watertight mesh; finally, mesh stitching operation is performed at the mesh boundaries of different components to generate the final global watertight surface mesh.

[0092] (2) Detailed introduction of each step:

[0093] Step 1: Model preparation and component segmentation;

[0094] The system receives the original CAD assembly model, which may contain various geometric defects. Based on geometric features (such as surface continuity and curvature abrupt changes), a predefined engineering semantic rule base (such as wings and fuselages), or user interactive input, the system decomposes the model into multiple relatively independent sub-components with more regular internal geometric features and potential defect types.

[0095] Execution Object and Logical Relationship: The object being processed is the entire CAD assembly. This step is fundamental to the model segmentation strategy, and its decomposition quality directly impacts the accuracy of AI recognition and the ease of repair. Part segmentation simplifies the geometric context within each sub-part, making it easier for the AI ​​model to learn and generalize.

[0096] Step 2: Data preparation and feature representation;

[0097] For each component sample (from historical projects or artificial constructions), experts annotate its geometric defects (gap edges, overlapping surfaces, hole rings, fragmented surfaces, etc.) and their type labels. Subsequently, the component's B-Rep model is converted into one or more deep learning-friendly representations.

[0098] 1. Graph Structure Representation: A property graph is constructed by treating the faces of a component as nodes and the shared edges between faces as edges. Node features can include the geometric properties of the face (area, curvature, normal), and edge features can include the included angle between two sides, length, etc.

[0099] 2. Hybrid Representation: Combining graph structure with locally densely sampled point clouds. Based on the graph structure, point cloud sampling is performed on each face or its boundary region, with the addition of local geometric details.

[0100] This step is the sample preparation stage for AI model training. Transforming unstructured geometric data into structured, semantically rich feature representations is crucial for the model's ability to effectively learn defect patterns. The pairing of labeled data (defect location and type) with feature representations constitutes the training samples.

[0101] Step 3: Training the AI ​​geometric defect recognition model;

[0102] This invention designs a multi-task hierarchical graph neural network (MT-HGNN) for aerospace geometric defects, instead of directly using a general GNN model. The customized network structure is as follows:

[0103] a) Multi-scale feature extraction layer:

[0104] Node-level encoder: A multilayer perceptron (MLP) is used to encode the basic geometric features (such as area, mean curvature, Gaussian curvature, and normal vector) of each patch node.

[0105] Edge encoder: Encodes the features of shared edges (such as the included angle between the two sides, the side length, and the relative curvature change).

[0106] Subgraph-level encoders: Aggregate node and edge information through graph convolutional layers (such as GATv2) to generate embedded representations of local subgraphs (such as neighborhoods around a vertex or an edge) to capture local context.

[0107] b) Hierarchical graph pooling and attention mechanisms:

[0108] An adaptive topological pooling layer is designed to perform hierarchical clustering of graph nodes based on the curvature characteristics and functional semantics of the surface (such as distinguishing between the high curvature leading edge of an airfoil and the smooth skin), generating graph representations at different levels of abstraction, thereby simultaneously capturing the distribution patterns of microscopic defects (such as tiny gaps) and macroscopic defects.

[0109] Introduce a domain-knowledge-guided attention mechanism. For example, assign higher attention weights to the edges of key aerodynamic regions such as the wing leading edge and wing-body docking zone, making the model pay more attention to these regions that have a significant impact on CFD results.

[0110] Domain knowledge refers to prior rules and experience in aerospace engineering. For example: Aerodynamic knowledge: Areas such as the leading edge of a wing, engine intake / exhaust ports, and wing-body junctions are highly aerodynamically sensitive, and geometric defects in these areas have a significant impact on CFD results. Manufacturing and process knowledge: Different manufacturing processes (such as sheet metal forming and composite material laying) have their typical defect types and tolerance ranges. Functional semantic knowledge: Certain areas are load-bearing structures, sealing surfaces, or moving surfaces, and have extremely high requirements for geometric integrity.

[0111] Specific Implementation of the Domain Knowledge-Based Attention Weight Guidance Mechanism: To enable the model to focus on key regions, this invention introduces a domain knowledge-based attention weight guidance mechanism in the hierarchical graph pooling and attention layers of the Multi-Task Hierarchical Graph Neural Network (MT-HGNN). The core of this mechanism lies in superimposing a prior weight bias determined by domain knowledge on top of the model's self-attention calculation, thereby guiding the model to pay more attention to key regions of engineering significance. Specifically, this includes:

[0112] 1. Construction of the domain knowledge rule base:

[0113] First, based on aerodynamic principles, structural design specifications, and manufacturing process manuals, a computable rule base is constructed. Each rule maps the geometric / topological conditions in the CAD model to an attention adjustment coefficient. For example:

[0114] Rule A (Aerodynamically Critical Area): The IF patch is located in the leading edge of the wing or the engine lip area (which can be determined by the parameter coordinates or labels of the patch). THEN Basic Attention Weighting Coefficient β = 1.5.

[0115] Rule B (Secondary Non-structural Area): IF panel belongs to non-load-bearing fairing or interior panel THENβ=0.8.

[0116] Rule C (High-risk areas for connections and gaps): IF panel is adjacent to more than two other components (at the boundary of complex assembly) THENβ=1.3.

[0117] 2. Integration of guidance mechanisms in attention computation:

[0118] Let a be the attention score of the i-th face to the j-th face obtained by standard computation in a graph neural network. ij This mechanism is adjusted in the following ways:

[0119] ;

[0120] in, The adjusted attention score is represented by λ, a trainable hyperparameter used to control the strength of domain knowledge guidance. This is introduced as an additive bias term. When both patches belong to critical regions (both with β values ​​greater than 1), this term is positive, significantly increasing their attention scores; conversely, it may decrease them if neither patch belongs to a critical region. β i and β j These are the patch-level prior attention modulation coefficients. They are two independent scalar values, representing the importance assessments of the i-th patch (node) and j-th patch (node) in the graph neural network based on domain knowledge, respectively.

[0121] This mechanism accelerates convergence by giving the model an expert-like perspective from the early stages of training, eliminating the need for extensive, inefficient trial and error to discover key areas and thus speeding up the training process.

[0122] Improve the accuracy of key area identification: The model allocates more computing resources to areas with zero tolerance for defects, such as aerodynamically sensitive areas and connection areas, which significantly improves the accuracy and recall of identifying tiny gaps and fragments in these areas.

[0123] Enhanced decision interpretability: The model’s focus is aligned with engineering consensus, and its identification results are more easily understood and trusted by domain experts, overcoming the obstacles of the traditional AI black box problem in engineering applications.

[0124] To achieve targeted intelligent recognition, this invention performs region division on the aircraft CAD model based on engineering semantics, mainly including two categories:

[0125] (1) Aerodynamic critical areas: These are areas whose geometric integrity directly determines the accuracy of aerodynamic simulation. For example, the leading and trailing edges of the wing and tail, the engine inlet lip, and the wing-body fairing. These areas have a low tolerance for defects such as gaps and step differences.

[0126] (2) Manufacturing process area: This refers to the area divided according to the actual production method of the part. Different areas have unique defect patterns and quality standards. For example, the skin of [part A] is marked as the composite material layup area, and its focus is on delamination and resin unevenness; the bracket of [part B] is marked as the precision casting area, and its focus is on shrinkage and deformation.

[0127] 2. Application of knowledge in the model:

[0128] The aforementioned regional division information is injected into the AI ​​model in two ways:

[0129] As input for attention guidance: In the attention mechanism, higher prior weight coefficients β are assigned to patches belonging to aerodynamically critical regions to guide the model to focus.

[0130] As input for feature engineering: Add a "manufacturing process type" feature label to each facet, enabling the model to understand defects in conjunction with the process context.

[0131] c) Multi-task learning output head:

[0132] The network endpoint is not a single classifier, but rather multiple specific task output headers connected in parallel:

[0133] Gap detection head: Classifies each edge in the image into two categories (gap / non-gap).

[0134] Fragmentation identification head: classifies each face node (whether it is a fragmented face or a normal face).

[0135] Hole Boundary Identification Head: Identifies and serializes the edges that constitute the hole boundary loop.

[0136] Defect severity assessment head: Regression predicts the impact coefficient (e.g., a score between 0 and 1) of each identified defect on the mesh generation quality, providing a basis for subsequent repair priority ranking.

[0137] This multi-task design enables the model to collaboratively learn the correlations between different defect types, thereby improving the overall recognition accuracy.

[0138] The gap detection head is an edge-level binary classifier. It receives features extracted by a graph neural network, judges each edge in the model, and outputs whether the edge is a gap. This accurately locates topological discontinuities. Gaps are the primary cause of watertightness in the model, and this output is a direct instruction to initiate topology repair.

[0139] The fragmented surface identification head is a face-level classifier. It receives features from each facet node and determines whether the face is fragmented (facets with excessively small area, abnormal aspect ratio, or no substantial contribution to geometry or fluid simulation). It identifies geometric noise that needs to be cleaned up or merged. The repair algorithm then removes or merges these faces based on the results to simplify the geometry and improve mesh quality.

[0140] The hole boundary recognition head is a boundary loop identification and serialization module. It analyzes edge features and connectivity to identify edge sequences that form closed loops but contain no faces, and labels these as holes. This accurately delineates the boundaries of missing surface areas. The output provides a clear, closed filling target for geometric restoration algorithms.

[0141] The defect severity assessment head is a regression or grading evaluator. For each identified defect (such as a crack or a hole), it outputs a numerical value (such as a score of 0-1) or a grade (such as high, medium, or low) representing its severity. This provides engineering prioritization for repairs. Severity assessment is based on predictions of downstream mesh generation success rates and quality impacts, enabling the system to prioritize the most critical defects and optimize the repair process.

[0142] Training strategy:

[0143] Loss Function: This invention employs a weighted multi-task loss function for end-to-end training of a multi-task hierarchical graph neural network. The total loss function is defined as follows:

[0144] .

[0145] The definitions of each sub-loss function are as follows:

[0146] For the loss function of the gap detection head, a weighted binary cross-entropy loss is used. Since the number of normal edges in the training set is much greater than that of gap edges, positive and negative samples are balanced by class weights, and the calculation formula is as follows:

[0147] ;

[0148] in, The true label for edge e (1 for gap, 0 for no gap). To predict probabilities for the model, is the weight of the positive sample, set as the ratio of the number of negative samples to the number of positive samples, and E is the set of all edges involved in the computation in the graph neural network.

[0149] For the loss of the facet recognition head, the weighted binary cross-entropy loss is also used, and the definition method is the same as... Similarly, weight The ratio of broken surface samples to normal surface samples is set.

[0150] For the hole boundary identification head, this task is a sequence labeling task, and either Connection Temporal Classification Loss (CTC Loss) or edge-level binary cross-entropy loss is used. This invention preferably uses edge-level binary cross-entropy loss, which performs binary classification on the edges constituting the hole boundary, with the loss form being the same as... Weight The ratio of the hole boundary edge to the normal edge is set.

[0151] For the loss of the defect severity assessment head, this task is a regression task, and the mean squared error loss is used:

[0152] ;

[0153] in, The severity score (0~1) of the defect as indicated by the experts. The severity score predicted by the model. This represents the total number of defects identified.

[0154] Weighting coefficient The setup method is as follows: To address the issues of inconsistent convergence speeds and dimensional differences among tasks in multi-task learning, this invention employs an adaptive loss weighting method based on task uncertainty or an inverse proportional weighting method based on sample size. In a preferred embodiment, the inverse proportional weighting method is used:

[0155] ;

[0156] in, This represents the number of valid samples for the corresponding task in the training set (the number of positive samples for classification tasks and the number of defect instances for regression tasks). This setting gives higher weights to tasks with smaller sample sizes, effectively alleviating class imbalance and improving the recall rate for identifying minority class defects. Here, z is the index variable for all tasks. This represents the total number of samples in the training set for the corresponding task.

[0157] In typical implementation scenarios, the range of values ​​for each weight can be: , , , And satisfy The specific values ​​can be dynamically calculated and determined based on the statistical distribution of the training set.

[0158] Domain-specific data augmentation: During training, in addition to conventional rotation and scaling, data augmentation is performed by simulating defect generation patterns unique to the aerospace field. For example, micron-level gaps are randomly introduced at surface joints, and irregular holes simulating process damage are generated in planar areas to improve the model's robustness to real engineering scenarios.

[0159] Step 4: Intelligent identification and location of defects in new components:

[0160] For a new component to be processed obtained after segmentation in step 1, the system automatically converts it into a feature representation consistent with the training data (as shown in the figure). This representation is then input into the AI ​​model trained in step 3. The model performs forward inference and outputs a structured recognition result.

[0161] Execution Objects and Logical Relationships: The output is a machine-readable defect report that clearly identifies which edges in the component are identified as gaps, which surfaces are identified as overlaps or fragments, and the location of suspected boundary loops of holes. This report accurately marks the spatial location and type of defects, providing direct and clear input guidance for targeted repair in step 5, replacing the traditional fully manual, experience-based visual search process.

[0162] Step 5: Phased geometric repair and global mesh stitching:

[0163] This step is the core repair process, strictly following the principle of topology first, then geometry, and is executed in three stages:

[0164] 1. B-Rep Level Gap Repair: Based on the gap markings output by AI, at the boundary representation (B-Rep) level of the original CAD, gap repair algorithms (such as boundary fusion, surface extension / trimming) are called to restore the topological connection of the marked adjacent surfaces, generating an intermediate B-Rep model that is topologically complete but may still have other defects.

[0165] 2. Discretization and Surface Mesh Repair: The B-Rep model obtained in the previous step is triangulated and discretized to generate a surface mesh. Based on the AI's marking of holes, overlaps, and fragmented surfaces, corresponding algorithms are executed on this mesh: a loop-finding algorithm is used to construct closed loops in the marked hole regions and fill the holes; marked overlapping patches are deleted or merged; marked fragmented surfaces are merged or removed. Subsequently, a surface wrapping algorithm (such as Alpha Wrap) is applied to generate a clean, watertight temporary closed surface mesh for the entire component, and then the portion outside the original boundary is trimmed from it to obtain a boundary-preserving watertight mesh for the component.

[0166] 3. Global Mesh Stitching: For all processed component-level watertight meshes, mesh stitching techniques such as node merging and edge swapping are used at their adjacent boundaries to seamlessly connect the component meshes, ultimately generating a complete, watertight, global surface mesh model suitable for subsequent simulations.

[0167] This step is crucial for transforming AI-powered intelligent recognition results into high-quality final geometric output. Through a rigorous algorithm, it assigns different types of defects to the most suitable repair stage, ensuring repair efficiency and result reliability. Ultimately, it achieves a fully (semi-)automated conversion from defective CAD to a clean computational mesh.

[0168] (3) Corresponding system composition:

[0169] Systems that implement the above methods include:

[0170] 1. Model Segmentation and Preprocessing Module: Responsible for decomposing the input CAD assembly into parts.

[0171] 2. Geometric Feature Extraction and Representation Module: Responsible for converting component geometry into multimodal features such as graph structures and point clouds.

[0172] 3. AI Geometric Defect Recognition Module: This module carries pre-trained AI models such as graph neural networks and provides defect recognition and labeling services.

[0173] 4. Staged geometric repair algorithm library: integrates core algorithms such as gap repair, discrete hole patching, surface covering, and mesh editing.

[0174] 5. Repair Process Driven and Mesh Stitching Module: Based on the AI ​​recognition results, automatically schedule and execute the corresponding repair algorithms, and manage the mesh stitching process between components.

[0175] The key innovation of this invention lies in explicitly encoding tacit expert knowledge in the aerospace field into the learning and reasoning process of the AI ​​model, specifically through the following three mechanisms:

[0176] Mechanism 1: Knowledge Embedding in Feature Engineering

[0177] In step 2: Data Preparation and Feature Representation, the geometric features we extract are not only general attributes, but also engineering features strongly correlated with aerodynamic performance and manufacturing processes. For example:

[0178] Aerodynamic sensitivity characteristics: Calculate the local angle of attack and the estimated flow separation risk coefficient of the patch location, and input these parameters as nodal features into the model so that the model knows to prioritize the defects in aerodynamically sensitive areas.

[0179] Process constraint features: Label the corresponding manufacturing process (such as sheet metal forming, composite material layup) for the patch, and use the allowable tolerance range of the process as a feature, so that the model can distinguish between acceptable process defects and fatal defects that must be repaired.

[0180] Mechanism Two: Knowledge Guidance in Attention Mechanisms

[0181] As described in the network structure section, in the attention layer of the GNN, prior attention weights are assigned to different functional regions using a predefined rule base. The rule base is derived from aerodynamic principles and design specifications, for example:

[0182] The IF region belongs to the wing leading edge. THEN attention weighting coefficient = 1.2;

[0183] The IF region belongs to the non-load-bearing fairing; THEN attention weighting coefficient = 0.8.

[0184] This gives the model an expert perspective from the early stages of training, accelerating convergence and improving the recognition accuracy of key regions.

[0185] Mechanism 3: Knowledge integration of loss function and evaluation criteria;

[0186] The optimization objective (loss function) of the model not only pursues classification accuracy, but also links it to the success rate of downstream CFD grid partitioning.

[0187] When labeling training data, not only is the existence of defects indicated, but experts also indicate the urgency level of their repair based on historical experience (e.g., causing mesh generation failure, causing poor mesh quality, or having negligible impact).

[0188] Through learning, the model can output an engineering severity score for defects, thereby enabling intelligent prioritization based on engineering impact and guiding the repair process to address the defects that have the greatest impact on simulation results first.

[0189] This invention proposes a new paradigm for automated processing of geometric defects in complex CAD models based on artificial intelligence and phased repair. The difference between this paradigm and traditional methods lies in:

[0190] 1. A pioneering integrated architecture of component-level AI recognition and phased repair: Traditional methods involve manual global inspection of the entire complex model followed by individual repair of identified issues. This invention utilizes intelligent component segmentation, parallel AI batch recognition of each component, and phased system repair based on defect type. This represents a fundamental shift in workflow, transforming a sequential, manual-dependent approach into a parallel, intelligently driven one.

[0191] 2. A multi-task hierarchical graph neural network (MT-HGNN) architecture for geometric defect identification in aerospace CAD. Its specific network components include: a multi-scale feature extraction layer, an attention mechanism incorporating domain knowledge weights, an adaptive topological pooling layer, and a multi-task output head for collaborative identification of gaps, surface fragments, holes, and assessment of defect severity.

[0192] 3. AI-Driven Intelligent Recognition and Localization Methods for Geometric Defects: This method transforms the domain knowledge of CAD geometric defect recognition, which heavily relies on expert experience, into a trainable and reusable AI model. Specific methods for encoding aerospace domain knowledge into the AI ​​model include: (a) embedding engineering features such as aerodynamic sensitivity and manufacturing tolerances into geometric features; (b) utilizing a predefined rule base to guide graph attention weights; and (c) integrating expert experience such as the urgency of defect repair into the model's loss function and output criteria.

[0193] 4. Phased Repair Engineering Process: A topology-first, geometry-second repair sequence: prioritize addressing gaps (restoring B-Rep topology), then tackle geometric issues (holes, overlaps, etc.). Combining B-Rep layer repair with discrete layer repair: fully leveraging the advantages of both representations. A strategy of independent component repair + overall stitching: reducing processing complexity and improving fault tolerance.

[0194] Example 2;

[0195] Based on Embodiment 1, Embodiment 2 of the present invention provides a method for intelligent component identification and segmentation of aircraft CAD models based on three-modal fusion to provide a detailed explanation of step 1 in Embodiment 1.

[0196] Example 2 can solve the problems of low efficiency, poor consistency and automation breakpoints caused by relying on manual component identification in the aircraft mesh generation process. Example 2 provides a method that can automatically and accurately identify aircraft components and directly output semantic grouping and feature geometry that meet the requirements of mesh generation engineering, so as to realize full-process automation from geometric model to computational mesh.

[0197] Example 2 provides a method for intelligent component recognition and segmentation of aircraft CAD models based on three-modal fusion. The method includes:

[0198] Step S1: Obtain the CAD model of the aircraft to be processed, wherein the CAD model of the aircraft has a boundary representation structure;

[0199] Step S2: Based on the engineering requirements of mesh generation, perform engineering semantic annotation on the faces in the boundary representation structure of the aircraft CAD model to obtain component semantic labels; extract the geometric parameters of each face from the boundary representation structure; generate a multi-view 2D rendering image for each annotated face; associate and store the multi-view 2D rendering image, the geometric parameters, and the component semantic labels to form training samples.

[0200] Step S3: Train a multimodal recognition model using the training samples. The multimodal recognition model is configured to fuse input image features and language semantic information to output semantic labels for components on the surface of the aircraft CAD model.

[0201] Step S4: Input the aircraft CAD model to be identified into the trained multimodal recognition model, identify the surfaces in the aircraft CAD model, and obtain the preliminary surface-part semantic label mapping relationship;

[0202] Step S5: Based on the preliminary face-part semantic label mapping relationship, for faces whose recognition confidence output by the multimodal recognition model is lower than a preset threshold or whose geometric area ratio is less than a set threshold, the geometric rule auxiliary module is activated to perform semantic completion to obtain the final part segmentation result.

[0203] Step S6: Based on the final component segmentation result, extract the key geometric feature lines corresponding to each component from the boundary representation structure of the aircraft CAD model;

[0204] Step S7: Convert the final component segmentation results and the extracted key geometric feature lines into a structured engineering semantic format for the target mesh generation software to read and utilize, and then output it.

[0205] This method, through step S1, limits the processing object to CAD models with boundary representation structures, ensuring the feasibility of all subsequent geometric operations (topology analysis, parameter extraction). By utilizing steps S2-S3 to construct a multimodal dataset oriented towards mesh engineering semantics and training a specialized model integrating visual, geometric, and linguistic information, the method gains the ability to understand engineering semantics, replacing manual identification by engineers. Steps S4-S5 employ a collaborative mechanism of AI recognition of large components and geometric rule completion of small features, utilizing AI to handle macroscopic patterns while using geometric rules to ensure detailed integrity, overcoming the limitations of single methods on complex engineering geometries. Steps S6-S7 automatically connect the semantic recognition results to the geometric kernel, extracting key geometric feature lines and converting them into a format directly readable by mesh software, thus completing the final step from recognition to application and eliminating manual conversion. This solves the end-to-end automation bottleneck problem in the aircraft CAD mesh generation process, where component recognition relies entirely on manual labor, resulting in extremely low efficiency, inconsistent results, and the inability to directly use the recognition results for subsequent engineering. It achieves a fully automated, high-precision processing flow from the original CAD model to mesh-generated ready engineering data, reducing the time required for manual work, which can take several weeks, to automated processing.

[0206] In step S2, the surfaces in the boundary representation structure of the aircraft CAD model are annotated with engineering semantics. Specifically, in the mesh preprocessing software environment, semantic labels are manually assigned to the surfaces contained in different components based on aerodynamic characteristics or structural functional differences.

[0207] Preferably, the geometric rule auxiliary module determines and corrects the semantic labels of the surface components based on the geometric topological relationships and parametric features of the aircraft CAD model, thereby forming the final component segmentation result.

[0208] Purely data-driven AI models are prone to low confidence or misjudgment when faced with small-sized, weakly textured, feature-fuzzy, or geometrically rare surfaces in the training data (such as slender wing trailing edges, tiny connecting protrusions, and hatches that smoothly transition into the main body). The above approach addresses this by introducing rule modules based on deterministic geometric knowledge to compensate for AI's shortcomings in such cases. General image segmentation models output pixel-level labels, which are difficult to automatically and error-free associate with the precise boundary representation (B-Rep) structure (faces, edges, vertices, and their topology) in CAD models. The above approach directly limits operations to the geometric and topological levels of the original CAD model, ensuring that any judgments and corrections are precisely applied to specific B-Rep surfaces, guaranteeing absolute geometric accuracy. By filling in gaps and correcting errors in the AI ​​recognition results, the final component segmentation results ensure that all geometric details affecting subsequent mesh generation and simulation analysis are covered, reducing the degradation of mesh quality or simulation accuracy caused by missing key small features. This approach eliminates reliance on a single AI model, creating a hybrid intelligence model that combines data-driven (AI) and knowledge-driven (rule-based) approaches. When faced with new aircraft models or those with unique geometries, even if the AI ​​model performs poorly initially, the method can still ensure the correct identification of basic and critical components through geometric rules, thus enhancing the method's generalization ability and practical stability.

[0209] Preferably, the multi-view 2D rendered image is generated in the following way:

[0210] For each surface to be processed, perform multi-angle wrapping along its normal vector, render it from each viewpoint, and generate a series of two-dimensional view images.

[0211] Specifically, for each facet, multi-angle rendering is performed along its normal vector. This simulates the process of viewing a geometric facet from various directions, ensuring that the generated image fully reflects the facet's shape characteristics. This solves the problem of generating richly informative 2D image data from a single, planar aircraft CAD model facet, allowing the visual model to learn from it. It provides the visual model with multi-angle, comprehensive observation data, enhancing the model's robustness in recognizing component faces under different viewing angles and lighting conditions.

[0212] The preferred training method for the multimodal recognition model is as follows:

[0213] A neural network architecture is constructed, comprising a visual encoder, a language understanding module, and a cross-modal attention fusion module. First, the visual encoding module, based on Vision Transformer (ViT), is responsible for extracting features from the input image and outputting a visual token. Second, the language encoding module, based on a pre-trained large model, is responsible for encoding text prompts into text tokens. Finally, the cross-modal fusion module achieves bidirectional deep interaction between the visual and language tokens through a cross-attention mechanism. After multiple rounds of interaction, the features of the two modalities are aligned and fused in a unified semantic space to form a joint feature representation.

[0214] The neural network architecture is trained under supervision by using the multi-view two-dimensional rendered images and corresponding component semantic label texts in the training samples as input.

[0215] The multimodal recognition model is optimized using a combination loss function, which includes the classification cross-entropy loss function, the segmentation cross-union ratio loss function, and the boundary weighted loss function.

[0216] This paper presents a clear technical framework for fusing multimodal information through an architecture consisting of a visual encoder, a language module, and a cross-modal attention fusion module. Supervised learning is performed using a pre-constructed dataset, enabling the model to learn the mapping from images to semantic labels. A combined loss function is employed to simultaneously optimize classification accuracy, segmentation region consistency, and boundary precision, ensuring the quality of the model's output from multiple dimensions. The paper addresses the core technical path of model implementation by solving the problem of how to specifically construct and train a dedicated recognition model capable of understanding image-semantic associations. An optimized and trained intelligent model capable of accurately associating surface images with component semantic labels has been implemented.

[0217] Preferably, step S4 specifically involves identifying large component surfaces in the aircraft CAD model; step S5 specifically involves semantic completion of local feature surfaces, connecting region surfaces, or low-confidence identification surfaces in the aircraft CAD model.

[0218] Steps S4 and S5 are defined as a hierarchical collaborative process: the first layer (AI) is responsible for large components (efficiently handling regularities); the second layer (rules) is responsible for local features, connecting surfaces, and other complex situations (accurately handling special characteristics). This design concept is an efficient way to solve engineering complexity problems. It realizes intelligent division of labor and collaboration in the recognition process, significantly improving the completeness of recognizing small and special features while ensuring the efficiency of large component recognition.

[0219] Preferably, the rules for determining the geometric topological relationships and parametric features of the aircraft CAD model include at least one of the following rules:

[0220] Rule 1: If a face is located at the geometric intersection of two identified large component faces, and the area of ​​the face is less than the area of ​​the adjacent large component face, then the component semantic label of the face is determined to be a connecting feature face or an edge face.

[0221] Rule 2: If a surface is directly adjacent to the main wing, and the angle between its normal vector and the normal vector of the main wing's reference plane is greater than or equal to 45°, and this surface is located on the outermost side of the wing span, with an area between 0.5% and 5% of the main wing's area, then it is classified as a winglet or wingtip. The area proportion rule can identify surface areas at boundaries (such as edges). The direction relationship rule can identify surface areas that are not parallel to the main airflow direction (such as doors). This provides quantifiable and verifiable rule examples, enabling the system to automatically identify specific types of geometric features such as the wing trailing edge and doors.

[0222] Preferably, the extraction of key geometric feature lines corresponding to each component includes:

[0223] For the components identified by the multimodal recognition model, by analyzing the curvature distribution and boundary topology of its constituent surfaces, a geometric algorithm is used to extract the key geometric feature lines of the components, including its boundary lines and feature contour lines.

[0224] For the feature surfaces identified by the geometric rule assistance module, the corresponding key geometric feature lines are extracted based on their geometric definitions and their relationship with adjacent surfaces.

[0225] For large components identified by AI, geometric algorithms such as curvature and topology analysis are used to automatically extract their boundaries and contours. For small features identified by rules, they are extracted directly based on their geometric definitions (such as the intersection of two faces). Both methods ensure that the output feature lines are accurately associated with the semantic labels of the components. This achieves automated connection from semantic segmentation to geometric feature extraction, outputting geometric data that is directly relied upon for operations such as mesh encryption and block division.

[0226] Preferably, in step S5, before or after forming the final component segmentation result, a post-processing step is also included:

[0227] The semantic labels of components are smoothed based on the adjacency relationship of surfaces to integrate the recognition results of the multimodal recognition model with the supplementary results of the geometric rule auxiliary module, thereby eliminating misjudgments of isolated surfaces; and the final component segmentation results are visualized and exported.

[0228] The label smoothing process is based on the adjacency relationship of faces, performing consistency checks and adjustments on the labels of adjacent faces to eliminate unreasonable abrupt changes or isolated points. The AI ​​results are fused with rule-based results, and visualization and export are supported, forming a complete closed loop from processing to delivery. Post-processing improves the topological consistency and visual smoothness of the segmentation results, and visualization and export functions make the results easy for engineers to verify and use.

[0229] The input to the multimodal recognition model includes natural language instructions; the method allows users to specify the type of part to be recognized or the segmentation requirements by inputting natural language instructions; the visual encoder adopts a Transformer-based architecture, the language understanding module adopts a pre-trained large language model, and the cross-modal attention fusion module adopts a cross-attention mechanism.

[0230] The model's input can include natural language commands, which are then translated into internal tasks that drive the model and geometry engine, achieving intelligent conversion from natural language to concrete operations. This enhances the system's usability and flexibility, allowing users to drive the system to perform specific tasks using natural language (such as segmenting all wing surfaces). The key modules in the architecture are concretized into corresponding selections, ensuring the feasibility of the solution: a visual encoder (ViT or other Transformer architectures), a language understanding module (QwenLM or other large models), and a fusion module (CrossAttn), ensuring the powerful capabilities of each module.

[0231] Preferably, in step S7, the structured engineering semantic format includes: a set of component geometric groups for the mesh generation software to recognize and call, a set of feature lines describing the key geometric features of the components, and optional mesh generation strategy parameters associated with preset components. By explicitly defining the output format as the structured engineering semantic format required by industrial software, including the component group set, feature line set, and mesh strategy parameters, the output of this method is no longer merely a display result, but rather instructions or data that can be directly read and applied by CAE software, thus achieving a closed loop of engineering value.

[0232] The one or more technical solutions provided in Embodiment 2 have at least the following technical effects or advantages:

[0233] (1) Full automation of component identification has been achieved, greatly improving the efficiency of mesh preprocessing;

[0234] Because this invention employs a visual-geometric-linguistic three-modal fusion recognition and hierarchical collaborative processing mechanism, it can automatically and accurately complete the identification and semantic segmentation of various aircraft components. The effect is that it shortens the process of identifying medium-sized passenger aircraft components, which originally relied entirely on manual identification and grouping by engineers and took 1-2 weeks, to automated processing, thus solving the primary efficiency bottleneck in the mesh generation process.

[0235] (2) Outputs engineering semantics and geometric features that can directly drive mesh generation;

[0236] Because the output of this invention is not a general image segmentation mask, but rather a face-part label mapping table and key geometric feature lines that are precisely associated with the original B-Rep model, the effect is that the recognition results can be directly converted into components, sets, or layers that mesh generation software can recognize, and key feature lines (such as wing leading / trailing edge lines) used for mesh partitioning and densification control are automatically extracted, achieving recognition-ready functionality and eliminating the large amount of manual conversion and interpretation work required in traditional methods.

[0237] (3) Improved the completeness of identification and engineering practicality of small-sized and weak-featured parts;

[0238] Because this invention employs a layered strategy of AI recognition for large components combined with geometric completion for small features, it intelligently completes small or inconspicuous components such as wing trailing edges and access hatches using geometric rules. The effect is a significant improvement in the overall segmentation integrity, ensuring that subsequent mesh generation covers all necessary geometric details and making the generated mesh more consistent with the geometric integrity requirements of high-fidelity simulation.

[0239] (4) Improve grid quality by supporting smart grid strategy presets through semantic understanding;

[0240] Because the model training of this invention incorporates knowledge of mesh engineering, it possesses engineering semantic understanding capabilities. The effect is that the system can automatically recommend or apply preset mesh generation strategies based on component semantics (such as the leading edge of an airfoil), thereby generating higher-quality meshes more suitable for computation in key aerodynamic regions.

[0241] (5) Possesses strong domain generalization ability and knowledge accumulation value;

[0242] Because the core framework of this invention is not dependent on any specific aircraft model and supports incremental learning, the benefits are significant: only a small amount of data from new aircraft models needs to be labeled and fine-tuned to quickly adapt to new aircraft models, greatly reducing technical maintenance and expansion costs. Simultaneously, the system can distill the mesh generation experience of senior engineers into reusable models and rules, forming core knowledge assets.

[0243] Please refer to Figure 5 , Figure 5This is a flowchart illustrating a method for intelligent component recognition and segmentation of aircraft CAD models based on three-modal fusion. Example 2 provides a method for intelligent component recognition and segmentation of aircraft CAD models based on three-modal fusion, the method comprising:

[0244] Step S1 - Model Acquisition and Format Confirmation: Acquire the CAD model of the aircraft to be processed, wherein the CAD model of the aircraft has a boundary representation structure;

[0245] Step S2 - Construction of a multimodal dataset oriented towards engineering semantics: Based on the engineering requirements of mesh partitioning, engineering semantic annotation is performed on the faces in the boundary representation structure of the aircraft CAD model to obtain component semantic labels; geometric parameters of each face are extracted from the boundary representation structure; for each labeled face, a multi-view 2D rendering image is generated; the multi-view 2D rendering image, the geometric parameters, and the component semantic labels are associated and stored to form training samples;

[0246] Step S3 - Multimodal Fusion Recognition Model Training: The multimodal recognition model is trained using the training samples. The multimodal recognition model is configured to fuse input image features and language semantic information to output semantic labels for components on the surface of the aircraft CAD model.

[0247] Step S4 - Initial identification of macroscopic components based on AI: Input the CAD model of the aircraft to be identified into the trained multimodal recognition model, identify the surfaces in the aircraft CAD model, and obtain the preliminary surface-component semantic label mapping relationship;

[0248] Step S5 - Collaborative completion of small features based on geometric rules: Based on the preliminary face-part semantic label mapping relationship, for faces whose recognition confidence output by the multimodal recognition model is lower than a preset threshold or whose geometric area ratio is less than a set threshold, the geometric rule auxiliary module is activated to perform semantic completion to obtain the final part segmentation result.

[0249] Step S6 - Automatic extraction of key geometric feature lines: Based on the final component segmentation results, extract the key geometric feature lines corresponding to each component from the boundary representation structure of the aircraft CAD model;

[0250] Step S7 - Structured Engineering Semantic Output: The final component segmentation results and extracted key geometric feature lines are converted into a structured engineering semantic format for the target mesh generation software to read and utilize, and then output.

[0251] Example 2 provides a fully automated and high-precision component identification and geometric feature extraction capability for mesh generation in computational fluid dynamics and finite element analysis, which is applicable to the digital simulation and manufacturing of complex equipment such as aviation, aerospace, and ships.

[0252] Example 2 provides a method for intelligent recognition, semantic segmentation, and geometric feature extraction of aircraft components based on a three-modal fusion of vision, geometry, and language. This invention addresses the inefficiencies, inconsistencies, and automation breakpoints caused by reliance on manual component recognition in the aircraft mesh generation process. Specifically, it includes:

[0253] (1) Eliminate the manual dependence in the component identification process and realize the automatic and accurate identification and semantic grouping of each component in the aircraft CAD model;

[0254] (2) Provide engineering semantic output that can be directly used for mesh generation, including component face sets and key feature lines (such as boundary lines and densification lines).

[0255] (3) Supports preset differentiated grid strategies for different components, enabling intelligent grid planning that is configured upon identification;

[0256] (4) Improve the automation level and processing efficiency of the entire mesh generation process, and provide reliable pre-processing tools for digital simulation.

[0257] Example 2 provides a joint modeling and processing method for the visual-geometric-linguistic three modalities. Its overall architecture comprises four parts: data construction, model training, inference and recognition, geometric processing, and mesh strategy generation. The specific steps are as follows:

[0258] (1) Data Construction: A Semantic Annotation System Oriented to Grid Requirements

[0259] 1) B-Rep Surface-Level Semantic Annotation: Using mesh generation software such as NNW-GridStar, and combining experience in aerodynamic and structural mesh generation, surface-level annotations are performed on aircraft CAD models represented in B-Rep format. This includes, but is not limited to: nose, forward fuselage, mid-fuselage, aft fuselage, upper wing surface, lower wing surface, vertical tail, horizontal tail, engine nacelle, etc. After annotation, each component is visualized in a different color for easy verification.

[0260] 2) Multi-view image generation: For each B-Rep face, perform 360-degree multi-view rendering around its normal vector to generate a series of main view images. Scaling, rotation, and lighting adjustments are supported to enhance view coverage and lighting robustness.

[0261] 3) Dataset Augmentation and Construction: Data diversity is enhanced using geometric transformations (rotation, translation, scaling), color perturbation, and noise injection. A structured multimodal dataset is constructed, with each data point including: B-Rep facet ID, multi-view image set, semantic labels, and original geometric parameters.

[0262] (2) Model training:

[0263] 1) Model Architecture Design: ViT (Vision Transformer) is used as the visual encoder to extract image features. CrossAttn (a cross-attention mechanism) is used to align visual features with linguistic features (from QwenLM). QwenLM serves as the language understanding and generation module, responsible for semantic mapping and prompt response.

[0264] 2) Supervised fine-tuning strategy: The input is the main view image of the face, the corresponding semantic label, and the structured prompt words (such as the face belongs to [category]).

[0265] 3) The loss function adopts a ternary combination: semantic cross-entropy loss (optimizes classification accuracy) + IoU loss (improves the overlap of segmented regions) + boundary weighted loss (enhances the accuracy of boundary pixel recognition).

[0266] The loss function is:

[0267] ;

[0268] in, For semantic cross-entropy loss, For IoU loss, For boundary-weighted loss, the weighting coefficients are initially set. , , Adjustments can be made based on subsequent experiments.

[0269] Semantic cross-entropy loss:

[0270] ;

[0271] In the formula, N is the total number of pixels, i.e., the total number of faces in the B-Rep model, C is the number of semantic categories of the component, and y i,c Let p be the true label of the i-th pixel belonging to category c. i,c Let be the probability predicted by the model that the i-th pixel belongs to category c.

[0272] The IoU loss is calculated by averaging the IoU for each part category:

[0273] ;

[0274] Boundary loss ( The initial setting is 0.3, which can be adjusted based on experiments. Boundary binary classification loss:

[0275] ;

[0276] In the formula, This represents the true boundary label of the i-th pixel. This indicates that the pixel is located inside the component. This indicates that the pixel is located on the geometric boundary of two different components; This represents the probability that a pixel predicted by the model is a boundary. Based on the topological relationships of the B-Rep model, when two adjacent faces belong to different components, their shared edge is marked as a boundary.

[0277] Loss for distance transformation:

[0278] ;

[0279] In the formula, The total number of boundary pixels, This represents the distance from pixel i to the nearest real geometric boundary. This represents the distance from pixel i predicted by the model to the boundary. This forces the predicted boundary lines to be closer to the actual geometric boundaries, improving the geometric usability of the segmentation results in CAD models.

[0280] 4) The training employs progressive fine-tuning, first pre-training on large-scale general image-text data, and then fine-tuning on domain image-text pairs labeled with semantic tags for each component of the aircraft. This approach can utilize prior knowledge of general visual language while enabling the model to accurately grasp the fine-grained semantics of the aircraft's unique components, thus improving domain adaptability and recognition accuracy.

[0281] (3) Reasoning and recognition:

[0282] 1) Input: The CAD model of the aircraft to be processed (B-Rep) and natural language prompts (e.g., please identify all fuselage surfaces).

[0283] 2) Layered recognition processing flow:

[0284] The first layer is the overall component recognition: the model extracts visual features face by face, combines them with the semantics of prompt words to make inferences, and prioritizes the identification of the main structural components, i.e., large components. The main structural components refer to the components with independent functions on the aircraft, such as the entire wing, the entire fuselage, and the tail. For simplicity, they will be referred to as large components in the following descriptions. The primary semantic labels of each B-Rep face are output to form a face-component mapping table.

[0285] The second layer is small component / feature geometric completion: For small surfaces (such as surfaces belonging to the trailing edge of an wing or local protrusions) whose confidence level is lower than the set threshold or ignored by the model in the previous round of recognition, the geometric rule assistance module is activated instead of relying directly on the AI ​​model for secondary recognition.

[0286] Widgets / features mainly include the following types:

[0287] Connecting feature surfaces or edge surfaces: Surfaces located at the junction of two large components (such as the upper and lower surfaces of an wing), such as the trailing edge and leading edge of an wing.

[0288] Featured surfaces with opening, closing, or maintenance functions: surfaces surrounded by large components and whose orientation differs significantly from the main body, such as hatches and maintenance access covers.

[0289] Localized raised or recessed surfaces: such as rivet heads, sensor mounting bases, oil filler caps, and other small geometric features.

[0290] Low-confidence recognition surfaces: Surfaces that AI models cannot recognize with high confidence due to reasons such as blurred geometric features and insufficient training data coverage.

[0291] 3) Working principle of the geometric rule auxiliary module:

[0292] Input: The set of large components identified by AI (such as the upper surface of the left wing and the lower surface of the left wing) and all the B-Rep surfaces contained therein.

[0293] Processing: Based on the geometric topology and parametric information of the CAD model, preset geometric rules are executed. For example, Rule 1: If a surface belongs to the upper or lower surface of the wing, or is located at the boundary between the two, at the leading edge of the wing, has high curvature, and its spanwise length is close to the entire wing span, it is classified as the leading edge of the wing. Rule 2: If a surface is located at the boundary between the upper and lower surfaces of the wing, and its area is less than 1% of the overall wing area, it is classified as the trailing edge of the wing. Rule 3: If a surface is directly adjacent to the main wing but its normal is significantly deviated from the main wing plane, and it is located at the outermost point of the wing spanwise, with an area between 0.5% and 5% of the main wing area, it is classified as a winglet or wingtip. Rule 4: If a surface is surrounded by a large fuselage component, and its normal is approximately perpendicular to the main fuselage direction, and its area is small, it may be an access hatch or door, and further determination is made based on the geometric characteristics of its adjacent surfaces. Rule 5: If a surface belongs to a large component of the engine nacelle, is located at the foremost end of the nacelle, and has a closed annular boundary and a smooth Gaussian positive curvature transition, then it is classified as an air intake lip.

[0294] Output: Add annotations to the corresponding part categories for facets that conform to the geometric rules, or create new subcategory labels for them.

[0295] 4) Post-processing: Label smoothing is performed based on face adjacency relationships, integrating AI recognition results with geometric rule supplementary results to avoid misjudgment of isolated faces. Component-level result visualization and export are supported.

[0296] (4) Geometric treatment:

[0297] 1) Key Geometric Feature Line Extraction: Based on the complete identification of all components (including large components identified by AI and small features supplemented by geometric rules), boundary lines, curvature extrema lines, and feature contour lines of each component are extracted using B-Rep topology and geometric information. For large components identified by AI, geometric algorithms (such as curvature-based edge detection and topological connectivity analysis) are used for automated extraction. For small features identified by the geometric rule-assisted module, feature lines are extracted directly using their geometric definitions (such as the trailing edge line, i.e., the intersection of the upper and lower wing surfaces), resulting in higher accuracy.

[0298] 2) Output Interface: Feature lines are output in a format supported by NNW-GridStar. Seamless integration with simulation workflows is supported, improving CAE analysis efficiency.

[0299] The visual encoder (ViT) can be replaced by other mainstream networks (such as ConvNeXt, ResNet), and this embodiment of the invention does not impose any restrictions on it. The language model (QwenLM) can be replaced by other large language models (such as GLM, LLaMA), and this embodiment of the invention does not impose any restrictions on it. The specific judgment conditions and parameters of the geometric rules can be adjusted according to the characteristics of different components. The output interface can be adapted to different target mesh generation software (such as ANSA, HyperMesh).

[0300] Please refer to Figure 6 , Figure 6 This is a schematic diagram of the architecture of an intelligent component recognition and segmentation system for aircraft CAD models based on three-modal fusion. Embodiment 2 of the present invention provides an intelligent component recognition and segmentation system for aircraft CAD models based on three-modal fusion. The system adopts a layered architecture and aims to serve the automated mesh generation in aircraft digital simulation, including:

[0301] The data construction layer, configured to build a multimodal training dataset oriented towards grid engineering semantics, includes:

[0302] Semantic annotation unit is used to perform engineering semantic annotation on the faces in the boundary representation structure of the aircraft CAD model and generate component semantic tags;

[0303] A geometric parameter extraction unit is used to extract the geometric parameters of each face from the boundary representation structure;

[0304] The multi-view rendering unit is used to generate multi-view 2D rendering surface images for each labeled surface;

[0305] The data management unit is used to associate and store the surface image, the corresponding geometric parameters, and the semantic labels of the components to form structured training samples;

[0306] A model training layer, connected to the data construction layer, is configured to train a multimodal recognition model using the multimodal training dataset. This layer includes:

[0307] The model architecture management unit is used to build and maintain a neural network architecture that includes a visual encoder, a language understanding module, and a cross-modal fusion module.

[0308] A training execution unit is used to perform supervised training on the neural network architecture by inputting the training samples;

[0309] An optimization unit is used to optimize the multimodal recognition model using a combined loss function;

[0310] The inference and recognition layer, connected to the model training layer, is configured to load the trained multimodal recognition model and perform intelligent recognition and segmentation on the input aircraft CAD model. This layer includes:

[0311] The first recognition sublayer is used to run the multimodal recognition model, perform preliminary recognition of the surfaces of the CAD model, and output preliminary surface-part semantic label mapping relationships;

[0312] The rule-assisted sublayer, as a geometric rule-assisted module, is used to determine, complete, and correct semantic labels for low-confidence or unrecognized surfaces output by the first recognition sublayer based on predefined geometric topology and parameterized rules.

[0313] The post-processing sub-layer is used to smooth and integrate the recognition results based on the face adjacency relationship, forming the final complete component segmentation result;

[0314] A geometry processing and output layer, connected to the inference and recognition layer, is configured to perform engineering processing and output of the recognition results. This layer includes:

[0315] The feature line extraction unit is used to automatically extract the key geometric feature lines corresponding to each component from the final component segmentation result based on the boundary representation structure of the CAD model;

[0316] The format conversion and output unit is used to convert the component segmentation results and key geometric feature lines into a structured engineering semantic format that can be read by the target mesh generation software and output it.

[0317] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0318] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for geometric defect identification and processing based on a large AI model, characterized in that, The method includes: Step 1: Obtain the CAD model of the aircraft to be processed, and divide the CAD model of the aircraft into multiple parts; Step 2: Obtain multiple labeled sample parts, and convert the geometric data of each sample part into a structured feature representation for processing by the deep learning model to obtain a training set. The structured feature representation includes at least one of the graph structure, point cloud, and feature vector of the sample part; the labeling includes the type and spatial location information of the geometric defects existing in the sample part. Step 3: Use the training set to train a multi-task hierarchical graph neural network model to obtain the trained multi-task hierarchical graph neural network model; Step 4: For any component obtained from the segmentation in Step 1, convert the geometric data of the component into the same structured feature representation as in Step 2, and input the converted data into the trained multi-task hierarchical graph neural network model; the trained multi-task hierarchical graph neural network model outputs the type and spatial location information of the geometric defects contained in the component. Step 5: Based on the geometric defect type and spatial location information output in Step 4, firstly, execute the topology repair algorithm for defects identified as gaps to restore the topological connectivity at that location; then, execute the corresponding geometric repair algorithm for defects identified as geometric anomalies, and perform mesh stitching on the repaired results of all components to generate the final surface mesh. In step 3, a data augmentation strategy that enhances domain knowledge is adopted during the training phase of the multi-task hierarchical graph neural network model. The data augmentation strategy includes simulating the defect generation pattern in the aerospace field, specifically: randomly introducing micron-level gaps at the surface splicing points, and / or generating irregular holes in the planar area to simulate process damage.

2. The method for geometric defect identification and processing based on an AI large model according to claim 1, characterized in that, Step 1 specifically includes: Step S1: Obtain the CAD model of the aircraft to be processed, wherein the CAD model of the aircraft has a boundary representation structure; Step S2: Based on the engineering requirements of mesh generation, perform engineering semantic annotation on the faces in the boundary representation structure of the aircraft CAD model to obtain component semantic labels; extract the geometric parameters of each face from the boundary representation structure; generate a multi-view 2D rendering image for each annotated face; associate and store the multi-view 2D rendering image, the geometric parameters, and the component semantic labels to form training samples. Step S3: Train a multimodal recognition model using the training samples. The multimodal recognition model is configured to fuse input image features and language semantic information to output semantic labels for components on the surface of the aircraft CAD model. Step S4: Input the aircraft CAD model to be identified into the trained multimodal recognition model, identify the surfaces in the aircraft CAD model, and obtain the preliminary surface-part semantic label mapping relationship; Step S5: Based on the preliminary face-part semantic label mapping relationship, for faces whose recognition confidence output by the multimodal recognition model is lower than a preset threshold or whose geometric area ratio is less than a set threshold, the geometric rule auxiliary module is activated to perform semantic completion to obtain the final part segmentation result. Step S6: Based on the final component segmentation result, extract the key geometric feature lines corresponding to each component from the boundary representation structure of the aircraft CAD model; Step S7: Convert the final component segmentation results and the extracted key geometric feature lines into a structured engineering semantic format for the target mesh generation software to read and utilize, and then output it.

3. The method for geometric defect identification and processing based on an AI large model according to claim 1, characterized in that, In step 2, the structured features are represented as a graph structure; wherein, the faces of the components are treated as nodes of the graph, and the shared edges between faces are treated as edges of the graph; the node features include the geometric properties of the faces, and the edge features include the geometric properties of the shared edges.

4. The method for geometric defect identification and processing based on an AI large model according to claim 1, characterized in that, In step 3, the multi-task hierarchical graph neural network model includes: A multi-scale feature extraction layer is used to encode geometric features at the node, edge, and subgraph levels. Hierarchical graph pooling and attention layers are used to perform hierarchical clustering of patches in different regions and assign attention weights; The multi-task output head includes a gap recognition head, a surface recognition head, a hole boundary recognition head, and a defect severity assessment head connected in parallel.

5. The method for geometric defect identification and processing based on an AI large model according to claim 4, characterized in that, In the hierarchical graph pooling and attention layer, an attention weight guidance mechanism based on domain knowledge is introduced; different prior attention weight coefficients are assigned to the facets according to the aerodynamic critical region or manufacturing process region where the facet is located.

6. The method for geometric defect identification and processing based on a large AI model according to claim 4, characterized in that, When training the multi-task hierarchical graph neural network model, a weighted multi-task loss function is used; in the weighted multi-task loss function, the weight of the loss term for each defect category is inversely proportional to the number of samples of that defect category in the training set.

7. The method for geometric defect identification and processing based on an AI large model according to claim 1, characterized in that, Step 5 specifically includes: Step 51: Based on the spatial location information of the gap, call the gap repair algorithm at the boundary representation level to restore the topological connection of adjacent surfaces and obtain the intermediate model after B-Rep level topological repair; Step 52: Discretize the intermediate model into a triangular shape to obtain a surface mesh; based on the spatial location information of holes, overlaps, and fragments, sequentially execute the hole-filling algorithm, the algorithm for deleting or merging overlapping patches, and the algorithm for fusing or removing fragments on the surface mesh; Step 53: Perform a mesh stitching operation at adjacent boundaries on the surface meshes corresponding to all components that have completed step 52 to generate the final surface mesh.

8. The method for geometric defect identification and processing based on an AI large model according to claim 1, characterized in that, In step 2, when converting the geometric data of the sample component into a structured feature representation, the extracted geometric features include aerodynamic sensitivity features and / or process constraint features.

9. The method for geometric defect identification and processing based on an AI large model according to claim 1, characterized in that, In step 4, the spatial location information output by the multi-task hierarchical graph neural network model is used to identify edges for gap defects, boundary loops for hole defects, and surfaces for overlapping or fragmented defects.

Images