A finite element intelligent modeling and interactive analysis integrated method based on structural semantic graph

Abstract

This invention discloses an integrated method for intelligent finite element modeling and interactive analysis based on structural semantic graphs. For structural engineering drawings, the method first performs entity analysis, initial semantic judgment of layers, and noise cleanup to determine adaptive tolerance. Then, it constructs a connected entity graph and completes candidate component grouping. A multimodal visual model, a graph-structured prediction model, and a large language model constraint inference module are introduced to generate a structural semantic graph with confidence. Consistency verification and a self-correction loop are used to improve the robustness of recognition and modeling. The structural semantic graph is mapped to a three-dimensional finite element analysis model, and consistency verification, local repair, and executability verification are performed on the generated model. An AI agent automatically generates and verifies the analysis script, executes it in a sandbox environment, and returns the results. The results are fed back into the structural semantic graph and visualization graphics, achieving a graph-model-result-graph closed loop. This invention can significantly reduce the threshold for structural modeling and analysis, and improve efficiency and traceability.

Description

Technical Field

[0001] This application relates to the field of digital technology in civil and structural engineering, and in particular to an integrated method for intelligent finite element modeling and interactive analysis based on structural semantic graphs. Background Technology

[0002] In structural engineering design, construction, and operation and maintenance, high-precision structural analysis models are a crucial foundation for load-bearing scheme comparison, performance evaluation, and rapid post-disaster assessment. However, current engineering practices generally employ a workflow of "drawings - manual modeling - analysis setup - solution calculation - post-processing." This traditional method suffers from the following prominent problems: 1) Modeling is highly dependent on human experience: Modelers need to manually extract component boundaries, node coordinates, and connection relationships from 2D drawings and repeatedly input them into the analysis software, which is time-consuming and prone to omissions, incorrect connections, and inconsistent parameters; 2) Heterogeneous drawing information and high noise interference: Drawings often contain non-structural entities such as annotations, text, dimension lines, and auxiliary lines; inconsistent layer naming and drawing habits across different projects; and problems such as loose entity endpoints, fragmented line segments, and unclosed component boundaries further exacerbate the difficulty of automatic recognition. 3) The analysis task setting threshold is high: static, modal, response spectrum, time history, pushover and other analyses involve professional configurations such as algorithm selection, convergence parameters, and output control, which are difficult for non-experts to complete accurately; 4) Lack of a closed loop of "drawing-model-result-drawing": Even if some drawing recognition or automatic modeling tools exist, they are often disconnected from the analysis engine and post-processing visualization; model modification and result display cannot be stably written back to the component semantic layer, resulting in poor traceability and reproducibility; 5) Lack of verifiable intelligent automation links: Simply relying on AI to generate analysis code or analysis configurations without strict code verification, execution isolation and exception self-correction mechanisms makes it difficult to reliably implement in engineering scenarios.

[0003] Therefore, there is an urgent need for an end-to-end automated technology solution that can connect "drawing analysis - semantic recognition - 3D finite element mechanical model generation - automatic analysis and calculation - result write-back." This solution effectively couples the geometric information and engineering semantics of structural engineering drawings without extensive manual intervention, automatically generating a 3D mechanical model (including necessary elements such as nodes, elements, material sections, boundaries, and loads) that meets the requirements of finite element solutions. It can also generate corresponding calculation task configurations and solution instructions based on user analysis needs, completing solution execution and result post-processing. Simultaneously, it stably maps analysis results such as nodal responses and component internal forces / stresses back to the original drawings and component semantic layers, forming a consistent closed loop of "drawing-model-result-drawing." This solution overcomes the problems of existing technologies, such as modeling relying on experience, model error-proneness and difficulty in reproduction, high threshold for analysis configuration, and unreliable calculation links, thereby improving the efficiency, reliability, and engineering usability of structural analysis model construction and evaluation. Summary of the Invention

[0004] To address the problems of insufficient robustness in structural drawing recognition, rule-dependent automatic modeling with weak generalization ability, high threshold for setting analysis tasks, and lack of a reliable execution and result feedback loop in existing technologies, this invention provides an integrated method for finite element intelligent modeling and interactive analysis based on structural semantic graphs.

[0005] Technical solution of the present invention: A method for integrating intelligent finite element modeling and interactive analysis based on structural semantic graphs includes the following steps: Step S1: Drawing input and entity analysis; Receive structural engineering drawing files, parse the structural engineering drawings to obtain a set of graphic entities, and extract the entity type, geometric information and the layer information of each graphic entity; Step S2: Initial semantic assessment and noise removal of layers; Based on the layer name, layer attributes, and preset filtering rules, entities (including annotations, text, and dimensions) in non-structural component layers are removed, and the remaining layers are classified by component type to obtain candidate entity sets for grid lines, beams, columns, walls, and slabs. Step S3: Determine the adaptive tolerance; An adaptive tolerance parameter is determined based on the scale statistics of the graphic entity set. The adaptive tolerance parameter is used for endpoint snapping, node merging, adjacent / intersecting relationship determination, and component group merging and splitting. "Adaptive" means that the tolerance parameter is automatically generated based on the scale statistics of the current drawing rather than using a fixed threshold. Step S4: Entity connectivity graph construction and component candidate grouping; A connected graph of entities is constructed using graphical entities as nodes and geometric relationships between entities that satisfy the conditions of intersection or adjacency as edges. Candidate groups of components are then extracted based on the connected components. Step S5: AI-enhanced component semantic recognition and structural semantic graph generation; For each candidate group of components, a structural semantic graph is constructed, with structural nodes as vertices and structural components as edges. An AI recognition model is introduced to predict and correct the component type, component boundary, connection topology and component attributes to obtain a structural semantic graph with confidence. The AI ​​recognition model includes: Multimodal visual model: used for visual recognition of images obtained from rendering structural engineering drawings, to assist in the identification of complex or special components; Structured prediction model: Input entity connectivity graph or structural semantic graph, comprehensively utilize layer text features, geometric features and topological features, output component categories and connection relationships; Constraint reasoning module: Based on a large language model, it is used to verify the consistency between the prediction results and the structural engineering constraints, and generate correction strategies when conflicts exist; Step S6: Graph-model mapping and finite element model generation; The structural semantic diagram in step S5 is mapped to the finite element analysis model data structure, generating three-dimensional coordinates, nodes, elements, material and section definitions, boundary conditions, mass and load information; among them, the floor height parameter is introduced into the components of the two-dimensional drawing and the Z coordinate is generated to realize the two-dimensional to three-dimensional mapping. Step S7: Model consistency verification and self-correction closed loop; The stability and solvability of the finite element analysis model generated in step S6 are checked. The check includes connection integrity, rationality of degree of freedom constraints, and legality of material and section parameters. When the check fails, the constraint reasoning module in step S5 is called to generate a repair operation and incrementally update the structural semantic diagram and finite element analysis model until the check passes or the preset iteration limit is reached. Step S8: Dialogue-based interactive analysis task orchestration; The system receives structural analysis requests from users in natural language, parses them into analysis task types and parameter slot sets, and generates corresponding finite element analysis scripts by an AI agent. Step S9: Code verification and secure execution; The generated finite element analysis script undergoes syntax checks, logic checks, parameter validity checks, and security checks. Once the checks pass, the finite element analysis engine is invoked to perform calculations in an isolated sandbox execution environment. Step S10: Result post-processing and model-graph re-feedback; The calculated nodal responses, component internal forces, or stress results are mapped back to the vertex and edge attributes of the structural semantic graph, and the visualization results, structural data files, and model files are output, realizing a closed loop of graph-model conversion.

[0006] Specifically, step S1 adapts to the differences in CAD drawings from different sources and completes the standardized input of drawings and the structured representation of graphic elements. Details are as follows: Receive structural engineering drawing files and establish a metadata set. The metadata set refers to basic configuration information related to drawing processing, modeling, and output, including unit system, coordinate datum, file version identifier, number of floors and floor height parameters or default strategy, default material and section parameter library index, and output file naming and version management rules.

[0007] The process involves first normalizing the units and coordinates of the DXF vector file; then parsing the primitive objects in the DXF vector file, constructing a set of graphic entities, and establishing a unique identifier for each entity.

[0008] For each graphic entity: extract geometric information and generate the corresponding geometric representation; extract the information of the layer to which it belongs and establish entity-layer association.

[0009] Perform preliminary quality control on the set of graphic entities and generate processing records for abnormal entities.

[0010] The final output includes a collection of graphic entities, an entity attribute table, and entity-layer association information.

[0011] Step S2 is as follows: The purpose of initial semantic judgment and noise removal of layers is to reduce the impact of differences between different engineering drawings on subsequent component identification and to distinguish between structural entities and non-structural noise entities. The processing procedure is as follows: Step S2.1: Extraction and normalization of semantic features of the layer; The graphic entity set is summarized by layer, the semantic features of the layers are extracted, and the layer names are normalized. The normalization process includes case unification, symbol and space standardization, and synonym mapping; the extracted features are then written into the layer feature table for use in subsequent steps.

[0012] Step S2.2 Preliminary classification of layer semantics based on rules and priors; Based on the layer name, layer attributes, and preset keyword or pattern library, the layers are classified according to component type, resulting in candidate entity sets for grid lines, beams, columns, walls, and slabs.

[0013] Step S2.3 Cleaning up unstructured noise layers and noise entities; Remove non-structural layers or entities and create a record for the removed objects.

[0014] After completing the above processing, the output includes a set of structure-related entities, a set of candidate entities, layer classification results, and a noise cleanup log.

[0015] In step S3, the adaptive tolerance establishment method is as follows: Based on the set of graphic entities obtained in step S1, extract the scale statistics of the drawing, including: median, quantile interval and interquartile range, and obtain representative scale indicators in a robust statistical manner.

[0016] An adaptive tolerance parameter set is generated based on the representative scale index.

[0017] Different tolerance sub-parameters are set for different candidate categories, and consistency checks are performed. The consistency check refers to determining whether the tolerance parameters match the current drawing scale and the geometric expression of the component based on the geometric feature statistics of the candidate category. If they do not match, the system will revert to the default range or trigger user confirmation.

[0018] The resulting tolerance system is provided for subsequent steps.

[0019] Step S4 is as follows: Step S4.1 Define the nodes, edges, and their attributes in the entity connected graph; Using graphical entities or their geometric primitives as nodes, edges are established based on the geometric relationships between entities to construct a connected entity graph. Node attributes include entity type, geometric information, and layer information; edge attributes include relationship type and connection strength.

[0020] Step S4.2: Edge generation mechanism based on coarse and fine judgment; Edge generation is achieved through a two-stage process of coarse and fine evaluation. In the coarse judgment stage, the bounding boxes of any two entities are expanded according to the adjacent judgment tolerance. If the expanded bounding boxes overlap or the distance is less than the threshold, they are included in the candidate entity pair. In the detailed judgment stage, a fine geometric judgment is performed on the candidate entity pairs. The judgment includes: the endpoint distance is less than the endpoint adsorption tolerance, the line segments or broken lines intersect, the projections overlap, the near collinearity and the spacing is less than the threshold, and the intersection or adjacency judgment of the discrete primitives of the curve. When the precise determination is valid, establish the corresponding connection and record the geometric evidence type.

[0021] Step S4.3 Differentiation constraints for connections within the same layer and across layers; A differentiated connection strategy is adopted for edges within the same layer and across layers. This means that edges within the same layer and across layers can coexist, but a differentiated strategy is adopted: edges within the same layer are used for merging fragments within a component, and a more lenient threshold or a higher connection weight is used; edges across layers are used to express the connection relationship between components, and a more stringent threshold or a lower connection weight is used to reduce false connections.

[0022] Step S4.4 Connectivity component extraction and component candidate group formation; Connected components are extracted from the entity connected graph to form candidate groups of components, which are then filtered and regularized.

[0023] Step S4.5: Structured output of candidate component groups; A structured description is generated for each candidate group of components and provided to subsequent steps.

[0024] Step S5 is used to transform the candidate entity set, adaptive tolerance parameter, entity connectivity graph, and component candidate group into a structural semantic graph. This invention employs a collaborative processing mechanism of "visual supplementary judgment—graph model primary judgment—constraint adjudication": Step S5.2 serves as the primary judgment path, performing semantic and topological predictions on the candidate group; Step S5.1 triggers visual supplementary judgment for low-confidence or high-conflict candidate groups; Step S5.3 performs constraint reasoning and conflict adjudication on the output results and writes the corrected results back to the structural semantic graph, thereby obtaining a structural semantic graph with candidate set, confidence level, and evidence source information.

[0025] Step S5 is as follows: Step S5.0 Construction and unified representation of structural semantic graph objects; Step S5.0.1 Initial structural semantic graph construction and field constraints; The preliminary judgment results of candidate entity categories obtained in step S2, the adaptive tolerance parameter set obtained in step S3, and the entity connectivity graph and component candidate groups obtained in step S4 are uniformly constructed into an initial structural semantic graph: Among them, the vertex set Represents structural nodes; Vertex attribute set This is a set of attribute fields for a vertex, including: a unique node identifier (ID), two-dimensional plane coordinates (x, y), three-dimensional coordinate placeholders or floor indexes and elevation information, a node type marker, a node source marker, and a node confidence score. The node type marker includes endpoints, intersections, connection points, and control points; the node source marker includes rule extraction, model inference, and manual correction.

[0026] edge set Indicates structural components; edge attribute set This is a set of attribute fields for an edge, including: unique component identifier (ID), component type, component geometric boundary representation, set of node IDs connected to the component, component principal direction, local coordinate information, component attribute fields, and component confidence level. The component type includes beams, columns, walls, floor slabs, grid lines, and others; the component attribute fields include material, section parameters, thickness, connection method, or boundary conditions.

[0027] Mapping set Used to record the mapping link of "original graphic entity - candidate group - structural component structural node".

[0028] Step S5.0.2 Explicitly represent uncertainty; To avoid error accumulation caused by early, single hard decisions, key fields in the structural semantic graph are explicitly represented using a "candidate set + confidence distribution" approach. These key fields include a candidate set of component types. Connectivity topology candidate set and attribute candidate set For each candidate key field, the candidate value, confidence level, source model, and version number are recorded. Before completing the consistency decision in step S5.3, low-scoring candidates are not directly deleted; only their candidate ranking is adjusted.

[0029] Step S5.0.3 Multi-source evidence fusion and confidence level merging; The overall confidence level for any field f (including component type, connection topology, and attributes). The confidence level after fusion is calculated as follows: in, This represents the confidence level given by the m-th source of evidence for field f. σ(·) represents the weight of the corresponding evidence source, k is the total number of evidence sources; the evidence sources include five categories: rule evidence, geometric evidence, visual evidence, graphical model evidence and constraint reasoning evidence; σ(·) is the Sigmoid function.

[0030] After fusion, the confidence scores for component type, connection topology, and attribute are obtained, and these are written into the structural semantic graph along with the evidence source.

[0031] Step S5.1 Multimodal visual recognition model-assisted interpretation; This step is used to perform visual supplementary judgment on complex candidate groups that are difficult to distinguish stably by rule-based methods and graph models.

[0032] In one embodiment of the present invention, the multimodal visual recognition model adopts the Alibaba Cloud Tongyi Qianwen open-source model Qwen2.5-VL-7B-Instruct, and combines it with structural engineering CAD rendering data for instruction fine-tuning or parameter fine-tuning, which is suitable for auxiliary recognition of local complex components in DXF renderings.

[0033] Step S5.1.1 Triggering conditions and applicable scope of visual recognition; Visual recognition is triggered when a candidate group meets any of the following conditions: (1) The highest posterior confidence of the current component type in the candidate group Less than the first threshold (Values ​​range from 0.70 to 0.80); (2) The candidate group satisfies the geometric complexity condition, which includes any one of the following: a. The candidate group contains no fewer than 12 original graphic entities; b. The proportion of overlapping entities within the smallest bounding rectangle of the candidate group Not less than 0.15; c. Near-closed boundary gap With boundary perimeter ratio Not greater than 0.05; d. There shall be no fewer than two sets of parallel double lines or composite lines, and the coefficient of variation of the line spacing shall not exceed 0.20; e. The proportion of the area of ​​the interfering entity to the area of ​​the circumscribed rectangle of the candidate group Not less than 0.20; (3) The candidate group satisfies the topological conflict condition, and the topological conflict condition is preferably determined by the conflict score. control: in, This indicates the proportion of beam ends that are not connected to vertical load-bearing member nodes. Indicates the proportion of discontinuity at the wall boundary. Indicates the proportion of isolated nodes in the candidate group; when A score ≥ 0.30 indicates a significant topological conflict. Candidate groups with a component type confidence score higher than the second threshold (0.85) and that do not meet the geometric complexity and topological conflict conditions are not triggered for visual recognition to reduce inference overhead.

[0034] Step S5.1.2 DXF rendering and localized input organization; For the candidate groups output in step S4, the DXF local regions are rendered as images, and a global context map, a local cropping map, and a target highlight map are generated. Simultaneously, text prompts are organized and input into the multimodal visual recognition model along with the aforementioned images. The model outputs candidate component categories, candidate boundaries or key points, and confidence scores in structured JSON format.

[0035] The text prompts preferably use a uniform template to reduce the impact of input differences on the recognition results. The prompts include: candidate group identifier, main layer name, number of entities, size of the bounding rectangle, main direction, prior information of candidate categories, and description of the task to be recognized.

[0036] Step S5.1.3 Output format and fusion strategy; The output of step S5.1 includes candidate component categories, candidate boundaries or key points, visual confidence level, and visual interpretation markers. If the visual output is consistent with the output of step S5.2, the overall confidence level is increased; if there is a conflict, multiple candidates are retained and submitted to step S5.3 for adjudication; if the confidence level of the visual output is lower than the threshold, it is archived only as supplementary evidence.

[0037] Step S5.2 Structured Deep Learning Prediction Model: Component Semantic Recognition and Topological Relationship Inference; This step is the main decision module for component semantic recognition and topological relationship inference. It takes the entity connectivity graph as input, jointly encodes layer text, geometric shape, and topological information, and predicts component category, instance affiliation, connection relationship, and end connection node.

[0038] Step S5.2.1 Input graph structure and feature construction; Step S5.2.1.1 Input graph definition; The entity-connected graph or its corresponding candidate subgraph obtained in step S4 is used as the input graph. Nodes represent graphic entities or geometric primitives, and edges represent the connection relationships between entities, with the connection strength attached.

[0039] Step S5.2.1.2 Node Feature and Edge Feature Encoding For each node, construct the entity type, geometric features, local topological features, and layer features, and generate the initial representation of the node.

[0040] For each side, construct distance, intersection, overlap, angle, cross-layer markers, and tolerance matching features, and generate edge features.

[0041] Step S5.2.2 Label system and training data construction; Step S5.2.2.1 Multi-granularity supervision labeling system; A multi-granularity labeling system is established, aligned one-to-one with the fields of the structural semantic graph. This labeling system is generated based on manually annotated data, rule recognition results, and user correction records, mapping original graphical entities, inter-entity relationships, component instance affiliation, and topological connection information to corresponding labels, including: Entity-level tags are used to supervise the semantic categories of components; Relationship-level labels are used to monitor the types of relationships or connection probabilities between entities; Instance-level tags are used to monitor the component instances to which an entity belongs; Topology-level tags are used to monitor component end nodes, plate boundary support relationships, and wall boundary continuity relationships.

[0042] Step S5.2.2.2 Training data sources and weakly supervised sample construction; The training data consists of real engineering drawing annotation data, programmatically synthesized data, weakly supervised data, and user correction logs. The weakly supervised data is generated based on the rule recognition results of steps S2 to S4: step S2 provides initial entity category candidates, step S3 provides an adaptive tolerance system, and step S4 provides entity connectivity graphs, edge relationships, and candidate group partitioning results; based on this, initial rule labels are generated using preset engineering rules.

[0043] The initial rule label is added to the training set as a weak label only if the following conditions are met: a. Rule confidence Preferred =0.80; b. The candidate samples meet the local topological consistency requirements; The local topological consistency is used to characterize the degree to which candidate samples meet engineering connection rules in the local structural topology, including beam end connection consistency, wall boundary continuity, and floor slab boundary support coordination. The determination method is as follows: based on the initial entity category judgment results, entity edge relationships, candidate group division results, and adaptive tolerance obtained in steps S2 to S4, a rule check is performed on the local connection relationships of the candidate samples. Specifically, beam end connection consistency refers to the existence of intersecting, overlapping, or endpoint adsorption relationships between beam ends and vertical load-bearing components such as columns and walls, or the distance from the beam end to the nearest vertical load-bearing component boundary or connection node is not greater than the connection judgment tolerance. Wall boundary continuity refers to the endpoint spacing between adjacent boundary segments belonging to the same wall candidate not being greater than the continuity tolerance, and the misalignment between adjacent boundary segments not being greater than the continuity tolerance. Floor slab boundary support coordination refers to the existence of a support correspondence between the floor slab boundary and supporting components such as beams and walls, and the distance between corresponding boundaries not being greater than the support coordination tolerance. When the above conditions are met, the candidate sample is determined to meet the local topological consistency requirements; otherwise, it is determined not to meet them. If there is no mismatch between the floor slab boundary and supporting components such as beams and walls, the candidate sample is deemed to meet the local topological consistency requirement; otherwise, it is deemed not to meet the requirement. c. Does not violate hard constraint rules; The hard constraint rules refer to the engineering rule constraints that must be met in component identification and topology determination, including connection rationality constraints, boundary continuity constraints, and component geometric rationality constraints. d. There is no strong conflict with the visual supplementary judgment result or the historical manual correction result in step S5.1.

[0044] Step S5.2.2.3 Perturbation enhancement; To improve the model's generalization ability to non-ideal engineering drawings, during the training phase, endpoint jitter, slight misalignment, line segment breakage, noise entity injection, layer name abbreviation replacement, symbol interference superposition, and scale perturbation are applied to the input drawings.

[0045] Step S5.2.3 Model Architecture and Multi-Task Output; Step S5.2.3.1 Fusion of graph encoder and multimodal features; The structured prediction model preferably employs a coding structure that concatenates an edge-aware graph neural network and a graph Transformer: Edge-aware graph neural network: First, initialize the node representation With edge features Input a two-layer edge-aware graph neural network, where i represents the node index and j represents the index of the node adjacent to node i. This represents the feature of the edge between node i and node j. Each hidden dimension is 256, and message passing units with residual connections and layer normalization are preferably used to model short-range local geometric connectivity.

[0046] Graph Transformer: The output of the graph neural network is then fed into a four-layer graph Transformer with a hidden dimension of 256, an attention head of 8, and a feedforward network dimension of 1024. Dropout=0.1 is used to model cross-candidate groups, cross-layers, and long-distance dependencies.

[0047] If step S5.1 has triggered visual supplementation, then the candidate group-level visual vectors output by the visual model will be... Text summary vectors of candidate groups A gated fusion layer is introduced and represented in the following manner with graph pooling. Fusion: in, This represents the learnable weight matrix used to calculate the gating coefficients. Indicates candidate group The gated fusion coefficient, || denotes vector concatenation. This represents element-wise multiplication. This represents the candidate group representation after fusion.

[0048] Step S5.2.3.2 Multi-task output header; After the shared graph encoder, set up multi-task output headers for node classification, edge relationship prediction, instance attribution, and topology prediction, respectively.

[0049] The node classification header outputs the probability distribution of an entity belonging to a beam, column, wall, slab, grid, or noise; the edge relationship header outputs the probability of belonging to the same component, being connected to the component, or being unrelated; the instance attribution header generates instance embedding vectors to group entities of the same component instance; and the topology prediction header outputs the association probability between the component end and the candidate node.

[0050] Step S5.2.3.3 Confidence estimation and calibration; Temperature scaling or order-preserving calibration is applied to the node classification, edge relationship prediction, and topology prediction outputs to improve the interpretability and comparability of the confidence levels. The calibrated probabilities are then used as the basis for the constraint decision in step S5.3 and subsequent manual clarification triggering.

[0051] Step S5.2.4 Training objectives, training process, and evaluation thresholds; Step S5.2.4.1 Loss function composition and topology consistency regularization; The overall training objective of the model is defined as follows: in, For node classification loss, For the loss of border relations, For instance attribution loss, For topology prediction loss, This is a topology consistency regularization term; to These are the weighting coefficients for each loss term, set according to the performance of the validation set and the convergence status.

[0052] The node classification loss uses cross-entropy loss: Where N represents the set of nodes, This indicates the total number of node classification categories; This represents the actual label of node i under category c. This represents the probability that node i is predicted to be of category c.

[0053] Edge relationship loss can be achieved using either cross-entropy loss or focus loss. The instance attribution loss uses contrastive loss: in, Represents a pair of instance entities. This represents a pair of heterogeneous instance entities, where m is the interval parameter. and Represents the instance embedding vectors of entities i and j.

[0054] Topology prediction loss is used to monitor the connection relationship between the end of a component and its nodes. Where T represents the set of component ends to be predicted, and V represents the set of candidate nodes. For real-world association tags, To predict the probability of association.

[0055] Topology consistency regularization is used to explicitly introduce structural engineering priors into the training process, and is defined as: in, This is a beam end connection constraint term used to constrain beam ends to preferentially connect to vertical load-bearing member nodes such as columns or walls; This is a wall continuity constraint term used to ensure that the wall boundary remains continuous. This is a floor slab boundary support coordination constraint, used to coordinate the floor slab boundary with the beam and wall support boundaries. , These are the weighting coefficients for the three constraint terms mentioned above, used to balance the contributions of different structural engineering priors to the overall regularization term; the calculation formulas are as follows: Indicates the assembly at the ends of the beam. This represents the set of vertical load-bearing member nodes consisting of columns or walls, and the ends of the constrained beams are preferentially connected to the vertical load-bearing member nodes. This indicates pairs of adjacent boundary segments belonging to the same wall candidate. This indicates the predicted probability that the two belong to the same wall. Indicates the distance of the boundary gap. This indicates a continuity tolerance, meaning this constraint restricts the continuity of the wall boundary. Represents the set of floor slab boundary segments. This represents the set of supporting boundaries formed by beams or walls, which constrains the floor slab boundaries to conform to the beam and wall profiles.

[0056] By combining the above loss functions, the model output not only pursues statistical classification accuracy, but also satisfies the topological rationality of structural engineering.

[0057] Step S5.2.4.2 Phased training and continuous learning; The training process includes three stages: pre-training, fine-tuning, and continuous learning. First, general patterns are learned on procedurally synthesized data and weakly supervised data. Then, fine-tuning is performed on manually labeled and corrected data, and incremental updates are made in conjunction with user correction logs.

[0058] Step S5.2.4.3 Evaluation indicator system and release threshold; Establish an evaluation index system that includes component categories, edge relationships, instance consistency, and critical connections and boundary quality. New models are only allowed to be released if they meet a preset threshold and are no lower than historical stable versions; otherwise, they are automatically rolled back.

[0059] Step S5.2.5 Output write-back and active learning closed loop; The output of step S5.2 is written back to the structural semantic graph, including candidate component types, instance attribution, candidate connection nodes, candidate boundary representations, and candidate attributes, and the confidence level and source information are recorded. For low-confidence or high-conflict regions, the system initiates interactive clarification; the user corrects the results and feeds them back into the training dataset, forming an active learning loop.

[0060] Step S5.3 Constraint reasoning and finite element instruction generation based on fine-tuning of open-source large language model; This step is used to apply engineering constraint decisions to the candidate results of steps S5.1 and S5.2, and to generate structured correction suggestions or correction instructions for the structural semantic graph and graph-module mapping process when there are conflicts, omissions or inconsistencies.

[0061] The constraint reasoning module preferably adopts an open-source large language model that has been fine-tuned using structural engineering rules, finite element modeling rules, and error correction samples.

[0062] Fine-tuning includes the following key points: 1) Construct instruction-response training samples for structural engineering scenarios. The input side includes structural semantic graph summary, visual supplementation results, graph model prediction results, conflict markers and verification context. The output side includes constraint adjudication results, correction suggestions or finite element instruction sequences. 2) The training samples are organized using a uniform structured template, enabling the model to learn the mapping relationship between fixed input fields and fixed output fields; 3) Introduce component type conflicts, connection relationship conflicts, missing parameters, inconsistent constraints, and script error repair samples into the training set to enhance the model's adaptability to adjudication and repair tasks. 4) Apply instruction set constraints and format constraints to the generated results, so that the output is preferably limited to the predefined set of finite element software commands and structured fields.

[0063] Specifically, the constraint reasoning module receives the structural semantic graph summary, the visual supplementary judgment result obtained in step 5.1, the graph model prediction result and conflict marker obtained in step 5.2, performs hard constraint verification on component type, connection topology and attribute candidate, and generates adjudication result and correction suggestion in the case of multiple candidate conflicts. When the structural semantic graph meets the minimum consistency requirement, the adjudicated structural semantic graph and the structured results required for its graph-model mapping are output for subsequent step S6. The minimum consistency requirement means that the structural semantic graph meets basic engineering modeling constraints, including: component types are identifiable, connection topology relationships are established, key attribute fields are not missing, and there are no conflicts preventing the generation of the finite element model.

[0064] Furthermore, when the consistency verification fails in subsequent steps, the constraint reasoning module generates a repair operation based on the verification report and error context.

[0065] Through this step, the multi-source outputs of steps S5.1 and S5.2 are uniformly adjudicated and transformed into engineering-executable structural semantic results.

[0066] Step S5.4: Solidify the structural semantic graph and output the external interface; The final semantic graph, candidate set, confidence level, evidence source, and correction record are solidified into structured data objects and assigned a unique version number.

[0067] The structural semantic graph and its mapping relationships are output to subsequent steps to support graph model mapping, consistency verification, task orchestration and script generation.

[0068] Step S6 is as follows: In step S6, the structural semantic graph is mapped to a finite element analysis model data structure to generate a finite element analysis model containing information on nodes, elements, materials and sections, boundary conditions, mass, and loads, and a traceable mapping relationship is established. Specifically: Step S6.1 Semantic object normalization and mapping benchmark determination; The input for this step is the structural semantic graph output from step S5, as well as project-level meta-information.

[0069] The structural nodes and structural components in the structural semantic graph are standardized and organized to form a set of nodes, a set of components, and a set of attributes. The set of attributes includes available information such as component type, geometric boundary representation, connection topology, material, cross section, or thickness.

[0070] The unit system, coordinate reference, and floor parameters are determined based on the project-level metadata, and these serve as the unified basis for subsequent coordinate mapping, floor expansion, and parameter completion, i.e., mapping reference parameters.

[0071] The output is a set of normalized semantic objects and mapping baseline parameters.

[0072] Step S6.2: 2D to 3D mapping and Z-coordinate generation; Based on the standardized semantic object set and floor parameters output in step S6.1, when the input is a two-dimensional drawing, three-dimensional coordinates are generated according to the floor height parameters and floor indexing rules: corresponding elevations are established for each floor, and the two-dimensional node coordinates belonging to the corresponding floor are mapped to three-dimensional node coordinates; if necessary, in-layer offsets are introduced to represent the elevation offsets of components such as beams and slabs. Geometric and topological relationships between floors are established for vertical components such as columns and walls, and same-floor geometric and topological relationships are established for horizontal components such as beams and slabs, while retaining the floor affiliation field. The output is a set of semantic nodes and components with three-dimensional coordinates and floor affiliation information.

[0073] Step S6.3 Finite element node generation, merging, and numbering; Based on the 3D semantic node set and adaptive tolerance parameters output in step S6.2, finite element nodes are generated according to the semantic node set. For duplicate nodes generated by endpoint adsorption or geometric nearest neighbor, node merging is performed according to the adaptive tolerance parameters to ensure consistent connection topology. A unique node number is assigned to each finite element node, and a bidirectional mapping table of "semantic node ID - finite element node number" is output.

[0074] Step S6.4: Element type mapping and element topology generation; Based on the component set obtained in step S6.2 and the finite element node set and node mapping table obtained in step S6.3, structural components are mapped to finite element elements according to component type, and element topology is generated: beam and column components are mapped to beam-column type elements, and wall and slab components are mapped to shell elements or equivalent elements; when component boundaries are missing or broken, they are completed or reconstructed based on semantic boundaries, connectivity relationships, and tolerance rules. A unique element number is assigned to each element, and a bidirectional mapping table of "semantic component ID - finite element element ID" is established, thereby obtaining the finite element element set, element topology relationships, and element mapping table.

[0075] Step S6.5 Binding of material, section, and component properties; Based on the finite element set obtained in step S6.4, and the parameter information from drawing information, project element information, default parameter library, or AI completion results, material and section or thickness object definitions are generated and bound to the corresponding elements. AI completion refers to the generation and completion of candidate material, section, or thickness parameters by a large language model based on component type, geometric information, and existing project parameters when parameters are missing. When material or section parameters are missing, they can be temporarily stored as a set of candidate parameters and filtered and corrected during subsequent consistency verification and self-correction. Material and section objects are managed using reusable IDs, recording their source, value range, and version identifier, thus obtaining a finite element set, material table, and section table with completed attribute binding.

[0076] Step S6.6 Generation of boundary conditions, mass and load information; Based on the finite element model object obtained in step S6.5, and the constraint and load information from user input, default strategy, or dialogue parsing results, boundary conditions, mass, and load information are generated in the finite element model: boundary conditions include foundation constraints, floor constraints, or connection constraints; mass is distributed by floor, node, or component; loads include dead load, live load, and the placeholder definition of dynamic action input; and the above information is written into the model data structure in a structured form, thereby obtaining a finite element model data structure containing boundary conditions, mass, and load information.

[0077] Step S6.7 Output channel planning and result index pre-binding; Based on the finite element model data structure obtained in step S6.6, and the mapping table established in steps S6.3 and S6.4, output channels are pre-planned and result indexes are established to support subsequent result feedback; the node responses and element responses that need to be recorded are defined, and the "output channel-semantic object ID-finite element number" is bound and recorded. The bound records and the bidirectional mapping table together constitute the model's traceable index system, thereby obtaining the output channel configuration and result index binding records.

[0078] Through the above steps, the data structure mapping and automatic generation from the structural semantic graph to the finite element analysis model are completed.

[0079] Step S7 is as follows: In step S7, for the finite element analysis model generated in step S6, in order to ensure that the model has engineering usability, numerical solvability and result reliability, a consistency check is performed on the finite element analysis model, and a self-correcting closed-loop update is triggered when the check fails.

[0080] Specifically, using the finite element analysis model and its mapping relationships as the verification object, a structured verification report is established, recording the problems found in each verification step, the identifiers of related objects, evidence summaries, and remediation suggestions; the consistency verification includes at least the following: Integrity check of node, element, material and section references: The integrity of the reference relationships of nodes, elements, materials and sections in the finite element model is checked to identify missing references, dangling references or unbound objects; Component connection integrity verification: Check the topological connectivity of the model and the connection relationships between components such as beams, columns, walls, and slabs to identify problems such as incorrect connections, missing connections, discontinuous boundaries, or incomplete supports; Degree of freedom constraint sufficiency and rationality verification: Check the boundary conditions and degree of freedom constraints to identify missing constraints, constraint conflicts, over-constraints, or abnormal situations that may lead to overall rigid body motion. Material and section parameter validity verification: Perform a validity check on material parameters, section parameters and their binding relationship with the corresponding elements to identify problems such as missing parameters, out-of-bounds parameters or mismatched parameter types; Overall model solvability verification: Perform a solvability check on the entire model, and perform stability checks and / or optional trial calculations as needed to determine whether the model meets the requirements for subsequent finite element solution.

[0081] When the verification fails, the constraint reasoning module described in step S5.3 is invoked. The verification report, error context, current structural semantic graph and finite element model summary are taken as input to generate a repair operation and perform local incremental updates on the structural semantic graph or finite element model.

[0082] Set an iterative control strategy for the verification, repair and update process, and record the iteration information.

[0083] Through step S7 above, the consistency verification and self-correcting update of the finite element analysis model are achieved.

[0084] Step S8 is as follows: In step S8, to enable users to submit structural analysis requirements in natural language and automatically convert them into executable finite element modeling and calculation instructions or scripts, the interactive analysis task orchestration is performed according to the following steps: Step S8.1 Dialogue input reception and context construction; Receive the user's natural language input for structural analysis and construct the current project context.

[0085] Step S8.2 Intent parsing and task type determination; Perform intent parsing on user input, extract key information including analysis type, analysis object, analysis direction, operating conditions and output requirements, and determine the analysis task type; if necessary, break down sub-tasks and establish task dependencies.

[0086] Step S8.3 Parameter slot extraction and completion; The system extracts a set of parameter slots corresponding to the task type from user input. These parameter slots refer to structured parameter fields in the analysis task, including load or excitation type, analysis direction, step size, damping parameters, convergence control parameters, output indicators, and output object range. Extraction refers to identifying the parameter expressions in the user input and mapping them to corresponding fields. When slots are missing, ambiguous, or have values ​​outside a reasonable range, the system completes or clarifies them based on project metadata and preset rules, and records the source of the completion.

[0087] Step S8.4 Task planning and execution strategy generation; A task plan and execution strategy are generated based on the task type and parameter slots. The task plan is used to organize the order and dependencies of subtasks, and the execution strategy is used to determine parameter calls, execution order, exception handling, and result output methods for subsequent script generation and execution.

[0088] Step S8.5 AI agent script generation and structured output constraints; Step S8.5 Generate a dedicated AI agent script for structural analysis and output structured constraints; In this step, the AI ​​agent is used to automatically generate finite element analysis scripts or instruction sequences based on the task plan generated in step S8.4. To ensure that the output results are verifiable, executable, and easy to check in subsequent safety checks, structured constraints are applied to the output of the AI ​​agent. The structured constraints include limiting the output fields, parameter names, parameter order, value types, and instruction templates, so that the generated results conform to the preset finite element instruction organization format. When the generated results have missing fields, inconsistent formats, or parameters exceeding limits, the abnormal items are marked and the previous steps are returned for completion, correction, or clarification.

[0089] In one embodiment of the present invention, the AI ​​agent adopts the Qwen3-235B-A22B open-source large language model and performs efficient parameter fine-tuning for structural analysis tasks. Specifically, a question-and-answer pair dataset of "natural language analysis requirements - finite element calculation instructions" is manually constructed to organize common task expressions, parameter combinations, and output requirements in structural analysis, forming supervised fine-tuning samples for script generation training; the dataset includes question-and-answer pairs for common finite element analysis tasks such as static analysis, modal analysis, response spectrum analysis, time history analysis, and pushover analysis, totaling approximately 100,000 samples.

[0090] During training, the LoRA fine-tuning method is preferably used to train the Qwen3-235B-A22B model, with a learning rate of 1×10^-5, a batch size of 512, and 3 training epochs. This allows the model to learn the mapping relationship between user natural language requirements, task plans, parameter slots, and finite element analysis instructions. After fine-tuning, the AI ​​agent can output a finite element analysis script or instruction sequence corresponding to the target analysis task based on the task plan generated in step S8.4, combined with the current project context, model state summary, and parameter slot information.

[0091] To ensure the output results are verifiable, executable, and facilitate subsequent security checks, this step applies structured constraints to the AI ​​agent's output. These structured constraints define output fields, parameter names, parameter order, value types, and instruction templates to ensure the generated results conform to a preset finite element instruction organization format. When the generated results contain missing fields, inconsistent formats, or parameters exceeding limits, anomalies are flagged, and the process returns to previous steps for completion, correction, or clarification. Through this method, AI-based finite element analysis script generation and structured output constraints are achieved.

[0092] Step S10 specifically involves receiving the node-level and element-level results output by the finite element analysis engine and standardizing them according to the analysis task type, working condition, time series, and output channel. Then, the semantic object and finite element object mapping relationship established in step S6 is invoked to locate and aggregate the result data. The located results are then written back to the node attributes and component attributes of the structural semantic diagram in a field-based format, along with metadata such as units, statistical methods, analysis task identifiers, and result version numbers. Based on this, a visualization output is generated from the updated structural semantic graph. This output is based on the nodes, components, and backfeed result fields in the structural semantic graph, forming a two-dimensional drawing annotation view and / or a three-dimensional structural display view. Key analysis results are displayed using color mapping, label annotation, or charts. At the same time, the backfeeded structural semantic data file, finite element model file, analysis script file, and optional running log and result index file are exported to support result reproduction, auditing, and subsequent reanalysis.

[0093] Compared with the prior art, the present invention has the following beneficial effects: (1) Significantly improved robustness of recognition: By introducing adaptive tolerance, connected graph and AI-assisted recognition, component grouping and semantic recognition can still be stably completed even under conditions such as loose endpoints, fragmented components and non-standard layer naming.

[0094] (2) Realize the closed loop of diagram-model conversion: Using the structural semantic graph as a unified intermediate layer, it supports the consistent export of structural three-dimensional spatial mechanical model data and structural calculation and analysis scripts, and supports the backfeeding of analysis results to the component semantic layer and visualization display.

[0095] (3) Lowering the analysis threshold: Users describe their analysis needs in natural language, and the AI ​​agent understands and generates finite element calculation scripts, covering a variety of analysis types.

[0096] (4) Trusted execution: The script is checked by a code verifier before execution and is executed in a sandbox environment to reduce errors and security risks.

[0097] (5) Improved interactive experience and engineering efficiency: Real-time communication streams back progress and results, making it easier for engineers to monitor the convergence status and iterate quickly.

[0098] (6) Supports multi-layer 3D modeling and visualization: It uses floor height and Z coordinate rules to complete the 2D to 3D mapping and provides 3D visualization and export capabilities. Attached Figure Description

[0099] Figure 1 This is a flowchart of the overall process of the present invention; Figure 2 This is a schematic diagram illustrating DXF parsing and layer classification in this invention; Figure 3 This is a schematic diagram of the adaptive tolerance and bounding box intersection detection of the present invention; Figure 4 This is a schematic diagram illustrating the entity connectivity graph construction and connectivity component extraction of the present invention; Figure 5 This is a schematic diagram of the structural semantic graph Gs data structure of the present invention; Figure 6 This is a schematic diagram of step S5 of the present invention; Figure 7 This is a schematic diagram illustrating the generation of two-dimensional to three-dimensional Z-coordinates and the generation of nodes / elements in this invention; Figure 8 This is a schematic diagram illustrating the AI ​​agent conversational task orchestration, script generation, verification, and sandbox execution of the present invention. Figure 9 This is a schematic diagram illustrating how the analysis results of this invention are fed back into the structural semantic graph and visualized. Detailed Implementation

[0100] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or modifications made by those skilled in the art to the drawing scale, number of building floors, structural system, finite element type, solving software, and model parameters without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0101] Example 1 This embodiment uses a reinforced concrete frame-shear wall structure office building as an example. The building has 6 floors, with a standard floor height of 3.6m, a first floor height of 4.2m, and a total building height of 22.2m. Input data includes: (1) One copy of each of the structural plane DXF files from layer 1 to layer 6, for a total of 6 copies; (2) Floor number and floor height parameter table; (3) Project-level default material library, section library, unit configuration and coordinate reference configuration; (4) Description of the structure analysis task of subsequent user input.

[0102] In this embodiment, each DXF file contains more than 20 layers, and the number of graphic entities in a single layer is 5,000-30,000. After receiving the input, the system sequentially performs drawing parsing, initial layer semantic judgment, noise cleanup, adaptive tolerance determination, entity connectivity graph construction, component candidate group extraction, AI-enhanced semantic recognition, structural semantic graph generation, graph-to-model mapping, finite element model consistency verification, natural language analysis task orchestration, solution script verification and execution, and result feedback display, ultimately forming a closed-loop processing chain of "drawing—structural semantic graph—finite element model—analysis results—drawing".

[0103] The final result generated in this embodiment includes: (1) One project-level structural semantic diagram file; (2) One copy of the structured data file for the finite element model; (3) One or more copies of the finite element analysis script file; (4) One model consistency verification report; (5) Calculate one execution log; (6) One or more copies of the nodal response, component internal force or stress result files; (7) One copy each of the graphical visualization file and the structured summary file after the results are fed back.

[0104] A method for integrating intelligent finite element modeling and interactive analysis based on structural semantic graphs includes the following steps: (e.g.) Figure 1 ) Step S1: Drawing input and entity analysis; In this embodiment, the computing device first receives six floor DXF files and establishes a project-level metadata record for each file. The metadata record includes basic configuration information related to drawing processing, modeling, and output, including unit system, coordinate reference, file version identifier, number of floors and floor height parameters or default strategy, default material and section parameter library index, and output file naming and version management rules. Subsequently, entity parsing is performed on each DXF file. The parsed objects include five types of graphic entities: LINE (straight line segment), LWPOLYLINE (lightweight polyline), POLYLINE (legacy polyline), CIRCLE (circle), and ARC (arc). For each graphic entity, the system generates a unique entity identifier (eid) and extracts the following fields: (1) Entity type; (2) The name of the layer to which it belongs; (3) Color, line type, and line width; (4) Bounding box coordinates; (5) Geometric parameters.

[0105] Among them, for LINE entities, extract the starting point, ending point, length, and direction angle; for LWPOLYLINE or POLYLINE entities, extract the vertex sequence, number of edge segments, total length, whether it is closed, area, perimeter, and local corner angle sequence; for CIRCLE entities, extract the center coordinates and radius; for ARC entities, extract the center, radius, start and end angles, and arc length.

[0106] Perform preliminary quality control on the set of graphic entities and form a processing record for abnormal entities. In this embodiment, the system uniformly converts the drawing unit to mm, and all coordinates are converted to a unified two-dimensional Cartesian coordinate system with the origin using the project-defined coordinate origin. For files with missing or inconsistent units, the system infers the unit based on the drawing frame scale, statistical value of the axis network spacing, and the size range of typical components; if the inference result is not unique, it is marked as a pending confirmation status and written into the abnormal record table.

[0107] After completion of parsing, the system forms a set of graphic entities for each layer of the drawing and aggregates them to obtain the total set of project-level entities . At the same time, the system outputs: (1) Entity attribute table; (2) Layer attribute table; (3) Entity-layer mapping table; (4) Abnormal parsing log.

[0108] Figure 2 is the schematic diagram of DXF parsing and layer classification of the present invention (keyword dictionary filtering and classification); Step S2: Preliminary judgment of layer semantics and noise cleaning; In this embodiment, the system normalizes the layer names. The normalization rules include: uniformly converting to lowercase, deleting leading and trailing spaces, unifying underscores and hyphens, deleting consecutive duplicate delimiters, and mapping common Chinese and English abbreviations to unified dictionary items.

[0109] After normalization, the system makes a preliminary judgment on layer semantics based on the keyword dictionary, layer attributes, and entity distribution characteristics. Specifically, those hitting beam, bm, girder, beam are given the beam candidate label; those hitting col, column, column are given the column candidate label; those hitting wall, shear, wall are given the wall candidate label; those hitting slab, floor, slab are given the slab candidate label; those hitting axis, grid, axis network are given the axis network candidate label; those hitting dim, text, note, anno, annotation, dimension are given the noise layer label.

[0110] In addition to name determination, the system also calculates the proportion of entities in a layer. A layer is marked as a noise-priority layer if it meets any of the following conditions: (1) The number of text, size, and leader line entities shall account for no less than 0.80%; (2) The proportion of short line segments with a length of less than 50 mm is not less than 0.70, and there are no continuous closed boundaries; (3) The layer name hits the noise keyword and the entity type does not conform to the main structure characteristics.

[0111] After initial layer assessment, the system performs noise cleanup. The cleanup targets include dimension lines, leader lines, text outlines, description boxes, title guide lines, and isolated decorative symbols. For each cleaned entity, the system records the entity identifier, its original layer, the cleanup reason code, the cleanup time, and whether it is recoverable, and writes the results to the noise cleanup log.

[0112] After this step, the system outputs a set of structurally relevant entities and a set of candidate beam entities. Candidate entity set of columns Candidate entity set for the wall Candidate entity set of board Candidate entity set of axis network .

[0113] Step S3: Adaptive tolerance determination (e.g.) Figure 3 ); In this embodiment, the system does not use a fixed tolerance, but rather relies on a set of structure-related entities. Automatically determine tolerance parameters. Specifically, the system starts from... Extract the median length of line-type entities Median of the longer side of the circumscribed rectangle of a closed boundary Median of the spacing between parallel double lines and the median of the smallest effective component size .

[0114] In this embodiment, the endpoint adsorption tolerance Node merging tolerance Nearest neighbor connectivity tolerance and boundary closure tolerance Generate them as follows: If the drawing dimensions are abnormal, causing the above parameters to exceed the preset range, the system will truncate them according to the upper and lower limits. The generated tolerance parameters are uniformly written into the project tolerance configuration table and uniformly called in subsequent processing.

[0115] Step S4: Entity connectivity graph construction and component candidate grouping (e.g.) Figure 4 ); In this embodiment, the system uses a set of structure-related entities. Each entity in the graph is treated as a node, and the geometric relationships between entities are used as edges to construct a connected entity graph. .in, For a set of nodes, Let it be the set of edges.

[0116] The system preferably adopts a "two-stage edge connection" strategy: first, it performs a coarse screening of the bounding rectangle of the entity, and only connects edges in stages if the bounding rectangle is already defined. Only when the expanded circumscribed rectangles overlap does the fine geometry judgment begin; then the endpoint distance, intersection type, overlap length, direction difference, and layer consistency are calculated. If the connection conditions are met, the connection is established.

[0117] In this embodiment, edge weight The normalized weighted form is defined as follows: in, For distance similarity, Scoring is based on the intersection relationship. Scoring is given for consistency of direction. To score layer consistency, Score based on overlap ratio. If Then the entity is considered and entity A valid connection exists.

[0118] After constructing the connected graph, the system extracts each connected component and uses it as the initial unit for the candidate component group. For connected components that are too large, a secondary split is performed; connected components that meet the conditions for distance, direction, and category are merged. After this step, the system generates a structured description record for each candidate group. This record includes a candidate group identifier, a list of entity identifiers, the bounding rectangle and center coordinates, the prior distribution of candidate categories, a local topological summary, a boundary complexity index, and internal edge weight statistics.

[0119] Step S5: AI-enhanced component semantic recognition and structural semantic graph generation; (e.g.) Figure 6 ) Step S5.0 Construction and unified representation of structural semantic graph objects; like Figure 5In this embodiment, the system constructs an initial structural semantic graph based on the output results of steps S2 to S4. .in, For a set of structural nodes, A collection of structural components. For a collection of node attributes, For a set of component attributes, It is a set of mapping relationships.

[0120] Node attributes include node identifier, two-dimensional coordinates, floor index, node type, source marker, and confidence level; component attributes include component identifier, component type candidate set, geometric boundary representation, connection node set, main direction, attribute candidate set, and confidence level.

[0121] In the initial state, component type, connection topology, and attribute fields can be stored in the form of a candidate set plus a probability distribution.

[0122] Step S5.1 Multimodal visual recognition model-assisted interpretation; In this embodiment, multimodal visual recognition is performed only on candidate groups that meet the triggering conditions. The triggering conditions include any one of the following: (1) Component type highest confidence level ; (2) Number of candidate group entities ; (3) The ratio of the area of ​​overlapping entities within a candidate group to the area of ​​its circumscribed rectangle ; (4) The ratio of near-closed gaps at the candidate group boundary ; (5) Topological conflict score of candidate groups .

[0123] The topology conflict score is defined as follows: In the formula, This refers to the proportion of beam ends that are not connected to vertical load-bearing member nodes. The proportion of discontinuity at the wall boundary. This represents the proportion of isolated nodes within the candidate group.

[0124] For candidate groups that meet the triggering conditions, the system crops the corresponding region from the original DXF and generates a global context map (1536×1536), a local candidate map (1024×1024), and a highlight mask map (1024×1024). Then, the layer name summary, the candidate group geometric summary, and the current candidate category list are organized into text prompts and input into the multimodal visual recognition model.

[0125] In this embodiment, the multimodal visual recognition model outputs structured results, including candidate component categories and their confidence levels, coordinates or bounding boxes of key boundary points, anomaly region markers, and model interpretation markers. These results are written into the structural semantic graph as visual evidence.

[0126] Step S5.2 Structured deep learning prediction model (main decision-making process); In this embodiment, component category identification and topological relationship inference are mainly accomplished by a structured deep learning prediction model. The system constructs a corresponding subgraph for each candidate group and extracts node features and edge features.

[0127] Node features include entity type encoding, geometric features, topological features, layer name text embedding, and layer attribute embedding; after concatenation, they are uniformly projected through linear mapping to obtain a 256-dimensional node representation. Edge features include endpoint distance, intersection type, overlap ratio, relative orientation angle, cross-layer label, and tolerance consistency score, which are mapped to obtain a 64-dimensional edge representation.

[0128] In this embodiment, the graph encoder employs a cascaded structure of a 2-layer edge-aware graph neural network and a 4-layer graph Transformer. The first two layers of the graph neural network are used to model local geometric connectivity, with each layer having a hidden dimension of 256; the latter four layers of the graph Transformer are used to model long-range dependencies, with each layer having a hidden dimension of 256, 8 attention heads, and a feedforward layer dimension of 1024.

[0129] The model outputs four types of results: (1) Component category distribution of nodes or candidate groups; (2) Distribution of inter-entity relationship categories; (3) The embedding of entity to component instance; (4) Component end-node association probability matrix.

[0130] Step S5.3: Constraint Reasoning and Conflict Resolution Process; In this embodiment, when the visual recognition result is inconsistent with the structured prediction model result, or when the model result violates engineering constraints, the system calls the constraint reasoning module to make a decision.

[0131] The constraint reasoning module performs the following rule checks: (1) One end of the beam should be connected to a column or wall node, and the proportion of the beam end that is suspended should not exceed 0.20; (2) The length of the continuous gap at the wall boundary shall not exceed ; (3) The maximum unsupported ratio between the slab boundary and the beam or wall boundary shall not exceed 0.30; (4) Column candidates should have closed or nearly closed boundaries and their aspect ratio should not exceed the preset upper limit.

[0132] After completing the adjudication according to the engineering rules, the system writes the final component type, boundary representation, set of connection nodes and attribute candidates back to the structural semantic graph, and records the final result, evidence source, confidence level of each evidence source and adjudication version number.

[0133] Step S5.4 Result of generating the structural semantic graph; After processing in steps S5.1 to S5.3, the system generates a project-level structural semantic graph. In this embodiment, the third layer identification results include 26 column components, 48 ​​beam components, 14 wall components, and 11 slab areas; each component corresponds to a unique component identifier, and the mapping relationship with the original entity identifier list is recorded.

[0134] Step S6: Graph-module mapping and finite element model generation; like Figure 7 In this embodiment, the system maps two-dimensional planar components into three-dimensional finite element model objects based on the structural semantic map and floor height parameters. Let the floor elevation of the k-th floor be... Then the columns and walls are arranged across floors, and the beams are located at... Plane, plate located The plane is offset downwards according to the plate thickness.

[0135] The system first generates finite element nodes: for columns, beams, walls, and slabs, nodes are generated at boundary corners, endpoints, turning points, opening boundary points, and intersections; if the 3D distance between nodes is not greater than the node merging tolerance... Then, the nodes will be merged and a unified node number will be generated.

[0136] Subsequently, the system generates unit objects. In this embodiment, beams and columns use beam-column elements, and walls and slabs use shell elements; each unit includes an element number, connection node number, material identifier, cross-sectional or thickness parameters, and semantic source component identifier.

[0137] For material and section parameters, the values ​​marked on the drawings should be used first; if they do not exist, candidate values ​​should be selected in turn according to the same type of components on the same floor, historical statistics of the same project, and the default parameter library.

[0138] After the mapping is completed, the system generates a finite element model data structure fe_model, which includes a node table, an element table, a material table, a section table, a boundary constraint table, a mass table, and a component-element mapping table semantic_to_fe_map.

[0139] Step S7: Model consistency verification and self-correction; In this embodiment, after the finite element model is generated, the system first performs a consistency check before entering the analysis and solution stage. The check includes the integrity of node, element, material, and section references, the integrity of component connection relationships, the sufficiency of degree-of-freedom constraints, the legality of material and section parameters, and the solvability of the overall model.

[0140] Specifically, the system checks the following metrics: a. The existence rate of nodes referenced by the unit should be 100%; b. The proportion of beam ends connecting to column or wall nodes should not be less than 0.80; c. The continuity ratio of the wall boundary should not be less than 0.90; d. The integrity rate of the edge support should not be less than 0.70; e. The scale of elements without assigned material or section parameters should be 0; f. The number of anomalies in the rigid body degrees of freedom of the model should be 0.

[0141] If any indicator fails to meet the requirements, the system generates a structured error report, fe_check_report, which includes the error type, error location, associated component identifier, associated node or unit identifier, possible cause, and suggested repair operations.

[0142] Step S8: Dialogue-based interactive analysis task orchestration; like Figure 8 In this embodiment, after the finite element model is verified, the user inputs their natural language analysis requirements.

[0143] The system performs structured parsing on the input, extracting task fields, including analysis type, analysis direction, damping ratio, output index, and output level.

[0144] If the user does not provide the necessary parameters, the system will complete them according to the default strategy.

[0145] Subsequently, the system generates an analysis task object (analysis_task) and automatically generates a corresponding finite element analysis script (analysis_script) based on the fe_model. The script includes a model loading section, a mass and boundary section, an analysis control section, an input section, a solution section, and a result recording section.

[0146] Step S9: Code verification and secure execution; In this embodiment, the system performs syntax verification, logic verification, parameter validity verification, and security verification sequentially before executing the analysis script. Syntax verification requires that the script commands are syntactically correct; logic verification requires that the analysis type matches the pre-solution process; parameter validity verification requires that the relevant parameters are within the legal range; and security verification requires that the script does not contain non-whitelisted paths or unauthorized commands.

[0147] If all verifications pass, the system sends the script to an isolated sandbox environment for execution. This sandbox environment restricts file access directories, system call permissions, and external network access permissions. During execution, the system periodically collects the solution progress and sends back the current stage, number of steps completed, solution time, and error alerts to the front end.

[0148] If the execution fails, the system generates an execution error report (exec_error_report) and submits it to the repair module for script correction or model rollback.

[0149] Step S10, Post-processing of results and Model-Graph reflow; (e.g.) Figure 9 ) After the solution is completed, the system reads the solution result file and maps the finite element calculation results back to the node objects and component objects in the structural semantic map according to the semantic_to_fe_map.

[0150] In this embodiment, the following result mapping is performed: (1) Map the maximum displacement of a floor to a floor semantic object; (2) Map the inter-story drift angle to the relationship object between adjacent floors; (3) Map the bottom shear force to the overall structural object; (4) Map the key internal forces or stress results of beam, column and wall units to the corresponding component objects.

[0151] After the results are written back, the system generates a visualization object, `result_overlay`. This object includes component coloring results, node displacement markers, floor statistics labels, and a component result pop-up index. The system supports color mapping based on result value ranges in the original drawing view; when a user clicks on any component, they can view the original element source, semantic attributes, finite element element number, and corresponding analysis results.

[0152] Through the above processing flow, this embodiment achieves a complete closed loop, starting from two-dimensional DXF structural drawings, automatically generating project-level structural semantic diagrams, three-dimensional finite element analysis models, structural analysis scripts, and result feedback into drawing display files. Compared to processing methods that rely solely on fixed rules, this invention improves the stability and traceability of component identification and model generation under complex engineering drawing conditions through adaptive tolerance, entity connectivity graphs, visual-assisted interpretation, structured graph model prediction, constraint reasoning adjudication, and consistency self-correction mechanisms.

[0153] In this embodiment, all types of core objects output by the system have unique identifiers and mapping relationships, including entity identifier eid, candidate group identifier gid, component identifier cid, node identifier nid, and finite element identifier fid, thereby realizing full-link tracking from the original drawing entity to the structural component, then to the finite element and the analysis results.

[0154] Example 2: Detailed Implementation of Adaptive Tolerance Establishment in Step S3 The standard floor plan (DXF) of a six-story reinforced concrete frame-shear wall residential structure is used as input. The unit of the drawing is mm. After parsing the graphic entity set in step S1, the geometric scale information of line segments, polylines, and arc entities related to the main structure is extracted. The geometric scale information includes line segment length, polyline side length, distance between entity endpoints, spacing of local parallel lines, and the size of the circumscribed rectangle of the candidate component group.

[0155] Robust statistical processing was performed on the above-mentioned scale samples, calculating the median, first quartile, third quartile, and interquartile range, and outliers that significantly deviated from the main structural scale were removed. For example, when the length of the boundary lines of the main components in the drawings is mainly distributed in the range of 300mm to 8000mm, and the length of some individual annotation leader lines or fragment short lines is less than 50mm, the scale corresponding to these abnormal short lines is excluded from the representative scale statistics.

[0156] In this embodiment, a representative scale index L is obtained based on the processed scale samples. ref Based on this, an adaptive tolerance parameter set is generated. Preferably, the endpoint adsorption tolerance is 0.02L. ref The node merging tolerance is set to 0.03L. ref The tolerance for determining adjacent relationships is set at 0.04L. ref The tolerance for merging segments within a component group is 0.05L. ref .

[0157] Furthermore, different tolerance sub-parameters are set for different candidate categories: for linear components such as beams and columns, smaller endpoint adsorption tolerance and node merging tolerance are preferred; for planar or boundary-type components such as walls and slabs, larger boundary closure judgment tolerance and segment merging tolerance are preferred.

[0158] A consistency check is then performed. This consistency check includes comparing the current tolerance parameter with the statistical characteristics of typical geometric features of the candidate categories. If, after tolerance processing, a candidate component exhibits an abnormally high merging ratio, numerous broken boundaries that should be connected, or obvious cross-component mis-adsorption, the current tolerance parameter is determined to be mismatched with the representation of that component category. In this case, the corresponding sub-parameter is reverted to a preset default range, or user confirmation is triggered in the interactive interface.

[0159] In this embodiment, the adaptive tolerance parameter set after verification is written into the project-level configuration file and serves as the unified input for entity connection generation in step S4, component boundary correction in step S5, and node merging processing in step S6. Through the above processing, the tolerance parameters can be automatically adjusted according to the drawing scale, reducing the problems of incorrect connections, missing connections, and excessive merging caused by fixed thresholds.

[0160] Key innovations of this invention Building upon the existing "keyword layer classification—connected graph grouping—rule modeling" process, this invention proposes a structural semantic graph (Gs) as a unified intermediate representation for graph-model conversion. Gs uses structural nodes as vertices and components such as beams, columns, walls, and panels as edges. At the semantic layer, it uniformly describes component types, geometric boundaries, connection topology, and engineering attributes, and introduces confidence and candidate set mechanisms to simultaneously carry rule recognition results and AI correction results. Furthermore, it establishes a bidirectional mapping and versioned record between semantic object IDs and finite element node or element IDs, thereby achieving consistent maintenance and traceable updates across the entire graph-model generation and model-graph backfeedback process. This "semantic graph-mapping table-version record" system constitutes the fundamental innovation of this invention for graph-model conversion.

[0161] To address the instability of rule-based methods under complex drawing conditions such as non-standard layer naming, imprecise endpoints, fragmented line segments, and overlapping walls and beams, this invention constructs a structured deep learning model that takes a connected entity graph or an initial semantic graph as input for component semantic recognition and topology correction. This model integrates layer text features, geometric morphology features, and topological neighborhood features, employing multi-task joint learning to output component categories, component instance affiliations, and connection relationship predictions, and outputs confidence scores for subsequent constraint verification and human-computer collaboration. During training, engineering noise enhancements such as breakage, jitter, fragmentation, layer name perturbations, and noise insertion are introduced, and continuous learning is combined with real project correction logs to improve generalization ability across projects and drawing conventions. This "graph structure representation - multi-task prediction - confidence-driven correction" component recognition framework constitutes one of the core innovations of this invention.

[0162] To enable conversational analysis to directly generate executable and verifiable finite element operation sequences, this invention introduces a dedicated large language model capability for finite element software into the constraint reasoning and script generation stages. Based on an open-source large language model, task-oriented fine-tuning is performed to construct an instruction set and syntactic constraints covering modeling, load, analysis control, output recording, and anomaly repair. A benchmark and high-quality fine-tuning data are also established, enabling the model to directly generate finite element modeling and calculation instructions based on the structural semantic graph and task parameters. Simultaneously, a self-correcting loop of "generation-verification-repair" is formed through structured prompts, a code verifier, and sandbox execution feedback, which, in conjunction with model consistency verification, achieves parameter completion, solvability filtering, and semantic rewrite reconstruction. This combination of "instruction set constraints - benchmark-driven fine-tuning - verification feedback self-correction" constitutes the key innovation of this invention in achieving engineering usability and reliability.

[0163] It should be noted that the number of floors, floor height, threshold range, tolerance value method, element type, and analysis task used in the embodiments of the present invention are merely illustrative examples and do not constitute a limitation on the present invention. The present invention is equally applicable to other building types, drawing sizes, analysis task types, and finite element solution platforms.

[0164] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A method integrating intelligent finite element modeling and interactive analysis based on structural semantic graphs, characterized in that, Includes the following steps: Step S1: Drawing input and entity analysis; Receive structural engineering drawing files, parse the structural engineering drawings to obtain a set of graphic entities, and extract the entity type, geometric information and the layer information of each graphic entity; Step S2: Initial semantic assessment and noise removal of layers; Based on the layer name, layer attributes, and preset filtering rules, entities in non-structural component layers are removed, and the remaining layers are classified by component type to obtain candidate entity sets for grid lines, beams, columns, walls, and slabs. Step S3: Determine the adaptive tolerance; The adaptive tolerance parameter is determined based on the scale statistics of the graphic entity set; the adaptive tolerance parameter is used for endpoint snapping, node merging, adjacent / intersecting relationship determination, and component group merging and splitting. "Adaptive" means that the tolerance parameter is automatically generated based on the scale statistics of the current drawing rather than using a fixed threshold. Step S4: Entity connectivity graph construction and component candidate grouping; A connected graph of entities is constructed using graphical entities as nodes and geometric relationships between entities that satisfy the conditions of intersection or adjacency as edges. Candidate groups of components are then extracted based on the connected components. Step S5: AI-enhanced component semantic recognition and structural semantic graph generation; For each candidate group of components, a structural semantic graph is constructed, with structural nodes as vertices and structural components as edges. An AI recognition model is introduced to predict and correct the component type, component boundary, connection topology and component attributes to obtain a structural semantic graph with confidence. The AI ​​recognition model includes: Multimodal visual model: used for visual recognition of images obtained from rendering structural engineering drawings, to assist in the identification of complex or special components; Structured prediction model: Input entity connectivity graph or structural semantic graph, comprehensively utilize layer text features, geometric features and topological features, output component categories and connection relationships; Constraint reasoning module: Based on a large language model, it is used to verify the consistency between the prediction results and the structural engineering constraints, and generate correction strategies when conflicts exist; Step S6: Graph-model mapping and finite element model generation; The structural semantic diagram in step S5 is mapped to the finite element analysis model data structure, generating three-dimensional coordinates, nodes, elements, material and section definitions, boundary conditions, mass and load information; among them, the floor height parameter is introduced into the components of the two-dimensional drawing and the Z coordinate is generated to realize the two-dimensional to three-dimensional mapping. Step S7: Model consistency verification and self-correction closed loop; The stability and solvability of the finite element analysis model generated in step S6 are checked. The check includes connection integrity, rationality of degree of freedom constraints, and legality of material and section parameters. When the check fails, the constraint reasoning module in step S5 is called to generate a repair operation and incrementally update the structural semantic diagram and finite element analysis model until the check passes or the preset iteration limit is reached. Step S8: Dialogue-based interactive analysis task orchestration; The system receives structural analysis requests from users in natural language input, combines the finite element analysis model data structure generated in step S6 with the current project context, parses it into analysis task type and parameter slot set, generates task plan and execution strategy, and the AI ​​agent generates finite element analysis script or instruction sequence corresponding to the analysis task. Step S9: Code verification and secure execution; The generated finite element analysis script undergoes syntax checks, logic checks, parameter validity checks, and security checks. Once the checks pass, the finite element analysis engine is invoked to perform calculations in an isolated sandbox execution environment. Step S10: Result post-processing and model-graph re-feedback; The calculated nodal responses, component internal forces, or stress results are mapped back to the vertex and edge attributes of the structural semantic graph, and the visualization results, structural data files, and model files are output, realizing a closed loop of graph-model conversion.

2. The method according to claim 1, characterized in that, Step S1 adapts to the differences in CAD drawings from different sources and completes the standardized input of drawings and the structured expression of graphic elements; Specifically as follows: Receive structural engineering drawing files and establish a metadata set; the metadata set refers to the basic configuration information related to drawing processing, modeling and output, including unit system, coordinate reference, file version identifier, number of floors and floor height parameters or default strategy, default material and section parameter library index, and output file naming and version management rules; The process involves first normalizing the units and coordinates of the DXF vector file; then parsing the primitive objects in the DXF vector file, constructing a set of graphic entities, and establishing a unique identifier for each entity. For each graphic entity: extract geometric information and generate the corresponding geometric representation; extract the information of its respective layer and establish entity-layer association; Perform preliminary quality control on the graphic entity set and generate processing records for abnormal entities; The final output includes a collection of graphic entities, an entity attribute table, and entity-layer association information.

3. The method according to claim 1, characterized in that, Step S2 is as follows: Step S2.1: Extraction and normalization of semantic features of the layer; The graphic entity set is summarized by layer, the semantic features of the layers are extracted, and the layer names are normalized. The normalization process includes case unification, symbol and space standardization, and synonym mapping. The extracted features are then written into the layer feature table for subsequent steps. Step S2.2 Preliminary classification of layer semantics based on rules and priors; Based on layer name, layer attributes, and preset keyword or pattern library, layers are classified according to component type to obtain grid candidate entity set, beam candidate entity set, column candidate entity set, wall candidate entity set, and slab candidate entity set; Step S2.3 Cleaning up unstructured noise layers and noise entities; Remove non-structural layers or entities and create a record for the removed objects; After completing the above processing, the output includes a set of structure-related entities, a set of candidate entities, layer classification results, and a noise cleanup log. In step S3, the adaptive tolerance establishment method is as follows: Based on the set of graphic entities obtained in step S1, extract the drawing scale statistics, including: median, quantile interval and interquartile range, and obtain representative scale indicators in a robust statistical manner. Generate an adaptive tolerance parameter set based on the representative scale index; Type-specific tolerance sub-parameters are set for different candidate categories, and consistency checks are performed. The consistency check refers to determining whether the tolerance parameters match the current drawing scale and the geometric expression of the component based on the geometric feature statistics of the candidate category. If they do not match, the system will revert to the default range or trigger user confirmation. The resulting tolerance system is provided for subsequent steps; Step S4 is as follows: Step S4.1 Define the nodes, edges, and their attributes in the entity connected graph; Using graphical entities or their geometric primitives as nodes, edges are established based on the geometric relationships between entities to construct a connected entity graph; node attributes include entity type, geometric information, and layer information; edge attributes include relationship type and connection strength. Step S4.2: Edge generation mechanism based on coarse and fine judgment; Edge generation is achieved through a two-stage process of coarse and fine evaluation. In the coarse judgment stage, the bounding boxes of any two entities are expanded according to the adjacent judgment tolerance. If the expanded bounding boxes overlap or the distance is less than the threshold, they are included in the candidate entity pair. In the detailed judgment stage, a fine geometric judgment is performed on the candidate entity pairs. The judgment includes: the endpoint distance is less than the endpoint adsorption tolerance, the line segments or broken lines intersect, the projections overlap, the near collinearity and the spacing is less than the threshold, and the intersection or adjacency judgment of the discrete primitives of the curve. When the precise determination is valid, establish the corresponding connection and record the geometric evidence type; Step S4.3 Differentiation constraints for connections within the same layer and across layers; A differentiated connection strategy is adopted for edges within the same layer and across layers. This means that edges within the same layer and across layers can coexist, but a differentiated strategy is adopted: edges within the same layer are used for merging fragments within a component, and a more lenient threshold or a higher connection weight is adopted; edges across layers are used to express the connection relationship between components, and a more stringent threshold or a lower connection weight is adopted. Step S4.4 Connectivity component extraction and component candidate group formation; Based on the entity connectivity graph, connected components are extracted to form candidate groups of components, which are then filtered and regularized. Step S4.5: Structured output of candidate component groups; A structured description is generated for each candidate group of components and provided to subsequent steps.

4. The method according to claim 1, characterized in that, Step S5 includes: Step S5.0 Construction and unified representation of structural semantic graph objects; Step S5.0.1 Initial structural semantic graph construction and field constraints; The preliminary judgment results of candidate entity categories obtained in step S2, the adaptive tolerance parameter set obtained in step S3, and the entity connectivity graph and component candidate groups obtained in step S4 are uniformly constructed into an initial structural semantic graph: Among them, the vertex set Represents structural nodes; Vertex attribute set The set of attribute fields for vertices includes: a unique node identifier (ID), two-dimensional plane coordinates (x, y), three-dimensional coordinate placeholders or floor indexes and elevation information, node type markers, node source markers, and node confidence scores; the node type markers include endpoints, intersections, connection points, and control points; the node source markers include rule extraction, model inference, and manual correction. edge set Indicates structural components; edge attribute set The set of attribute fields for the edge includes: unique component ID, component type, component geometric boundary representation, set of node IDs connected to the component, component main direction, local coordinate information, component attribute fields, and component confidence level; the component type includes beam, column, wall, floor slab, grid, and others; the component attribute fields include material, section parameters, thickness, connection method, or boundary conditions; Mapping set Used to record the mapping link of "original graphic entity - candidate group - structural component structural node"; Step S5.0.2 Explicitly represent uncertainty; Key fields in the structural semantic graph are explicitly represented using a "candidate set + confidence distribution" approach; the key fields include a candidate set of component types. Connectivity topology candidate set and attribute candidate set For each candidate key field, the candidate value, confidence level, source model, and version number are recorded. Before completing the consistency decision in step S5.3, low-scoring candidates are not directly deleted; their candidate ranking is only adjusted. Step S5.0.3 Multi-source evidence fusion and confidence level merging; Overall confidence level for any field f The confidence level after fusion is calculated as follows: in, This represents the confidence level given by the m-th source of evidence for field f. σ(·) represents the weight of the corresponding evidence source, k is the total number of evidence sources; the evidence sources include five categories: rule evidence, geometric evidence, visual evidence, graphical model evidence, and constrained reasoning evidence; σ(·) is the Sigmoid function; After fusion, the confidence scores for component type, connection topology, and attribute are obtained, and these are written into the structural semantic graph along with the evidence source.

5. The method according to claim 4, characterized in that, Step S5 includes: Step S5.1 Multimodal visual recognition model-assisted interpretation; Step S5.1.1 Triggering conditions and applicable scope of visual recognition; Visual recognition is triggered when a candidate group meets any of the following conditions: (1) The highest posterior confidence of the current component type in the candidate group Less than the first threshold , The value ranges from 0.70 to 0.80; (2) The candidate group satisfies the geometric complexity condition, which includes any one of the following: a. The candidate group contains no fewer than 12 original graphic entities; b. The proportion of overlapping entities within the smallest bounding rectangle of the candidate group Not less than 0.15; c. Near-closed boundary gap With boundary perimeter ratio Not greater than 0.05; d. There shall be no fewer than two sets of parallel double lines or composite lines, and the coefficient of variation of the line spacing shall not exceed 0.20; e. The proportion of the area of ​​the interfering entity to the area of ​​the circumscribed rectangle of the candidate group Not less than 0.20; (3) The candidate group satisfies the topological conflict condition, which is determined by the conflict score. control: in, This indicates the proportion of beam ends that are not connected to vertical load-bearing member nodes. Indicates the proportion of discontinuity at the wall boundary. Indicates the proportion of isolated nodes in the candidate group; when A score ≥ 0.30 indicates a significant topological conflict. For candidate groups whose component type confidence is higher than the second threshold but do not meet the geometric complexity and topological conflict conditions, visual recognition is not triggered to reduce inference overhead. The second threshold is 0.

85. Step S5.1.2 DXF rendering and localized input organization; For the candidate group output in step S4, the DXF local region is rendered as an image, and a global context map, a local cropping map, and a target highlight map are generated; at the same time, text prompts are organized and input into the multimodal visual recognition model along with the above images; the model outputs component category candidates, boundary or key point candidates, and confidence scores in a structured JSON format. The text prompts are organized using a uniform template to reduce the impact of input differences on the recognition results. The prompts include: candidate group identifier, main layer name, number of entities, size of the bounding rectangle, main direction, prior information of candidate categories, and description of the task to be recognized. Step S5.1.3 Output format and fusion strategy; The output of step S5.1 includes component category candidates, boundary or key point candidates, visual confidence level, and visual interpretation markers. If the visual output is consistent with the output of step S5.2, the overall confidence level is increased. If there is a conflict, multiple candidates are retained and submitted to step S5.3 for adjudication. If the confidence level of the visual output is lower than the threshold, it is archived only as supplementary evidence.

6. The method according to claim 5, characterized in that, Step S5 includes: Step S5.2 Structured Deep Learning Prediction Model: Component Semantic Recognition and Topological Relationship Inference; Using entity connectivity graphs as input, the system jointly encodes layer text, geometry, and topology information to predict component categories, instance affiliations, connectivity relationships, and end connection nodes. Step S5.2.1 Input graph structure and feature construction; Step S5.2.1.1 Input graph definition; The entity-connected graph or its corresponding candidate subgraph obtained in step S4 is used as the input graph. Nodes represent graphic entities or geometric primitives, and edges represent the connection relationships between entities, with the connection strength attached. Step S5.2.1.2 Node feature and edge feature encoding; For each node, construct its entity type, geometric features, local topological features, and layer features, and generate an initial representation of the node; For each edge, construct distance, intersection, overlap, angle, cross-layer label, and tolerance matching features, and generate edge features; Step S5.2.2 Label system and training data construction; Step S5.2.2.1 Multi-granularity supervision labeling system; A multi-granularity labeling system is established, aligned one-to-one with the fields of the structural semantic graph. This system is generated based on manually labeled data, rule recognition results, and user correction records, mapping the original graphical entities, inter-entity relationships, component instance affiliation, and topological connection information to corresponding labels, including: Entity-level tags are used to supervise the semantic categories of components; Relationship-level labels are used to monitor the types of relationships or connection probabilities between entities; Instance-level tags are used to monitor the component instances to which an entity belongs; Topology-level tags are used to monitor component end nodes, plate boundary support relationships, and wall boundary continuity relationships; Step S5.2.2.2 Training data sources and weakly supervised sample construction; The training data consists of real engineering drawing annotation data, programmatically synthesized data, weakly supervised data, and user correction logs; wherein, the weakly supervised data is generated based on the rule recognition results of steps S2 to S4: step S2 provides initial entity category candidates, step S3 provides an adaptive tolerance system, and step S4 provides entity connectivity graphs, edge relationships, and candidate group partitioning results; on this basis, initial rule labels are generated using preset engineering rules; The initial rule label is added to the training set as a weak label only if the following conditions are met: a. Rule confidence ,in =0.80; b. The candidate samples meet the local topological consistency requirements; c. Does not violate hard constraint rules, which refer to engineering rule constraints that must be met in component identification and topology determination, including connection rationality constraints, boundary continuity constraints, and component geometric rationality constraints; d. There is no strong conflict with the visual supplementary judgment result or the historical manual correction result in step S5.1; Step S5.2.2.3 Perturbation enhancement; During the training phase, endpoint jittering, slight misalignment, line segment breakage, noise entity injection, layer name abbreviation replacement, symbol interference overlay, and scale perturbation are applied to the input image. Step S5.2.3 Model Architecture and Multi-Task Output; Step S5.2.3.1 Fusion of graph encoder and multimodal features; The structured prediction model employs a coding structure that concatenates an edge-aware graph neural network and a graph Transformer: Edge-aware graph neural network: First, initialize the node representation With edge features Input a two-layer edge-aware graph neural network, where i represents the node index and j represents the index of the node adjacent to node i. The features of the edge between node i and node j are represented; each hidden dimension is 256, and message passing units with residual connections and layer normalization are preferred to model short-range local geometric connectivity. Graph Transformer: Subsequently, the output of the graph neural network is fed into a four-layer graph Transformer, with each layer having a hidden dimension of 256, an attention head of 8, and a feedforward network dimension of 1024. Dropout=0.1 is used to model cross-candidate groups, cross-layers, and long-distance dependencies. If step S5.1 has triggered visual supplementation, then the candidate group-level visual vectors output by the visual model will be... Text summary vectors of candidate groups A gated fusion layer is introduced and represented in the following manner with graph pooling. Fusion: in, This represents the learnable weight matrix used to calculate the gating coefficients. Indicates candidate group The gated fusion coefficient, || denotes vector concatenation. This represents element-wise multiplication. This represents the candidate group representation after fusion; Step S5.2.3.2 Multi-task output header; After the shared graph encoder, set up a multi-task output header for node classification, edge relationship prediction, instance attribution, and topology prediction, respectively. The node classification header outputs the probability distribution of an entity belonging to a beam, column, wall, slab, grid, or noise; the edge relationship header outputs the probability of belonging to the same component, being connected to the component, or being unrelated; the instance attribution header generates instance embedding vectors to group entities of the same component instance; and the topology prediction header outputs the association probability between the component end and the candidate node. Step S5.2.3.3 Confidence estimation and calibration; Temperature scaling or order-preserving calibration is applied to the node classification, edge relationship prediction, and topology prediction outputs to improve the interpretability and comparability of the confidence level, and the calibrated probabilities are used as the basis for the constraint decision in step S5.3 and subsequent manual clarification triggering. Step S5.2.4 Training objectives, training process, and evaluation thresholds; Step S5.2.4.1 Loss function composition and topology consistency regularization; The overall training objective of the model is defined as follows: in, For node classification loss, For the loss of border relations, For instance attribution loss, For topology prediction loss, This is a topology consistency regularization term; to These are the weighting coefficients for each loss term, set according to the performance of the validation set and the convergence status; The node classification loss uses cross-entropy loss: Where N represents the set of nodes, This indicates the total number of node classification categories; This represents the actual label of node i under category c. This represents the probability that node i is predicted to be of category c; Edge relationship loss can be achieved using either cross-entropy loss or focus loss. The instance attribution loss uses contrastive loss: in, Represents a pair of instance entities. This represents a pair of heterogeneous instance entities, where m is the interval parameter. and Represents the instance embedding vectors of entities i and j; Topology prediction loss is used to monitor the connection relationship between the end of a component and its nodes. Where T represents the set of component ends to be predicted, and V represents the set of candidate nodes. For real-world association tags, To predict the probability of association; Topology consistency regularization is used to explicitly introduce structural engineering priors into the training process, and is defined as: in, This is a beam end connection constraint term used to constrain beam ends to preferentially connect to vertical load-bearing member nodes such as columns or walls; This is a wall continuity constraint term used to ensure that the wall boundary remains continuous. This is a floor slab boundary support coordination constraint, used to coordinate the floor slab boundary with the beam and wall support boundaries; , These are the weighting coefficients for the three constraint terms mentioned above, used to balance the contributions of different structural engineering priors to the overall regularization term; the calculation formulas are as follows: Indicates the assembly at the ends of the beam. This represents the set of vertical load-bearing member nodes consisting of columns or walls, and the ends of the constrained beams are preferentially connected to the vertical load-bearing member nodes. This indicates pairs of adjacent boundary segments belonging to the same wall candidate. This indicates the predicted probability that the two belong to the same wall. Indicates the distance of the boundary gap. This indicates a continuity tolerance, meaning this constraint restricts the continuity of the wall boundary. Represents the set of floor slab boundary segments. This represents the set of supporting boundaries formed by beams or walls, which constrains the floor slab boundaries to conform to the beam and wall profiles. Step S5.2.4.2 Phased training and continuous learning; The training process includes three stages: pre-training, fine-tuning, and continuous learning. First, general patterns are learned on procedurally synthesized data and weakly supervised data. Then, fine-tuning is performed on manually labeled and corrected data, and incremental updates are made in conjunction with user correction logs. Step S5.2.4.3 Evaluation indicator system and release threshold; Establish an evaluation index system that includes component categories, edge relationships, instance consistency, and critical connections and boundary quality; release is only permitted when a new model meets the preset threshold and is no lower than the historical stable version; otherwise, it will be automatically rolled back. Step S5.2.5 Output write-back and active learning closed loop; The output of step S5.2 is written back to the structural semantic graph, including candidate component types, instance attribution, candidate connection nodes, candidate boundary representations, and candidate attributes, and the confidence level and source information are recorded. For low-confidence or high-conflict regions, the system initiates interactive clarification. The user corrects the results and feeds them back into the training dataset, forming an active learning loop.

7. The method according to claim 5, characterized in that, Step S5 includes: Step S5.3 Constraint reasoning and finite element instruction generation based on fine-tuning of open-source large language model; This step is used to implement engineering constraint adjudication on the candidate results of steps S5.1 and S5.2, and generate structured correction suggestions or correction instructions for the structural semantic graph and graph-module mapping process when there are conflicts, omissions or inconsistencies. The constraint reasoning module adopts an open-source large language model that has been fine-tuned using structural engineering rules, finite element software instruction sets, and error correction samples; the fine-tuning includes the following key points: 1) Construct instruction-response training samples for structural engineering scenarios. The input side includes structural semantic graph summary, visual supplementation results, graph model prediction results, conflict markers and verification context. The output side includes constraint adjudication results, correction suggestions or finite element instruction sequences. 2) The training samples are organized using a uniform structured template, enabling the model to learn the mapping relationship between fixed input fields and fixed output fields; 3) Introduce component type conflicts, connection relationship conflicts, missing parameters, inconsistent constraints, and script error repair samples into the training set to enhance the model's adaptability to adjudication and repair tasks. 4) Apply instruction set constraints and format constraints to the generated results to limit the output to the predefined set of finite element software commands and structured fields; Specifically, the constraint reasoning module receives the structural semantic graph summary, the visual supplementary judgment result obtained in step 5.1, the graph model prediction result and conflict marker obtained in step 5.2, performs hard constraint verification on component type, connection topology and attribute candidate, and generates adjudication result and correction suggestion in the case of multiple candidate conflicts. When the structural semantic graph meets the minimum consistency requirement, the constraint reasoning module converts it into a sequence of finite element software instructions for subsequent steps to call. When the consistency check fails in subsequent steps, the constraint reasoning module generates a repair operation based on the check report and error context. Through this step, the multi-source outputs of steps S5.1 and S5.2 are uniformly adjudicated and transformed into engineering-executable structural semantic results; Step S5.4: Solidify the structural semantic graph and output the external interface; The final structural semantic graph, candidate set, confidence level, evidence source, and correction record are solidified into structured data objects and assigned a unique version number; The structural semantic graph and its mapping relationships are output to subsequent steps to support graph model mapping, consistency verification, task orchestration and script generation.

8. The method according to claim 1, characterized in that, Step S6 is as follows: In step S6, the structural semantic graph is mapped to a finite element analysis model data structure to generate a finite element analysis model containing information on nodes, elements, materials and sections, boundary conditions, mass, and loads, and a traceable mapping relationship is established; specifically as follows: Step S6.1 Semantic object normalization and mapping benchmark determination; The input for this step is the structural semantic graph output from step S5, as well as project-level meta-information; The structural nodes and structural components in the structural semantic graph are standardized and organized to form a node set, a component set, and an attribute set; the attribute set includes component type, geometric boundary representation, connection topology, and available information such as material, cross section, or thickness; The unit system, coordinate reference, and floor parameters are determined based on the project-level metadata, and these serve as the unified basis for subsequent coordinate mapping, floor expansion, and parameter completion, i.e., mapping reference parameters. The output is a set of normalized semantic objects and mapping baseline parameters; Step S6.2: 2D to 3D mapping and Z-coordinate generation; Based on the standardized semantic object set and floor parameters output in step S6.1, when the input is a two-dimensional drawing, three-dimensional coordinates are generated according to the floor height parameters and floor indexing rules: corresponding elevations are established for each floor, and the two-dimensional node coordinates belonging to the corresponding floor are mapped to three-dimensional node coordinates; an in-layer offset is introduced to represent the elevation offset of beam and slab components; geometric and topological relationships between floors are established for vertical components such as columns and walls, and geometric and topological relationships within the same floor are established for horizontal components such as beams and slabs, while retaining the floor affiliation field; the output is a set of semantic nodes and components with three-dimensional coordinates and floor affiliation information. Step S6.3 Finite element node generation, merging, and numbering; Based on the three-dimensional semantic node set and adaptive tolerance parameters output in step S6.2, finite element nodes are generated according to the semantic node set; for duplicate nodes generated by endpoint adsorption or geometric nearest neighbor, node merging is performed according to the adaptive tolerance parameters to ensure the consistency of the connection topology; Assign a unique node number to each finite element node and output a bidirectional mapping table of "semantic node ID - finite element node number"; Step S6.4: Element type mapping and element topology generation; Based on the component set obtained in step S6.2 and the finite element node set and node mapping table obtained in step S6.3, the structural components are mapped to finite element elements according to the component type and the element topology is generated: beam and column components are mapped to beam and column type elements, and wall and plate components are mapped to shell elements or equivalent elements; when there are missing or broken component boundaries, they are completed or reconstructed based on semantic boundaries, connectivity relationships and tolerance rules; a unique element number is assigned to each element, and a bidirectional mapping table of "semantic component ID - finite element element ID" is established, thereby obtaining the finite element element set, element topology relationship and element mapping table; Step S6.5 Binding of material, section, and component properties; Based on the finite element set obtained in step S6.4, and the parameter information in the drawing information, project element information, default parameter library or AI completion results, material and section or thickness object definitions are generated and bound to the corresponding elements; when material parameters or section parameters are missing, they can be temporarily stored in the form of a candidate parameter set, and filtered and corrected in the subsequent consistency verification and self-correction process; material and section objects are managed with reusable IDs, and their source, value range and version identifier are recorded, thereby obtaining the finite element set, material table and section table with completed attribute binding; Step S6.6 Generation of boundary conditions, mass and load information; Based on the finite element model object obtained in step S6.5, and the constraint and load information in the user input, default strategy, or dialogue parsing results, boundary conditions, mass, and load information are generated in the finite element model: boundary conditions include foundation constraints, floor constraints, or connection constraints; mass is distributed by floor, node, or component; loads include dead load, live load, and the placeholder definition of dynamic action input; and the above information is written into the model data structure in a structured form, thereby obtaining a finite element model data structure containing boundary conditions, mass, and load information. Step S6.7 Output channel planning and result index pre-binding; Based on the finite element model data structure obtained in step S6.6 and the mapping table established in steps S6.3 and S6.4, output channels are pre-planned and result indexes are established to support subsequent result feedback; the node responses and element responses that need to be recorded are defined, and "output channel - semantic object ID - finite element number" is bound and recorded; the bound records and the bidirectional mapping table together constitute the traceable index system of the model, thereby obtaining the output channel configuration and result index binding records; Through the above steps, the data structure mapping and automatic generation from the structural semantic graph to the finite element analysis model are completed; Step S7 is as follows: In step S7, for the finite element analysis model generated in step S6, in order to ensure that the model has engineering usability, numerical solvability and result reliability, a consistency check is performed on the finite element analysis model, and a self-correcting closed-loop update is triggered when the check fails. Specifically, using the finite element analysis model and its mapping relationships as the verification object, a structured verification report is established, recording the problems found in each verification step, the identifiers of related objects, evidence summaries, and remediation suggestions; the consistency verification includes at least the following: Integrity check of node, element, material and section references: The integrity of the reference relationships of nodes, elements, materials and sections in the finite element model is checked to identify missing references, dangling references or unbound objects; Component connection integrity verification: Check the topological connectivity of the model and the connection relationships between beams, columns, walls and slabs to identify problems such as incorrect connections, missing connections, discontinuous boundaries or incomplete supports; Verification of the sufficiency and rationality of degree-of-freedom constraints: Check the boundary conditions and degree-of-freedom constraints to identify missing constraints, constraint conflicts, over-constraints, or abnormal situations that may lead to the motion of the overall rigid body. Material and section parameter validity verification: Perform a validity check on material parameters, section parameters and their binding relationship with the corresponding elements to identify problems such as missing parameters, out-of-bounds parameters or mismatched parameter types; Overall model solvability verification: Perform a solvability check on the entire model, and perform stability checks and / or optional trial calculations as needed to determine whether the model meets the requirements for subsequent finite element solution; When the verification fails, the constraint reasoning module described in step S5.3 is invoked, and the verification report, error context, current structural semantic graph and finite element model summary are taken as input to generate a repair operation and perform local incremental updates on the structural semantic graph or finite element model. Set an iterative control strategy for the verification, repair, and update processes, and record the iteration information; Step S8 is as follows: In step S8, to enable users to submit structural analysis requirements in natural language and automatically convert them into executable finite element modeling and calculation instructions or scripts, the interactive analysis task orchestration is performed according to the following steps: Step S8.1 Dialogue input reception and context construction; Receive structural analysis requirements from the user's natural language input and construct the current project context; Step S8.2 Intent parsing and task type determination; Perform intent parsing on user input, extract key information including analysis type, analysis object, analysis direction, operating conditions and output requirements, and determine the analysis task type; if necessary, break down sub-tasks and establish task dependencies; Step S8.3 Parameter slot extraction and completion; The system extracts a set of parameter slots corresponding to the task type from user input. These parameter slots refer to structured parameter fields in the analysis task, including load or excitation type, analysis direction, step size, damping parameters, convergence control parameters, output indicators, and output object range. Extraction refers to identifying and mapping the parameter expressions in the user input to corresponding fields. When slots are missing, ambiguous, or their values ​​exceed a reasonable range, the system completes or clarifies them based on project metadata and preset rules, and records the source of the completion. Step S8.4 Task planning and execution strategy generation; A task plan and execution strategy are generated based on the task type and parameter slots. The task plan is used to organize the order and dependencies of subtasks, and the execution strategy is used to determine the parameter calls, execution order, exception handling, and result output methods for subsequent script generation and execution. Step S8.5 AI agent script generation and structured output constraints; In this step, the AI ​​agent is used to automatically generate finite element analysis scripts or instruction sequences based on the task plan generated in step S8.

4. To ensure that the output results are verifiable, executable, and easy to check in subsequent safety checks, structured constraints are applied to the output of the AI ​​agent. The structured constraints include limiting the output fields, parameter names, parameter order, value types, and instruction templates, so that the generated results conform to the preset finite element instruction organization format. When the generated results have missing fields, inconsistent formats, or parameters exceeding limits, the abnormal items are marked and the previous steps are returned for completion, correction, or clarification.

9. The method according to claim 8, characterized in that, In step S8, The AI ​​agent uses the Qwen3-235B-A22B open-source large language model and performs parameter fine-tuning for structural analysis tasks. Specifically, a question-and-answer pair dataset of "natural language analysis requirements - finite element calculation instructions" is manually constructed to organize common task expressions, parameter combinations, and output requirements in structural analysis, forming supervised fine-tuning samples for script generation training. The dataset includes question-and-answer pairs for common finite element analysis tasks such as static analysis, modal analysis, response spectrum analysis, time history analysis, and pushover analysis. During training, the Qwen3-235B-A22B model was trained using the LoRA fine-tuning method, with a learning rate of 1×10^-5, a batch size of 512, and 3 training epochs. This enabled the model to learn the corresponding mapping relationship between user natural language requirements, task plans, parameter slots, and finite element analysis instructions. After fine-tuning, the AI ​​agent can output a finite element analysis script or instruction sequence corresponding to the target analysis task based on the task plan generated in step S8.4, combined with the current project context, model state summary and parameter slot information. Apply structured constraints to the output of the AI ​​agent; the structured constraints refer to limiting the output fields, parameter names, parameter order, value types and instruction templates so that the generated results conform to the preset finite element instruction organization format; when the generated results have missing fields, inconsistent formats or parameters exceeding limits, mark the abnormal items and return to the previous steps for completion, correction or clarification.

10. The method according to claim 1, characterized in that, Step S10 specifically involves receiving the node layer results and element layer results output by the finite element analysis engine, and standardizing and organizing them according to the analysis task type, working condition, time series, and output channel. Then, the semantic object and finite element object mapping relationship established in step S6 is invoked to locate and aggregate the result data, and the located results are written back to the node attributes and component attributes of the structural semantic graph in the form of fields, while adding metadata such as units, statistical methods, analysis task identifiers and result version numbers. Based on this, a visualization output is generated from the updated structural semantic graph. This output is based on the nodes, components, and backfeed result fields in the structural semantic graph, forming a two-dimensional drawing annotation view and / or a three-dimensional structural display view. Key analysis results are displayed using color mapping, label annotation, or charts. At the same time, the backfeeded structural semantic data file, finite element model file, analysis script file, and optional running log and result index file are exported to support result reproduction, auditing, and subsequent reanalysis.

Images