Steel bridge automatic finite element modeling method fusing three-dimensional point cloud and intelligent graph recognition

Abstract

The invention provides a steel bridge automatic finite element modeling method fusing three-dimensional point cloud and intelligent graph recognition, and the method comprises a bridge key construction external information automatic extraction method based on a point cloud model, and the method comprises the steps: carrying out the principal component analysis and transformation of a bridge point cloud principal axis; and secondly, carrying out secondary segmentation on the bridge structural member by using an adaptive threshold algorithm and a region growing RANSAC algorithm, and finally, realizing information extraction of a key carrier surface through a minimization loss function and parameterization expression, and accurately extracting external key parameters. On the other hand, a structure internal information extraction method based on deep learning and drawings is provided. According to the method, geometric dimension measurement and automatic finite element modeling can be automatically completed, reliable digital twin model support is provided for bridge safety assessment, and the method has high engineering application value.

Description

Technical Field

[0001] This invention relates to the field of automatic finite element modeling of bridges, and in particular to a method for automatic finite element modeling and performance evaluation of steel bridges that integrates 3D point cloud and intelligent image recognition technology. Background Art

[0002] Bridges, as a crucial component of transportation infrastructure, play a vital role in promoting regional economic development and facilitating social connectivity. However, with increasing service life, the actual stress state of bridge structures is prone to degradation. This degradation not only reduces the structure's load-bearing capacity but may also pose a potential threat to its safety. Finite element models are an important tool for assessing the service status of bridges. However, existing finite element modeling methods suffer from limitations such as low modeling efficiency and susceptibility to human factors. Point cloud technology, as a key technology for bridge digital research, has been hampered by the lack of automated finite element modeling based on point cloud models. Automated finite element modeling using point cloud models still faces numerous challenges in practical applications:

[0003] First, complex bridge structures typically contain multiple structural components with diverse and complex cross-sectional forms. Conventional algorithms using point cloud models struggle to effectively extract the dimensions of complex bridge structures, making in-depth mining and extraction of point cloud model information relatively difficult.

[0004] Secondly, point cloud models cannot directly reflect the internal structural information of bridges, and it is difficult to guarantee the accuracy of bridge structure finite element models directly established from point cloud models.

[0005] In summary, while laser point cloud technology can effectively acquire geometric information of in-service bridge structures, automatically building complex bridge finite element models based on point cloud models has always been a challenge in the industry. The limitation to expanding the application of automated modeling methods to large bridge structures lies in the difficulty of accurately obtaining the internal and external dimensions of the structure. Existing traditional point cloud models struggle to handle bridge structures with irregular boundaries and to reflect hollow cross-sections within the structure, making it difficult to directly and automatically build complex bridge structure finite element models from point cloud models, thus hindering accurate and effective bridge performance evaluation. Summary of the Invention

[0006] This invention proposes an automatic finite element modeling method for steel bridges that integrates 3D point cloud and intelligent drawing recognition. The aim is to extract key information from the bridge's geometric model to support bridge safety assessment and subsequent mechanical analysis. By combining an automatic extraction method for external information of key bridge components from point cloud models with a method for extracting internal structural information based on deep learning and drawings, this method can automatically complete geometric dimension measurement and mechanical model construction, providing reliable data support for bridge safety assessment and mechanical analysis, and possessing high engineering application value.

[0007] To achieve the above-mentioned technical objectives, the present invention employs the following technical means:

[0008] An automated finite element modeling method for steel bridges, integrating 3D point cloud and intelligent map recognition, includes the following steps:

[0009] S1. A principal component analysis-based method for transforming the principal axis of bridge point clouds is used to preprocess point cloud data, which facilitates subsequent extraction of projection density.

[0010] S2. Utilize a physics-based, knowledge-driven fast segmentation algorithm for key bridge components to achieve automatic extraction and information acquisition of key components for the entire bridge.

[0011] S3. The method of parametric expression of cross sections and minimization of loss function is used to automatically fit the key cross sections of the external tower columns, main beams and stay cables of bridge components, so as to accurately extract the key external parameters.

[0012] S4. Based on deep learning and text recognition, automatically extract the drawings of key external sections of bridge components obtained in step S3, and realize the automatic extraction and classification of structural information.

[0013] S5. Based on the identification results of step S4, construct a hierarchical structure tree of key section drawings to realize the automatic classification of key section drawings.

[0014] S6. The bridge internal cross-sectional information is automatically extracted by using a fine-tuned large language model, so as to realize the intelligent extraction and analysis of the internal structural dimensions of the bridge and obtain the internal features of the bridge.

[0015] S7. Based on the above-mentioned key external parameters and internal characteristics of the bridge, automatically establish a finite element model according to the structural characteristics of the bridge.

[0016] Beneficial Effects: Accurate segmentation of key bridge structural components is achieved. Traditional point cloud technology still faces many challenges in practical applications, especially the problem of massive point cloud data and difficulty in accurate segmentation. This invention proposes an automatic extraction method combining principal component analysis, adaptive thresholding algorithm, and region growing RANSAC algorithm, successfully achieving accurate segmentation of key bridge structural components. Based on a segmentation strategy of projection density and adaptive thresholding, this method can accurately extract key components such as bridge towers, cables, and main beams according to the structural characteristics of the bridge, providing a reliable foundation for modeling.

[0017] This invention achieves efficient extraction of internal information of key sections: addressing the challenge of automatically extracting internal parameters of key section drawings, this invention proposes an automatic extraction method for internal information of key sections based on fine-tuning a large language model, aiming to improve the efficiency of information extraction.

[0018] In an optional embodiment, the principal component analysis transformation bridge point cloud principal axis method described in step S1 specifically includes the following sub-steps:

[0019] S101. Principal component analysis (PCA) is used for point cloud principal axis extraction. The principle of this method is to calculate the discreteness of the points and identify the directions with the greatest discreteness as principal axis directions. Specifically, the covariance matrix of the point cloud is calculated and analyzed. For a point cloud P = {p...} i =(x i ,y i ,z i )∈R 3 The covariance matrix of i∈[1,N]} is expressed as:

[0020]

[0021] Where X = {x1, x2, ..., x} N}、Y={y1,y2,…,y N Z = {z1, z2, ..., z} N};

[0022] S102. Calculate the centroid coordinates X of the point cloud along the X-axis. c Then, by expanding forward and backward, the effective interval [X] is determined. c -Δ,X c +Δ], in actual calculations, Δ is set to 2.5 meters, and valid point cloud data within this interval is extracted;

[0023] S103. Cut the point cloud along the X-axis at intervals of Δx and extract the point with the largest absolute y-coordinate value;

[0024] S104. Filter the extracted points and use the principal component analysis algorithm to calculate the three principal axes. Take the point with the smallest X-axis coordinate in the original point cloud as the origin, establish a coordinate system, and map the denoised point cloud into the new coordinate system to complete the coordinate transformation.

[0025] Beneficial effect: The purpose of point cloud coordinate transformation is to make its vertical extension direction parallel to the coordinate axis, which facilitates the subsequent extraction of projection density.

[0026] In an optional embodiment, the physics-based fast segmentation algorithm for key bridge components described in step S2 specifically includes the following sub-steps:

[0027] S201. After obtaining the point cloud in the correct coordinate system, the main beam of the bridge is segmented first through the projection density and adaptive threshold algorithm. The algorithm steps are as follows: take the point cloud in the correct coordinate system as input, slice along the Y-axis with a slice thickness of Δy, calculate the total number of points distributed within each slice, use this number as the projection density of the slice, and smooth it to obtain a one-dimensional continuous array smooth density. Then, use the DBSCAN clustering algorithm to cluster the smooth data. Subsequently, find the minimum value among the maximum values ​​of all categories, and multiply this value by the magnification factor as the threshold.

[0028] S202. Extracting the main beam alignment of a bridge based on the RANSAC algorithm of region growing, the steps are as follows:

[0029] S2021. Extract all local maxima in the smoothed projected density array and filter out the local maxima that are greater than the threshold; calculate the local minima on both sides of the filtered maxima, and set the region enclosed by the x-coordinates corresponding to these two minima as the initial region; merge the adjacent initial regions to form the effective region; obtain the extraction result.

[0030] S2022. Extract the projection density along the Z-axis in each slice interval along the X-axis direction, and extract the local regions corresponding to the two maximum values ​​of the projection density. Cluster the local regions according to their z-coordinates to obtain the upper and lower planes of the main beam.

[0031] S2023. Automatically solve the main beam alignment parameters through algorithms;

[0032] S203. Combine the main beam alignment and region growth algorithm to quickly divide the bridge superstructure, bridge deck system and substructure;

[0033] S204. Further combining the bridge's structural characteristics and projection density information, a secondary segmentation is achieved, dividing the superstructure into bridge towers and cable stays, and the bridge deck system into main beams, crossbeams, and diagonal braces, ultimately accurately segmenting the key components of the entire bridge.

[0034] Beneficial Results: This method demonstrates significant accuracy in detecting various bridge components, surpassing existing advanced segmentation algorithms based on bridge structural characteristics. Furthermore, it exhibits superior segmentation accuracy compared to learning-based advanced algorithms, further demonstrating the effectiveness of the proposed method.

[0035] In an optional embodiment, the automatic fitting step of key sections such as external tower columns, main beams, and stay cables of bridge components described in step S3 specifically includes the following sub-steps:

[0036] S301. The rectangular cross-section parameterized model includes position parameters (X, Y, θ) and shape parameters (W, H). For rectangle fitting, the model is fitted by minimizing the loss function loss_rec, which is the sum of the squares of the distances from each point to its nearest rectangular boundary. Each point is rotated relative to the center of the rectangle, and then the shortest distance to the rotated rectangular boundary is calculated. For any point (x, y, θ), the model is modified accordingly. i ,y i Its coordinates after rotating around the center (X, Y) of the rectangle by an angle θ are (x, y). r ,y r The calculation formula is as follows:

[0037] x r =(x i -X)×cos(θ)-(y i -Y)×sin(θ)+X

[0038] y r =(x i -X)×sin(θ)+(y i -Y)×cos(θ)+Y

[0039] The formula for calculating the loss function loss_rec is as follows:

[0040] loss_rec=min(||x r -X|-w / 2|,||y r -Y|-h / 2|)

[0041] S302. The difference between the circular cross-section parameterization algorithm and the rectangular cross-section algorithm is that when fitting the sample points to a circle, the least squares method is directly used to calculate the parameters. The goal of the least squares method optimization is to minimize the sum of the squares of the distances from all points to the center of the circle.

[0042] S303. For different components, advance along their main axis with a certain slice step size and perform cross-section fitting. The direction of the bridge tower is the Z-axis, and the direction of the cable is the local PCA direction obtained from the clustering.

[0043] S304. The parametric characterization calculation method for cables differs from that for bridge towers in that their forward direction is no longer the Z-axis, but the local PCA direction obtained from clustering. Before calculation, coordinate transformation is performed based on their local PCA direction.

[0044] In an optional embodiment, the automatic extraction of key cross-sectional drawings based on deep learning and character recognition in step S4 specifically includes the following sub-steps:

[0045] S401. The object detection process is as follows: input drawing -- split into multi-page images -- detect drawing -- detect drawing name and section symbols based on drawing. This project selects the YOLOv5 network framework. The program runs based on the open-source YOLOv5 model on GitHub. The deep learning framework is PyTorch.

[0046] S402. Extraction of graphic elements: A total of 20 sets of construction drawings were collected, with 280 valid images used as the dataset for detecting drawings and 840 valid images used as the dataset for detecting drawing names and section symbols. During the dataset creation process, Labelme was used to annotate and generate JSON annotation files, which were then converted into VOC format data for training.

[0047] S403. Use Chinese character OCR recognition to extract and identify relevant text information in key graphics. The text to be identified includes three categories: the text corresponding to the graphic title, the numbers corresponding to the section symbols, and the text contained in the actual graphic. After detecting the text, perform regular expression matching to extract the text that meets the predetermined rules. The text that meets the predetermined rules includes: text composed entirely of Chinese characters, text composed of Chinese characters and English letters, and text composed of Chinese characters, English letters, and numbers.

[0048] Beneficial Effects: This invention improves the automation level of internal information acquisition. Addressing the difficulty of acquiring internal bridge information, this paper proposes an automatic extraction method based on deep learning and drawing information. This method combines OCR technology with the number of drawing layers to automatically classify key cross-section drawings. Then, by fine-tuning a large language model, it achieves intelligent extraction of internal cross-section information, significantly improving the efficiency of internal information acquisition.

[0049] In an optional embodiment, the construction of the hierarchical structure tree of the key section drawings in step S5 specifically includes the following sub-steps:

[0050] S501. The semantic segmentation of point clouds includes categories such as steel beams, tower columns, bridge piers and cables. Drawings with structural categories in their titles are designated as first-level drawings, and the corresponding structural categories are called root nodes.

[0051] S502. After determining the first-level drawings, use them as a basis to find the second-level drawings. The search rule is: if a non-structural class name, such as segment A, appears in the determined first-level drawings, then record it under the corresponding first-level drawings to form a leaf node.

[0052] S503. Classify drawings whose titles contain the text of leaf nodes as second-level drawings and find third-level drawings, and so on, to form a hierarchical structure of drawings.

[0053] Beneficial effects: By constructing a hierarchical structure based on the text in the drawings, the drawings can be automatically categorized. This results in a clearer hierarchical structure and makes it easier to match the segmented point cloud models with the drawings.

[0054] In an optional embodiment, step S6, the automatic extraction of internal cross-sectional information from the fine-tuned large language model, specifically includes the following sub-steps:

[0055] S601. Based on bridge drawing dataset images, high-quality pseudo-labels are automatically generated using Gemini, a closed-source LLM developed by the Google team.

[0056] S602. Use a few-sample prompting strategy to assist LLM in generating more accurate labels by using a small amount of manually annotated drawing information. Evaluate the effectiveness of the pseudo-annotation method by comparing pseudo-annotation data with manually annotated data.

[0057] S603. Design three types of prompts: (1) Create a drawing analysis dataset; (2) Extract geometric and dimensional information; (3) Generate a structured description; The prompt consists of four parts: roles and tasks, thought process, output format, and precautions. It adopts an iterative mechanism that includes human feedback, called the Prompt Development Loop (PrDC). The steps of PrDC are as follows:

[0058] S6031. Assign roles to architects or structural engineers and provide general instructions for identifying component types, shapes, and dimensions;

[0059] S6032. Define specific task instructions to clarify how to identify dimensioning and geometric symbols;

[0060] S6033. Apply the thinking chain method to break down complex parsing tasks into multiple steps, and use contextual information to assist LLM in accurate parsing.

[0061] S6034. Evaluate the accuracy of LLM analysis, correct errors, and fine-tune the model so that it can automatically recognize the geometry, dimensions and material information in the drawings and generate descriptions in structured JSON format or natural language, including component type, geometric features and dimensional details.

[0062] S604. Integrate the text annotations in the drawings into the prompts, and convert the structured JSON format data into a string format as the target output of the model. Subsequently, fine-tuning is performed using quantized low-rank adaptive tuning. The fine-tuning steps are as follows:

[0063] S6041. Use computer vision technology to extract text, annotations and geometric information from drawings and images, and convert them into an input format that the model can process;

[0064] S6042. The tokenizer of the model is responsible for encoding the extracted information into a sequence of integer tokens, while the basic Llama model with a 4-bit weight format is responsible for parsing and generating the structural description.

[0065] S6043. During the fine-tuning process, the LoRA parameter weights are continuously optimized through forward and backward propagation. A quantization-aware training strategy is used to dynamically switch between 32-bit and 4-bit weights to achieve gradient calculation and weight update.

[0066] S6044. The fine-tuning process introduces supervised fine-tuning and reinforcement learning based on human feedback to enhance the model's ability to understand and describe the structural shapes and dimensions in the drawings.

[0067] S6045. The finely tuned model can generate a structured JSON format or natural language description based on the drawing information, including component type, geometric features and dimensional details.

[0068] Beneficial Effects: This invention achieves efficient extraction of internal information from key cross-sections. Addressing the challenge of automatically extracting internal parameters from key cross-section drawings, it proposes an automatic extraction method for internal information from key cross-sections based on fine-tuning a large language model, aiming to improve the efficiency of information extraction. The method is characterized by using the open-source model GLM4 as its core component, offering advantages such as low cost and strong compliance. It achieves efficient extraction of internal information from key cross-sections through fine-tuning the large language model.

[0069] Automatic extraction and classification of drawing information was achieved: A method for automated drawing search using deep learning networks was proposed, enabling rapid location of component drawings requiring measurement. The advantages of this method are that it significantly reduces manual screening workload, quickly obtains internal dimensional information, and achieves automatic drawing classification and retrieval through graphic element extraction, text recognition, and the construction of a hierarchical structure tree.

[0070] In an optional embodiment, the specific steps of automatically establishing the finite element model according to the bridge structural characteristics in step S7 are as follows:

[0071] S701. Establish a finite element model of the main beam. The main beam is discretized using three-dimensional solid elements to accurately describe the spatial structure of the bridge deck and web. Subsequently, establish a bridge tower model on the basis of the main beam. The bridge tower is modeled using solid elements or beam elements to ensure that its load-bearing capacity and stiffness characteristics can be accurately reflected.

[0072] S702. Arrange the stay cables according to the bridge design parameters. The stay cables are simulated using cable elements, and a connection is established between the bridge tower and the main beam to correctly describe the cable force transmission mechanism.

[0073] S703. Establish a pier model. Piers are usually modeled using solid elements or beam elements and connected to the main beam through support elements to accurately simulate the bridge support conditions.

[0074] S704. The main beam and the bridge tower are connected by rigid connection or contact element to simulate their stress relationship. The stay cables are connected to the main beam and the bridge tower through cable element. The pier and the main beam are supported by elastic support element or rigid constraint.

[0075] Beneficial effects: This modeling method realizes automated modeling of complex bridge structures, ensures the refinement and accuracy of the finite element model, and provides a reliable numerical basis for subsequent structural health monitoring and dynamic analysis.

[0076] In summary, this invention advances the application level of bridge digital twin technology: through verification using a real steel pedestrian bridge, the study demonstrates that the proposed method can effectively combine point cloud data, design drawings, and physical knowledge to rapidly construct a bridge finite element model. This provides innovative ideas and solutions for the development of bridge digital twin technology, showcasing its significant value in engineering applications. Attached Figure Description

[0077] Figure 1 This is the basic idea behind the automatic finite element modeling method for steel bridges that integrates 3D point cloud and intelligent map recognition proposed in this invention.

[0078] Figure 2 This invention provides an automatic method for extracting external information of key bridge structures based on point cloud models.

[0079] Figure 3 This invention provides a method for extracting the principal axis of a point cloud.

[0080] Figure 4 A physics-based algorithm for fast segmentation of key bridge components;

[0081] Figure 5 This is the X-axis projected density map;

[0082] Figure 6 The automatic solution process for the proposed main beam parameters;

[0083] Figure 7 The main beam parameterization results are shown in the diagram, including (a) the calculation principle and (b) the main beam parameter results.

[0084] Figure 8 The results show the segmentation of the bridge's superstructure, substructure, and deck system.

[0085] Figure 9 Projected density distribution map of the bridge tower area and cable area;

[0086] Figure 10 A parameterization method for tower columns;

[0087] Figure 11 The parametric fitting results for the cable-stayed bridge are as follows: (a) the initial segment; (b) the main beam segment; and (c) the inclined segment.

[0088] Figure 12 A flowchart of a two-step method for rapidly obtaining the internal dimensions of structures based on deep learning and large language models;

[0089] Figure 13 This is a flowchart of a deep learning-based automatic drawing recognition method.

[0090] Figure 14 A schematic diagram of the target to be identified;

[0091] Figure 15 The results of text recognition on the drawings are as follows: (a) schematic diagram of main beam segment division; (b) text extraction results of main beam segment division; (c) YOLO drawing name and section symbol detection results.

[0092] Figure 16 This is a schematic diagram of the structural hierarchy tree (with the main beam as the root node);

[0093] Figure 17 A flowchart for fine-tuning a large language model;

[0094] Figure 18 Methods for generating training data based on closed-source models: (a) label generation methods; (b) rapid development cycle;

[0095] Figure 19 This document outlines the fine-tuning and evaluation process using an open-source large language model based on a quantized low-rank adaptive method.

[0096] Figure 20 For the pedestrian bridge: (a) Site view; (b) Sensor deployment diagram;

[0097] Figure 21 For key bridge components: (a) component division; (b) main beam stage division;

[0098] Figure 22 Visualization of parameter fitting for key bridge components: (a) main girder; (b) crossbeam bracing; (c) stay cables;

[0099] Figure 23 Search results for the main beam hierarchy tree;

[0100] Figure 24 Measurement results for key parts of the bridge;

[0101] Figure 25The following are the results of internal dimension identification for the large model: (a) Main beam segment A; (b) Main beam segment B; (c) Main beam segment C; (d) Main beam segment D; (e) Main beam segment E; (f) Main beam segment F; (g) Crossbeam; (h) Diagonal brace; (i) Main tower A; (j) Main tower B; (k) Main tower C; (j) Main tower B; (l) Pier;

[0102] Figure 26 The results of automatic finite element modeling;

[0103] Figure 27 The results of static load measurements and finite element analysis of the bridge are as follows: (a) Condition 1; (b) Condition 2.

[0104] Figure 28 For the measured and finite element acceleration of bridge impact load: (a) measuring point 1; (b) measuring point 2; (c) measuring point 3;

[0105] Figure 29 For the measured and finite element acceleration of the bridge under impact load: (a) measuring point 1; (b) measuring point 2; (c) measuring point 3. Detailed Implementation

[0106] To address this, this paper proposes a self-dynamic evaluation method for bridges without design information based on point cloud geometric models. Guided by visual measured information, it integrates 3D point cloud and intelligent map recognition technology to achieve automatic finite element modeling and performance evaluation of steel bridges. This paper addresses the challenges of modeling bridges without design information by proposing two innovative contributions:

[0107] (1) To address the problem that traditional point cloud models cannot handle bridge structures with irregular boundaries, a segmentation strategy based on projection density and adaptive threshold is proposed. Semantic segmentation is achieved based on the characteristics of the bridge structure, and key components such as bridge towers, cables, and main beams are accurately extracted.

[0108] (2) To address the issue that point cloud models cannot reflect the hollow cross-sections inside structures, an automatic drawing search method is proposed. This method uses intelligent technology to assist manual acquisition of component internal dimensions, significantly reducing the workload of manual screening and achieving accurate construction of geometric models. Compared with traditional methods, the technology presented in this paper can automatically complete geometric dimension measurement and mechanical model construction, providing reliable data support for bridge safety assessment and mechanical analysis, and has high engineering application value.

[0109] refer to Figure 1 This method mainly consists of three parts: an automatic extraction method for external information of key bridge components based on point cloud models, an automatic classification method for key cross-section drawings based on deep learning, and an automatic extraction method for internal information of key cross-sections based on fine-tuned large language models. The automatic extraction method for external information of key bridge components based on point cloud models corresponds to steps S1-S3, as follows: Figure 2As shown, steps S4-S5 correspond to the method for automatic classification and text recognition of key cross-section drawings based on deep learning and OCR recognition technology, and step S6 corresponds to the method for automatic extraction of internal information of key cross-sections based on fine-tuning a large language model. The automatic finite element modeling and performance evaluation method for existing damaged hollow slab beam bridges based on visual data includes the following steps:

[0110] Step S1 specifically includes:

[0111] (1) As Figure 3 Principal Component Analysis (PCA) was used to extract the principal axes of the point cloud. This method calculates the dispersion (variance) of the points and identifies the directions with the highest dispersion as the principal axes. Specifically, the covariance matrix of the point cloud is calculated for the point cloud P = {p...} i =(x i ,y i ,z i )∈R 3 For i ∈ [1, N], its covariance matrix can be expressed as:

[0112]

[0113] Where X = {x1, x2, ..., x} N}、Y={y1,y2,…,y N Z = {z1, z2, ..., z} N}

[0114] (2) Calculate the centroid coordinates X of the point cloud along the X-axis. c Then, by expanding forward and backward, the effective interval [X] is determined. c -Δ,X c +Δ], in actual calculations, the length of the Δ segment is set to 2.5 meters. The effective point cloud data P within this segment is then extracted. s ={P si =(x i ,y i ,z i )∈R 3 i = [1, K], x i ∈[x c -Δ,x c +Δ]}.

[0115] (3) Cut the point cloud along the X-axis at intervals of Δx (0.01m in the calculation) and extract the points with the largest absolute y-coordinate values.

[0116] (4) The extracted points are filtered and the principal component analysis (PCA) algorithm is used to calculate the three principal axes. The point with the smallest X-axis coordinate in the original point cloud is used as the origin to establish a coordinate system, and the denoised point cloud is mapped to the new coordinate system to complete the coordinate transformation.

[0117] Step S2 specifically includes:

[0118] (1) As Figure 4 After obtaining the point cloud in the correct coordinate system, the main beam of the bridge is segmented using a projection density and adaptive threshold algorithm. The algorithm steps are as follows: Using the point cloud in the correct coordinate system as input, and with a slice thickness of Δy (as mentioned later in this paper, the slice thickness along the Y-axis is Δy = 0.01m), slice along the Y-axis. Calculate the total number of points distributed within each slice, and use this number as the projection density of that slice. Smooth the slice to obtain a one-dimensional continuous array, smooth density, and then use the DBSCAN clustering algorithm to cluster the smoothed data. Finally, find the minimum value among the maximum values ​​of all categories, and multiply this value by a magnification factor to obtain the threshold.

[0119] (2) The main beam alignment of the bridge is extracted based on the RANSAC algorithm of region growing. The steps are as follows:

[0120] Step 1. As follows Figure 5 Extract all local maxima in the smoothed projected density array and filter out the local maxima that are greater than a threshold; calculate the local minima on both sides of the filtered maxima, and set the region enclosed by the x-coordinates corresponding to these two minima as the initial region; merge the adjacent initial regions to form the effective region; and obtain the extraction result.

[0121] Step 2. Extract the projection density along the Z-axis in each slice interval (the same interval as the slice projection) along the X-axis, and extract the local regions corresponding to the two maximum values ​​of the projection density. Cluster the regions according to their z-coordinates to obtain the upper and lower planes of the main beam.

[0122] Step 3. (as follows) Figure 6 Based on the algorithm proposed in this paper, the automatic solution of the main beam alignment parameters is realized. The specific running logic of the program is as follows:

[0123] ① Project the upper (or lower) plane of the main beam onto the XOZ plane to obtain the point cloud model S={s i =(x i ,z i )∈R 3 ,i=[1,K]};

[0124] ② Select points with smaller X values ​​to form the initial point set S ini ={s j =(xj , z j ), 0 < x j <ini, s j ∈ S} (In the actual calculation process, ini takes 1.5), and the RANSAC algorithm is used to fit a straight line to the initial point to obtain the initial straight line L.

[0125] ③ Advance along the X-axis and continuously add new points S new , and calculate the point S new The distance d between the initial straight line L. If d < thre, then S new is regarded as an inlier and incorporated into the inlier set I; if d > thre, then temporarily mark the point S new as a mutation point. If among the subsequent n (n takes 30 in the actual calculation process) points added, 2 / 3n points have a continuously small increase in distance (appearance of an inclined plane) or are all within the range of [d - thre, d + thre] (appearance of a new platform), then S new is determined as a mutation point, otherwise it is regarded as a noise point;

[0126] ④ Stop advancing along the X-axis after obtaining the mutation point, and refit the straight line L' for the inlier set I at this time to obtain the representation of the main girder before the mutation point;

[0127] ⑤ Let the coordinates of the mutation point be (X c , Z c ). Delete all points in the point cloud model S whose abscissa is less than Xc, and set the X coordinate of the mutation point to 0 to form a new point cloud model S = {s i = (x i - X c , z i ) ∈ R 3 , i = [1, K']};

[0128] ⑥ Repeat steps (2) to (5) until the parametric representation of the main girder is calculated, and this representation can be regarded as the linear shape of the main girder.

[0129] S203, such as Figure 8 , combine the main girder line type and the region growing algorithm to quickly divide the upper structure, bridge deck system, and lower structure of the bridge;

[0130] S204, such as Figure 9 , further combine the bridge structure characteristics and projection density information to achieve secondary segmentation, divide the upper structure into bridge towers and stay cables, divide the bridge deck system into main girders, cross beams, and diagonal braces, and finally accurately segment the key components of the entire bridge.

[0131] Step S3 specifically includes:

[0132] (1) The parametric model of the rectangular cross-section includes position parameters (X, Y, θ) and shape parameters (W, H). For rectangle fitting, the model is fitted by minimizing the loss function `loss_rec`, which is chosen as the sum of squared distances from each point to its nearest rectangular boundary. To handle rectangle rotation, each point is rotated relative to the rectangle's center, and then its shortest distance to the rotated rectangle's boundary is calculated. For any point (x, y, θ)... i ,y i Its coordinates after rotating around the center (X, Y) of the rectangle by an angle θ are (x, y). r ,y r The calculation formula is as follows:

[0133] x r =(x i -X)×cos(θ)-(y i -Y)×sin(θ)+x

[0134] y r =(x i -X)×sin(θ)+(y i -Y)×cos(θ)+Y

[0135] The formula for calculating the loss function loss_rec is:

[0136] loss_rec=min(||x r -X|-w / 2|,||y r -Y|-h / 2|)

[0137] (2) The circular cross-section parameterization algorithm is similar to that of the rectangular cross-section. The difference is that when fitting the sample points to a circle, the least squares method is used directly to calculate the parameters. The goal of the least squares method is to minimize the sum of the squares of the distances from all points to the center of the circle.

[0138] (3) Figure 10 For different components, the cross-section is fitted along the main axis with a certain slice step size. The direction of the bridge tower is the Z-axis, and the direction of the cable is the local PCA direction of the cluster.

[0139] (4) Figure 11 The parameterized representation calculation method of the cable is similar to that of the bridge tower. The difference is that its forward direction is no longer the Z-axis, but the local PCA direction obtained by clustering. Before the calculation, the coordinate transformation is performed according to its local PCA direction.

[0140] Internal information extraction steps are as follows Figure 12 As shown, step S4 specifically includes:

[0141] (1) As Figure 13The object detection process is as follows: input drawing -- split into multi-page images -- detect drawing -- detect drawing name and section symbols based on the drawing. This project uses the YOLOv5 network framework, the program runs on the open-source YOLOv5 model from GitHub, and the deep learning framework is PyTorch.

[0142] (2) Figure 14 For the extraction of graphic elements, a total of 20 sets of construction drawings were collected, with 280 valid images used as the dataset for detecting drawings and 840 valid images used as the dataset for detecting drawing names and section symbols. During the creation of the dataset, Labelme was used to annotate and generate JSON annotation files, which were then converted into VOC format data for training.

[0143] (3) Figure 15 Significant progress has been made in OCR recognition technology for Chinese characters, which can be used to extract and recognize relevant text information in key graphics. The text to be recognized includes three categories: text corresponding to the drawing title, numbers corresponding to section symbols, and text contained in the actual drawing (excluding the drawing title). After text detection, regular expression matching is performed to extract text that meets predetermined rules. Text that meets the predetermined rules includes: text composed entirely of Chinese characters, text composed of Chinese characters and English letters, and text composed of Chinese characters, English letters, and numbers.

[0144] Step S5 specifically includes:

[0145] (1) As Figure 16 The semantic segmentation of point clouds mainly includes steel beams, tower columns, bridge piers and cables (hereinafter collectively referred to as structural classes). Drawings with structural classes in their titles are taken as first-level drawings, and the corresponding structural classes are called root nodes. This makes the hierarchical structure of the drawings clearer and the segmented point cloud models easier to match with the drawings.

[0146] (2) After determining the first-level drawings, the second-level drawings are searched based on them. The search rule is: if a non-structural class name, such as segment A, appears in the first-level drawings, it is recorded under the corresponding first-level drawings to form a leaf node.

[0147] (3) Classify drawings whose titles contain the text of leaf nodes as second-level drawings and find third-level drawings, and so on, to form a hierarchical structure of drawings.

[0148] Step S6 specifically includes:

[0149] (1) As Figure 17This paper proposes a pseudo-labeling method based on LLM (Liquid Least Merger), which reduces the burden of manual labeling by automatically generating high-quality labels. Based on bridge drawing dataset images, high-quality pseudo-labels are automatically generated using Gemini (Gemini Team Anil et al., 2024), a closed-source LLM developed by the Google team.

[0150] (2) To further verify the quality of the pseudo-labels, we used a few-sample hint strategy, using a small amount of manually annotated drawing information to assist LLM in generating more accurate labels. The effectiveness of the pseudo-labeling method was evaluated by comparing the pseudo-labeled data with manually annotated data.

[0151] (3) Figure 18 To enable closed-source LLM to parse drawings and generate structural shape and dimension descriptions, this paper designs three types of prompts: (1) creating a drawing parsing dataset; (2) extracting geometric and dimensional information; and (3) generating a structured description. The prompts consist of four parts: roles and tasks, thought process, output format, and precautions. An iterative mechanism incorporating human feedback is employed, called the Prompt Development Cycle (PrDC). The steps of PrDC are as follows:

[0152] Step 1. Assign roles of architect or structural engineer, providing general instructions for identifying component types, shapes, and dimensions;

[0153] Step 2. Define specific task instructions, clarifying how to identify dimensioning and geometric symbols;

[0154] Step 3. Apply the Chain-of-Thought method to break down the complex parsing task into multiple steps, and use contextual information (such as paper type, annotation standards, etc.) to assist LLM in accurate parsing;

[0155] Step 4. Evaluate the accuracy of the LLM's analysis and correct errors. The fine-tuned model can automatically recognize the geometry, dimensions, and material information in the drawings, and generate descriptions in structured JSON format or natural language, including component types, geometric features, and dimensional details.

[0156] (4) Figure 19 The text annotations from the drawings are integrated into the prompts, and the structured JSON data is converted into a string format as the model's target output. Subsequently, Quantized Low-Rank Adaptation (QLoRA) is used for fine-tuning. The fine-tuning steps are as follows:

[0157] Step 1. Use computer vision technology to extract text, annotations, and geometric information from the drawings and images, and then convert them...

[0158] Transform it into an input format that the model can process;

[0159] Step 2. The model's word segmenter is responsible for encoding the extracted information into a sequence of integer tokens, using a 4-bit weight.

[0160] The underlying Llama model of the reformatted structure is responsible for parsing and generating the structure description;

[0161] Step 3. During fine-tuning, the LoRA parameter weights are continuously optimized through forward and backward propagation. A quantization-aware training strategy is used to dynamically switch between 32-bit and 4-bit weights to achieve gradient...

[0162] Calculation and weight update;

[0163] Step 4. The fine-tuning process introduces Supervised Fine-Tuning (SFT) and Reinforcement Learning with Human Feedback (RLHF) to...

[0164] Enhance the model's ability to understand and describe the structural shapes and dimensions in the drawings;

[0165] Step 5. After fine-tuning, the model can generate a structured JSON format or natural language description based on the drawing information, including component type, geometric features and dimensional details.

[0166] The specific steps of S7 are as follows.

[0167] (1) Establish a finite element model of the main beam. The main beam is discretized using three-dimensional solid elements (such as SOLID elements) to accurately describe the spatial structure of the bridge deck and web. Subsequently, a bridge tower model is established on the basis of the main beam. The bridge tower is modeled using solid elements or beam elements to ensure that its load-bearing capacity and stiffness characteristics can be accurately reflected.

[0168] (2) Arrange the stay cables according to the bridge design parameters. The stay cables are simulated using cable elements (such as LINK or CABLE elements), and a connection is established between the bridge tower and the main beam to correctly describe the cable force transmission mechanism.

[0169] (3) Establish a pier model. Piers are usually modeled using solid elements or beam elements and connected to the main beam through support elements to accurately simulate the bridge support conditions.

[0170] (4) The stress relationship between the main girder and the bridge tower is simulated using rigid connections or contact elements. The stay cables are connected to the main girder and the bridge tower through cable elements, while the support relationship between the piers and the main girder is achieved using elastic support elements or rigid constraints. This modeling method realizes automated modeling of complex bridge structures, improves the efficiency and accuracy of finite element models, and provides a reliable numerical basis for structural performance evaluation.

[0171] The following uses a real bridge scenario to illustrate the solution of this embodiment.

[0172] like Figure 20 As shown, the pedestrian bridge studied in this study is a cable-stayed bridge. The main bridge is a mast-type cable-stayed bridge with a main span of 37.6m. The bridge deck uses orthotropic plates, with a total width of 5m. The main pier towers use circular steel tube sections with a diameter of 0.48m and a height of 10m above the bridge deck. The intermediate towers also use circular steel tube sections, with the height and diameter decreasing sequentially. The main girder of the cable-stayed bridge is a double-sided box girder, with the side box sections having a height and width of 0.6m. Several load conditions were tested on the bridge, including static load and impact vibration tests under pedestrian jumping conditions. The static load test included two conditions: Condition 1 applied a static load of approximately 700N at each measuring point, and Condition 2 applied a load of approximately 1200N at measuring points 3, 4, and 5. Seven PCB piezoelectric accelerometers and industrial cameras were evenly spaced on both the east and west sides of the pedestrian bridge to collect the dynamic response of the structure under the three test conditions. The spacing between the sensors was 3m.

[0173] Figure 21 (a) The non-gray components are those in the final parametric geometric representation. Due to the bridge's symmetry, only the data for the right side of the bridge is shown. The red components represent the main beams, labeled G, and the segment division of the main beams is as follows: Figure 21 As shown in (b), from left to right are G1, G2, ..., G... 11 The pink area represents the bridge towers; the towers in the upper part of the diagram are T1, T2, ..., T6 from left to right. The green area represents the cables; the cables are C1, C2, ..., C6 from left to right. 10 The blue areas represent the crossbeams, numbered B1, B2, ..., V from left to right. 13 The cyan lines represent diagonal braces, while the yellow diagonal braces, from left to right, are L1, L2, ..., L. 12 .

[0174] Figure 22 The results show the fitting of key bridge sections extracted from point cloud data, including the geometric features of the main beam, crossbeam bracing, and stay cables. Figure 22 (a) shows the cross-sectional fitting results of the main beam. Through the analysis of high-precision point cloud data, the linear characteristics of the main beam were accurately identified. Figure 22(b) focuses on the crossbeam bracing area and effectively restores its spatial layout and connection relationship by using the fitted cross-sectional profile; Figure 22 (c) The key section fitting results of the stay cables are presented. The connection positions and intersection relationships of the stay cables are identified using circle center fitting and line fitting methods, and the intersection points of the fitting are marked with "Fitted intersection". The relevant variables and values ​​of the parameter fitting process are shown in Table 3.

[0175] Figure 23 This paper presents the calculated, measured, and error analysis results of the structural dimensions of key bridge components extracted from point cloud data, covering the dimensional features of main beams, towers, stay cables, crossbeams, and diagonal braces. Overall, the error between the calculated and measured values ​​for the main beam is generally controlled within ±20 mm, demonstrating the high accuracy of the point cloud extraction method in dimensional measurement. The errors in the radius and height of the tower are also small, especially in the height direction, where the accuracy is controlled within ±7 mm. The radius error of the stay cables does not exceed ±2 mm. The extraction results for crossbeams and diagonal braces also show high accuracy, with the errors in the height and width of the crossbeam section both within ±1.1 mm, and the radius error of the diagonal braces being 0.5 mm. The overall results indicate that the proposed method can effectively extract the dimensional features of bridge main beams, towers, stay cables, crossbeams, and diagonal braces, with small errors compared to manual measurements. This significantly improves the efficiency of external structural dimension measurement and provides data support for subsequent automated finite element modeling.

[0176] After constructing the hierarchical structure tree, relevant files can be generated. Finally, by searching for the corresponding hierarchical files based on the name of each structure, drawings can be obtained layer by layer, thus completing the automatic search for drawings. Figure 24 The search results for the keyword "main beam" demonstrate the automatic classification and key section extraction process of bridge drawing information based on a hierarchical structure. First, the main beam, as the top-level structural object, includes an overall structural schematic and dimensional information, and its segmentation information is further refined through "finite segment division of the main beam." This hierarchical division allows different structural regions of the main beam to be organized independently, providing a foundation for subsequent modeling and analysis. Based on the main beam segmentation, it is further divided into various key nodes (nodes AF), each containing corresponding cross-sectional dimensions, internal structural information, and sectional views. The bottom-level structure stores the key section information of each node and organizes it in the form of independent image files, achieving automatic classification and rapid indexing of structural cross-sectional data. This hierarchical tree search method can effectively parse bridge structural drawing data, providing an efficient information management solution for the geometric modeling and finite element analysis of the main beam and other structural components.

[0177] Figure 25This paper presents the results of internal dimension identification for large structural components based on Large Language Modeling (LLM). The figures cover different structural sections, including main beams (a–f), crossbeams (g), diagonal braces (h), and various tower column sections (i–l). Each sub-figure compares manually described internal structural information with the output generated by LLM. LLM accurately identifies key geometric parameters such as shape, internal width, internal height, outer diameter, and edge thickness, covering various cross-sectional forms including rectangular, box-shaped, and circular. The results demonstrate that this method has high applicability in the extraction and interpretation of internal dimensions from engineering drawings, providing effective support for automated structural assessment and digital twin modeling.

[0178] Figure 26 This paper showcases a finite element model automatically generated based on point cloud data and bridge drawing information. The specific modeling process is as follows: First, a finite element model of the main girder is established. The main girder is discretized using three-dimensional solid elements (such as SOLID elements) to accurately describe the spatial structure of the bridge deck and web. Next, a bridge tower model is built on the foundation of the main girder. The bridge tower is modeled using solid elements or beam elements to ensure its load-bearing capacity and stiffness characteristics are accurately reflected. Then, the stay cables are arranged according to the bridge design parameters. The stay cables are simulated using cable elements (such as LINK or CABLE elements), and a connection is established between the bridge tower and the main girder to correctly describe the cable force transmission mechanism. After completing the superstructure modeling, a pier model is further established. The piers are typically modeled using solid elements or beam elements and connected to the main girder through support elements to accurately simulate the bridge's support conditions. The connection relationships of each part of the finite element model are as follows: the main girder and bridge tower use rigid connections or contact elements to simulate their force relationship; the stay cables are connected to the main girder and bridge tower through cable elements; and the piers and main girder use elastic support elements or rigid constraints to achieve the support relationship. This modeling method enables automated modeling of complex bridge structures, ensuring the refinement and accuracy of the finite element model and providing a reliable numerical basis for subsequent structural health monitoring and dynamic analysis.

[0179] Figure 27 (a) Shows the measured data and finite element simulation results for working condition one; Figure 27 (b) then displays the measured data and finite element simulation results for working condition two. From Figure 27 The comparison results show that the displacement variation trends of the finite element simulation results and the measured data under different loading conditions are basically consistent, and the error is controlled within a reasonable range. Compared with the measured data, the errors of the method in this paper and the measured data are X% and X%, respectively, verifying the accuracy of the established finite element model in static analysis.

[0180] This section presents the results of the structural dynamic analysis, primarily the structural acceleration response under hammer impact. Considering the structural symmetry, test points 2, 3, and 4 were selected as the hammer impact vibration test points. The acceleration at each test point was compared to the measured acceleration when the hammer struck it. Figure 28 (b), (c), and (d) respectively show the acceleration signals collected at the corresponding measuring points at positions 2, 3, and 4 under the action of a human-induced impact load. Figure 28 As shown, the structural response calculated by the method proposed in this paper is closer to the measured response and can more accurately reflect the dynamic response of the structure, further verifying the effectiveness of the model established in this paper in calculating the dynamic response.

[0181] Through the detailed experimental design and result analysis described above, the effectiveness of the proposed automatic finite element modeling method for steel bridges, which integrates 3D point cloud and intelligent image recognition, has been verified. This method not only solves the problem of effectively extracting external structural information based on point cloud models, but also significantly improves the accuracy of internal structural dimension extraction by combining intelligent image recognition technology, thus facilitating automated bridge modeling. Furthermore, it demonstrates broad application potential in the field of safety assessment for large and complex bridge structures.

[0182] In summary, the successful demonstration of the specific embodiments demonstrates that the proposed solution is feasible and effective. Experiments have shown that combining point cloud models and intelligent image recognition technology can significantly improve the efficiency of automated modeling of complex bridge structures, which is of great significance for simulating the stress conditions of existing structures.

[0183] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An automatic finite element modeling method for steel bridges integrating 3D point cloud and intelligent map recognition, characterized in that, Includes the following steps: S1. A principal component analysis-based method for transforming the principal axis of bridge point clouds is used to preprocess point cloud data, which facilitates subsequent extraction of projection density. S2. Utilize a physics-based, knowledge-driven fast segmentation algorithm for key bridge components to achieve automatic extraction and information acquisition of key components for the entire bridge. S3. The method of parametric expression of cross sections and minimization of loss function is used to automatically fit the key cross sections of the external tower columns, main beams and stay cables of bridge components, so as to accurately extract the key external parameters of the bridge. S4. Based on deep learning and text recognition, the drawings of the external tower columns, main beams and key sections of the stay cables of the bridge components obtained in step S3 are automatically extracted to realize the automatic extraction and classification of structural information. S5. Based on the identification results of step S4, construct a hierarchical structure tree of key section drawings to realize the automatic classification of key section drawings. S6. The bridge internal cross-sectional information is automatically extracted by using a fine-tuned large language model, so as to realize the intelligent extraction and analysis of the internal structural dimensions of the bridge and obtain the internal features of the bridge. S7. Combining the external key parameters of the bridge described in step S3 and the internal features of the bridge obtained in step S6, an automatic finite element model is established according to the structural characteristics of the bridge.

2. The automatic finite element modeling method for steel bridges integrating 3D point cloud and intelligent map recognition as described in claim 1, characterized in that, The principal component analysis transformation bridge point cloud principal axis method described in step S1 specifically includes the following sub-steps: S101. Principal component analysis (PCA) is used for point cloud principal axis extraction. The principle of this method is to calculate the discreteness of the points and identify the directions with the greatest discreteness as principal axis directions. Specifically, the covariance matrix of the point cloud is calculated and analyzed. For a point cloud P = {p...} i =(x i ,u i ,z i )∈R 3 The covariance matrix of i∈[1,N]} is expressed as: where X={x1,x2,...,x N }、Y={y1,y2,…,y N }、Z={z1,z2,...,z N }; S102. Calculate the centroid coordinates X of the point cloud along the X-axis. c Then, by expanding forward and backward, the effective interval [X] is determined. c -Δ,X c +Δ], in actual calculations, Δ is set to 2.5 meters, and valid point cloud data within this interval is extracted; S103. Cut the point cloud along the X-axis at intervals of Δx and extract the point with the largest absolute y-coordinate value; S104. Filter the extracted points and use the principal component analysis algorithm to calculate the three principal axes. Take the point with the smallest X-axis coordinate in the original point cloud as the origin, establish a coordinate system, and map the denoised point cloud into the new coordinate system to complete the coordinate transformation.

3. The automatic finite element modeling method for steel bridges integrating 3D point cloud and intelligent map recognition as described in claim 1, characterized in that, The physics-based fast segmentation algorithm for key bridge components described in step S2 specifically includes the following sub-steps: S201. After obtaining the point cloud in the correct coordinate system, the main beam of the bridge is segmented first through the projection density and adaptive threshold algorithm. The algorithm steps are as follows: take the point cloud in the correct coordinate system as input, slice along the Y-axis with a slice thickness of Δy, calculate the total number of points distributed within each slice, use this number as the projection density of the slice, and smooth it to obtain a one-dimensional continuous array smooth density. Then, use the DBSCAN clustering algorithm to cluster the smooth data. Subsequently, find the minimum value among the maximum values ​​of all categories, and multiply this value by the magnification factor as the threshold. S202. Extracting the main beam alignment of a bridge based on the RANSAC algorithm of region growing, the steps are as follows: S2021. Extract all local maxima in the smoothed projected density array and filter out the local maxima that are greater than the threshold; calculate the local minima on both sides of the filtered maxima, and set the region enclosed by the x-coordinates corresponding to these two minima as the initial region; merge the adjacent initial regions to form the effective region and obtain the extraction result. S2022. Extract the projection density along the Z-axis in each slice interval along the X-axis direction, and extract the local regions corresponding to the two maximum values ​​of the projection density. Cluster the local regions according to their z-coordinates to obtain the upper and lower planes of the main beam. S2023. Automatically solve the main beam alignment parameters through algorithms; S203. Combine the main beam alignment and region growth algorithm to quickly divide the bridge superstructure, bridge deck system and substructure; S204. Further combining the bridge's structural characteristics and projection density information, a secondary segmentation is achieved, dividing the superstructure into bridge towers and cable stays, and the bridge deck system into main beams, crossbeams, and diagonal braces, ultimately accurately segmenting the key components of the entire bridge.

4. The automatic finite element modeling method for steel bridges integrating 3D point cloud and intelligent map recognition as described in claim 1, characterized in that, Step S3, which involves automatically fitting the key sections of the bridge components—external towers, main beams, and stay cables—specifically includes the following sub-steps: S301. The rectangular cross-section parameterized model includes position parameters (X, Y, θ) and shape parameters (W, H). For rectangle fitting, the model is fitted by minimizing the loss function loss_rec, which is the sum of the squares of the distances from each point to its nearest rectangular boundary. Each point is rotated relative to the center of the rectangle, and then the shortest distance to the rotated rectangular boundary is calculated. For any point (x, y, θ), the model is modified accordingly. i ,y i Its coordinates after rotating around the center (X, Y) of the rectangle by an angle θ are (x, y). r ,y r The calculation formula is as follows: x r =(x i -X)×cos(θ)-(y i -Y)×sin(θ)+X and r =(x i -X)×sin(θ)+(y i -Y)×cos(θ)+Y The formula for calculating the loss function loss_rec is: loss_rec=min(||x r -X|-w / 2|,||y r -Y|-h / 2|) S302. The difference between the circular cross-section parameterization algorithm and the rectangular cross-section algorithm is that when fitting the sample points to a circle, the least squares method is directly used to calculate the parameters. The goal of the least squares method optimization is to minimize the sum of the squares of the distances from all points to the center of the circle. S303. For different components, advance along their main axis with a certain slice step size and perform cross-section fitting. The direction of the bridge tower is the Z-axis, and the direction of the cable is the local PCA direction obtained from the clustering. S304. The parametric characterization calculation method for cables differs from that for bridge towers in that their forward direction is no longer the Z-axis, but the local PCA direction obtained from clustering. Before calculation, coordinate transformation is performed based on their local PCA direction.

5. The automatic finite element modeling method for steel bridges integrating 3D point cloud and intelligent map recognition as described in claim 1, characterized in that, Step S4, which involves automatically extracting key cross-sectional drawings based on deep learning and character recognition, specifically includes the following sub-steps: S401. The object detection process is as follows: Input drawing -- split into multi-page images -- detect drawing -- detect drawing name and section symbols based on the drawing. This project selects the YOLOv5 network framework, and the program runs based on the open-source YOLOv5 model on GitHub. The deep learning framework is PyTorch. S402. Extraction of graphic elements: A total of 20 sets of construction drawings were collected, with 280 valid images used as the dataset for detecting drawings and 840 valid images used as the dataset for detecting drawing names and section symbols. During the dataset creation process, Labelme was used to annotate and generate JSON annotation files, which were then converted into VOC format data for training. S403. Use Chinese character OCR recognition to extract and recognize relevant text information in key graphics. The text to be recognized includes three categories: the text corresponding to the graphic title, the numbers corresponding to the sectioning symbols, and the text contained in the actual graphic. After detecting the text, regular expression matching is performed to extract the text that meets the predetermined rules. The text that meets the predetermined rules includes: text composed entirely of Chinese characters, text composed of Chinese characters and English letters, and text composed of Chinese characters, English letters and numbers.

6. The automatic finite element modeling method for steel bridges integrating 3D point cloud and intelligent map recognition as described in claim 1, characterized in that, Step S5, which involves constructing the hierarchical structure tree of the key section drawings, specifically includes the following sub-steps: S501. The semantic segmentation of point clouds includes categories such as steel beams, tower columns, bridge piers and cables. Drawings with structural categories in their titles are designated as first-level drawings, and the corresponding structural categories are called root nodes. S502. After determining the first-level drawings, use them as a basis to find the second-level drawings. The search rule is: if a non-structural class name, such as segment A, appears in the determined first-level drawings, then record it under the corresponding first-level drawings to form a leaf node. S503. Classify drawings whose titles contain the text of leaf nodes as second-level drawings and find third-level drawings, and so on, to form a hierarchical structure of drawings.

7. The automatic finite element modeling method for steel bridges integrating 3D point cloud and intelligent map recognition as described in claim 1, characterized in that, In step S6, the automatic extraction of internal cross-sectional information from the fine-tuned large language model specifically includes the following sub-steps: S601. Based on bridge drawing dataset images, high-quality pseudo-labels are automatically generated using Gemini, a closed-source LLM developed by the Google team. S602. Use a few-sample prompting strategy to assist LLM in generating more accurate labels by using a small amount of manually annotated drawing information. Evaluate the effectiveness of the pseudo-annotation method by comparing pseudo-annotation data with manually annotated data. S603. Design three types of prompts: (1) Create a drawing analysis dataset; (2) Extract geometric and dimensional information; (3) Generate a structured description; The prompt consists of four parts: roles and tasks, thought process, output format, and precautions. It adopts an iterative mechanism that includes human feedback, called the Prompt Development Loop (PrDC). The steps of PrDC are as follows: S6031. Assign roles to architects or structural engineers and provide general instructions for identifying component types, shapes, and dimensions; S6032. Define specific task instructions to clarify how to identify dimensioning and geometric symbols; S6033. Apply the thinking chain method to break down complex parsing tasks into multiple steps, and use contextual information to assist LLM in accurate parsing. S6034. Evaluate the accuracy of LLM parsing by evaluating contextual information, correct errors, and fine-tune the model so that it can automatically identify the geometry, dimensions and material information in the drawings and generate descriptions in structured JSON format or natural language, including component type, geometric features and dimensional details. S604. Integrate the text annotations in the drawings into the prompts, and convert the structured JSON format data into a string format as the target output of the model. Subsequently, fine-tuning is performed using quantized low-rank adaptive tuning. The fine-tuning steps are as follows: S6041. Use computer vision technology to extract text, annotations and geometric information from drawings and images, and convert them into an input format that the model can process; S6042. The tokenizer of the model is responsible for encoding the extracted information into a sequence of integer tokens, while the basic Llama model with a 4-bit weight format is responsible for parsing and generating the structural description. S6043. During the fine-tuning process, the LoRA parameter weights are continuously optimized through forward and backward propagation. A quantization-aware training strategy is used to dynamically switch between 32-bit and 4-bit weights to achieve gradient calculation and weight update. S6044. The fine-tuning process introduces supervised fine-tuning and reinforcement learning based on human feedback to enhance the model's ability to understand and describe the structural shapes and dimensions in the drawings. S6045. The finely tuned model can generate a structured JSON format or natural language description based on the drawing information, including component type, geometric features and dimensional details.

8. The automatic finite element modeling method for steel bridges integrating 3D point cloud and intelligent map recognition as described in claim 1, characterized in that, In step S7, the specific steps for automatically establishing the finite element model according to the structural characteristics of the bridge are as follows: S701. Establish a finite element model of the main beam. The main beam is discretized using three-dimensional solid elements to accurately describe the spatial structure of the bridge deck and web. Subsequently, establish a bridge tower model on the basis of the main beam. The bridge tower is modeled using solid elements or beam elements to ensure that its load-bearing capacity and stiffness characteristics can be accurately reflected. S702. Arrange the stay cables according to the bridge design parameters. The stay cables are simulated using cable elements, and a connection is established between the bridge tower and the main beam to correctly describe the cable force transmission mechanism. S703. Establish a pier model. Piers are usually modeled using solid elements or beam elements and connected to the main beam through support elements to accurately simulate the bridge support conditions. S704. The main beam and the bridge tower are connected by rigid connection or contact element to simulate their stress relationship. The stay cables are connected to the main beam and the bridge tower through cable element. The pier and the main beam are supported by elastic support element or rigid constraint.

Images