Bridge vortex-induced vibration sensitive pneumatic structure detection method and system based on three-dimensional laser scanning

Abstract

The invention discloses a bridge vortex-induced vibration sensitive pneumatic structure detection method and system based on three-dimensional laser scanning, and belongs to the technical field of bridge engineering and structural health monitoring. Three-dimensional point cloud data of a bridge structure are obtained through three-dimensional laser scanning, and geometric features are extracted after preprocessing; a pre-trained machine learning model is used for identifying the type of the component, shape comparison with a standard bridge type database is carried out, and whether the part has sensitive aerodynamic characteristics easily causing vortex vibration or not is comprehensively judged; for the sensitive characteristics, vortex-induced vibration risk parameters are calculated based on geometric parameters and compared with the inherent dynamic characteristics of the bridge to determine the risk level; for non-sensitive features, the evaluation is terminated. The method is mainly used for rapid evaluation of vortex-induced vibration risks in bridge design, construction and operation stages, provides technical support for bridge safety monitoring and maintenance, effectively prevents vortex-induced vibration accidents, and improves bridge durability and operation safety.

Description

Technical Field

[0001] This invention relates to bridge engineering and structural health monitoring. More specifically, this invention relates to a method and system for detecting bridge vortex-sensitive aerodynamic structures based on three-dimensional laser scanning. Background Technology

[0002] In bridge engineering, vortex-induced vibration (VIN) is a finite-amplitude vibration caused by periodic vortex shedding when airflow passes over specific structural parts of a bridge. When the vibration reaches certain conditions, it can affect driving comfort, structural fatigue life, and even operational safety. Therefore, accurately identifying aerodynamically sensitive structural features that may trigger VIN is crucial during the bridge design and maintenance phases.

[0003] For a long time, the identification and evaluation of aerodynamic structures of bridges susceptible to vortex-induced vibration have mainly relied on aerodynamic shape analysis during the design phase, wind tunnel scaling tests, and periodic manual inspections after construction. These methods face several inherent difficulties in practical applications.

[0004] Firstly, in the inspection of existing bridges, traditional methods mainly rely on manual surveying and contact measurement. Engineers obtain the macroscopic shape of the structure through visual observation and simple measuring tools, but it is difficult to accurately and comprehensively capture the microscopic geometric features and spatial attitude of the components. For example, the specific shape, angular curvature, and protruding dimensions of auxiliary components such as railings, maintenance vehicle tracks, crash barriers, and main beam flanges are all key aerodynamic factors affecting vortex shedding characteristics. Manual measurement methods are inefficient, have insufficient data density, and are easily constrained by the subjective experience of the measurer and on-site conditions, leading to deviations or even omissions in the quantitative description of complex three-dimensional geometric features. This means that some aerodynamically sensitive parts that were not fully considered in the design or that have newly emerged due to deformation or damage during operation may be overlooked.

[0005] Secondly, in the sensitive feature identification stage, existing methods largely rely on engineers' personal experience and understanding of the aerodynamic characteristics of standard bridge types. This typically requires comparing on-site observations with existing wind tunnel test databases or typical aerodynamic shapes in design specifications. However, the actual geometry of components in real bridges, especially older bridges that have been in operation for many years, may differ from the original design drawings or have undergone unexpected changes for various reasons. This judgment method, relying on standard models and subjective experience, lacks an objective and systematic analysis process of the actual bridge's current three-dimensional geometry, making it difficult to form standardized and repeatable identification conclusions. For complex composite components or non-standard designs, accurately determining whether they possess sensitive aerodynamic features prone to vortex-induced vibration is particularly difficult, easily leading to misjudgments or omissions.

[0006] Furthermore, even if potentially aerodynamically sensitive areas are identified during the risk assessment phase, accurately assessing vortex-induced vibration risk based on their actual geometric parameters remains a challenge. Vortex-induced vibration is closely related to the structure's own dynamic characteristics (such as frequency and damping). Traditional methods convert the geometric parameters of identified sensitive features into specific vortex-induced force parameters and correlate them with the bridge's inherent dynamic characteristics. This process is cumbersome and often requires complex computational fluid dynamics simulations or further wind tunnel testing, which is difficult to apply widely in routine inspections and rapid screening. Therefore, the risk assessment phase often becomes disconnected from the initial detection and identification phases, making it impossible to promptly screen and prioritize multiple potential risk points after on-site inspection, thus affecting assessment efficiency and the development of targeted measures.

[0007] The root cause of these problems lies in the inadequacy of accurately acquiring and structurally analyzing the actual three-dimensional geometric information of the bridge structure, as well as the insufficient integration with aerodynamic judgment. Overcoming these difficulties requires exploring how to more effectively combine high-precision three-dimensional geometric data of actual bridges with aerodynamic sensitivity judgment logic based on component type and geometric characteristics, and rapid risk estimation methods, thereby improving the objectivity, accuracy, and efficiency of inspection and assessment. Summary of the Invention

[0008] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0009] Another objective of this invention is to provide a method and system for detecting bridge vortex-sensitive aerodynamic structures based on three-dimensional laser scanning. This addresses the technical problems of traditional bridge vortex-sensitive detection, which mainly relies on manual surveys and wind tunnel tests. These methods are characterized by low efficiency, strong subjectivity, difficulty in accurately capturing microscopic geometric features through manual measurement leading to missed detection of aerodynamically sensitive parts, and a lack of systematic evaluation processes that hinder rapid screening of risk points.

[0010] To achieve these objectives and other advantages of the present invention, a method for detecting bridge vortex-sensitive aerodynamic structures based on three-dimensional laser scanning is provided, comprising the following steps: Step 1: Obtain 3D point cloud data of the bridge structure, preprocess the 3D point cloud data and extract the geometric features of the bridge; Step 2: Based on the extracted geometric features, a pre-trained machine learning model is used to identify the component types of each part of the bridge, and the geometric features are compared with the models in the standard bridge type database. Based on the identified component types and corresponding geometric features, it is determined whether the part is a sensitive aerodynamic feature that is prone to vortex-induced vibration. Step 3: For parts identified as having sensitive aerodynamic characteristics, calculate their vortex-induced vibration risk parameters based on their geometric parameters, and compare these risk parameters with the inherent dynamic characteristics of the bridge structure to determine their vortex-induced vibration risk level; for parts identified as having non-sensitive aerodynamic characteristics, terminate the vortex-induced vibration risk assessment process for those parts.

[0011] Preferably, in the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, the specific process of acquiring the three-dimensional point cloud data of the bridge structure in step one is as follows: A UAV and a ground-based vehicle platform work collaboratively, wherein the UAV is equipped with a lidar to perform top-down and side-view scans of the bridge deck and superstructure, while the ground-based vehicle platform performs bottom-up and multi-view scans of the bridge piers and substructure; during the scanning process, GPS, IMU, and odometer data are acquired simultaneously, and multi-source data fusion and real-time point cloud stitching are performed using a graph-optimized Simultaneous Localization and Mapping (SLAM) algorithm to generate a three-dimensional point cloud model; The synchronous positioning and mapping process uses bridge semantic features as constraint factors for pose graph optimization. Specifically, it identifies the bridge structural semantic features in the point cloud data, calculates the difference between the measured 3D coordinates of the semantic features and the theoretical coordinates in the preset bridge BIM model to obtain coordinate residuals, adds the coordinate residuals as constraint edges to the pose graph, and optimizes them together with constraint factors from GPS, IMU, and odometer to obtain 3D point cloud data of the bridge structure. Semantic features include cable anchor points, support centers, expansion joint outlines, or specific segments of crash barriers.

[0012] Preferably, in the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, step one, extracting the geometric features of the bridge, includes the following steps: S11. Perform noise filtering and descratizing preprocessing on the three-dimensional point cloud data of the bridge structure to obtain the preprocessed three-dimensional point cloud data. S12. Perform principal axis analysis on the preprocessed 3D point cloud data and align the bridge deck centerline with the X-axis or Y-axis of the global coordinate system. Then, slice the point cloud along the longitudinal direction of the bridge at preset equal intervals to generate a series of transverse point cloud slices. S13. For each horizontal point cloud slice, if the point cloud slice is a bridge deck area point cloud, identify the sudden change in point cloud density and the abrupt change in normal vector, and extract its horizontal hollow structure outline; if the point cloud slice is a pier area point cloud, extract the point cloud of its top plane through the Random Sampling Consensus (RANSAC) plane fitting algorithm, and use it as the pier cap area point cloud. S14. Based on the point cloud of the bridge deck area, the outline of the hollow structure, and the point cloud of the pier cap area, the geometric features of the bridge are extracted by calculating the spatial coordinates and topological relationships of the point cloud. These features include the curvature and sharpness of the box girder cross-section edge, the protrusion height and width of the auxiliary components relative to the surface of the main beam, and the aperture size and boundary shape of the hollow structure outline.

[0013] Preferably, in the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, in step two, the pre-trained machine learning model adopts an online incremental learning mechanism. The execution steps of the online incremental learning mechanism include: S21. When the confidence level of the model in recognizing a certain component in the bridge point cloud is lower than the preset threshold, the point cloud data slice corresponding to the component and its context feature vector are automatically stored in a sample pool to be verified. S22. Receive the operator's annotation and approval of the component types in the sample pool to be verified through the user interface; S23. Using the approved sample data and annotations, perform a round of parameter fine-tuning on the classifier of the machine learning model and update the model weights.

[0014] Preferably, in the bridge vortex-induced vibration sensitive aerodynamic structure detection method based on three-dimensional laser scanning, step three, the specific process of calculating the vortex-induced vibration risk parameters based on its geometric parameters, includes: S31. Based on the point cloud data of the parts identified as sensitive aerodynamic features, a closed three-dimensional curved surface mesh model of the parts is generated using the Poisson reconstruction algorithm. S32. Using the three-dimensional curved surface mesh model as the boundary, generate an unstructured mesh for computational fluid dynamics simulation around it; S33. Using the measured average wind speed and wind direction angle at the location of the bridge as the simulation inlet boundary conditions, perform large eddy simulation to obtain the flow field data around the bridge. S34. Perform a fast Fourier transform on the velocity vector sequence of specific monitoring points in the wake region of the structure in the simulation results, and extract the peak frequency of the transformed energy spectrum as the vortex-induced vibration risk parameter, i.e., the main vortex shedding frequency.

[0015] Preferably, in the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, step two, determining whether the part is a sensitive aerodynamic feature that easily induces vortex-induced vibration based on the identified component type and corresponding geometric features, specifically includes: A sensitive aerodynamic feature rule base is constructed, which defines the influence patterns of different component types on airflow under different geometric parameter thresholds. The influence patterns include the blunting of the guide angle of streamlined box girder, the protruding features of maintenance vehicle track or crash barrier, and the depth-to-width ratio of the cavity in the opening or grid structure. Based on the shape comparison results, the identified component types and their quantified geometric feature parameters are input into the rule base for matching. If the matching result matches any preset geometric feature pattern that may cause periodic vortex shedding, the part is determined to be a sensitive aerodynamic feature; otherwise, the part is determined to be a non-sensitive aerodynamic feature.

[0016] Preferably, in the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, step three, comparing the risk parameter with the inherent dynamic characteristics of the bridge structure to determine its vortex-induced vibration risk level, specifically includes: Obtain the natural frequencies f of the bridge structure in multiple modes. i and corresponding mode shape; The calculated main vortex shedding frequency f v With each order of natural frequency f i Compare and calculate the normalized frequency ratio η. i =f v / f i ; Define a risk level determination matrix that simultaneously considers the frequency ratio η i The degree of proximity and the energy proportion of the corresponding mode shape in the sensitive aerodynamic feature parts; When a certain mode exists, its frequency is higher than η i When the location falls within the locked range and exhibits significant displacement within that mode shape, it is classified as a high-risk level; when the frequency ratio is η... i If the risk level is close to but not completely within the locked-in area, it is classified as medium risk; otherwise, it is classified as low risk.

[0017] Preferably, the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning further includes: Step 4: Collect actual vibration response data of the bridge in a natural wind field by using vibration sensors installed on the sensitive aerodynamic feature parts of the bridge; perform spectrum analysis on the vibration response data to extract the actual main vibration frequency; The actual principal vibration frequency is compared with the calculated principal eddy shedding frequency. If the deviation between the two exceeds the preset tolerance, the turbulence model parameters or boundary condition settings in the computational fluid dynamics simulation are automatically calibrated, and the risk parameters are recalculated to optimize the accuracy of subsequent assessments.

[0018] The present invention also provides a system for implementing the above-described method for detecting bridge vortex-sensitive aerodynamic structures based on three-dimensional laser scanning, comprising: The data acquisition module is used to acquire three-dimensional point cloud data of the bridge structure, preprocess the three-dimensional point cloud data and extract the geometric features of the bridge. The identification module is used to identify the component type of each part of the bridge based on the extracted geometric features and a pre-trained machine learning model. It also compares the geometric features with the model in the standard bridge type database and, based on the identified component type and corresponding geometric features, determines whether the part is a sensitive aerodynamic feature that is prone to vortex-induced vibration. The assessment module is used to calculate the vortex-induced vibration risk parameters of parts identified as having sensitive aerodynamic characteristics based on their geometric parameters, and compare these risk parameters with the inherent dynamic characteristics of the bridge structure to determine its vortex-induced vibration risk level; for parts identified as having non-sensitive aerodynamic characteristics, the vortex-induced vibration risk assessment process for that part is terminated.

[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0020] The present invention has at least the following beneficial effects: 1. This invention provides a method for detecting bridge vortex-induced vibration-sensitive aerodynamic structures based on three-dimensional laser scanning. Through a systematic approach, it achieves rapid and objective assessment of bridge vortex-induced vibration risk. First, it utilizes three-dimensional laser scanning to acquire detailed point cloud data of the bridge structure and extract key geometric features, overcoming the limitation of traditional manual inspection in capturing microscopic geometric defects. Then, a pre-trained machine learning model identifies component types and compares their shapes with a standard bridge type database to comprehensively determine whether a location possesses a sensitive aerodynamic feature, effectively reducing misjudgments and omissions caused by reliance on subjective experience. For sensitive features, vortex-induced vibration risk parameters are further calculated based on geometric parameters and compared with the bridge's inherent dynamic characteristics to determine the risk level; for non-sensitive features, the assessment is terminated, optimizing resource allocation. This invention achieves full automation from data acquisition to risk assessment, significantly improving detection efficiency and accuracy, reducing detection costs and risks. It is applicable to all stages of bridge design, construction, and operation, providing reliable technical support for bridge safety monitoring and maintenance, helping to prevent vortex-induced vibration accidents and enhance bridge durability.

[0021] 2. This invention further optimizes the acquisition method of 3D point cloud data. By coordinating UAVs and ground-based vehicle platforms, it achieves multi-view coverage scanning of the superstructure and substructure of bridges, avoiding the blind spot problem of traditional single-platform scanning. Combining GPS, IMU, and odometer data, and employing a graph-optimized SLAM algorithm for multi-source data fusion and real-time point cloud stitching, the accuracy and completeness of the point cloud model are significantly improved. In particular, by using bridge semantic features (such as cable anchor points and support centers) as constraint factors for pose graph optimization, and utilizing the residuals between measured coordinates and theoretical coordinates of the BIM model for optimization, the accumulated error is effectively reduced, enhancing the reliability and geometric consistency of the point cloud data. This high-precision data acquisition method lays a solid foundation for subsequent feature extraction and risk analysis, improves the automation and intelligence level of the entire inspection process, and reduces dependence on external environment and manual operation.

[0022] 3. This invention details the specific steps for geometric feature extraction, including noise filtering, principal axis analysis, point cloud slicing, and region-specific feature extraction, ensuring the accuracy and comprehensiveness of geometric parameter calculations. Preprocessing removes noise and impurities, improving point cloud data quality. Principal axis analysis aligns the bridge deck's centerline with the global coordinate system, facilitating subsequent standardization. Equal-spacing slicing along the bridge's longitudinal direction enables refined analysis of local features. For the bridge deck area, abrupt changes in point cloud density and normal vector mutations are identified to extract the hollow structure contour. For the pier area, the RANSAC plane fitting algorithm is used to extract the point cloud of the pier cap area, enhancing the robustness of feature extraction. The finally extracted geometric features (such as the curvature of the box girder edge and the protruding dimensions of auxiliary components) provide key parameters for sensitive aerodynamic feature identification. This method is highly automated, reduces human error, and improves the consistency and repeatability of feature extraction.

[0023] 4. This invention introduces an online incremental learning mechanism using a pre-trained machine learning model, enabling the system to dynamically adapt to complex and variable components in actual bridges. When the model's confidence level falls below a threshold, low-confidence samples are automatically stored in a verification pool and manually labeled and verified through a user interface. The verified data is then used to fine-tune the model's parameters. This mechanism effectively solves the problem of insufficient generalization ability when facing unseen or uniquely shaped components, continuously improving recognition accuracy and adaptability. Online learning avoids the high cost of retraining traditional models, achieving incremental optimization and enhancing the system's robustness and flexibility in practical applications. Through a human-machine collaborative feedback loop, the system can gradually accumulate domain knowledge, improve the recognition accuracy of diverse bridge components, and reduce the risk of misjudgment.

[0024] 5. This invention clarifies the calculation process of vortex-induced vibration risk parameters. A closed three-dimensional curved surface mesh is generated based on point cloud data of sensitive areas, and the main vortex shedding frequency is extracted through computational fluid dynamics simulation. The Poisson reconstruction algorithm ensures the integrity and accuracy of the geometric model, and the generated unstructured mesh adapts to complex shapes, improving simulation accuracy. Using large eddy simulation combined with measured wind speed and direction angle as boundary conditions, the characteristics of the flow field around the vortex can be realistically reflected, thus accurately capturing vortex shedding phenomena. By performing a fast Fourier transform on the velocity vector sequence in the wake region, the peak frequency of the energy spectrum is extracted as a risk parameter, realizing a direct conversion from geometric data to aerodynamic risk parameters. This method replaces the traditional approach that relies on wind tunnel tests or simplified formulas, is unaffected by scaling effects and model manufacturing errors, provides a more scientific and reliable basis for risk assessment, and significantly improves the accuracy and practicality of the analysis.

[0025] 6. This invention constructs a rule base for sensitive aerodynamic features, transforming aerodynamic expertise into executable judgment logic. This enables objective and systematic identification of sensitive features. The rule base defines the influence patterns of different component types on airflow under specific geometric parameter thresholds, such as guide angle blunting, protruding body features, and the aspect ratio of hollow cavities, covering common vortex-induced vibrations. Combining shape comparison results, the identified component types and their quantified geometric parameters are input into the rule base for matching, automatically determining whether a feature is sensitive. This method reduces reliance on engineers' personal experience, improving the consistency and repeatability of judgments. The scalability of the rule base allows for updates and additions of new rules at any time, adapting to different bridge types and advancements in aerodynamic research, enhancing the system's adaptability and comprehensiveness, and effectively reducing the probability of missed and false judgments.

[0026] 7. This invention calculates the frequency ratio by comparing the main eddy current shedding frequency with the bridge's natural frequencies, and then comprehensively assesses the risk level by combining this with the modal energy ratio. The defined risk level judgment matrix considers both frequency matching and the significant displacement of the modal shape at sensitive locations, achieving multi-dimensional risk analysis. When the frequency ratio falls within the locked range and the modal displacement at the location is significant, it is classified as high risk; close to the locked range, it is classified as medium risk; otherwise, it is classified as low risk. This classification method is scientific and reasonable, accurately identifying potentially high-risk locations and providing a clear basis for subsequent maintenance decisions. Through quantitative analysis and matrix judgment, it avoids the limitations of single-frequency comparison in traditional methods, improves the comprehensiveness and reliability of risk assessment, and helps optimize monitoring resources and take targeted measures.

[0027] 8. This invention adds a calibration step based on actual vibration response. Vibration sensors are deployed to collect data from the natural wind field, and spectral analysis is performed to extract the actual principal vibration frequency. This actual principal vibration frequency is compared with the calculated principal eddy shedding frequency. If the deviation exceeds a preset tolerance, the turbulence model parameters or boundary condition settings in the computational fluid dynamics simulation are automatically calibrated, and risk parameters are recalculated to optimize the accuracy of subsequent assessments. This feedback mechanism effectively solves the problem of discrepancies between the simulation model and the real environment, improving the accuracy and adaptability of risk assessment. Through continuous data acquisition and model optimization, the system can gradually approximate real working conditions, enhancing the reliability of long-term assessments. This invention achieves a close integration of theory and practice, providing dynamic correction capabilities for bridge health monitoring and improving the practicality and reliability of the entire system.

[0028] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description.

[0030] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0031] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0032] This invention provides a method for detecting bridge vortex-sensitive aerodynamic structures based on three-dimensional laser scanning, which includes the following steps: Step 1: Obtain 3D point cloud data of the bridge structure, preprocess the 3D point cloud data and extract the geometric features of the bridge; Step 2: Based on the extracted geometric features, a pre-trained machine learning model is used to identify the component types of each part of the bridge, and the geometric features are compared with the models in the standard bridge type database. Based on the identified component types and corresponding geometric features, it is determined whether the part is a sensitive aerodynamic feature that is prone to vortex-induced vibration. Step 3: For parts identified as having sensitive aerodynamic characteristics, calculate their vortex-induced vibration risk parameters based on their geometric parameters, and compare these risk parameters with the inherent dynamic characteristics of the bridge structure to determine their vortex-induced vibration risk level. For parts identified as having non-sensitive aerodynamic characteristics, terminate the vortex-induced vibration risk assessment process for those parts. Simultaneously, record the component type, geometric parameters (such as cross-sectional dimensions and relative position), and the assessment results, storing them in the bridge health monitoring database for subsequent state comparison and trend analysis.

[0033] This technical solution provides a method for detecting bridge vortex-sensitive aerodynamic structures based on three-dimensional laser scanning. This method solves the technical problems of traditional bridge vortex-sensitive detection methods, which rely on manual surveys and wind tunnel tests, resulting in low efficiency, strong subjectivity, difficulty in accurately capturing microscopic geometric features, and a lack of systematic evaluation processes and rapid risk point screening. Traditional methods suffer from low efficiency and susceptibility to experience-based biases, failing to comprehensively quantify structural details; wind tunnel tests are costly, time-consuming, and limited by scaling effects, making it difficult to reflect the three-dimensional geometric variations of actual bridges. This invention achieves objective and rapid detection of bridge aerodynamically sensitive features through automated data acquisition, intelligent identification, and simulation analysis.

[0034] In the specific implementation of the above technical solution, the three-dimensional point cloud data refers to the dense set of three-dimensional coordinate points on the bridge surface obtained by LiDAR scanning. Geometric features include parameters such as the curvature and sharpness of the box girder cross-section edge, the protrusion height and width of the auxiliary components relative to the main beam surface, and the aperture size and boundary shape of the hollow structure outline. Sensitive aerodynamic features refer to structural shapes that are prone to periodic vortex shedding, such as the blunting of the guide angle of the streamlined box girder, the protruding features of the maintenance vehicle track or the guardrail, and the abnormal depth-to-width ratio of the hollow cavity in the opening or grid structure. The vortex-induced vibration risk parameter refers to the main vortex shedding frequency extracted by computational fluid dynamics simulation. The inherent dynamic characteristics of the bridge include the natural frequencies and corresponding mode shapes of the bridge under multiple modes.

[0035] The aforementioned technical solution first employs a collaborative operation between a drone and a ground-based vehicle platform for 3D laser scanning. The drone, equipped with a lidar, performs top-down and side-view scans of the bridge deck and superstructure, while the ground-based vehicle platform performs bottom-up and multi-view scans of the piers and substructure, ensuring full coverage without blind spots. During the scanning process, GPS, IMU, and odometer data are simultaneously acquired. A graph-optimized synchronous positioning and mapping algorithm is used for multi-source data fusion and real-time point cloud stitching. Specifically, the measured 3D coordinates of bridge semantic features, such as cable anchor points, support centers, expansion joint contours, or specific segments of crash barriers, are compared with the theoretical coordinates of the pre-set bridge BIM model to obtain coordinate residuals. These residuals serve as constraint factors for pose graph optimization and are used in conjunction with GPS, IMU, and odometer data to improve point cloud accuracy and geometric consistency. In the point cloud preprocessing stage, noise filtering and de-emphasis are performed. Subsequently, principal axis analysis aligns the bridge deck's central axis with the global coordinate system, and point cloud slices are generated along the bridge's longitudinal direction at pre-set equal intervals. For each slice, the outline of the lateral hollow structure in the bridge deck area is extracted by identifying the abrupt changes in point cloud density and the abrupt changes in normal vector. In the pier area, the point cloud of the pier cap area is extracted by the random sampling consistency plane fitting algorithm. Finally, the geometric features are calculated based on the spatial coordinates and topological relationships of the point cloud.

[0036] Subsequently, a pre-trained machine learning model was used to identify component types. The model employed an online incremental learning mechanism, automatically storing the corresponding point cloud slices and contextual feature vectors into a verification sample pool when the identification confidence level fell below a threshold. User annotations and verifications were received via a user interface, and the model classifier parameters were fine-tuned using the verified samples, updating weights to improve identification adaptability. Simultaneously, geometric features were compared with a standard bridge type database, combined with a constructed sensitive aerodynamic feature rule base. This rule base defined the influence patterns of different component types on airflow under geometric parameter thresholds. Whether a feature was sensitive was determined by matching the component type with quantified geometric parameters. For sensitive features, a closed three-dimensional curved surface mesh model was generated using the Poisson reconstruction algorithm. This model served as the boundary for generating an unstructured mesh for computational fluid dynamics simulation. Measured average wind speed and wind direction angle were used as the simulation inlet boundary conditions to perform large eddy simulation to obtain flow field data around the structure. A fast Fourier transform was performed on the velocity vector sequence of monitoring points within the structure's wake region, and the peak frequency of the energy spectrum was extracted as the main vortex shedding frequency. The risk parameter is compared with the bridge's inherent dynamic characteristics to calculate the normalized frequency ratio of each natural frequency to the main vortex shedding frequency. This ratio, combined with the risk level determination matrix, considers both the closeness of the frequency ratios and the energy proportion of the corresponding vibration mode in sensitive areas to determine high, medium, or low risk levels. Furthermore, actual vibration response data from natural wind fields is collected by deploying vibration sensors. Spectral analysis is performed to extract the actual principal vibration frequency, which is then compared with the calculated main vortex shedding frequency. If the deviation exceeds the tolerance, the simulation parameters are automatically calibrated and the risk parameter is recalculated to optimize the accuracy of the assessment.

[0037] The above-mentioned technical solution realizes the fully automated detection of bridge vortex-sensitive aerodynamic structures. It captures microscopic geometric features through high-precision three-dimensional scanning, overcoming the subjectivity and omissions of manual measurement. The combination of machine learning and rule base improves the objectivity and consistency of identification. Computational fluid dynamics simulation provides scientific risk parameters, and calibration with actual vibration data enhances the reliability of the results. The overall method significantly improves detection efficiency and accuracy, reduces labor costs and risks, and is applicable to all stages of bridge design, construction and operation, providing reliable technical support for bridge safety monitoring and maintenance.

[0038] In another technical solution, the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, in step one, the specific process of acquiring the three-dimensional point cloud data of the bridge structure is as follows: A drone and a ground-based vehicle platform work collaboratively. The drone, equipped with a lidar, performs top-down and side-view scans of the bridge deck and superstructure, while the ground-based vehicle platform performs bottom-up and multi-view scans of the piers and substructure. During the scanning process, GPS, IMU, and odometer data are acquired simultaneously, and multi-source data fusion and real-time point cloud stitching are performed using a graph-optimized Simultaneous Localization and Mapping (SLAM) algorithm to generate a three-dimensional point cloud model. The lidar on the drone must meet the requirement of a point cloud density ≥ 200 points / cm². 2 The ranging accuracy must be ±5mm, and the scanning rate ≥1 million points / second; the LiDAR mounted on the ground vehicle platform must meet the requirement of a point cloud density ≥150 points / cm². 2 The ranging accuracy is ±3mm, and the scanning rate is ≥2 million points / second. The equipment selection should be combined with the detection range: for long-span bridges, long-range (range ≥200m) lidar should be preferred, while for medium and small-span bridges, medium and short-range (range 50-100m) lidar can be selected. In the synchronous positioning and mapping process, the semantic features of the bridge are used as constraint factors for pose graph optimization. Specifically, the semantic features of the bridge structure in the point cloud data are identified, and the measured 3D coordinates of the semantic features are subtracted from the theoretical coordinates in the preset bridge BIM model to obtain coordinate residuals. These coordinate residuals are added as constraint edges to the pose graph and optimized together with constraint factors from GPS, IMU, and odometer to obtain 3D point cloud data of the bridge structure. Semantic features include cable anchor points, bearing centers, expansion joint outlines, or specific segments of crash barriers. In the synchronous positioning and mapping process, the semantic features (cable anchor points, bearing centers, etc.) of the initially stitched point clouds from each platform are first identified. Then, the measured 3D coordinates of the semantic features are subtracted from the theoretical coordinates in the preset bridge BIM model to obtain coordinate residuals. These coordinate residuals are added as constraint edges to the pose graph and optimized together with constraint factors from GPS, IMU, and odometer to achieve accurate point cloud stitching. If a bridge has been in operation for more than 5 years or has significant deformation (such as settlement or displacement ≥50mm), the preset bridge BIM model must be updated first: the BIM model coordinates must be corrected based on historical inspection data (such as settlement observation and displacement monitoring data) to ensure that the theoretical coordinates of the BIM model deviate from the current state of the actual bridge by ≤10mm; if there is no historical inspection data, the coordinates of at least 5 key semantic features (such as the top of the bridge tower and the center of the support) must be measured using a total station to correct the BIM model.

[0039] When using 3D laser scanning technology for bridge inspection, a fundamental and critical challenge lies in how to efficiently, completely, and accurately acquire a comprehensive 3D point cloud model of the entire bridge. Traditional scanning methods often rely on a single scanning platform, such as using only ground-based fixed scanning or vehicle-mounted mobile scanning. This approach is easily limited by viewpoint occlusion, often resulting in missing data areas for large bridges, such as the underside of the bridge deck, the cable-stayed area, and the tops of high piers. Furthermore, when fusing and stitching multi-source data from multiple scans, it typically relies mainly on sensor data such as GPS, inertial measurement units (IMUs), and odometers. However, in environments where satellite signals are obstructed, such as under bridges, the accuracy of GPS drops significantly, and IMUs themselves have accumulated errors. This leads to stitching misalignments or overall deviations in the final generated point cloud model. This lack of accuracy in the basic data directly impacts all subsequent geometric analysis and safety assessment stages.

[0040] To address the aforementioned issues, this technical solution proposes a refined modeling method based on collaborative scanning and semantic assistance. This method employs a collaborative operation between a UAV and a ground-based vehicle platform. The UAV performs aerial scans of the bridge deck and superstructure from both above and below, effectively covering areas such as the top surface of the beams, stay cables, and suspenders. The ground-based vehicle platform then performs supplementary scans from below, including upward and multi-view views of the substructure such as piers, box girder bottom plates, and supports, thus eliminating blind spots in physical space. At the data fusion level, it goes beyond simply importing scan data from various platforms into the software; it involves deep sensor fusion. During the scanning process, raw data from the Global Positioning System (GPS), Inertial Measurement Unit (INS), and odometer are recorded simultaneously, and a graph-optimized simultaneous positioning and mapping algorithm is used as the core technical framework for real-time calculation and point cloud stitching. This invention incorporates the inherent, identifiable semantic features of the bridge itself as strong constraints into the pose graph optimization model. Specifically, the system identifies semantic feature points with clear geometric definitions and stable structural positions from the real-time point cloud, such as cable anchor points, support centers, the outlines of specific types of expansion joints, or regular segments of crash barriers. Then, the measured 3D coordinates of these feature points in the current scanning coordinate system are subtracted from the corresponding theoretical design coordinates in the pre-imported bridge information model to obtain coordinate residuals. This coordinate residual is constructed as a new constraint edge and added to the pose graph of the synchronous positioning and mapping process, participating in the optimization iteration calculation along with constraint factors from the Global Positioning System, Inertial Measurement Unit, and Odometer. In this way, the known and stable geometric relationships of the bridge structure itself are used to correct and constrain trajectory drift and point cloud misalignment caused by sensor errors, thereby generating a more accurate and consistent 3D point cloud model in the global coordinate system.

[0041] In the above technical solutions, collaborative operation refers to multiple platforms equipped with scanning equipment cooperating to collect data according to a pre-defined spatial division of labor. Ground-based vehicle platforms typically refer to mobile vehicles equipped with devices such as lidar, satellite positioning receivers, and inertial navigation systems. Semantic features specifically refer to key points or contours in the point cloud that can be identified and correspond to specific structural elements in the bridge information model, possessing precise theoretical coordinates; they are not merely geometric points but also carry structural identity information. Pose graph optimization is a commonly used backend optimization method in simultaneous localization and mapping (SMR) technology. It represents the pose (position and orientation) of the robot (or scanning platform) at different times as nodes in the graph, and various sensor measurements or constraints as edges connecting these nodes. By minimizing the sum of errors of all edges, the node pose is optimized, thereby obtaining a globally consistent map. Coordinate residuals refer to the vector difference between the measured coordinates of corresponding feature points and the theoretical coordinates of the bridge information model, intuitively reflecting the degree of deviation between the current point cloud data and the actual structure in absolute position.

[0042] The above technical solution first requires task planning, allocating scanning areas and paths to both the UAV and the ground vehicle to ensure comprehensive coverage and sufficient overlap between their scanning areas for subsequent fusion. During data acquisition, each platform operates independently, but all sensor data must be accurately timestamped. In the data processing stage, preliminary point cloud stitching and trajectory calculation are performed within each platform. Next comes the global optimization stage, which is the core of the technical implementation. The algorithm automatically or semi-automatically extracts predefined semantic features from the point clouds of each platform and performs coarse and fine registration with the bridge information model to establish accurate correspondences. Then, a large pose graph is constructed, where nodes represent the poses of each platform at different times, and edges include various constraints: adjacent pose constraints generated by pre-integration of the platform's internal inertial measurement unit, absolute position constraints provided by the satellite positioning system when the signal is good, relative motion constraints provided by the odometry, and the most critical new constraint—coordinate residual constraints generated by semantic feature matching. The optimizer strives to minimize the sum of squared errors of all these constraints. Because the theoretical coordinates provided by the bridge information model have high absolute accuracy and stability, these semantic constraints act like "pillars" driven into the optimization process, effectively correcting the cumulative drift caused by relying solely on satellite positioning systems and inertial measurement units. This is particularly crucial in the under-bridge area where satellite positioning system signals fail. Through iterative optimization, better trajectories for each scanning platform are obtained, and a globally consistent 3D point cloud model with significantly improved accuracy is generated.

[0043] The aforementioned technical solution fundamentally solves the problem of completeness in the acquisition of 3D data for complex bridge structures through multi-platform collaboration, ensuring no blind spots in scanning. More importantly, by introducing bridge semantic features as optimization constraints, it cleverly utilizes the known precise geometric information of the structure itself to correct the inherent errors of the sensor system. This significantly improves the absolute coordinate accuracy and internal consistency of the data at the source of production—the 3D point cloud reconstruction stage. The obtained point cloud model is not only morphologically complete, but its spatial correspondence with the actual bridge structure is also more accurate. This lays a reliable data foundation for subsequent fine geometric feature extraction, component identification, and aerodynamic characteristic analysis, avoiding the risk of distorted conclusions in all subsequent analyses and evaluations due to deviations in the basic data. Overall, it improves the quality and reliability of bridge digital inspection work.

[0044] In another technical solution, the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, step one, extracting the geometric features of the bridge includes the following steps: S11. Perform noise filtering and descratizing preprocessing on the three-dimensional point cloud data of the bridge structure to obtain the preprocessed three-dimensional point cloud data. S12. Perform principal axis analysis on the preprocessed 3D point cloud data to align the bridge deck centerline with the X-axis or Y-axis of the global coordinate system. Then, slice the point cloud along the longitudinal direction of the bridge at a preset equal interval (the preset equal interval needs to be determined in conjunction with the bridge span and component size, and the general range is 0.2-1m; for large-span bridges (span ≥ 100m) or complex components (such as irregular box girders), the interval should be 0.2-0.5m; for small-to-medium span bridges (span < 100m) or regular components (such as rectangular piers), the interval can be 0.5-1m), generating a series of transverse point cloud slices. S13. Region division based on Z-coordinate range and semantic feature labels of point cloud: The Z-coordinate range of the bridge deck area point cloud is ±1m of the bridge design deck elevation, and includes semantic features such as expansion joints and crash barriers; the Z-coordinate range of the pier area point cloud is from the top surface of the foundation to the bottom surface of the bridge deck, and includes semantic features such as pier caps and supports; for each transverse point cloud slice, if the point cloud slice is the bridge deck area point cloud, the sudden change in point cloud density and the abrupt change in normal vector are identified, and its transverse hollow structure outline is extracted; if the point cloud slice is the pier area point cloud, the point cloud of its top plane is extracted by the Random Sampling Consensus (RANSAC) plane fitting algorithm, which is used as the pier cap area point cloud; the hollow structure outline refers to the internal cavity enclosed by the web, top plate and bottom plate in the box girder section, or the transverse boundary shape of the grid-type component, which is identified by the abrupt change in point cloud density and the discontinuity of normal vector; S14. Based on the point cloud of the bridge deck area, the outline of the hollow structure, and the point cloud of the pier cap area, the geometric features of the bridge are extracted by calculating the spatial coordinates and topological relationships of the point cloud. These features include the curvature and sharpness of the box girder cross-section edge, the protrusion height and width of the auxiliary components relative to the surface of the main beam, and the aperture size and boundary shape of the hollow structure outline.

[0045] After obtaining the raw 3D point cloud data of a bridge, a key technical challenge is how to accurately and efficiently extract geometric features for aerodynamic analysis from this massive and noisy collection of spatial points. The raw point cloud typically contains noise from vegetation, vehicles, and other environmental sources, and the scanning process itself introduces noise, which severely impacts the accuracy of subsequent analyses. Even more challenging is how to correlate the disordered point cloud with the structural logic of the bridge. For example, how to automatically distinguish between the bridge deck area and the pier area, further identify the outline of the box girder from the bridge deck point cloud, separate the web and top and bottom plates from the point cloud inside the hollow structure, and accurately calculate subtle geometric parameters such as edge curvature and component protrusion height. Traditional methods often rely on manual interactive segmentation and measurement in point cloud software. This process is extremely time-consuming and highly dependent on the operator's experience, leading to inconsistent and poorly repeatable feature extraction results, making it difficult to meet the needs of engineering applications.

[0046] To address the aforementioned issues, this technical solution provides a structured automatic geometric feature extraction process. First, the original point cloud is preprocessed by using a statistical filtering algorithm to identify and remove isolated noise points far from the main point cloud group, providing a clean data foundation for subsequent analysis. Then, principal axis analysis is performed on the overall point cloud to calculate the principal axis representing the main extension direction of the bridge. This principal axis is then aligned with the X-axis or Y-axis of the global coordinate system. This step achieves spatial standardization of the point cloud data, laying the foundation for subsequent regular slicing. After alignment, the system automatically generates a series of transverse slices perpendicular to the principal axis along the longitudinal direction of the bridge at preset fixed intervals. These slices transform the massive 3D point cloud problem into a series of more easily processed 2D profile problems.

[0047] For each transverse point cloud slice, the system first determines its structural region. If it is identified as a bridge deck region, the algorithm focuses on analyzing abrupt changes in point cloud density and point normal vectors. These abrupt changes typically correspond to the boundary positions of the box girder cross-section from solid to hollow, or from top plate to web plate, allowing for precise extraction of the hollow structural outline of the box girder cross-section. If the point cloud slice is identified as belonging to the pier region, a random sampling consensus algorithm is used to perform planar fitting on this point cloud, robustly identifying the top planar point cloud of the pier cap and separating it as a representative of the pier cap region. Based on successful region identification and preliminary segmentation of the slices, the system further extracts key geometric features in batches by calculating the spatial coordinates and topological relationships of the point cloud. These features include the curvature and sharpness of the box girder cross-section edges, reflecting the streamlined nature of the aerodynamic shape; the protrusion height and width of auxiliary components such as crash barriers and maintenance vehicle tracks relative to the main beam surface, which are important parameters for determining whether vortex shedding will occur; and the aperture size and boundary morphology calculated from the hollow structural outline, used to assess the aerodynamic stability of the internal cavity.

[0048] In the above technical solutions, noise filtering refers to the process of removing scattered points unrelated to the bridge structure from the point cloud. Principal axis analysis refers to calculating the main direction of the point cloud data distribution using an algorithm and then performing coordinate rotation and alignment based on this direction. Point cloud slicing refers to cutting a three-dimensional point cloud with a series of parallel planes to obtain a series of two-dimensional point cloud profiles. Abrupt changes in point cloud density refer to a rapid change in the number of points per unit volume or area, usually indicating the boundary between different structures. Abrupt changes in normal vectors refer to the discontinuity of the local surface orientation of the point cloud, usually occurring at edges or corners. The hollow structure profile refers to the closed boundary formed by the internal cavity of the box girder on the sliced ​​plane. The random sampling consensus algorithm is an iterative method that can robustly estimate the parameters of a mathematical model from data containing a large number of outliers. Topological relationships describe the spatial relationships such as connections and adjacencies between points in the point cloud. Curvature is a mathematical quantity describing the degree of bending of a curve or surface. Sharpness refers to the sharpness of an edge.

[0049] In the above technical solution, statistical filtering is used in the preprocessing stage. The average distance from each point to all its nearest neighbors is calculated. Assuming these distances form a Gaussian distribution throughout the point cloud, points with average distances exceeding the standard range are removed. Principal axis analysis is achieved by eigenvalue decomposition of the point cloud covariance matrix; the eigenvector corresponding to the largest eigenvalue is the principal axis direction. After aligning the coordinate system, the slice spacing can be set according to the required modeling accuracy, typically between several centimeters and tens of centimeters. For the point cloud within each slice, density-based clustering analysis is first performed to separate the bridge deck point cloud from potentially mixed-in noise points. In the bridge deck point cloud, the k-nearest neighbor normal vector of each point is calculated, and its variance is statistically analyzed. Regions with significantly increased variance are identified as boundary feature regions. For hollow structures, the internal cavity region is separated from the external solid region by combining point cloud density differences and region growing algorithms, and its contour line is extracted. In the pier region, a random sampling consensus algorithm randomly samples a small number of points to fit the planar model. Then, the number of points within the entire slice that conform to the plane is counted. Through multiple iterations, the planar model with the most internal points is selected as the top surface of the pier cap. The final calculation of geometric parameters is based on these segmented point cloud subsets. Curvature is estimated by fitting a local quadratic surface and calculating the principal curvature. The salient height is achieved by calculating the distance between the point cloud of the auxiliary components and the point cloud of the main beam structure in the normal vector direction.

[0050] This technical solution significantly improves the automation and reliability of geometric feature extraction. Systematic preprocessing and spatial standardization provide a high-quality data foundation for subsequent analysis. Automated slicing and region recognition replace tedious manual operations, greatly improving processing efficiency and making it possible to extract comprehensive features of the entire bridge with millimeter-level precision. Rule-based feature extraction ensures the consistency and objectivity of the results; repeatable results can be obtained by different operators or at different times. Accurately extracted parameters such as curvature and protrusion height provide reliable data input for subsequent intelligent judgment of aerodynamic sensitive features, ensuring the accuracy and credibility of the entire vortex-induced vibration detection method's conclusions from the outset.

[0051] In another technical solution, the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, in step two, the pre-trained machine learning model adopts an online incremental learning mechanism. The execution steps of the online incremental learning mechanism include: S21. When the model's confidence level in recognizing a component in the bridge point cloud is lower than a preset threshold (typically 0.6-0.8, set based on historical recognition accuracy ≥90% and engineering error tolerance ≤5%, and can be dynamically adjusted according to the importance of the bridge; the threshold for Class I highway bridges should be 0.7-0.8, and the threshold for Class II and lower highway bridges can be 0.6-0.7), the point cloud data slice corresponding to the component and its context feature vector are automatically stored in a sample pool to be verified; the context feature vector includes the geometric features (curvature, size, normal vector distribution) of the component's point cloud slice, its spatial position in the bridge's global coordinate system, and its connection relationship with adjacent components (such as distance, angle, and other topological attributes). S22. Receive the operator's annotation and approval of the component types in the sample pool to be verified through the user interface; S23. Using the approved sample data and annotations, perform a round of parameter fine-tuning on the classifier of the machine learning model, update the model weights, and after fine-tuning, use a test set (containing 1000+ sets of point cloud data of different components) to verify the model recognition accuracy. The accuracy improvement must be ≥3% and there must be no overfitting (the difference between the accuracy of the test set and the accuracy of the training set ≤5%). If this is not met, readjust the fine-tuning parameters (such as learning rate and number of iterations) and train again.

[0052] A common challenge in the application of intelligent bridge component recognition technology is that the reliability of pre-trained machine learning models decreases significantly when faced with bridge components not fully covered in the training data or that differ from standard forms. Given the diverse types of bridges in my country, and the subtle morphological variations that can occur in the same type of component across different bridges, coupled with damage or modifications during operation, building a universal initial model that covers all situations is virtually impossible. The traditional solution involves periodically collecting large amounts of new samples and retraining the model offline. This process is computationally intensive and time-consuming, resulting in severely delayed model updates. This makes it unable to meet the immediate accuracy requirements of on-site inspections, thus affecting the continuity and reliability of subsequent aerodynamic feature assessments.

[0053] To address the issue of insufficient model adaptability, this technical solution introduces an online incremental learning mechanism. The core of this mechanism lies in endowing the machine learning model with the ability to continuously evolve and self-optimize during practical applications. When the model performs type identification on a component in a scanned point cloud, the system simultaneously calculates a confidence score reflecting the degree of certainty in this identification. Once this score falls below a preset threshold, indicating that the model's judgment at this point is not very confident, the system will not simply give a potentially incorrect conclusion. Instead, it automatically packages the corresponding point cloud data slice and its related contextual feature vectors into a dedicated verification sample pool. This sample pool, acting as a buffer and interactive intermediary, is presented to professional operators through a user interface. Based on their professional knowledge and experience, operators review these questionable samples and provide correct component type labels and verifications. These high-quality new samples and their labels, verified manually, then become valuable learning materials for the model. The system uses these few new samples to perform a rapid round of parameter fine-tuning on the original machine learning classifier, updating the model weights only with a few iterations, rather than starting from scratch. This process enables the model to absorb new feature patterns without forgetting previously learned knowledge, thereby gradually adapting to the shapes of components that were previously difficult to identify.

[0054] In the above technical solutions, the online incremental learning mechanism specifically refers to a learning paradigm where, after deployment, the machine learning model can continuously update its parameters using incoming new sample data without undergoing full retraining. Recognition confidence is a probabilistic estimate of the model's output correctness, typically represented by a value between zero and one; a lower value indicates greater uncertainty. The verification sample pool is a temporary data storage area on the storage medium specifically used to store the original data and features corresponding to the model's low-confidence recognition results. The context feature vector not only contains the geometric information of the component point cloud slice itself but may also include environmental information such as the component's relative position in the overall structure and its relationship with surrounding components. The user interface is a graphical software interface designed for operators, used to display questionable point cloud images and receive manual annotation instructions. Parameter fine-tuning refers to training and optimizing the model with a limited number of steps using new sample data at a small learning rate, based on existing model parameters, aiming to achieve smooth knowledge integration.

[0055] In the above technical solution, a confidence estimation module needs to be integrated after the initial model output layer. This module can be based on the probability distribution entropy value output by the model's softmax layer or through a dedicated confidence branch network. The preset confidence threshold needs to be set empirically based on the actual application's requirements for balancing accuracy and recall. When a low confidence condition is triggered, the system automatically extracts the point cloud slices of the current component and calls the feature extraction network to generate its high-dimensional feature vector. Both are then serialized and stored together with the corresponding low-confidence recognition result in the database of samples to be verified. The user interface displays the samples to be processed in the pool in the form of a task list. Operators can rotate and scale the point cloud slices to assist in judgment and select the correct label from the predefined component type list for confirmation. The system background periodically checks whether newly approved samples have accumulated to a certain batch. Once the conditions are met, a round of fine-tuning is initiated. The fine-tuning process loads the latest model weights, fixes most layers of the feature extraction backbone network, mainly targeting the classifier layer, and trains it using a new set of approved samples at a learning rate much lower than the initial training rate. During training, a portion of the original training data is mixed in to prevent catastrophic forgetting. After fine-tuning, the new model weights seamlessly replace the old ones, and subsequent recognition tasks immediately use the optimized model, thus achieving dynamic evolution of the model's recognition capabilities.

[0056] The aforementioned technical solution significantly enhances the robustness and lifespan of the entire detection system in complex real-world environments, breaking the limitations of traditional static model performance fixation. This allows the system to continuously learn from challenging cases through human-machine collaboration, becoming increasingly intelligent with use. This progressive learning mechanism avoids the heavy burden of large-scale data collection and periodic centralized training, making model updates lightweight, timely, and cost-effective. It effectively integrates expert domain knowledge into the model optimization process, improving the model's accuracy in identifying specific bridges or special components. This ensures the input quality of the subsequent crucial step of aerodynamic sensitivity feature judgment, providing solid technical support for the long-term, automated, and highly reliable operation of bridge vortex-induced vibration detection.

[0057] In another technical solution, the bridge vortex-induced vibration sensitive aerodynamic structure detection method based on three-dimensional laser scanning, step three, specifically the process of calculating its vortex-induced vibration risk parameters based on its geometric parameters, includes: S31. Based on the point cloud data of the parts identified as sensitive aerodynamic features, a closed three-dimensional curved surface mesh model of the parts is generated using the Poisson reconstruction algorithm. S32. Using the 3D curved surface mesh model as the boundary, generate an unstructured mesh for computational fluid dynamics simulation around it; refine the mesh in the area around sensitive parts (distance ≤ 5 times the component size) (mesh size ≤ 1 / 10 of the minimum component size), and use a coarse mesh in non-sensitive areas (mesh size ≤ 1 / 5 of the minimum component size) to reduce computational resource consumption; if the number of sensitive parts is ≥ 5, use parallel computing technology (based on GPU cluster, allocating 1 GPU core to every 2 sensitive parts) to perform large eddy simulation, and control the simulation time of a single sensitive part to within 30 minutes; if the number of sensitive parts is < 5, use single CPU core computing, and control the total time to within 2 hours; S33. Using the measured average wind speed and wind direction angle at the location of the bridge as the simulation inlet boundary conditions, perform large eddy simulation to obtain the flow field data around the bridge. S34. Perform a fast Fourier transform on the velocity vector sequence of specific monitoring points in the wake region of the structure in the simulation results, and extract the peak frequency of the transformed energy spectrum as the vortex-induced vibration risk parameter, i.e., the main vortex shedding frequency.

[0058] In bridge vortex-induced vibration risk assessment, accurately calculating vortex-induced vibration risk parameters based on the identified sensitive aerodynamic features and their geometric shapes is a key challenge. Traditional methods often rely on simplified empirical formulas or steady-state flow field simulations, which struggle to accurately capture the unsteady, separated flow phenomena generated by complex three-dimensional geometries in real wind fields, particularly the formation, development, and shedding processes of vortices. Furthermore, the conversion from discrete point cloud data to computational meshes suitable for computational fluid dynamics simulations is typically cumbersome and requires manual intervention, making it difficult to guarantee the quality of the generated meshes. This directly impacts the accuracy of flow field simulations, leading to unreliable calculations of key risk parameters such as vortex shedding frequency.

[0059] To address the aforementioned issues, this technical solution provides a refined numerical simulation process from point cloud to flow field to risk parameters, starting with high-precision point cloud data of areas identified as sensitive aerodynamic features. First, the Poisson reconstruction algorithm is used to reconstruct surfaces from these spatially discrete points. This algorithm, based on the normal vector information of the point cloud, implicitly fits a smooth and closed three-dimensional surface mesh model by solving a Poisson equation. This generated surface mesh completely describes the physical boundary of the area's contact with water (air), ensuring geometric fidelity. Next, using this closed surface mesh as the boundary, an unstructured mesh for computational fluid dynamics simulation is generated in the surrounding spatial region. The unstructured mesh is composed of tetrahedral, pyramidal, and other polyhedral elements, which can flexibly and densely fill the flow field region around complex geometries, especially in areas with drastic gradient changes such as the wake of structures, providing a foundation for capturing fine vortex structures. Then, the measured average wind speed and prevailing wind direction angle at the bridge location are used as the entry boundary conditions for the simulation calculation, and the large eddy simulation method is employed to perform transient flow field calculations. Large eddy simulation (LES) can directly analyze large-scale eddy structures in the flow field, while simulating the effects of small-scale eddies using sub-grid-scale models. This strategy achieves a good balance between computational cost and accuracy, making it very suitable for simulating eddy shedding phenomena in flow. After the transient simulation reaches statistical steady state, specific monitoring points are set in the wake region of the structure, and the velocity vectors of these points are recorded as a sequence of changes over time. Finally, a fast Fourier transform is performed on this velocity time series to convert the time-domain signal to the frequency domain. The frequency corresponding to the most significant peak is identified from the resulting energy spectrum, and this frequency is extracted as the eddy-induced vibration risk parameter—the main eddy shedding frequency.

[0060] In the aforementioned technical solutions, Poisson reconstruction is an algorithm for generating closed surface meshes from discrete point clouds. It reconstructs the surface by treating the normal vector field of the point cloud as a sample of the gradient field of an indicator function and solving the Poisson equation satisfied by this indicator function. A closed 3D surface mesh refers to a set of triangular or quadrilateral facets without cracks, completely enclosing a volume; it is a prerequisite for flow field calculations. Unstructured meshes are meshes composed of units of various shapes and sizes, with no fixed rules for node connections. Compared to structured meshes, they are better suited to complex geometries. Computational fluid dynamics simulation is a technique that uses numerical methods to solve a system of partial differential equations controlling fluid motion on a computer to simulate flow fields. Large eddy simulation (LES) is a turbulence simulation method that directly simulates the motion of large-scale eddies, while using a model to close eddies smaller than the mesh scale. Inlet boundary conditions are flow parameters set at the inlet of the computational domain, such as velocity magnitude and direction. Flow field data around an obstacle is a collection of calculated physical quantities such as velocity and pressure at various points in space. The Fast Fourier Transform (FFT) is an efficient algorithm used to decompose a time-series signal into a superposition of sinusoidal waves of different frequencies, thereby analyzing its frequency components. The energy spectrum is a graphical representation of the signal power distribution with frequency. The main vortex shedding frequency is the dominant frequency of the vortices periodically shedding in the wake of a fluid flowing around a bluff body; it is the direct cause of vortex-induced vibrations in structures.

[0061] Poisson reconstruction first requires robustly estimating the normal vector of each point from the point cloud, typically achieved through principal component analysis of the local point cloud. Then, an octree structure in three-dimensional space is constructed to discretize the scalar field defined by the Poisson equation. Using the point cloud and its normal vector information as source terms, numerical methods such as the conjugate gradient method are used to solve this large linear system, resulting in a three-dimensional scalar field. Finally, the moving cube algorithm is used to extract isosurfaces from this scalar field, forming a triangular mesh surface. For mesh generation, the reconstructed surface mesh first needs to be repaired and smoothed to eliminate potential defects such as self-intersections and small cracks. Then, the size of the computational domain is defined; this domain should be large enough to eliminate the interference of the far-field boundary on the near-wall flow. A boundary layer mesh is set in the region near the structure surface, and the height of the first mesh layer needs to be carefully set according to the flow Reynolds number to meet the requirements of the turbulence model. The mesh generator automatically fills the spatial region according to the surface boundary, generating a volume mesh dominated by tetrahedrons, and performs local refinement in the wake region. When performing large eddy simulation, appropriate inlet turbulence intensity, outlet boundary conditions, and wall functions need to be set. The simulation was performed on a high-performance computing cluster, requiring a sufficient number of physical time steps until the statistics at the monitoring points stabilized. Before performing a Fast Fourier Transform on the velocity signals at the monitoring points, potential linear trends needed to be removed, and an appropriate window function was applied to reduce spectral leakage. Finally, the frequency corresponding to the global maximum value was automatically found from the energy spectrum as the main eddy current shedding frequency.

[0062] The aforementioned technical solution significantly improves the scientific rigor and accuracy of vortex-induced vibration risk parameter calculation. By directly generating conformal computational boundaries from the point cloud through Poisson reconstruction, it avoids the geometric distortions that may be introduced by traditional manual modeling, ensuring consistency between the numerical model and the real structure. The strategy of combining unstructured meshes with large eddy simulation (LES) enables high-resolution capture of unsteady, three-dimensional vortex structures generated by complex geometric flow, thus more realistically reflecting the physical process of vortex shedding. Spectral analysis based on Fast Fourier Transform (FFT) provides an objective and quantitative basis for extracting vortex shedding frequencies. The entire process is highly automated, reducing uncertainties caused by human intervention and providing reliable and accurate input data for subsequent comparison of risk parameters with structural dynamic characteristics and final assessment of vortex-induced vibration risk levels, making the risk assessment conclusions more credible.

[0063] In another technical solution, the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, in step two, determining whether the part is a sensitive aerodynamic feature that easily induces vortex-induced vibration based on the identified component type and corresponding geometric features specifically includes: A sensitive aerodynamic feature rule base is constructed, defining the influence patterns of different component types on airflow under different geometric parameter thresholds. Influence patterns include the blunting of the guide angle of streamlined box girders, the protruding features of maintenance vehicle tracks or crash barriers, and the depth-to-width ratio of cavities in open or grid structures. The rule base is stored according to bridge type: 1. Cable-stayed / suspension bridges: A guide angle curvature radius > 1.5 times the design value for streamlined box girders is a sensitive feature; a protruding height of maintenance vehicle tracks > 0.25m and a width-to-height ratio < 0.6 are sensitive features; 2. Beam bridges: A depth-to-width ratio of the hollow cavity in rectangular box girders > 2.0 is a sensitive feature; a protruding height of crash barriers > 0.3m and a width-to-height ratio < 0.5 are sensitive features; 3. Arch bridges: An edge sharpness of the arch rib section > 60° (sharpness = 90° - edge angle) is a sensitive feature; a protruding height of the hanger anchor point > 0.15m is a sensitive feature. The rule base can be dynamically updated by importing wind tunnel test data for new bridge types. Based on the shape comparison results, the identified component types and their quantified geometric feature parameters are input into the rule base for matching. If the matching result matches any preset geometric feature pattern that may cause periodic vortex shedding, the part is determined to be a sensitive aerodynamic feature; otherwise, the part is determined to be a non-sensitive aerodynamic feature.

[0064] A significant technical challenge in assessing the aerodynamic sensitivity of bridges is the lack of objective and systematic criteria. Traditional methods rely primarily on subjective judgments by experts based on wind tunnel test databases and personal experience. This approach is prone to inconsistencies and poor repeatability when dealing with complex bridge structures, especially those components that deviate from standard designs or undergo unexpected deformation. Different experts may hold differing opinions on whether the same geometric feature constitutes a risk, and it is difficult to clearly quantify the degree of guide angle blunting or the size of protrusions sufficient to trigger significant periodic vortex shedding, leading to considerable uncertainty and arbitrariness in the assessment conclusions.

[0065] Based on the technical shortcomings analyzed above, this technical solution aims to achieve the objectification and standardized identification of sensitive aerodynamic features by constructing a structured rule knowledge base. This approach first requires the construction of a sensitive aerodynamic feature rule base. This rule base is not a simple set of cases, but a knowledge system that defines the logical relationship between different component types and their key geometric parameter thresholds. It explicitly stipulates that when one or more geometric parameters of a certain type of component meet specific conditions, it is considered to conform to a certain geometric feature pattern that may affect airflow patterns and lead to periodic vortex shedding. For example, the rule base defines a "guide angle blunting" judgment rule for components such as "streamlined box girders," which may be expressed as "when the measured value of the radius of curvature of the box girder edge is greater than a certain multiple of its design value, it is judged as guide angle blunting"; it defines a "protruding body feature" judgment rule for auxiliary components such as "maintenance vehicle tracks" or "crash barriers," which may be expressed as "when the ratio of the protruding height to the width of the component relative to the surface of the main beam exceeds a certain preset threshold, it is judged as a protruding body feature with significant aerodynamic effects"; and it defines a "hollow cavity aerodynamic risk" judgment rule for components with openings or grilles, which may be expressed as "when the ratio of the depth to the width of the opening or cavity falls within a certain numerical range, it is judged that its depth-to-width ratio may cause vortex-induced vibration."

[0066] In practice, once the system identifies the component type of a bridge section using a machine learning model and quantifies its specific geometric parameters through a geometric feature extraction process, this information—namely, "what kind of component it is" and its "specific geometric parameter values"—is input into a rule base for automatic matching and retrieval. The system compares the extracted actual geometric parameters one by one with the parameter thresholds of various influence modes set for that type of component in the rule base. If the actual parameters meet any of the preset geometric feature modes in the rule base that could lead to periodic vortex shedding, the system automatically determines that the section is a sensitive aerodynamic feature; if none of the rules are triggered, the section is determined to be a non-sensitive aerodynamic feature. This judgment process is entirely driven by preset rules and quantified data, eliminating subjective assumptions.

[0067] In the aforementioned technical solution, the sensitive aerodynamic feature rule base is a structured database or knowledge base that stores the correspondence between various bridge component types and geometric conditions that may induce vortex-induced vibration. Component type refers to the classification of bridge components, such as main beams, railings, and maintenance vehicle tracks. Geometric parameter thresholds are critical values ​​or ranges set for a specific geometric feature of a particular component type; when measured parameters exceed or fall within this range, the corresponding aerodynamic sensitivity condition is considered triggered. Influence modes describe the physical mechanisms by which specific geometric features alter the surrounding airflow structure and may induce vortex shedding. For example, blunting of the guide angle disrupts streamlines, promoting flow separation; protruding features generate wake vortex streets behind them; and specific aspect ratios of cavities may lead to fluid resonance within the cavity. Shape comparison refers to the process of comparing the measured geometric shape of a component with an ideal model in a standard database to identify significant differences. Periodic vortex shedding refers to the phenomenon where, as fluid flows around an object, vortices are alternately and regularly generated and shed on both sides of its tail.

[0068] The first and crucial step in constructing the rule base for the aforementioned technical solution is to systematically organize and summarize historical wind tunnel test data, computational fluid dynamics simulation results, and proven bridge wind-induced vibration case studies. Based on these authoritative data sources, aerodynamic sensitivity thresholds for key geometric parameters of various components are determined through data analysis. For example, for streamlined box girders, by analyzing wind tunnel test results from numerous different guide angle models, the critical radius of curvature at which aerodynamic performance begins to significantly deteriorate is statistically determined. The rule base can be represented in the software as a relational data table or a set of XML configuration files. Each rule explicitly records the component type, the relevant geometric parameter name, the judgment conditions (e.g., greater than, less than, within a range), and the corresponding threshold. During actual matching and judgment, the system loads all applicable judgment rules from the rule base based on the input component type. Then, like executing an SQL query or conditional statement, the extracted actual geometric parameter values ​​are substituted into the conditions of these rules for logical judgment. This process is automatic and fast. In order to achieve a comprehensive judgment based on the shape comparison results, some threshold settings in the rule base may take into account the allowable deviation from the standard model, or introduce the difference obtained from the shape comparison as a weighting factor or auxiliary judgment condition in the judgment logic.

[0069] The aforementioned technical solution establishes an objective benchmark for judging sensitive aerodynamic characteristics, significantly improving the consistency and repeatability of assessment results. It transforms implicit expert knowledge into explicit, actionable judgment rules, ensuring consistent conclusions when different operators or at different times assess the same bridge, provided the input data is identical. This reduces interpretation fluctuations caused by differences in individual experience. The quantitative standards of the rule base clearly define the geometric parameters and their critical states that play a key role in aerodynamic sensitivity, making the judgment process evidence-based and the conclusions more convincing. This standardized automated judgment process also greatly improves analysis efficiency, enabling rapid screening of numerous components on the bridge and quickly locating genuine potential risk points. This saves valuable time and resources for subsequent targeted and detailed assessments, making the entire inspection process more rigorous and efficient.

[0070] In another technical solution, the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning, in step three, compares the risk parameter with the inherent dynamic characteristics of the bridge structure to determine its vortex-induced vibration risk level, specifically including: Obtain the natural frequencies f of the bridge structure in multiple modes. i and corresponding mode shape; The calculated main vortex shedding frequency f v With each order of natural frequency f i Compare and calculate the normalized frequency ratio η. i =f v / f i ; Define a risk level determination matrix, which uses frequency ratio η i The degree of proximity (the locking range is set at 0.9-1.1, based on the relevant provisions of the vortex-induced vibration frequency locking in the "Code for Wind Resistance Design of Highway Bridges" (JTG / T3360-01-2018), and adjusted in combination with the characteristics of different bridge types; the locking range for suspension bridges and cable-stayed bridges remains at 0.9-1.1, while that for beam bridges can be relaxed to 0.85-1.15) and the energy ratio of the corresponding vibration mode in the sensitive aerodynamic feature parts are the core indicators; When a certain mode exists, its frequency is higher than η i When the location falls within the locked range and exhibits significant displacement within that mode shape, it is classified as a high-risk level; when the frequency ratio is η... i If the risk level is close to but not completely within the locked-in area, it is classified as medium risk; otherwise, it is classified as low risk. Frequency ratio η i The locking range is generally 0.9 to 1.1; for flexible bridges or specific cross-sectional shapes, it can be adjusted to 0.85–1.15 based on wind tunnel test data.

[0071] A common challenge in bridge vortex-induced vibration risk assessment is the one-sidedness of the assessment. Traditional methods typically focus on comparing the proximity of vortex shedding frequencies to the structure's natural frequencies, a phenomenon known as "frequency locking." However, this method has significant limitations. Even if the frequency ratio falls within the locking range, if the energy proportion of that vibration mode in sensitive aerodynamic features is low—meaning the vibration amplitude at that location is small—the actual probability of large-scale vortex-induced vibration remains low. Conversely, in cases where frequency ratios are close but not completely locked, if that location participates significantly in the corresponding vibration mode, there may still be a considerable risk. Judging solely based on a single frequency parameter can easily lead to underestimating high-risk situations or over-warning low-risk situations, failing to accurately reflect the true vortex-induced vibration risk.

[0072] Based on the technical shortcomings analyzed above, this technical solution provides a multi-dimensional method for determining the risk level of vortex-induced vibration. This method first requires obtaining the natural frequencies and corresponding mode shapes of the bridge structure under multiple modes. These dynamic characteristic parameters can be obtained through finite element model calculations or on-site environmental vibration tests. Simultaneously, the main vortex shedding frequency calculated for specific sensitive aerodynamic features is obtained from the aforementioned computational fluid dynamics simulation. Next, the ratio of the main vortex shedding frequency to each natural frequency is calculated, i.e., the normalized frequency ratio. The innovation of this invention lies in defining and applying a risk level determination matrix. This matrix no longer relies solely on the frequency ratio dimension, but simultaneously considers two key factors: first, the closeness of the frequency ratio to the theoretical or empirical locking range; and second, the energy proportion or displacement significance of the sensitive aerodynamic feature in the corresponding mode shape, reflecting the active participation of this part in that mode of vibration.

[0073] Based on this two-factor judgment matrix, the risk level classification is more refined and reasonable. When a certain mode exists whose frequency ratio clearly falls within the vortex-induced vibration locking range, and the identified sensitive component exhibits a significant displacement response in this mode, meaning that the component is a key participant in this mode, the system classifies it as a high-risk level. When the frequency ratio is close to the locking range but not completely within it, and the component still has a considerable participation in the corresponding mode, it is classified as a medium-risk level, indicating a need for attention. If the frequency ratio is far from the locking range, or even if the frequency ratio is close but the component hardly participates in vibration in the relevant mode, it is classified as a low-risk level.

[0074] In the above technical solutions, the natural frequency is the inherent free vibration frequency of the structure, determined by its mass and stiffness distribution. The mode shape is the structural deformation pattern corresponding to the natural frequency, describing the relative displacement relationship of each point when the structure vibrates at a specific frequency. The main vortex shedding frequency is the most significant frequency at which periodic vortices are generated and shed in the wake of the airflow passing over a blunt body. The frequency ratio is the ratio of the vortex shedding frequency to the structure's natural frequency, and is a key parameter for determining whether frequency lock-up has occurred. The risk level assessment matrix is ​​a logical judgment framework that comprehensively assesses the risk level based on both frequency matching degree and mode shape participation. The lock-up interval refers to a specific range of the frequency ratio, typically around 1.0, within which strong vortex-induced resonance is prone to occur. Energy proportion or displacement significance are indicators that quantify the intensity of vibration in a sensitive part within a certain mode shape, and can be assessed by the magnitude of the displacement value of the mode shape vector on the coordinates of that part or the proportion of modal kinetic energy contribution.

[0075] The specific implementation of the above technical solution is as follows: First, complete structural dynamic characteristic data needs to be prepared, including the natural frequencies and mode shape vectors of the first few orders. Mode shape data usually needs to be normalized. The main vortex shedding frequency comes from previous flow field simulation results. After calculating the frequency ratio, the specific range of the locking interval needs to be determined. This range can be determined according to the bridge type and code recommendations, for example, it may be set to 0.9 to 1.1. For the assessment of mode shape participation, the absolute displacement value of the node where the sensitive aerodynamic feature is located in the normalized mode shape vector can be calculated, or the ratio of the modal kinetic energy of the local structure at that order to the total modal kinetic energy can be calculated. The system will traverse each order of modes, calculating its frequency ratio and the mode shape participation index of that part for each order. Then, these two values ​​are input into the preset judgment logic. The judgment logic can be a series of nested conditional judgment statements. For example, a high-risk condition might be defined as: the frequency ratio is between 0.95 and 1.05, and the mode shape displacement of that part is greater than a certain percentage threshold of the maximum mode shape displacement. The criteria for medium risk might be defined as a frequency ratio between 0.8 and 0.95 or 1.05 and 1.2, and a modal displacement exceeding a lower threshold. Locations not meeting these criteria are classified as low risk. This process is automated, outputting a risk level based on the dynamic response for each identified sensitive location.

[0076] The aforementioned technical solution significantly improves the accuracy and engineering guidance value of vortex-induced vibration risk level assessment. By introducing the key dimension of modal participation, it closely integrates the analysis of simple "airflow excitation characteristics" with "structural dynamic response characteristics," ensuring that risk assessment is based on the physical foundation of the actual vibration behavior of the structure. This comprehensive approach effectively avoids misjudgments that may arise from frequency comparison alone, enabling more accurate identification of truly dangerous conditions that could lead to significant vibrations. It also provides a reasonable low-risk criterion for conditions that appear to have similar frequencies but have limited actual risk. This allows maintenance management departments to make more targeted decisions, prioritizing limited resources on addressing truly high-risk areas, thereby improving operational efficiency while ensuring safety.

[0077] In another technical solution, the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning further includes: Step 4: Collect actual vibration response data of the bridge's sensitive aerodynamic features using vibration sensors deployed in a natural wind field; perform spectral analysis on the vibration response data to extract the actual principal vibration frequency; the vibration sensors are piezoelectric accelerometers (range ±5g, sensitivity 100mV / g), and the number deployed is determined based on the component length: one sensor is deployed for components ≤10m in length (located at the midpoint of the component), and one sensor is deployed every 10m for components >10m in length; the installation location is preferably selected on flat surfaces of the component (such as the top surface of the crash barrier, the side of the box girder); if it is a high-altitude location (height >10m), use a drone to mount installation tools for assistance, with an installation error ≤2mm; the sensor data is uploaded to the data acquisition terminal in real time via a wireless transmission module (transmission distance ≥500m); The actual principal vibration frequency is compared with the calculated principal eddy shedding frequency. If the deviation exceeds a preset tolerance (typically 3%-5%, based on a comprehensive setting of ±2% vibration sensor accuracy and ±3% simulation model error; for critical components such as main beams and cables, a tolerance of 3% is recommended; for auxiliary components such as guardrails, a tolerance of 5%), the turbulence model parameters or boundary condition settings in the computational fluid dynamics simulation are automatically calibrated, and the risk parameters are recalculated to optimize the accuracy of subsequent assessments. A parameter optimization method based on sensitivity analysis is adopted: using the measured principal vibration frequency as the target, a response surface model of CFD parameters (such as turbulence intensity, length scale, and subgrid model constants) is constructed, and the parameters are iteratively adjusted using a sequential quadratic programming algorithm until the deviation between the simulation and measured frequencies is ≤5%.

[0078] In the numerical simulation and evaluation of bridge vortex-induced vibration, an inherent challenge is the discrepancy between the computational fluid dynamics simulation results and the actual behavior of the structure in real natural wind fields. This discrepancy may stem from the selection of the turbulence model in the simulation, the failure of the inlet boundary conditions to fully reflect the complex characteristics of the actual wind field, and the simplification of the computational grid and physical model itself. The traditional approach relies on engineers' experience to manually adjust simulation parameters. This process is not only inefficient but also highly subjective, lacking objective calibration criteria. Consequently, the long-term accuracy and reliability of risk parameters based on simulation predictions are difficult to guarantee, affecting the credibility of subsequent risk assessment conclusions.

[0079] Based on the technical problems analyzed above, this technical solution provides a simulation model self-calibration method based on measured data feedback. This method, after calculating the main vortex shedding frequency of sensitive aerodynamic feature parts through numerical simulation and completing the initial risk assessment, does not terminate the process but adds a physical verification and optimization step. Vibration sensors, such as accelerometers, are deployed on the bridge sections identified as sensitive aerodynamic features. These sensors continuously collect vibration response data of these sections under actual wind loads under natural wind field conditions. By performing spectral analysis on the collected vibration signals, the frequency with the most concentrated vibration energy in the real environment can be extracted, i.e., the actual main vibration frequency. Subsequently, the system automatically compares this measured main vibration frequency with the main vortex shedding frequency predicted by the preceding computational fluid dynamics simulation. The system presets a frequency deviation tolerance value. If the comparison finds that the deviation between the two frequencies exceeds this tolerance, it indicates that the current simulation model settings differ significantly from the actual situation. At this point, the system will not wait for manual intervention but will automatically initiate a calibration process. This process adaptively adjusts key parameters in the computational fluid dynamics simulation, such as coefficients of the turbulence model or turbulence intensity and integral scale in the inlet boundary conditions. It then re-executes the flow field simulation using the adjusted parameters to calculate the new main vortex shedding frequency. This process can be viewed as a "calibration" of the virtual model using measured data, ensuring that the next prediction is closer to reality.

[0080] In the above technical solutions, vibration sensors refer to physical devices used to measure structural vibration acceleration, velocity, or displacement, such as piezoelectric accelerometers. Actual vibration response data is a sequence of physical quantities generated by the structure under real environmental excitation, recorded by the sensor, that change over time. Spectrum analysis is a signal processing technique that converts time-domain vibration signals to the frequency domain to analyze their frequency components; commonly used methods are based on Fast Fourier Transform. The actual dominant vibration frequency is the frequency component with the largest amplitude identified from the measured vibration signal spectrum, reflecting the main response frequency of the structure in a real wind field. The dominant vortex shedding frequency is the frequency at which vortices shed when airflow passes over the structure, predicted through computational fluid dynamics simulation. Deviation tolerance is a pre-set threshold range that allows for differences between simulated and measured values. Automatic calibration refers to the process where the system automatically adjusts simulation parameters and recalculates according to preset logic based on deviation comparison results. Turbulence models are mathematical equations used in computational fluid dynamics to simulate turbulence effects, such as the k-epsilon model or the subgrid-scale model in large eddy simulation. Boundary condition settings refer to the parameters used in computational fluid dynamics simulations to define the flow state at the boundary of the computational domain, such as the inlet wind speed profile and turbulence characteristics.

[0081] The specific implementation of the above technical solution is as follows: First, vibration sensors need to be installed at the locations identified as sensitive aerodynamic features. Vibration time-history signals are continuously or triggered by a data acquisition system. After preprocessing such as amplification and filtering, the signals undergo a fast Fourier transform to obtain a power spectral density map. The algorithm automatically searches for significant peaks in the power spectrum and uses their frequencies as the actual principal vibration frequencies. The system maintains a list of key parameters, which are pre-identified simulation settings sensitive to the calculated vortex shedding frequency and adjustable within a reasonable range. When the system detects that the frequency deviation exceeds the limit, it initiates an optimization algorithm. One feasible strategy is to construct a simple surrogate model or use a gradient-based optimization method, using the measured principal vibration frequency as the target, to solve in reverse for the parameter combination that makes the simulated frequency approach the target value. For example, the system might slightly increase the inlet turbulence intensity or adjust the constants of the subgrid model in the large eddy simulation, and then submit a new simulation calculation task. After recalculation, the new principal vortex shedding frequency is compared with the measured value again. If the tolerance requirement is still not met, the iteration process may continue for a limited number of times, or the optimal improvement result may be recorded. The goal of the entire calibration process is to narrow the gap between simulation and reality, rather than to perform an infinite number of precise matches.

[0082] The aforementioned technical solution establishes a closed-loop evaluation and optimization system, significantly improving the long-term prediction accuracy of the numerical simulation model and the overall robustness of the method. By introducing objective measured data as a calibration benchmark, it reduces reliance on manual parameter tuning, enabling the simulation model to learn and adapt to the specific bridge's actual environment. This data-driven feedback mechanism effectively corrects systematic errors caused by model simplification or improper initial parameter settings, making risk assessments based on the updated model closer to reality. It provides a dynamic and evolutionary analytical tool for long-term bridge condition monitoring and safety early warning, enhancing the reliability and practical value of this detection method in engineering practice.

[0083] The present invention also provides a system for implementing the above-described method for detecting bridge vortex-sensitive aerodynamic structures based on three-dimensional laser scanning, comprising: The data acquisition module is used to acquire three-dimensional point cloud data of the bridge structure, preprocess the three-dimensional point cloud data and extract the geometric features of the bridge. The identification module is used to identify the component type of each part of the bridge based on the extracted geometric features and a pre-trained machine learning model. It also compares the geometric features with the model in the standard bridge type database and, based on the identified component type and corresponding geometric features, determines whether the part is a sensitive aerodynamic feature that is prone to vortex-induced vibration. The assessment module is used to calculate the vortex-induced vibration risk parameters of parts identified as having sensitive aerodynamic characteristics based on their geometric parameters, and compare these risk parameters with the inherent dynamic characteristics of the bridge structure to determine its vortex-induced vibration risk level; for parts identified as having non-sensitive aerodynamic characteristics, the vortex-induced vibration risk assessment process for that part is terminated.

[0084] In the above technical solution, the present invention provides a bridge vortex-vibration sensitive aerodynamic structure detection system. This system is designed as an integrated software platform, the core of which consists of three collaborative functional modules: a data acquisition module, an identification module, and an evaluation module. These modules are connected sequentially to form a complete automated processing flow.

[0085] The data acquisition module is responsible for digitizing the bridge's 3D information. This module controls a drone and a ground-based vehicle platform to work together to acquire the bridge's raw 3D point cloud data. The identification module's core task is to determine aerodynamic sensitivity based on geometric features. This module receives geometric feature data from the data acquisition module and calls a pre-trained machine learning model. This model can automatically identify the component type of each part of the bridge based on point cloud features, such as determining whether a certain part is a main beam, a crash barrier, or a maintenance vehicle track. The assessment module is specifically responsible for performing quantitative risk analysis on sensitive areas. For areas identified as having sensitive aerodynamic features by the identification module, this module initiates a sophisticated numerical simulation process. Through the close cooperation of these three modules, the entire system achieves fully automated processing from raw data to risk assessment conclusions.

[0086] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for detecting bridge vortex-sensitive aerodynamic structures based on three-dimensional laser scanning.

[0087] This invention implements its automated detection process using a computer-readable storage medium. In practical applications, this medium can be a solid-state drive (SSD) that stores a computer program capable of implementing the detection method. This program contains a series of processor-executable instructions organized in a specific logical order to form a complete processing flow.

[0088] Example: Detection of vortex-sensitive aerodynamic structures of a long-span cable-stayed bridge based on 3D laser scanning Background: A long-span cable-stayed bridge, after years of operation, recently exhibited slight vibrations under specific wind conditions, suspected to be vortex-induced vibration. Traditional manual inspections struggle to accurately capture microscopic geometric features, and wind tunnel testing is costly and time-consuming. This invention employs a rapid and objective detection method for vortex-induced vibration-sensitive aerodynamic structures. In this embodiment, the bridge is a double-tower steel box girder cable-stayed bridge with a main span of 480 meters, a girder height of 3.5 meters, and a bridge width of 35 meters. The finite element model of the structure yielded a fundamental frequency of 1.0 Hz, consistent with the measured frequency.

[0089] Step 100: Data Acquisition and Preprocessing S101. Multi-platform collaborative scanning: A drone equipped with a lidar system is used to perform overhead and side-view scans of the bridge deck and superstructure (main beams, cables). Simultaneously, a ground-based vehicle platform is used to perform overhead and multi-view scans of the piers and substructure. GPS, IMU, and odometer data are acquired concurrently during the scanning process. Collaborative operations must meet the following environmental conditions: wind speed ≤ 5 m / s (to avoid drone instability), visibility ≥ 100 m (to ensure lidar scanning accuracy), and no strong electromagnetic interference (to avoid GPS / IMU signal abnormalities). If there are high-voltage power lines under the bridge (≤ 50 m from the scanning path) or strong winds on the river (wind speed > 5 m / s), the drone must use a low-altitude detour (≥ 10 m from the high-voltage lines), and the ground-based vehicle platform must suspend operations until the wind speed drops to a safe level before resuming. Within narrow bridge openings (width < 5 m), the ground-based vehicle platform is replaced with a portable laser scanning device. S102. Semantic Feature-Assisted SLAM Optimization: Identify semantic features in the point cloud (such as cable anchor points and support centers), calculate the difference between the measured coordinates and the theoretical coordinates of the preset BIM model, and use the coordinate residuals as constraint factors. Integrate these residuals into the graph-based SLAM algorithm, and optimize them together with GPS, IMU and other data to generate a high-precision 3D point cloud model.

[0090] S103. Point Cloud Preprocessing and Feature Extraction: Noise filtering and clutter removal preprocessing are performed on the point cloud. The bridge deck centerline is aligned with the global coordinate system X-axis through principal axis analysis, generating transverse point cloud slices at 0.5-meter intervals along the bridge's longitudinal direction. For the point cloud slices in the bridge deck area, abrupt changes in point cloud density and abrupt changes in normal vectors are identified to extract the hollow structure contour of the box girder. For the point cloud slices in the pier area, the RANSAC plane fitting algorithm is used to extract the point cloud of the pier cap area. Based on the spatial coordinates and topological relationships of the point cloud, geometric features are extracted, including the edge curvature of the box girder section, the protruding height and width of the crash barrier, and the aperture size of the hollow structure.

[0091] Step 200: Sensitive aerodynamic feature identification S201. Machine Learning Recognition and Shape Comparison: A pre-trained machine learning model (based on the PointNet++ architecture) is used to identify component types (such as main beams, crash barriers, and maintenance tracks). The extracted geometric features are compared with models in a standard bridge type database to calculate geometric errors and shape similarity.

[0092] S202, Online Incremental Learning Mechanism: If the model's confidence level in identifying a certain maintenance track is below a threshold (0.7), the point cloud slice and its contextual feature vector are automatically stored in the verification sample pool. Through the user interface, operators verify that the component is a "maintenance track" and label it with the correct type. The verified samples are used to fine-tune the model classifier parameters and update the model weights.

[0093] S203. Rule Base Matching and Judgment: Construct a sensitive aerodynamic feature rule base, defining rules such as: "Streamlined box girder guide angle curvature radius > 1.5 times the design value" as a sensitive feature; "Bumper guardrail protrusion height > 0.3 meters and width-to-height ratio < 0.5" as a sensitive feature. Identified bumper guardrail geometric parameters (protrusion height 0.35 meters, width-to-height ratio 0.4) are input into the rule base. If a "protrusion feature" pattern is matched, the part is determined to be a sensitive aerodynamic feature. When multiple rules are triggered simultaneously, they are sorted by risk severity level (vortex-induced vibration risk > flutter risk > buffeting risk); if rules of the same level conflict, the rule with the largest deviation of geometric parameters from the standard value is used as the judgment basis. The rule base is constructed based on the aerodynamic shape requirements in the "Specifications for Wind-Resistant Design of Highway Bridges" (JTG / T3360-01-2018), publicly available wind tunnel test databases (such as the TN series benchmark models), and historical CFD cases from this project team.

[0094] Step 300: Risk Assessment and Calibration S301. Calculation of Vortex-Induced Vibration Risk Parameters: Based on point cloud data of sensitive areas (crash barrier), a closed three-dimensional curved surface mesh model is generated using the Poisson reconstruction algorithm. Using this mesh as the boundary, an unstructured mesh for CFD simulation is generated and locally refined. The measured average wind speed (12 m / s) and wind direction angle (90°) at the bridge are used as inlet boundary conditions, and large eddy simulation is performed to obtain the flow field data around the bridge. A fast Fourier transform is performed on the velocity sequence of monitoring points in the wake region, and the peak frequency of the energy spectrum is extracted as the main vortex shedding frequency (f). v =1.2Hz).

[0095] S302. Risk Level Determination: Obtain the inherent dynamic characteristics of the bridge: the first three natural frequencies (f1=1.0Hz, f2=1.5Hz, f3=2.0Hz) and their corresponding mode shapes.

[0096] Calculate the frequency ratio: η1 = f v / f1=1.2, η2=0.8, η3=0.6.

[0097] A risk level determination matrix was defined, considering both frequency ratio proximity (locked-in range 0.9-1.1) and the energy proportion of the mode shape in sensitive areas (significant displacement threshold >5%). The significance of mode shape displacement is calculated by the ratio of the normalized mode shape vector's nodal displacement value at that location to the global maximum displacement value, with a threshold of 5%—meaning that a mode shape's displacement amplitude at that location is considered significant when it accounts for ≥5% of the total displacement amplitude). Results: η1 falls within the locked-in range, and the guardrail's displacement proportion in the first mode shape is 8%, classifying it as a high-risk level.

[0098] S303. Actual Vibration Data Calibration: Vibration sensors are installed on the crash barrier to collect actual vibration response data in the natural wind field. Spectral analysis is used to extract the actual principal vibration frequency (f).actual =1.15Hz). This is related to the simulated main eddy current shedding frequency (f). v =1.2 Hz) compared to the previous value, the deviation exceeded the tolerance (5%), so the turbulence model parameters in the CFD simulation were automatically calibrated (k-epsilon model constants were adjusted), and f was recalculated. v =1.14Hz, optimizing the accuracy of subsequent evaluations.

[0099] S304. The system generates an inspection report: the guardrail is marked as a high-risk area, and it is recommended to install a deflector or optimize its shape.

[0100] The entire process, from data acquisition (including collaborative scanning by UAVs and ground vehicle platforms), data preprocessing (denoising, stitching, and slicing), feature extraction, sensitive identification, to risk assessment (including simulation calculation), is fully automated, taking only 8 hours in total (3 hours for data acquisition, 2 hours for data preprocessing, 1 hour for feature extraction and sensitive identification, and 2 hours for risk assessment). The traditional method (2 days of manual inspection + 2 weeks of wind tunnel testing) takes about 16 days in total, and this invention improves efficiency by about 48 times.

[0101] The results of this example demonstrate the high efficiency, accuracy, and practicality of this invention in bridge vortex-induced vibration detection, highlighting its innovation in data acquisition, processing, identification, and evaluation, and providing reliable technical support for bridge safety monitoring.

[0102] This invention replaces manual subjective judgment with 3D laser scanning, machine learning recognition, and rule-based matching, solving the problems of low efficiency and easy omissions in traditional detection methods. Multi-platform collaborative scanning and semantic feature-assisted SLAM optimization ensure point cloud accuracy and overcome blind spots in single-platform scanning and cumulative sensor errors. An online incremental learning mechanism continuously optimizes the model, adapting to complex bridge components and improving long-term detection reliability. CFD simulation and Fast Fourier Transform directly extract risk parameters from unknown point clouds, combined with a multi-dimensional judgment matrix, to provide accurate risk levels. Vibration sensor feedback calibration of the simulation model achieves a combination of theory and practice, enhancing the credibility of the results.

[0103] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0104] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A method for detecting bridge vortex-sensitive aerodynamic structures based on three-dimensional laser scanning, characterized in that, Includes the following steps: Step 1: Obtain 3D point cloud data of the bridge structure, preprocess the 3D point cloud data and extract the geometric features of the bridge; Step 2: Based on the extracted geometric features, a pre-trained machine learning model is used to identify the component types of each part of the bridge, and the geometric features are compared with the models in the standard bridge type database. Based on the identified component types and corresponding geometric features, it is determined whether the part is a sensitive aerodynamic feature that is prone to vortex-induced vibration. Step 3: For parts identified as having sensitive aerodynamic characteristics, calculate their vortex-induced vibration risk parameters based on their geometric parameters, and compare these risk parameters with the inherent dynamic characteristics of the bridge structure to determine their vortex-induced vibration risk level. For parts that are determined to be non-sensitive aerodynamic features, the vortex-induced vibration risk assessment process for those parts will be terminated.

2. The bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning as described in claim 1, characterized in that, In step one, the specific process of acquiring the three-dimensional point cloud data of the bridge structure is as follows: a UAV and a ground vehicle platform work together. The UAV is equipped with a lidar to perform top-down and side-view scans of the bridge deck and superstructure, while the ground vehicle platform performs bottom-up and multi-view scans of the bridge piers and substructure. During the scanning process, GPS, IMU, and odometer data are acquired simultaneously, and multi-source data fusion and real-time point cloud stitching are performed through graph-optimized synchronous positioning and mapping algorithms to generate a three-dimensional point cloud model. The synchronous positioning and mapping process uses bridge semantic features as constraint factors for pose graph optimization. Specifically, it identifies the bridge structural semantic features in the point cloud data, calculates the difference between the measured 3D coordinates of the semantic features and the theoretical coordinates in the preset bridge BIM model to obtain coordinate residuals, adds the coordinate residuals as constraint edges to the pose graph, and optimizes them together with constraint factors from GPS, IMU, and odometer to obtain 3D point cloud data of the bridge structure. Semantic features include cable anchor points, support centers, expansion joint outlines, or specific segments of crash barriers.

3. The bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning as described in claim 2, characterized in that, Step one involves extracting the geometric features of the bridge, including the following steps: S11. Perform noise filtering and descratizing preprocessing on the three-dimensional point cloud data of the bridge structure to obtain the preprocessed three-dimensional point cloud data. S12. Perform principal axis analysis on the preprocessed 3D point cloud data and align the bridge deck centerline with the X-axis or Y-axis of the global coordinate system. Then, slice the point cloud along the longitudinal direction of the bridge at preset equal intervals to generate a series of transverse point cloud slices. S13. For each horizontal point cloud slice, if the point cloud slice is a bridge deck area point cloud, identify the sudden change in point cloud density and the sudden change in normal vector, and extract its horizontal hollow structure outline; if the point cloud slice is a pier area point cloud, extract the point cloud of its top plane through the random sampling consistency plane fitting algorithm, and use it as the pier cap area point cloud. S14. Based on the point cloud of the bridge deck area, the outline of the hollow structure, and the point cloud of the pier cap area, the geometric features of the bridge are extracted by calculating the spatial coordinates and topological relationships of the point cloud. These features include the curvature and sharpness of the box girder cross-section edge, the protrusion height and width of the auxiliary components relative to the surface of the main beam, and the aperture size and boundary shape of the hollow structure outline.

4. The bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning as described in claim 1, characterized in that, In step two, the pre-trained machine learning model employs an online incremental learning mechanism; The execution steps of the online incremental learning mechanism include: S21. When the confidence level of the model in recognizing a certain component in the bridge point cloud is lower than the preset threshold, the point cloud data slice corresponding to the component and its context feature vector are automatically stored in a sample pool to be verified. S22. Receive the operator's annotation and approval of the component types in the sample pool to be verified through the user interface; S23. Using the approved sample data and annotations, perform a round of parameter fine-tuning on the classifier of the machine learning model and update the model weights.

5. The bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning as described in claim 1, characterized in that, Step three, the specific process of calculating its vortex-induced vibration risk parameters based on its geometric parameters, includes: S31. Based on the point cloud data of the parts identified as sensitive aerodynamic features, a closed three-dimensional curved surface mesh model of the parts is generated using the Poisson reconstruction algorithm. S32. Using the three-dimensional curved surface mesh model as the boundary, generate an unstructured mesh for computational fluid dynamics simulation around it; S33. Using the measured average wind speed and wind direction angle at the location of the bridge as the simulation inlet boundary conditions, perform large eddy simulation to obtain the flow field data around the bridge. S34. Perform a fast Fourier transform on the velocity vector sequence of specific monitoring points in the wake region of the structure in the simulation results, and extract the peak frequency of the transformed energy spectrum as the vortex-induced vibration risk parameter, i.e., the main vortex shedding frequency.

6. The bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning as described in claim 1, characterized in that, In step two, based on the identified component type and corresponding geometric features, determining whether the location is a sensitive aerodynamic feature prone to vortex-induced vibration specifically includes: A sensitive aerodynamic feature rule base is constructed, which defines the influence patterns of different component types on airflow under different geometric parameter thresholds. The influence patterns include the blunting of the guide angle of streamlined box girder, the protruding features of maintenance vehicle track or crash barrier, and the depth-to-width ratio of the cavity in the opening or grid structure. Based on the shape comparison results, the identified component types and their quantified geometric feature parameters are input into the rule base for matching. If the matching result matches any preset geometric feature pattern that may cause periodic vortex shedding, the part is determined to be a sensitive aerodynamic feature; otherwise, the part is determined to be a non-sensitive aerodynamic feature.

7. The bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning as described in claim 1, characterized in that, Step three involves comparing this risk parameter with the inherent dynamic characteristics of the bridge structure to determine its vortex-induced vibration risk level. Specifically, this includes: Obtain the natural frequencies f of the bridge structure in multiple modes. i and corresponding mode shape; The calculated main vortex shedding frequency f v With each natural frequency f i Compare and calculate the normalized frequency ratio η. i =f v / f i ; Define a risk level determination matrix that simultaneously considers the frequency ratio η i The degree of proximity and the energy proportion of the corresponding mode shape in the sensitive aerodynamic feature parts; When a certain mode exists, its frequency is higher than η i When the location falls within the locked range and exhibits significant displacement within that mode shape, it is classified as a high-risk level; when the frequency ratio is η... i If the risk level is close to but not completely within the locked-in area, it is classified as medium risk; otherwise, it is classified as low risk.

8. The bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning as described in claim 7, characterized in that, Also includes: Step 4: Collect actual vibration response data of the bridge in a natural wind field using vibration sensors installed on the sensitive aerodynamic features of the bridge. Perform spectral analysis on the vibration response data to extract the actual principal vibration frequencies; The actual principal vibration frequency is compared with the calculated principal eddy shedding frequency. If the deviation between the two exceeds the preset tolerance, the turbulence model parameters or boundary condition settings in the computational fluid dynamics simulation are automatically calibrated, and the risk parameters are recalculated to optimize the accuracy of subsequent assessments.

9. A system for implementing the bridge vortex-sensitive aerodynamic structure detection method based on three-dimensional laser scanning as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire three-dimensional point cloud data of the bridge structure, preprocess the three-dimensional point cloud data and extract the geometric features of the bridge. The identification module is used to identify the component type of each part of the bridge based on the extracted geometric features and a pre-trained machine learning model. It also compares the geometric features with the model in the standard bridge type database and, based on the identified component type and corresponding geometric features, determines whether the part is a sensitive aerodynamic feature that is prone to vortex-induced vibration. The assessment module is used to calculate the vortex-induced vibration risk parameters of parts identified as having sensitive aerodynamic characteristics based on their geometric parameters, and compare the risk parameters with the inherent dynamic characteristics of the bridge structure to determine its vortex-induced vibration risk level. For parts that are determined to be non-sensitive aerodynamic features, the vortex-induced vibration risk assessment process for those parts will be terminated.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.