A High-Precision Erection Method and System for Simply Supported Box Girder Based on AI Visual Perception
By introducing AI visual perception technology into the box girder erection process, combined with the precise layout of the pad stone reference points and box girder feature points, and utilizing the improved YOLOv11 network and lidar data fusion, high-precision erection of simply supported box girders was achieved. This solved the problems of difficult precision control and lack of data traceability in traditional methods, achieving millimeter-level measurement accuracy and full-process data management.
Patent Information
- Application Number
- CN202611607119.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-16
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional box girder erection relies on manual experience, making precision control difficult and data traceability impossible. This makes it hard to meet the digital and intelligent requirements of modern railway engineering. Furthermore, the accuracy of visual inspection is limited by the imaging quality of small targets at long distances, making it difficult to achieve millimeter-level erection accuracy.
A high-precision erection method for simply supported box girders based on AI visual perception is adopted. By setting up reference points for pad stones and feature points at the bottom of the box girder, combined with a radar-based monitoring terminal, an improved YOLOv11 network, adaptive homomorphic filtering, and Zernike matrix sub-pixel positioning, sub-pixel-level positioning and three-dimensional deviation calculation of the target are achieved. Data fusion is performed by combining LiDAR data to establish a closed-loop management of the entire process.
It achieves high-precision control of box girder erection, with measurement accuracy reaching ΔX, ΔY≤±2mm, and ΔZ≤±1mm. It solves problems such as beam end misalignment, non-linearity, and support deviation in traditional methods, and realizes full-process data traceability and resource saving.
Smart Images

Figure CN123109075A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bridge engineering construction technology, and in particular to a high-precision erection method and system for simply supported box girders based on AI visual perception. Background Technology
[0002] With the rapid development of high-speed railways, intercity railways, and urban railways, precast simply supported box girders have been widely used in bridge engineering due to their superior structural performance and high construction efficiency. Traditional box girder erection methods mainly rely on manual observation and walkie-talkie command, which presents the following prominent problems: alignment control depends on the experience of construction personnel, lacking scientific quantitative control methods; measurement data during construction cannot be effectively traced, and quality control lacks data support; and the level of informatization is low, making it difficult to meet the needs of modern railway engineering's digital and intelligent construction.
[0003] In recent years, some studies have attempted to introduce target detection technology into the field of box girder construction monitoring. For example, some studies have used an improved YOLOv4 model to identify the construction status of box girders, while others have used digital twin technology to construct a collaborative control model between the bridge erecting machine and the girder. However, existing technologies still have shortcomings: first, the accuracy of visual detection is limited by the imaging quality of small targets at long distances, making it difficult to meet the millimeter-level erection accuracy requirements; second, a single vision or positioning system cannot simultaneously guarantee the accuracy of deviation measurement in the three-dimensional directions of lateral, longitudinal, and elevation; and third, there is a lack of complete closed-loop data management from the re-measurement of the bearing pads and simulated assembly to the control of girder placement.
[0004] To address the aforementioned issues, there is an urgent need to develop an intelligent erection scheme for box girders that enables high-precision erection, full-process traceability, and reusable equipment. Therefore, this paper proposes a high-precision erection method and system for simply supported box girders based on AI visual perception. Summary of the Invention
[0005] The main objective of this invention is to provide a high-precision erection method and system for simply supported box girders based on AI visual perception, aiming to solve the technical problems of traditional box girder erection relying on manual experience, difficulty in precision control, and lack of data traceability.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] In a first aspect, the present invention provides a high-precision erection method for simply supported box girders based on AI visual perception, comprising the following steps:
[0008] Step S1: Set up the reference points for the pad stones and the feature points at the bottom of the box girder.
[0009] Specifically, based on the pad stones delivered by the substructure construction unit, and with the center line of the support as a reference, cross lines are marked on the pad stones to ensure that the parallelism error between the cross lines and the axis is less than the first threshold; the intersection of the cross lines is the pad stone reference point, and the position error of the center point is controlled within the second threshold; the three-dimensional coordinates of the reference point are verified using a total station.
[0010] Specifically, during the prefabrication of the box girder, feature points are arranged along the centerline of the supports. Each girder bottom has at least four feature points, located longitudinally on the outer side of each support, with a distance from the support center point not less than the third threshold. The feature points are formed by welding square iron sheets with sides of six centimeters and a thickness of two millimeters to the corresponding positions on the bottom formwork of the box girder, creating grooves. After demolding, visual targets containing concentric circle coding strips are pasted into the grooves.
[0011] Through the above layout, a spatial positioning benchmark is established with the pad stone reference point and the characteristic point at the bottom of the beam as the core.
[0012] Step S2, Installation and Data Acquisition of the Ravis Monitoring Terminal
[0013] Specifically, the radar-based monitoring terminal equipment, which integrates industrial cameras and lidar, will be installed at the reference points of the corresponding pad stones and leveled for "target alignment measurement at the bottom of the box girder".
[0014] Specifically, during the initial erection, the installation of the bridge erecting machine's radar-based AI edge computing workstation and the deployment of the signal transmission network also need to be completed. During the erection process, the radar-based monitoring terminal collects real-time image data of the target at the bottom of the box girder and lidar distance data.
[0015] Step S3, AI visual recognition and sub-pixel level positioning
[0016] Specifically, the AI visual recognition algorithm uses an improved YOLOv11 network to detect and coarsely locate targets in acquired images. This network replaces standard convolution with full-dimensional dynamic convolution. ODConv can dynamically adjust the convolution kernel along four dimensions: kernel space, spatial location, input channel, and output channel. Specifically, let the input feature map be X, and the output feature map be Y, represented as:
[0017]
[0018] Among them, W i For the i-th convolutional kernel, α wi ,α fi ,α ci ,α si These are four sets of dynamic attention scalars, representing the kernel space dimension, spatial location, input channel, and output channel, respectively, generated from input features through lightweight attention branches.
[0019] The network further introduces a coordinate attention mechanism, performing global average pooling of the feature maps along the height and width directions to generate orientation-aware features:
[0020]
[0021] After concatenation, the vectors undergo a 1×1 convolution transformation and are then split into two attention vectors g. h and g w Final weighted output:
[0022]
[0023] Where σ(⋅) is the Sigmoid activation function.
[0024] Specifically, during model training, normalized Gaussian Wasserstein distance is used as the bounding box loss function to enhance sensitivity to small target position offsets. After detecting the target region, adaptive homomorphic filtering is used to perform illumination equalization processing on the ROI region, i.e., for the image: Take the logarithm and perform a Fourier transform, then pass the sample through a frequency domain Gaussian high-pass filter. Enhance the reflection component to eliminate shadow and glare interference. Then, use an improved Zernike moment for sub-pixel edge localization: the Zernike moment is defined as:
[0025]
[0026] Among them, V nm (r,θ)=R nm (r)e jmθ For Zernike polynomials, R nm (r) is a real-valued radial polynomial.
[0027] Based on the ideal step edge model, through four main moments Z 00 Z 11 Z 20 Z 31 Solving for edge parameters including background grayscale h, step amplitude k, edge-to-center distance l, and rotation angle ϕ yields the sub-pixel edge point positions. , where (x c ,y c () is the pixel center coordinate, which enables sub-pixel level precise positioning of the target center.
[0028] Step S4, eccentricity error compensation based on concentric circle coding bands
[0029] Specifically, the target is designed as a concentric circle coding strip structure, comprising an outer circle and an inner circle. Because the camera optical axis is not perpendicular to the target plane, the target image suffers from elliptical distortion, resulting in an eccentricity error between its geometric center and the actual projection point of the spatial circle center. This invention, based on the conic section theory of spatial circular projection, utilizes the deviation between the outer ellipse center c1 and the inner ellipse center c2, through a compensation formula: Calculate the true projection point o of the target space center in the image.
[0030] Step S5: Data fusion and 3D deviation calculation
[0031] Specifically, after obtaining the sub-pixel coordinates of the target's true projection point, the image point coordinates are then determined using the camera intrinsic parameter matrix K, combined with the depth distance d synchronously acquired by the lidar. Convert to normalized camera coordinate system direction vector The three-dimensional coordinates of the target in the camera coordinate system are Using the three-dimensional coordinates of the four targets as known conditions, a rigid body transformation model is established. The rotation matrix R and translation vector t are solved using the SVD decomposition method or the ICP algorithm to obtain the real-time pose of the box girder coordinate system relative to the pad reference coordinate system. The real-time pose is compared with the target pose to decompose the box girder into real-time deviations in the longitudinal ΔX, transverse ΔY, and elevation ΔZ directions. The system measurement error is controlled to be no greater than the fourth threshold in the longitudinal ΔX and transverse ΔY directions, and no greater than the fifth threshold in the elevation ΔZ direction.
[0032] Step S6, Beam Drop Control and Data Closed Loop
[0033] Specifically, the operator controls the girder erecting machine to adjust the box girder's posture based on the real-time deviation data displayed on the monitoring screen. Fine-tuning is done in stages until the deviations in each direction meet the design thresholds before the girder is finally lowered. Data from the entire process, including the re-measurement of the bearing pads, box girder acceptance data, girder lowering deviation data, and process adjustment data, are uploaded to the railway engineering management platform in real time. The platform automatically stores and determines the compliance status, generating a data report containing deviation curves and process replays, forming a complete digital archive and achieving full traceability.
[0034] Preferably, the reference points of the pad stones are marked with durable markers and protected, and a verification log is established, with verification and confirmation conducted again before each hole is erected.
[0035] Preferably, the improved YOLOv11 network employs ODConv full-dimensional dynamic convolution, and its dynamic attention scalar is generated from the input features through a lightweight attention branch.
[0036] Preferably, in the adaptive homomorphic filtering, D(u,v) is the frequency domain radius, and γ H ,γ L These represent the high-frequency gain and low-frequency gain, respectively, with D0 being the cutoff frequency.
[0037] Preferably, in the eccentricity error compensation model, λ=R2 / R1 is the ratio of the inner radius R2 of the target to the outer radius R1.
[0038] Preferably, before lowering the beam, a three-dimensional model is generated based on the measured data of the pad stone and the box girder, and the assembly is simulated. The longitudinal and transverse deviations of the support center, the elevation deviation of the beam end, and the thickness of the support mortar are calculated. The optimal beam lowering position is obtained by adjusting the tolerances to guide the on-site erection.
[0039] Preferably, the first threshold is 0.5 / 1000.
[0040] Preferably, the second threshold is ±0.5 mm.
[0041] Preferably, the third threshold is 480 mm.
[0042] Preferably, the fourth threshold is ±2mm.
[0043] Preferably, the fifth threshold is ±1 mm.
[0044] Secondly, the present invention also provides a high-precision erection system for simply supported box girders based on AI visual perception, for implementing the above method, including:
[0045] The Leishi monitoring terminal integrates an industrial camera and a lidar. It is installed at the reference point of the corresponding pad stone and leveled to collect target images with concentric circle coding bands and lidar distance data.
[0046] The AI edge computing workstation has an improved YOLOv11 detection network, an adaptive homomorphic filtering module, a Zernike matrix subpixel positioning module, an eccentricity error compensation module, and a radar-visual fusion solution module built in it, which are used to process multi-source data in real time and solve the three-dimensional spatial deviation of the box girder.
[0047] The monitoring screen is used to display real-time deviation data and guide the operator in adjusting the beam drop.
[0048] The data transmission module, including an industrial-grade wireless bridge and a gigabit industrial switch, is used to upload real-time data to the railway engineering management platform.
[0049] The visual target is a circular positioning plate containing concentric circle coding strips, which is pasted into the feature point grooves at the bottom of the box girder to serve as the AI recognition target.
[0050] Preferably, the railway engineering management platform includes a box girder transportation and erection management module, which supports the creation and application approval of erection tasks according to the hole position dimension, automatic storage and judgment of data throughout the process, and generation of data reports containing deviation curves and process playback, which are signed and confirmed by both the technical and supervisory parties.
[0051] Beneficial effects
[0052] Compared with the prior art, the present invention has the following beneficial effects:
[0053] This invention constructs a high-precision spatial positioning system with the crosshair reference point of the bearing pad and the concentric circle coded target at the bottom of the beam as the core. Through the precise layout of the bearing pad reference point and the feature point at the bottom of the beam, combined with AI sub-pixel level visual positioning and radar vision fusion three-dimensional calculation, the measurement accuracy of ΔX, ΔY≤±2mm and ΔZ≤±1mm is achieved, effectively solving the problems of beam end misalignment, non-linearity and support deviation in traditional methods.
[0054] This invention systematically solves the complex problems of long-distance small target detection, uneven lighting, and projection distortion caused by viewing angle tilt in box girder erection scenarios through a series of algorithmic innovations, including an improved YOLOv11 network (containing ODConv dynamic convolution and coordinate attention mechanism), adaptive homomorphic filtering, Zernike moment subpixel edge localization, and an eccentricity error compensation model based on concentric circle coding bands. It can work stably under all-weather conditions.
[0055] The radar-visual monitoring terminal equipment of this invention can be transferred as a whole to the next span after the box girder is in place and accepted; the entire system adopts digital and paperless management, which effectively reduces resource consumption and construction costs.
[0056] This invention relies on a railway engineering management platform to achieve closed-loop data and quality traceability throughout the entire process. From the re-measurement of the bearing pad, the acceptance of the box girder, the simulated assembly, the deviation of the girder placement and the process adjustment, the data of the entire process is automatically collected, uploaded in real time, automatically judged and permanently stored. It supports video playback and data curve analysis, and completely solves the problems of no record, difficulty in traceability and difficulty in review in traditional construction. Attached Figure Description
[0057] Figure 1 This is a flowchart illustrating the high-precision erection method for simply supported box girders based on AI visual perception, as described in this invention.
[0058] Figure 2 This is a schematic diagram showing the location of the feature points of the pad stone in the system of the present invention;
[0059] Figure 3 This is a schematic diagram of the layout of feature points of the box girder in the system of the present invention;
[0060] Figure 4 This is a schematic diagram of the arrangement of feature points at the bottom of the box girder in the system of the present invention. Detailed Implementation
[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0062] Example 1: Construction Preparation
[0063] Before construction, the project's technical manager organized the workers to carefully understand the design intent and master the requirements of the construction design drawings and relevant technical standards. The project team formulated a plan for the use of surveying instruments and promptly allocated surveying instruments calibrated by qualified professional institutions, including high-precision total stations and levels. The coordinates and elevation measurements of the substructure construction unit were carefully calculated and verified to ensure the accuracy of the data.
[0064] Example 2: Re-measurement of foundation stones and layout of feature points
[0065] See Figure 2 As shown, after the substructure construction unit handed over the pad stones, the erection unit verified and accepted the location of the bolt holes of the pad stone supports, the center coordinates of the supports, the elevation, and the location of the installation reference point of the radar monitoring terminal.
[0066] Using the support centerline as a reference, mark crosshairs on the bearing pad, ensuring the parallelism error between the crosshairs and the axis is less than 0.5 / 1000. The intersection of the crosshairs is the bearing pad reference point, with the center point position error controlled within ±0.5mm. Use a total station to verify the three-dimensional coordinates of the reference point; the measured values are used for system correction. The reference point is marked with durable markers, protected against collisions, displacement, and contamination, and a verification log is established.
[0067] Example 3: Box Girder Acceptance and Feature Point Layout
[0068] See Figures 3-4 As shown, during the prefabrication of the box girder, feature points are arranged along the centerline of the supports. A total of 4 feature points are set on the bottom of each girder, placed on the longitudinal outer side of each support. The distance between the longitudinal outer feature point of the support and the center point of the support is 480mm. A square iron sheet with a side length of 6 cm and a thickness of 2 mm is welded at the corresponding position of the bottom formwork of the box girder. After the box girder is demolded, it forms a groove for pasting the visual target.
[0069] Before the box girder is erected, the construction unit and the erection unit jointly inspect and hand over the finished precast beams to be delivered, ensuring that the target positioning error is ≤±1mm. Unqualified box girders are strictly prohibited from leaving the site. Before erection, visual targets containing concentric circle coded strips are pasted into the grooves as targets for identification by the radar monitoring terminal.
[0070] Example 4: System Initialization and Setup Task Creation
[0071] The IT administrator authorizes personnel from various units who need to use the box girder transportation and erection management module through the personnel and organization management module. Once authorized, the module becomes active, and personnel information must be synchronized after logging into the platform for normal access.
[0072] Before the box girder is erected, the erection task application and approval are carried out on the platform. The task is created according to the dimensions of the erection hole position. When applying for the task, the acceptance records of the pad stone and the box girder of the relevant hole position must be linked.
[0073] Example 5: Installation of RaidVision Monitoring Terminal
[0074] Before the box girder is erected and lowered, the radar-based monitoring terminal equipment must be installed. For the first erection, the installation of the radar-based AI edge computing workstation on the bridge erecting machine and the deployment of the signal transmission network must also be completed.
[0075] Four radar-guided monitoring terminal devices are used for "target alignment measurement at the bottom of the box girder". They are installed at the reference points of the corresponding pad stones and are precisely leveled. The equipment undergoes third-party verification before being put into use.
[0076] Example 6: Implementation of AI Visual Recognition Algorithm
[0077] During the beam lowering process, the Leishi monitoring terminal collects video data, deviation data, and distance data in real time and transmits them to the Leishi AI edge computing workstation. The workstation runs the AI visual recognition algorithm according to the following process:
[0078] S61: Object Detection
[0079] An improved YOLOv11 network was used to perform real-time detection and coarse localization of targets with four concentric circular encoded bands at the bottom of a beam. This network introduces full-dimensional dynamic convolution (ODConv) into the YOLOv11 backbone to replace standard convolution. Let the input feature map be X, and the output feature map be Y. The formula for calculating the output feature map of ODConv is:
[0080]
[0081] Among them, W i For the i-th convolutional kernel, α wi ,α fi ,α ci ,α si These are four sets of dynamic attention scalars, representing the kernel space dimension, spatial location, input channel, and output channel, respectively.
[0082] Four dynamic attention scalars are generated from input features through lightweight attention branches. The convolution kernel weights can be dynamically adjusted along four dimensions: kernel space dimension, spatial location, input channel, and output channel, which significantly improves the feature extraction capability for small-sized targets.
[0083] The network also introduces a coordinate attention mechanism, which performs global average pooling on the feature map along the height and width directions to generate orientation-aware features:
[0084]
[0085] After concatenation, the vectors undergo a 1×1 convolution transformation and are then split into two attention vectors g. h and g w Final weighted output:
[0086]
[0087] Where σ(⋅) is the Sigmoid activation function.
[0088] In model training, Normalized Gaussian Wasserstein distance (NWD) is used as the bounding box loss function to enhance sensitivity to small target position offsets and reduce the false negative rate. The formula for calculating the Normalized Gaussian Wasserstein distance is:
[0089]
[0090] Where C is a normalization constant associated with the dataset, and W2 2 For the closed-form solution of the second-order Wasserstein distance, B1 and B2 represent two bounding boxes, namely: B1 is the predicted bounding box, which is the target position output by the detection network; B2 is the ground truth bounding box, which is the actual position of the target in the labeled data; N1 and N2 are the distributions obtained by modeling the bounding boxes B1 and B2 as two-dimensional Gaussian distributions.
[0091] S62: Enhanced Adaptive Homomorphic Filtering
[0092] For the detected target ROI region, adaptive homomorphic filtering is used for illumination equalization. The image is modeled as the product of the illumination component and the reflection component. After taking the logarithm and performing a Fourier transform, the result is passed through a designed frequency-domain Gaussian high-pass filter. Filtering is performed in the frequency domain, where D(u,v) is the frequency domain radius, and γ H ,γ L These represent the high-frequency gain and low-frequency gain, respectively, with D0 being the cutoff frequency. After filtering and exponential restoration, an enhanced image with uniform illumination is obtained, effectively eliminating shadows and reflections.
[0093] S63: Zernike Subpixel Edge Positioning
[0094] For the enhanced target region, sub-pixel-level edge localization is achieved using improved Zernike moments. The nth-order m-fold Zernike moment of the image function f(x,y) is defined as:
[0095]
[0096] Among them, V nm (r,θ)=R nm (r)e jmθ For Zernike polynomials, R nm (r) is a real-valued radial polynomial.
[0097] Based on the ideal step edge model (background grayscale h, step amplitude k, distance from edge to center l, rotation angle φ), the relationship between the four main moments and the edge parameters is as follows:
[0098]
[0099] The sub-pixel edge point positions are obtained by solving the above system of equations. , where (x c ,y c () is the pixel center coordinate, which enables sub-pixel level precise positioning of the target center.
[0100] S64: Eccentricity Error Compensation
[0101] The target has a concentric circular coding band structure, with an outer radius of R1 and an inner radius of R2, and the radius ratio λ = R2 / R1. Because the camera optical axis is not perpendicular to the target plane, the target is projected as an ellipse in the image, with both the outer ellipse center c1 and the inner ellipse center c2 deviating from the true projection point of the spatial center. Based on the conic section theory of spatial circular projection, the eccentricity error compensation formula is: [Formula for true projection point]. The above compensation effectively eliminates the center positioning deviation caused by the tilt of the viewing angle.
[0102] S65: Rayvision Fusion 3D Solving
[0103] Obtain the subpixel coordinates of the four target true projection points Then, the depth and distance d acquired synchronously by the lidar are combined. i Using the camera intrinsic parameter matrix K, the normalized camera coordinate system direction vector is calculated. The three-dimensional coordinates of target i in the camera coordinate system are: Establish a rigid body transformation model. The rotation matrix R and translation vector t are solved using the SVD decomposition method, thus obtaining the real-time pose of the box girder relative to the pad datum. Where, P w,i The target is represented by three-dimensional coordinates. The deviations in the longitudinal direction (ΔX), lateral direction (ΔY), and elevation (ΔZ) are decomposed and controlled stably within ΔX and ΔY ≤ ±2mm, and ΔZ ≤ ±1mm.
[0104] Example 7: Simulated Assembly
[0105] After the remeasured data of the bearing pad and box girder are entered, the corresponding box girder is simulated and erected in the system. The system generates a three-dimensional model based on the measured data, obtains the longitudinal and transverse deviations of the support center, the elevation deviation of the beam end, and the thickness of the support mortar, and adjusts the tolerances according to the relevant data to finally obtain the optimal assembly data to guide the on-site erection. The simulated assembly realizes virtual erection and risk prediction before the beam is lowered, avoiding repeated alignment and adjustments.
[0106] Example 8: Beam Drop Control
[0107] The operator controls the girder erecting machine to adjust the position of the box girder based on the offset displayed on the monitoring screen. The girder lowering process is carried out in stages with fine adjustments: first, a coarse adjustment is made to ensure the position is in place, then a fine adjustment is made until the deviation is ≤±2mm. The girder can only be lowered after all points are confirmed to be qualified. Technicians, surveyors, and supervisors jointly confirm the deviation data, and a "no girder lowering if not up to standard" system is implemented.
[0108] Example 9: Gravitational Potential Energy Recovery for Power Generation
[0109] During the uniform descent of the beam, a dedicated kinetic energy recovery system serves as the core carrier for energy conversion. Through a mechanical transmission device, it simultaneously transfers the gravitational potential energy and linear kinetic energy of the falling beam to the generator, driving the generator rotor to rotate at high speed. Utilizing the principle of electromagnetic induction, it converts the mechanical kinetic energy into alternating current (AC). The system is equipped with voltage stabilization and rectification modules to adjust the unstable current and voltage generated during power generation in real time, converting fluctuating AC into stable DC, thus realizing the recovery and utilization of energy from the beam's descent.
[0110] Example 10: Process Replay and Data Evaluation
[0111] After the monitoring terminal equipment collects data in real time, it uploads it to the platform according to the task association. The platform automatically judges the data for compliance according to relevant specifications. After the erection of each box girder span is completed, a data report is automatically generated, including deviation curves, process playback, and compliance judgment, which is signed and confirmed by both the technical and supervisory parties. All data is archived monthly and backed up by span, realizing traceability, review, and analysis of the entire erection process.
[0112] Example 11: Quality Control
[0113] Benchmark point accuracy control: Before the benchmark points for the pad stones are laid out, the surveying engineer will provide specific technical instructions to the operators; the cross lines of the pad stones will be double-checked using ink lines and a total station to ensure that the parallelism error is less than 0.5 / 1000, and the center position error of the benchmark points will be strictly controlled within ±0.5mm; the benchmark points will be marked with durable marks and protected, and will be checked and confirmed again before each span is erected; the size of the iron plates, welding position, and groove depth of the characteristic points of the box girder will be checked for each span to ensure that the target positioning error is ≤±1mm, and unqualified box girders will be strictly prohibited from leaving the factory.
[0114] System measurement accuracy control: The laser vision monitoring terminal undergoes third-party verification before entering the site, and can only be used after passing the verification; the system accuracy is calibrated once every 5 box girders are erected on site to ensure that ΔX, ΔY≤±2mm, ΔZ≤±1mm; the cleanliness of the camera lens and lidar probe is checked regularly to avoid dust and water mist affecting the recognition accuracy, and a re-inspection is carried out immediately after severe weather; the AI visual recognition model is updated and optimized regularly to ensure that the target recognition accuracy rate is ≥98.5%; an equipment operation log is established to record calibration time, accuracy data, and fault handling.
[0115] Process data control: The entire process data is automatically collected and uploaded in real time, and the platform automatically stores and judges the data. Manual modification of data is strictly prohibited. The system automatically issues warnings for deviations exceeding limits, failure to set up within time limits, and missing data. Technical personnel respond within 10 minutes and complete the handling and recording within 30 minutes. A closed-loop data management mechanism is established, and all data is archived monthly and backed up by hole.
[0116] Acceptance and verification: A five-level acceptance system is implemented, which includes "base stone acceptance → box girder acceptance → system acceptance → girder placement acceptance → finished product acceptance". The technical supervisor conducts parallel random checks on the benchmark points, feature points, simulated assembly data, and girder placement deviations throughout the entire process, with a sampling rate of no less than 30%. After the erection is completed, the support mortar construction, linear verification, and elevation re-measurement are carried out in a timely manner to form complete acceptance data.
[0117] Example 12: Engineering Case
[0118] The Nantong-Ningbo High-Speed Railway Station Front Section X Project, undertaken by China Railway Beijing Engineering Bureau, has a main line length of 24.269km. It includes the prefabrication and erection of 1024 double-track simply supported box girders for the Cixi Grand Bridge (sections A to D) and the approach bridge in the shallow water area on the south bank of the Hangzhou Bay Bridge. The project employed the aforementioned AI-based visual perception-based high-precision erection method and system for simply supported box girders during the girder erection. The erection accuracy, efficiency, and quality all met the design and specification requirements, with no deviations exceeding standards or rework occurring, proving the feasibility and effectiveness of this technical solution.
[0119] The AI-based visual perception-based high-precision erection method and system for simply supported box girders provided by this invention are applicable to engineering projects such as high-speed railways, intercity railways, and urban railways that use prefabricated simply supported box girders for erection. In particular, it has wide applicability and promotion value for high-speed railway bridge projects with high erection accuracy requirements and strict alignment control.
[0120] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A high-precision erection method for simply supported box girders based on AI visual perception, characterized in that, Includes the following steps: S1: Set up reference points on the pad stones and set up visual target feature points with concentric circle coding bands at the bottom of the box girder to establish a spatial positioning reference. S2: A radar-based monitoring terminal integrating an industrial camera and lidar is used to collect target images and distance data; S3: Target detection based on an improved YOLOv11 network. This network adopts full-dimensional dynamic convolution and coordinate attention mechanisms, and uses normalized Gaussian Wasserstein distance as the bounding box loss function. After enhancing the detection region with adaptive homomorphic filtering, sub-pixel edge localization is performed using improved Zernike moments to extract the target center. S4: Based on the deviation of the inner and outer ellipse centers of the concentric circle coding band, establish an eccentricity error compensation model and calculate the true projection point of the target space circle center in the image. S5: Using the lidar ranging value as an absolute scale constraint, the three-dimensional coordinates of the target are calculated in conjunction with the sub-pixel center coordinates, and the longitudinal, lateral and elevation deviations of the box girder relative to the pad stone reference are solved through rigid body transformation. S6: Adjust the box girder posture according to the deviation data until it meets the threshold, then lower the girder and upload the entire process data to generate a digital archive.
2. The high-precision erection method for simply supported box girders based on AI visual perception according to claim 1, characterized in that, The reference point is determined by marking a crosshair, and the parallelism error between the crosshair and the axis is less than a first threshold, and the position error of the center point is less than a second threshold.
3. The high-precision erection method for simply supported box girders based on AI visual perception according to claim 1, characterized in that, The bottom of the box girder has no fewer than four feature points, which are located on the longitudinal outer side of each support and the distance between each feature point and the center point of the support is no less than the third threshold.
4. The high-precision erection method for simply supported box girders based on AI visual perception according to claim 1, characterized in that, In the eccentricity error compensation model, the true center projection point o of the target is determined by the outer ellipse center c1 and the inner ellipse center c2 through the formula: Calculate, where λ is the ratio of the inner and outer radii.
5. The high-precision erection method for simply supported box girders based on AI visual perception according to claim 1, characterized in that, The adaptive homomorphic filtering employs a frequency-domain Gaussian high-pass filter. The frequency domain of the image after logarithmic transformation is filtered, and then exponentially restored to obtain an enhanced image with uniform illumination, where D(u,v) is the frequency domain radius, and γ... H ,γ L These represent the high-frequency gain and low-frequency gain, respectively, with D0 being the cutoff frequency.
6. The high-precision erection method for simply supported box girders based on AI visual perception according to claim 1, characterized in that, The subpixel edge localization is solved using a Zernike moment step edge model, which includes edge parameters such as background grayscale h, step amplitude k, edge-to-center distance l, and rotation angle ϕ, to obtain the subpixel edge point position. , where (x c ,y c () represents the center coordinates of the pixel.
7. The high-precision erection method for simply supported box girders based on AI visual perception according to claim 1, characterized in that, Before the beam is lowered, a three-dimensional model is generated based on the measured data of the pad stone and the box girder to simulate assembly, predict deviations, and recommend the optimal beam lowering position.
8. The high-precision erection method for simply supported box girders based on AI visual perception according to claim 1, characterized in that, The longitudinal, lateral, and elevation deviations after the beam is lowered shall meet the following conditions: the longitudinal deviation ΔX and the lateral deviation ΔY shall not exceed the fourth threshold, and the elevation deviation ΔZ shall not exceed the fifth threshold.
9. A high-precision erection system for simply supported box girders based on AI visual perception, used to implement the method described in any one of 1-8, characterized in that, include: The Leishi monitoring terminal integrates an industrial camera and a lidar to acquire target images and distance data containing concentric circle coded bands. The AI edge computing workstation has an improved YOLOv11 detection network, an adaptive homomorphic filtering module, a Zernike matrix subpixel positioning module, an eccentricity error compensation module, and a radar-visual fusion solution module built in it, which are used to solve the three-dimensional spatial deviation of the box girder. The monitoring screen is used to display deviation data to guide the beam dropping operation; The data transmission module is used to upload data to the platform in real time. And the reference point set at the pad stone and the feature point set at the bottom of the box girder.
10. The high-precision erection system for simply supported box girders based on AI visual perception according to claim 9, characterized in that, The radar monitoring terminals are installed at the reference points of the corresponding pad stones and are leveled. The platform is used to store the data of the whole process, automatically determine the qualification, and generate data reports containing deviation curves and process playback.
