A method for hoisting a steel anchor box with variable spacing of cable pylons of a cross-sea cable-stayed bridge

Abstract

This invention discloses a method for hoisting steel anchor boxes with variable spacing for cable-stayed bridge towers across the sea, belonging to the field of bridge engineering. To address the challenges of short working windows at sea and the low precision, efficiency, and slowness of traditional methods, this invention constructs an integrated intelligent hoisting system. By deploying visual markers and positioning anchor points on the tower columns and steel anchor boxes, and combining differential satellite positioning, inertial measurement units, and other multi-source sensors, the precise pose of the steel anchor boxes is calculated in real time under a unified spatiotemporal reference. Artificial intelligence algorithms are used to predict the dynamic trajectory of sea conditions, the ship's hull, and the hoisting process, and a model predictive controller generates crane linkage compensation commands to dynamically suppress wind and wave disturbances. During the near-field installation phase, the system performs trajectory fine-tuning and anti-collision control based on a segment gap database; finally, a three-way hydraulic jack completes millimeter-level precise alignment and locking. This invention achieves an integrated closed loop of perception, prediction, control, and fine-tuning, significantly improving installation accuracy and efficiency, and shortening the time spent at sea.

Description

Technical Field

[0001] This invention relates to the field of bridge engineering construction technology, specifically to a method for hoisting steel anchor boxes with variable spacing for cable-stayed bridge towers across the sea, and particularly to an integrated on-site multi-source collaborative positioning, precise hoisting, and three-way jack fine-tuning installation method. Background Technology

[0002] This invention relates to the field of bridge engineering construction technology, and particularly to the hoisting and installation of steel anchor boxes for cable-stayed bridge towers across the sea. Construction across the sea is significantly affected by external factors such as wind speed, wave height, and visibility, resulting in a short working window. The steel anchor box segments are large in mass and have high rigidity, with non-equidistant spacing between installation surfaces and strict requirements for alignment accuracy. Current practices often involve total station measurement followed by manual commands for adjustment. This disconnect between measurement and crane operations leads to long feedback loops, frequent repeated lowering and retraction, alignment collisions, and waiting at height, making it difficult to consistently achieve millimeter-level installation accuracy. This results in lost working windows and increases both construction time and cost.

[0003] Existing technologies have attempted to use single-type sensors for construction monitoring, such as visual cameras for attitude recognition, UWB / RTK for single-unit positioning, and heave compensation or sway control for reducing local disturbances. However, these solutions are mostly used for condition monitoring or localized control, failing to form real-time data fusion and action closed-loop under a single reference for a multi-coordinate system encompassing the tower, ship crane, and steel anchor box. Furthermore, the lack of a unified time reference and integrated coordinate calculation between different data sources makes them prone to cumulative errors and hysteresis under sea state disturbances.

[0004] In terms of hoisting kinematic control, some studies have introduced sway suppression algorithms or feedforward compensation, but they typically do not combine sea state and hull six-degree-of-freedom response for short-time trajectory prediction, nor do they link the prediction results with model predictive control for coordinated optimization of slewing, luffing, and lifting. During the approach to the installation port, there is insufficient parameterized management of the nominal clearances of different segments and the guide surface normal vectors, and a lack of approach control and collision avoidance limiting strategies for non-equidistant clearances, resulting in insufficient alignment efficiency and safety margin in the final 0.3 m.

[0005] In the fine-tuning stage, current methods still rely mainly on manual labor or single-point jacks, lacking a closed-loop fine-tuning system composed of three-way hydraulic jacks, position / mechanical sensors, and visual residuals. Furthermore, the fine-tuning sequence and control precision lack unified standards, making it difficult to stably achieve residual errors of ≤1 mm under marine disturbance conditions. In addition, safety strategies linked to environmental thresholds, equipment health, and time synchronization status, as well as the full-process data retention mechanism, are not yet sound, affecting quality traceability and cross-segment reuse.

[0006] In summary, the installation of steel anchor boxes for cable-stayed bridge towers across the sea still lacks an integrated construction method that achieves real-time data fusion, machine vision recognition, and artificial intelligence algorithm prediction in multiple coordinate systems, combined with hoisting motion compensation, approach stage gap control, and millimeter-level fine-tuning of three-way hydraulic jacks, and establishes a safety control and data archiving framework linked to environmental thresholds and system health. To address these technical issues, a hoisting method for steel anchor boxes of cable-stayed bridge towers across the sea is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a method for hoisting steel anchor boxes with variable spacing for cable-stayed bridge towers across the sea, aiming to solve the technical problems of traditional offshore hoisting operations being greatly affected by sea conditions, relying on manual observation, experiencing feedback lag, and suffering from low accuracy and efficiency. This invention constructs an intelligent hoisting and installation system integrating multi-source sensing, dynamic prediction, closed-loop control, and precise fine-tuning, achieving high-precision, high-efficiency, and high-safety automated installation of steel anchor boxes throughout the entire hoisting process, especially in complex marine environments.

[0008] To achieve the above objectives, the present invention discloses a method for hoisting variable-spacing steel anchor boxes for cable-stayed bridge towers across the sea, applicable to the installation of the first section and all subsequent continuous sections. The technical solution includes the following steps:

[0009] The multi-source sensing system is constructed with a unified spatiotemporal reference. First, a unified tower reference coordinate system is established in the tower installation area, and relevant benchmarks are measured. To achieve high-precision positioning, ultra-wideband (UWB) positioning anchor points are deployed on the outer surface of the tower at predetermined non-coplanar positions. Simultaneously, machine vision targets are affixed to key areas on the outer surface of the tower installation opening and the steel anchor box segment to be hoisted. On the deck of the floating crane vessel, the core equipment for the operation, and at the crane's slewing base, a differential satellite positioning (RTK-GNSS) reference station and rover, a high-precision inertial measurement unit (IMU), and industrial cameras are installed. All sensors and control units are synchronized via a unified clock protocol (such as IEEE 1588), and the coordinate systems of each sensor are uniformly calibrated, thereby constructing a multi-source sensing network with a unified spatiotemporal reference.

[0010] Real-time pose fusion calculation. During the hoisting process, the system synchronously acquires machine vision images from industrial cameras, ranging data from ultra-wideband positioning anchor points, absolute position data from differential satellite positioning, and attitude and acceleration data from inertial measurement units at a sampling frequency no less than the preset frequency. Using machine vision algorithms such as perspective-n-point (PnP) pose calculation, the six-degree-of-freedom pose of the steel anchor box relative to the tower reference coordinate system is obtained. Subsequently, multi-source data fusion algorithms such as extended Kalman filter (EKF) are used to fuse the visual calculation results with ultra-wideband ranging, satellite positioning, and inertial measurement data to accurately calculate the real-time deviation vector between the hook, steel anchor box components, and target installation pose, providing precise input for subsequent control.

[0011] Dynamic Trajectory Prediction and Compensation Control. To proactively suppress environmental disturbances such as sea state and wind, this invention introduces short-time trajectory prediction based on artificial intelligence algorithms. This algorithm uses time-series data such as sea state, wind speed, ship motion, sling swing, and crane drive status as input to predict the future trajectory of the steel anchor box. Based on the prediction results, a Model Predictive Controller (MPC) is constructed with the objective functions of minimizing real-time pose deviation, limiting sling swing angle, and ensuring control smoothness. The controller generates coordinated control commands for the crane's slewing, luffing, and hoisting mechanisms through optimization, and sends these commands to the crane for execution, thereby achieving dynamic compensation and precise control against environmental disturbances. Furthermore, heave can be calculated based on the vertical displacement and acceleration data from the ship's inertial measurement unit, and equal-amplitude, reverse-biased compensation commands can be superimposed on the hoisting mechanism to achieve proactive heave compensation.

[0012] Near-field guidance and collision avoidance. During the descent of the steel anchor box and its approach to the installation opening, a segment gap database is introduced to ensure safe and precise docking. This database pre-records geometric information such as the nominal gap, guide surface normal vector, and allowable deviation for each segment. After entering the preset approach buffer zone, the system executes refined construction feedback adjustment strategies based on the database information, including gap sensing, trajectory micro-correction, and collision avoidance limiting. For example, it employs normal approach and tangential limiting control, and can trigger a force-limiting retreat strategy based on signals from contact sensors.

[0013] Millimeter-level fine-tuning and locking. After the steel anchor box is supported by the support frame of the installation mechanism, the crane maintains a low speed and gradually unloads most of the load. At this time, the three-way hydraulic jack system in the installation mechanism starts to work. Based on the residual deviations fed back by high-precision three-way displacement sensors, contact sensors, and vision systems, the system generates fine-tuning commands. The fine-tuning process is executed step by step according to a preset sequence (such as adjusting the posture first and then the translation, and adjusting the horizontal first and then the vertical) until the residual deviations in each direction are no more than 1 mm. After confirming that the accuracy requirements are met, temporary locking and quality confirmation are completed, and an updated comparison report of the current three-dimensional structural model can be generated.

[0014] Full-process monitoring and data archiving. This method continuously monitors the working environment (wind speed, wave height, visibility), equipment status, and control parameters throughout the entire hoisting process. The system is equipped with corresponding safety thresholds. When any parameter exceeds the threshold, or when any critical sensor malfunctions or time synchronization fails, the corresponding safety strategy will be automatically triggered, such as stopping hoisting, safe evacuation, or prompting manual takeover. Simultaneously, all key data throughout the process, including position and trajectory, environmental parameters, control commands, and quality inspection results, are digitally recorded and archived for quality traceability, final acceptance, and operation and maintenance handover.

[0015] Compared with existing technologies, this invention has the following advantages: 1) By constructing a multi-source sensing system with unified spatiotemporal reference and adopting a multi-source data fusion algorithm, high-precision, high-frequency real-time acquisition of the steel anchor box's position and attitude in a dynamic marine environment is achieved; 2) An innovative strategy combining artificial intelligence prediction algorithms and model predictive control is introduced to achieve active compensation for environmental disturbances such as wind and waves and precise control of the hoisting trajectory; 3) A closed-loop automated control system is integrated from remote dynamic hoisting to near-field guidance and finally to millimeter-level hydraulic fine-tuning, significantly improving installation accuracy and operational efficiency; 4) A comprehensive safety monitoring and data archiving mechanism greatly enhances the safety of high-risk marine operations and the traceability of engineering quality. Attached Figure Description

[0016] Figure 1 A schematic diagram of a method for hoisting variable-spacing steel anchor boxes for a cross-sea cable-stayed bridge tower;

[0017] Figure 2 Schematic diagram of sea state-hull-lifting linkage prediction and compensation method;

[0018] Figure 3 Schematic diagram of a three-dimensional hydraulic fine-tuning system and method; Detailed Implementation

[0019] Example 1: The entire process of precise hoisting and millimeter-level fine-tuning of continuous segments of steel anchor boxes for cable-stayed bridge towers in a marine environment.

[0020] I. Project Overview and Construction Boundary Conditions

[0021] This embodiment uses the installation of steel anchor boxes for the main tower of a cross-sea cable-stayed bridge as an example. The main tower of the bridge adopts a steel-concrete composite structure, with the tower top at an elevation of +225.5 m relative to sea level and the construction platform height in the tower area at +18.0 m. The main tower crossbeam is 8.5 m high, and the steel anchor boxes adopt a segmented design, with the geometric dimensions of a single segment being 4.2 m × 3.6 m × 2.8 m, and the typical segment self-weight being 26.5 t (including stiffening ribs and inner lining components). Because the anchorage angle and spacing of the stay cables at different heights of the tower vary with the structural stress requirements, the nominal gaps between the segments exhibit a non-equidistant distribution characteristic, and the specific parameters are shown in Table 1.

[0022] Table 1. Nominal clearance and guide surface parameters between steel anchor box segments (excerpt)

[0023] Segment numbering Nominal clearance (mm) Guide surface normal vector n Permissible deviation (mm) Section 1 (Base Section) 32 (0， 0， 1) ±3 Section 2 36 (0.017， 0， 0.99985) ±3 Section 3 34 (-0.012， 0.006， 0.99985) ±3 Section 4 37 (0.020， -0.004， 0.99978) ±3 Section 5 35 (-0.008， 0.010， 0.99982) ±3

[0024] The offshore lifting operation utilizes a 600-ton floating crane and a deck barge working in tandem. The floating crane has a rated lifting capacity of 600 tons, a maximum lifting speed of 12 m / min, and a slewing radius covering the tower installation area. Based on engineering safety regulations and sea condition statistics, the operational meteorological and sea condition control indicators are set as follows: wind speed ≤ 12 m / s (at a 10 m height measuring point), significant wave height ≤ 1.5 m, and visibility ≥ 800 m. These thresholds are consistent with the safety linkage strategy. The construction window is mainly concentrated between April and October each year, with each window typically lasting 4-6 hours, requiring extremely high operational efficiency.

[0025] II. Equipment Layout and Multi-Source Sensor Configuration

[0026] The tower reference coordinate system is established and synchronized with time. A tower reference coordinate system C_T is established in the tower installation area, with the origin at the geometric center of the first steel anchor box installation surface. The Z-axis points vertically upward along the tower, the X-axis points longitudinally along the bridge, and the Y-axis is determined according to the right-hand rule. The entire system uses IEEE 1588-2008 (Precision Clock Protocol PTP) for time synchronization, with clock drift controlled within ±200 ns. The controller uses a rugged industrial control computer (8-core CPU, real-time Linux operating system, scheduling capability 1kHz) to meet the requirement that the end-to-end calculation delay does not exceed the set value.

[0027] Ultra-wideband (UWB) positioning anchor point deployment. Six UWB anchor points were deployed at non-coplanar locations around the tower, with a spatial distribution that met the geometric requirements of covering all four quadrants of the tower and ensuring that no three points were collinear. The anchor point installation heights were +20.5 m, +22.0 m, +19.8 m, +21.5 m, +20.0 m, and +21.8 m, forming a three-dimensional positioning network. The three-dimensional coordinates of each anchor point were measured using RTK-GNSS (Real-time Dynamic Differential Satellite Positioning) and stored in the system database, achieving a positioning accuracy of ±1 cm in the plane and ±2 cm in elevation. The geometrical accuracy degradation factor (GDOP) of the construction area was evaluated as 2.4, meeting the requirement of ≤ set threshold. Integrated UWB tags were fixed near the floating crane hook, with an update rate of 100 Hz and a single distance measurement standard deviation ≤7 cm.

[0028] Differential satellite positioning and inertial measurement unit configuration. An RTK-GNSS receiver is installed on both the floating crane's slewing base and the deck. The base station is fixed to the construction trestle on the shore, transmitting differential correction data in real time via a radio data link. Simultaneously, a high-precision inertial measurement unit (IMU) is installed on the slewing base, with gyroscope zero-bias stability of 3° / h, accelerometer zero-bias ≤0.5mg, and a sampling frequency of 200 Hz. The IMU is used to capture the ship's six degrees of freedom motion (pitch, roll, yaw, heave, surge, and sway), providing high-frequency inertial increments for subsequent heave compensation and trajectory prediction.

[0029] Machine vision system setup. Four industrial cameras with a resolution of 1920×1080 pixels and a frame rate of 60 fps are positioned at the end of the boom. A global shutter is used to avoid motion blur, and the lens focal length is 8 mm. The field of view covers the steel anchor box mounting surface and the tower column mounting opening area. The camera axis of view intersection angle is set to 120°~150° to form a multi-view stereo vision network. Twelve black and white coded visual marker targets are affixed to the non-stress-critical areas around the tower column mounting opening, the four corners of the steel anchor box to be hoisted, and the perimeter of the mounting surface. Each target has a side length of 200 mm and is made of anti-glare and salt spray-resistant materials. The intrinsic parameter calibration (focal length, principal point, distortion coefficient) using the Zhang Zhengyou calibration method and the extrinsic parameter calibration (rotation matrix, translation vector) based on rigid targets are completed, with image point reprojection error ≤0.35 pixels. The entire link is connected to the controller via gigabit Ethernet, with end-to-end communication latency ≤5 ms.

[0030] The installation mechanism is equipped with a three-way hydraulic jack configuration. A support frame for the installation mechanism is installed at the tower column mounting opening. Three independent servo hydraulic jacks are mounted on the frame, each with a stroke of 80 mm, a displacement resolution of 0.01 mm, and a maximum thrust of 300 kN. Each jack is equipped with an LVDT (Linear Variable Differential Transformer) displacement sensor and a pressure-sensitive pad-type force sensor. The displacement sensor has a resolution of 0.01 mm, and the force sensor has a range of 0~350 kN and an accuracy of ±0.5%FS. Replaceable rigid pads (20 mm thick, tolerance ±0.02 mm) are installed between the jack reaction points and the main beam to ensure repeatability and contact stiffness. In addition, eight contact sensors are arranged around the mounting surface to detect the contact state and contact force distribution between the steel anchor box and the tower column.

[0031] III. Real-time Sensing and Multi-Source Data Fusion Algorithms

[0032] Machine vision pose calculation. A pose calculation algorithm based on perspective-n-point (PnP) is adopted. First, the acquired images are preprocessed with distortion correction and illumination equalization. Then, the corner detection algorithm of the coded target is used to identify the target position, and the corner positioning accuracy is improved by sub-pixel fitting. The Efficient Perspective-n-Point (EPnP) algorithm combined with RANSAC robust estimation is used to solve the pose (rotation matrix R and translation vector t) of the steel anchor box coordinate system C_B relative to the tower reference coordinate system C_T, and output six degrees of freedom parameters: translation (x, y, z) and attitude (roll, pitch, yaw). The single-frame processing latency is ≤12 ms, which meets the real-time requirements.

[0033] Extended Kalman filter fusion solution. Construct the state vector x=[p x p y p z v x v y v z , q w , q x q y , q z b a b g ] T Where p represents position, v represents velocity, q represents attitude quaternion, and b_a and b_g represent the accelerometer and gyroscope zero biases, respectively. An extended Kalman filter (EKF) framework is employed, fusing machine vision pose measurement, UWB ranging (distance between the tag and six anchor points), RTK baseline solution, and IMU inertial increment. The filtering period is set to 20 ms to meet the requirement that the filtering period should not exceed the set period.

[0034] The EKF prediction step propagates state based on IMU inertial increments, and the update step sequentially fuses visual pose, UWB ranging, and RTK position observations. The observation equations are as follows:

[0035] Visual observation: z_vision = [x, y, z, roll, pitch, yaw]^T = h_vision(x) + v_vision

[0036] UWB observation: z_uwb_i = ||p - p_anchor_i|| + v_uwb_i, i=1, 2,..., 6

[0037] RTK observation: z_rtk = [p_x, p_y, p_z]^T = h_rtk(x) + v_rtk

[0038] Where v represents the measurement noise of each sensor. The observation equations are linearized using the Jacobian matrix, the Kalman gain is calculated, and the state estimate and covariance matrix are updated. The fused output real-time deviation vector ΔX=[Δx, Δy, Δz, Δφ, Δθ, Δψ, Δẋ, Δẏ, Δż]^T serves as the sole input for subsequent control and proximity strategies.

[0039] Virtual data construction and validation. Under typical sea state conditions (significant wave height 1.2 m, period 6.5 s), the statistical characteristics of the floating crane hull's six-DOF response are: Heave amplitude ±0.22 m, Pitch amplitude ±0.6°, and Roll amplitude ±0.5°. Against this background, the statistical characteristics of the EKF output within a 1-second time window are: Δz mean +9.6 mm, standard deviation 4.2 mm; Δx and Δy standard deviations 3.8 mm and 4.1 mm, respectively; attitude residuals (roll / pitch / yaw) standard deviations 0.042°, 0.038°, and 0.051°, respectively. This virtual data sequence is used for the subsequent training and validation of the artificial intelligence algorithm prediction model, as well as the parameter tuning of the model predictive controller.

[0040] IV. Artificial Intelligence Algorithm Trajectory Prediction and Model Predictive Control

[0041] Short-term trajectory prediction model architecture. A Long Short-Term Memory (LSTM) network is used as the artificial intelligence algorithm prediction model. The network structure is a two-layer LSTM with 128 and 64 hidden layer units respectively. The input dimension is 18, including: wind speed (1D), ship's six-DOF motion (6D), sling tension (4D, four main slings), visual residual (3D, Δx / Δy / Δz), UWB residual (1D, average ranging error), and crane drive feedback (3D, slewing angular velocity, luffing velocity, and hoisting speed). The time window length is set to 10 seconds, and the sampling frequency is 10 Hz, meaning the input sequence length is 100 time steps. The output is the relative pose prediction sequence ΔX̂(t+0.1s, t+0.2s, ..., t+1.5s) for the next 1.5 seconds, a total of 15 time steps.

[0042] Training Dataset Construction and Pre-training Strategy. The training dataset consists of three parts: historical construction data (60%), outdoor full-scale component loading test data (20%), and numerical simulation synthetic data (20%). Historical construction data is derived from steel anchor box hoisting records over the past 8 working days, with a cumulative effective duration of 7.1 hours, a sampling frequency of 10 Hz, and approximately 256,000 sample points. Test data was collected through hoisting tests conducted at a land-based test site using a 1:1 steel anchor box model under different wind speeds (414 m / s) and hoisting heights (1040 m). Simulation data is generated by establishing a coupled model of the floating crane, sling, and steel anchor box using multibody dynamics software, considering different sea states (significant wave height 0.5–2.0 m) and wind speed combinations.

[0043] Data augmentation strategies include: illumination variation (adjusting image brightness and contrast ±20%), local occlusion (randomly occluding 10%–30% of the target area), viewpoint perturbation (simulating camera installation deviation ±5°), and noise injection (superimposing Gaussian white noise into sensor data). The training loss function uses a weighted Huber loss, assigning weights of 0.6 and 0.4 to position prediction error and pose prediction error, respectively, to balance the impact of the two types of errors. The Adam optimizer is used with an initial learning rate of 1 × 10^-3, a batch size of 64, and training for 50 epochs. On the validation set, the root mean square error (RMSE) of Δz in the 1-second prediction time domain is 2.9 mm, and the RMSEs of Δx and Δy are 2.4 mm and 2.7 mm, respectively, meeting engineering accuracy requirements.

[0044] Online adaptive mechanism. After the pre-trained model is deployed online, only the weights of the last fully connected layer are made available for online fine-tuning. The trigger condition is that the mean residual between the predicted and measured values ​​exceeds 0.6 mm for one consecutive minute. A recursive least squares (RLS) algorithm is used for closed-loop fine-tuning of the output layer weights, with a forgetting factor set to 0.98 to adapt to changes in sea state and ship response characteristics. After online adaptation, the RMSE of the 1-second prediction Δz can be further reduced to 2.6 mm, improving the model's generalization ability and robustness.

[0045] Active heave compensation algorithm. Utilizing the heave displacement estimated by the rotating base IMU. (t) and vertical acceleration (t), through quadratic integration and drift suppression filtering (high-pass filter cutoff frequency 0.01 Hz), the constant amplitude reverse compensation amount u_h(t) is obtained. (t). This compensation is superimposed on the control channel of the hoisting mechanism to suppress vertical velocity during the approach to the installation opening. The goal is to maintain the vertical velocity of the steel anchor box relative to the tower column within ≤5 mm / s when within 0.3 m of the installation opening. The active heave compensation bandwidth covers 0.05~0.22 Hz, effectively suppressing low-frequency heave motion caused by swell.

[0046] Model predictive control solution. A discrete state-space model based on linearized crane kinematics and the Dubins swing approximation is constructed. The control variables u = [ω_slew, v_luff, v_hoist]^T, representing the slewing angular velocity, luffing velocity, and hoisting velocity, respectively. The rolling time domain N = 20, with a sampling period of 50 ms, i.e., the prediction time domain is 1 second. The objective function is:

[0047]

[0048] Where W_e is the state deviation weight matrix and W_u is the control increment weight matrix. The constraints include:

[0049]

[0050]

[0051]

[0052]

[0053]

[0054] The constrained optimization problem is solved using the Sequential Quadratic Programming (SQP) algorithm, with a single solution time of ≤50 ms. Control commands are sent to the crane PLC in a dual-channel redundant manner, with an update frequency of 100 Hz. Each command frame includes a frame ID, timestamp, and CRC checksum.

[0055] V. Segmental gap database and proximity control strategy

[0056] Segment gap database structure. A segment gap database is established to record the key parameters of each segment. Database fields include: segment number, nominal gap g_n, guide surface normal vector n, allowable deviation ±δ_g (standard value 3 mm), contact risk threshold F_lim (set to 50 kN), and minimum normal approach speed v_n,max (set to 4 mm / s). The database is written to the controller's read-only memory before operation to prevent accidental modification. Taking segment 3 as an example, the nominal gap is 34 mm, the guide surface normal vector n = (-0.012, 0.006, 0.99985), the allowable deviation is ±3 mm, and the final residual gap measured at the joint is 35.2 mm, which meets the database constraints.

[0057] Hierarchical control of the proximity buffer. The proximity process is divided into three buffers:

[0058] Buffer 1 (Δz∈[0.30 m, 0.12 m]): The safety weight coefficient is added to the MPC objective function, the state deviation weight is increased by 20%, and the control increment penalty is increased by 10% to improve trajectory smoothness and safety margin.

[0059] Buffer Zone 2 (Δz∈[0.12 m, 0.03 m]): Superimposed normal approximation and tangential limiting control. Calculate the angle ∠(n, z_B) between the normal of the steel anchor box mounting surface and the normal of the guide surface. If the angle > 0.2°, attitude correction is performed first; subsequently, the normal velocity v_n linearly decreases from 3.5 mm / s to 1.5 mm / s, and the tangential velocity v_t is limited to ≤ 2.0 mm / s.

[0060] Approach zone (Δz≤0.03 m): Forced vertical velocity ≤5 mm / s, swing angle ≤0.5°, start the installation mechanism to prepare for bearing. Contact / pressure sensors monitor the contact force in real time. When F≥F_lim is detected, the force limiting and yielding strategy is triggered, the crane reverses and lifts 5 mm and replans the approach path.

[0061] VI. Installation mechanism load-bearing capacity and millimeter-level fine adjustment of the three-way jack

[0062] Loading and unloading process. When Δz ≤ 30 mm, the steel anchor box descends to the three-point support position of the support frame, and the crane maintains a low-speed follow-up state. The load is gradually unloaded from the crane to the support frame at a rate of 10% / min until 90% of the load is borne by the support frame, and the crane only maintains 10% load to maintain system stability. At this time, the residual deviation statistics of the EKF fusion output are: Δx = -1.8 mm, Δy = +2.7 mm, Δz = +3.6 mm; attitude residuals roll = -0.042°, pitch = +0.031°, yaw = -0.015°.

[0063] The three-way jack fine-tuning sequence and command generation. The controller generates fine-tuning commands based on the LVDT displacement sensor and visual residual, executing them in the order of "attitude first, then translation; horizontal first, then vertical":

[0064] 1. Posture adjustment phase:

[0065] Roll correction -0.042°: Jack 1 elongates +0.46 mm, Jack 2 shortens -0.41 mm, Jack 3 remains unchanged;

[0066] Pitch correction +0.031°: Jack 1 extends +0.28 mm, Jack 3 shortens -0.27 mm, Jack 2 remains unchanged.

[0067] 2. Translation and Adjustment Stage:

[0068] X-axis correction -1.8 mm: Jacks 1 and 2 extend synchronously by +0.9 mm, while jack 3 remains unchanged;

[0069] Y-axis correction +2.7 mm: Jack 1 elongates +1.3 mm, Jack 3 shortens -1.4 mm, Jack 2 remains unchanged.

[0070] 3. Vertical adjustment stage:

[0071] Z-axis correction -3.6 mm: Jack shortening in steps 1, 2, and 3 -1.2 mm.

[0072] After each step, a retest was performed at 0.5-second intervals, and the residual deviation was confirmed using a combination of a vision system and an LVDT sensor. After the above three-stage fine-tuning, the final residual deviations were: Δx = -0.3 mm, Δy = +0.5 mm, and Δz = -0.4 mm; the attitude residuals were all ≤0.01°, meeting the requirement that the residual deviation in any direction should not exceed 1 mm. Temporary locking was then completed, and the steel anchor box was fixed by a combination of high-strength bolt pre-tightening and positioning welding.

[0073] Quality verification and 3D structural model update. A rapid verification process combining vibration excitation and visual measurement was performed after temporary locking. The four corners of the steel anchor box were tapped with a light hammer (excitation amplitude approximately 0.2 g, frequency 2050 Hz), and the anchor box's pose response was monitored using a vision system to confirm stable contact and the absence of abnormal vibration modes. Simultaneously, the contact sensor output was checked to verify that the contact pressure distribution was uniform and within the allowable range (3080 kN). A "Segment Installation Quality Confirmation Form" was generated, recording the final pose parameters, residual deviation, contact force distribution, and temporary locking status. The actual installation pose data of the current segment was updated to the 3D structural model, generating a comparison report to provide a precise alignment reference for the installation of the next segment.

[0074] VII. Security Collaboration Strategy and Anomaly Handling Mechanism

[0075] Environmental threshold linkage control. The environmental monitoring system acquires wind speed (10 m high ultrasonic anemometer), significant wave height (shipborne wave radar), and visibility (laser visibility meter) in real time. When any of the following conditions are met: wind speed > 12 m / s, significant wave height > 1.5 m, or visibility < 800 m, the system prevents the issuance of a "descent" control command, displays evacuation path suggestions on the human-machine interface, and prompts the operator to raise the steel anchor box back to a safe height and evacuate the work area. The environmental parameter judgment cycle is 5 seconds; evacuation is triggered only after two consecutive cycles of exceeding the limit to avoid false alarms caused by instantaneous disturbances.

[0076] Sensor health and time synchronization monitoring. The system continuously monitors the number of visible UWB anchors, PTP time synchronization lock status, and self-test results of critical sensors. When the number of visible UWB anchors is less than 4, PTP lockout occurs (clock drift > 2 μs), or self-test failure of critical sensors (camera, IMU, LVDT) occurs, the system switches to a safe shutdown policy within 100 ms, freezing current control commands and allowing only reverse unloading and crane lifting actions. All safety events are logged, including timestamps, trigger conditions, and system response measures, for post-event analysis and improvement.

[0077] Degradation Mode and Manual Takeover. In abnormal situations, the system supports degradation mode operation. For example, when PTP synchronization is lost, the AI ​​prediction module is frozen, MPC switches to a conservative parameter set (weight coefficient adjustment, constraint tightening), no new "descent" commands are issued, only minor attitude corrections and tangential micro-movements are allowed, and the operator is prompted to prepare for manual takeover. When the abnormal conditions are restored (e.g., PTP re-locks) and remain stable for 45 seconds, the system automatically exits degradation mode and resumes normal operation.

[0078] VIII. Data Archiving and Digital Transfer

[0079] Full-process data archiving. Operational data is archived to a local database at a frequency of 10 Hz. Recorded content includes: pose X(t), control command u(t), environmental quantities (wind speed, wave height, visibility), raw sensor data (visual image summary, UWB ranging, IMU inertial increment), quality inspection results (residual deviation, contact force distribution), and alarm events (trigger time, type, response measures). The database uses an append-only write mode, and each record is accompanied by a SHA-256 digital signature to ensure data immutability.

[0080] Segment-level digital archive generation. Upon completion of each segment installation, the system automatically generates an electronic archive, including a PDF-format "Segment Installation Quality Confirmation Form" and a CSV-format time-series data file. The archive content covers: segment number, hoisting start and end times, environmental condition statistics, posture trajectory curves, control command history, AI prediction error statistics, fine-tuning process records, and final quality assessment results. After all segments are installed, a digital handover package is formed, corresponding to "segment—trajectory—control—quality," for data traceability during final acceptance and operation and maintenance phases.

[0081] IX. Engineering Effects and Comparative Analysis

[0082] This embodiment was applied in the installation of steel anchor boxes for the main tower of a cross-sea cable-stayed bridge, completing the hoisting and fine-tuning of 12 segments. The results are compared with a control segment in an adjacent construction window of the same bridge that used the traditional method of "total station + manual adjustment," as follows:

[0083] Approach zone time: reduced from 18.5 min / section using the traditional method to 7.4 min / section, an efficiency improvement of 60%;

[0084] Number of repeated lowering / retractions: reduced from 2.1 times / section in the traditional method to 0 times, avoiding repetitive work;

[0085] The final residual pose error was stably controlled to ≤1 mm, while the average error of the traditional method was 3.2 mm.

[0086] The crane's swing angle has been reduced from 1.6° using traditional methods to 0.5°, improving operational safety.

[0087] Total offshore operation time: The total time for the 12 segments has been reduced by approximately 134 minutes, significantly reducing the risk of "missing the window".

[0088] The above results verify the significant advantages of the method of the present invention in improving hoisting accuracy, shortening operation time, and reducing construction risks.

[0089] Example 2: Parameter Tuning and Robust Operation in a Low-Wind, High-Inrush Environment at Night

[0090] This embodiment focuses on the hoisting operation of steel anchor boxes under nighttime sea conditions. The nighttime environmental characteristics are: wind speed 68 m / s, significant wave height 1.2-1.6 m, and visibility ≥1500 m (relying on artificial lighting systems). Since the energy of swells at night is mainly concentrated in the low-frequency range (0.05~0.15 Hz), higher requirements are placed on the heave compensation algorithm.

[0091] Parameter adjustment strategy. To suppress low-frequency heave caused by surges, the active heave compensation bandwidth is extended from the standard configuration of 0.15 Hz to 0.22 Hz, achieved by adjusting the high-pass filter parameters. In the MPC objective function, the weight of the safety term is increased by 30%, and the control increment penalty is increased by 20% to enhance trajectory smoothness and disturbance resistance. The upper limit of the vertical velocity in the approach zone is reduced from 5 mm / s to 4 mm / s to further improve the safety margin.

[0092] Online adaptive training. The LSTM prediction model was fine-tuned online using 40 minutes of real-world data within a daily window. The learning rate was set to 1×10^-4, the batch size to 64, and only the weights of the last layer were updated to avoid overfitting. After adaptive training, the RMSE of Δz in the 1-second prediction time domain decreased from the initial 3.4 mm to 2.6 mm, and the prediction accuracy improved by approximately 24%.

[0093] Implementation Results. During the nighttime installation of section 6, no contact alarms occurred within the approach zone of 0.3 m. The final residual deviations were Δx = +0.6 mm, Δy = -0.4 mm, and Δz = +0.7 mm, with attitude residuals all ≤0.01°, meeting the 1 mm accuracy target. The operation time increased slightly by 12% compared to the standard daytime conditions, but it was still significantly better than the traditional process, verifying the robustness of this method under complex sea conditions.

[0094] Example 3: Verification of Abnormal Situation Handling and Degradation Strategy

[0095] Handling PTP Time Synchronization Loss. During operation in the approach zone of segment 8, electromagnetic interference caused PTP time synchronization to lose lock, with clock drift exceeding 2 μs. The system detected the anomaly within 80 ms and triggered a degraded mode: freezing the AI ​​prediction module output, switching MPC to a conservative parameter set (increasing state deviation weight by 50% and reducing the upper limit of control increment by 30%), not issuing new "descent" commands, and only allowing minor attitude corrections and tangential micro-motions. Simultaneously, the human-machine interface prompted the operator to prepare for manual takeover. After 45 seconds, PTP re-locked, the system automatically exited degraded mode, and resumed normal operation. The final residual error for this segment was ≤1 mm, verifying the effectiveness of the degraded strategy.

[0096] Insufficient UWB visible anchor points: Due to tower wall obstruction, the number of visible UWB anchor points dropped to 3 at a certain moment, below the normal operating threshold of 4. The system maintained a vision + IMU open-loop compensation mode, limiting the crane's movement speed (slewing, luffing, and hoisting speeds all reduced by 50%) and swing angle (≤0.3°), relying on vision and IMU fusion to maintain pose estimation accuracy. After 5 seconds, as the steel anchor box position adjusted, the number of visible anchor points recovered to ≥4, and the system automatically returned to normal operating mode. No contact force limit alarm was triggered throughout the process, verifying the redundancy and fault tolerance of the multi-source fusion algorithm.

[0097] Example 4: Establishment and Verification of Segmental Gap Database

[0098] Based on the detailed design drawings and statistical analysis of tower wall manufacturing tolerances, a segment gap database was established. Segments 2, 4, 7, and 11 were selected as sampling segments, and the geometric dimensions of the steel anchor boxes and the normal vector of the guide surfaces were measured using a coordinate measuring machine from the manufacturer, with an accuracy of ±0.05 mm. Simultaneously, a finite element simulation software was used to establish a tower column-steel anchor box contact model to verify the nominal gap and contact stress distribution, thus validating the rationality of the database parameters.

[0099] After the first two sections were installed, the measured residual clearance data (35.0 mm for the first section and 36.8 mm for the second section) was written back to the database to fine-tune the nominal clearance parameters (adjustment not exceeding ±1 mm) for optimizing the proximity control strategy of subsequent sections. Through the above closed-loop correction mechanism, the proximity control accuracy of subsequent sections was further improved, and the average proximity time was reduced by 8%.

[0100] Example 5: Hardware and Software Scheduling and Real-Time Guarantee

[0101] The system adopts a hierarchical scheduling architecture, with each functional module running at a fixed cycle to ensure that end-to-end control latency meets requirements. The specific scheduling cycle is as follows:

[0102] Visual processing: 16.7 ms (60 fps)

[0103] Multi-source fusion: 20 ms (EKF filter period)

[0104] AI prediction: 50 ms (LSTM forward inference)

[0105] MPC solution time: 50 ms (SQP optimized algorithm)

[0106] Jack control: 10 ms (servo control cycle)

[0107] Executor management: ≤5 ms (instruction packaging and distribution)

[0108] The end-to-end control latency is ≤80 ms, meeting the requirement of not exceeding the set latency. The above scheduling is implemented under the RT-Preempt real-time kernel, with a peak CPU utilization of 62% and a peak memory utilization of 1.8 GB, leaving sufficient margin to cope with sudden computing loads.

[0109] Example 6: First Section Installation and Reuse of Reference Standards for Subsequent Sections

[0110] After the first steel anchor box is installed, the mounting surface of the first section and the reference coordinate system of the tower column are used as the alignment reference for subsequent sections. Before the installation of each section, a rapid recalibration process for camera extrinsic parameters is performed: using the marker target on the mounting surface of the first section as known control points, the camera extrinsic parameters (rotation matrix and translation vector) are re-solved using the PnP algorithm, with a calibration time of ≤3 min. At the same time, a UWB health check is performed, including anchor-tag ranging accuracy verification (standard deviation ≤7 cm) and GDOP evaluation (≤2.6), to ensure that the positioning network performance meets the requirements.

[0111] The aforementioned rapid recalibration and health check mechanism ensures the continuity and reliability of each segment's installation, avoiding cumulative errors caused by reference drift or sensor performance degradation. Experimental results show that during the installation of the 12 segments, the camera extrinsic parameter drift was ≤0.3 mm, and the UWB positioning accuracy remained stable, verifying the effectiveness of the reference reuse strategy.

[0112] Example 7: Engineering Implementation and Quality Acceptance

[0113] Workflow coordination. Each segment follows a single-line process: "Marine transportation → Spreading gear connection → Benchmark verification → Precision offshore lifting → Proximity control → Installation mechanism bearing → Millimeter-level fine-tuning → Temporary locking → Quality confirmation → Data archiving." Electronic workflow sheets are used between processes to ensure that each step is completed and passes quality inspection before proceeding to the next.

[0114] Acceptance Criteria and Results. According to the project quality plan, the acceptance criteria for the steel anchor box installation quality are: positional error ≤ 1mm, guide surface normal angle ≤ 0.05°, uniform contact force distribution within limits (30~80 kN), and the quality of the anti-corrosion coating and welds is inspected according to established standards. All 12 segments met the above acceptance criteria, achieving a 100% first-time acceptance pass rate. No rework or repairs were required, verifying the engineering reliability and practical value of the method of this invention.

[0115] Example 8: Comparison and Performance Verification of Artificial Intelligence Algorithm Models

[0116] To verify the performance of different artificial intelligence algorithms in predicting the hoisting trajectory of steel anchor boxes, this embodiment compares three models: LSTM, GRU (Gated Recurrent Unit), and TCN (Temporal Convolutional Network).

[0117] Table 2 Comparison of Prediction Performance of Different AI Models

[0118] Model Parameters Training time (min) 1s Prediction Δz RMSE (mm) 1s Prediction Δx RMSE (mm) 1s Prediction Δy RMSE (mm) Inference time (ms) LSTM 128K 45 2.9 2.4 2.7 48 GRU 96K 38 3.1 2.6 2.8 42 TCN 156K 52 2.7 2.3 2.6 55

[0119] The results show that TCN is slightly better than LSTM in prediction accuracy, but has a longer inference time; GRU has the fewest parameters and the fastest inference speed, but its accuracy is slightly lower. Considering both accuracy and real-time requirements, this invention prefers LSTM as the default model, while also supporting model replacement according to specific engineering needs, demonstrating the flexibility and scalability of the method.

[0120] Example 9: Standardization of Multi-Source Data Interfaces and Data Security Verification

[0121] The system adopts standardized data interfaces and security mechanisms. The communication protocol uses TCP over Gigabit Ethernet, and the message structure uses Protocol Buffers for serialization. The fields include: timestamp (64-bit integer, nanosecond precision), pose data (6 floating-point numbers), control commands (3 floating-point numbers), environmental variables (3 floating-point numbers), sensor status (Boolean array), and checksum (SHA-256 hash value).

[0122] Data transmission employs TLS 1.3 encryption and two-way certificate authentication, with access control based on a whitelist policy. In data security verification experiments, simulating man-in-the-middle attacks, replay attacks, and data tampering attacks, the system correctly detected and rejected illegal data packets, verifying the effectiveness of the data security mechanism. End-to-end message transmission latency is ≤5 ms, and the packet loss rate is <0.01%, meeting real-time control requirements.

[0123] Through the detailed description of the above nine embodiments, the technical solutions of the present invention, "A method for hoisting steel anchor boxes with variable spacing for cable-stayed bridge towers across the sea," are fully demonstrated in terms of real-time perception, intelligent prediction, precise control, safety assurance, and data management, providing an operable implementation path and verification basis for engineering applications.

Claims

1. A method for hoisting variable-spacing steel anchor boxes for cable-stayed bridge towers across the sea, characterized in that, Includes the following steps: (1) Construction of multi-source sensing system and unification of spatiotemporal reference: Establish a reference coordinate system for the tower column, and deploy ultra-wideband positioning anchor points, machine vision target markers, differential satellite positioning stations, inertial measurement units and industrial cameras on the tower column and floating crane to complete system calibration and clock synchronization; (2) Real-time pose fusion calculation: Collect and fuse machine vision, ultra-wideband, satellite positioning and inertial measurement data to accurately calculate the real-time six-degree-of-freedom pose deviation of the steel anchor box relative to the tower column installation port; (3) Dynamic trajectory prediction and compensation control: Based on artificial intelligence algorithms, predict the short-term evolution trend of sea conditions, ship motion and hoisting status, and combine model prediction control to generate linkage compensation commands for crane rotation, luffing and lifting to actively suppress environmental disturbances; (4) Near-field guidance and collision avoidance: During the hoisting approach phase, the trajectory micro-correction and collision avoidance amplitude limit control are performed based on the preset segment gap database; (5) Millimeter-level fine adjustment and locking: After the steel anchor box is carried by the installation mechanism, the millimeter-level precise alignment is performed based on the real-time residual deviation through the three-way hydraulic jack system, and temporary locking is completed; (6) Full-process monitoring and data archiving: The working environment, equipment status and control process are continuously monitored, safety threshold linkage protection is set, and the full-process data is digitally recorded and archived.

2. The method according to claim 1, characterized in that: The construction of the multi-source sensing system in step (1) specifically includes: setting up ultra-wideband positioning anchor points at non-coplanar positions around the tower column to ensure positioning geometric accuracy; setting up machine vision target markers at the installation openings of the steel anchor box and the tower column, and configuring industrial cameras with cross-view angles for accurate visual measurement.

3. The method according to claim 1, characterized in that: The real-time pose fusion calculation in step (2) uses the extended Kalman filter algorithm to fuse and calculate the pose information provided by machine vision, ultra-wideband ranging information, and differential satellite positioning and inertial measurement data to obtain high-precision, high-frequency pose deviation estimation.

4. The method according to claim 1, characterized in that: The dynamic trajectory prediction and compensation control in step (3) uses a long short-term memory network (LSTM) for trajectory prediction and model predictive control (MPC) to solve for the optimal control command. The command aims to minimize the pose deviation and sling swing angle and is subject to the performance and safety constraints of the crane.

5. The method according to claim 1, characterized in that: Between steps (3) and (4), an active heave compensation step is further included, which is to superimpose equal-amplitude and opposite-axis compensation commands on the hoisting mechanism based on the vertical displacement and acceleration of the hull measured by the inertial measurement unit, so as to suppress the vertical fluctuations caused by the waves.

6. The method according to claim 1, characterized in that: The segment gap database in step (4) records the nominal gap, guide surface normal vector and allowable deviation of each segment; the near-field guidance control executes the normal approximation and tangential amplitude limiting strategy based on this database, and triggers safety retreat when the contact force exceeds the limit.

7. The method according to claim 1, characterized in that: The three-way hydraulic jack system in step (5) consists of three independent hydraulic jacks with high-precision displacement and force sensors, forming a three-point leveling mechanism to achieve six-degree-of-freedom fine adjustment of the steel anchor box.

8. The method according to claim 1, characterized in that: The millimeter-level fine-tuning in step (5) is performed step by step in the order of first adjusting the posture, then adjusting the translation, first the horizontal and then the vertical, until the residual deviation in each direction is less than the preset threshold, and then jointly verified by visual and contact sensors.

9. The method according to claim 1, characterized in that: The safety threshold linkage protection in step (6) includes environmental thresholds such as wind speed, effective wave height and visibility, as well as system thresholds such as sensor health and time synchronization status; when any threshold exceeds the limit, the system automatically triggers the corresponding safety strategy.

10. The method according to claim 1, characterized in that: After the installation of the first section or any segment is completed, its final installation position is used as the alignment reference for subsequent segments, and the sensor system is quickly recalibrated before the installation of the next segment. After the installation of each segment is completed, vibration verification is performed to verify the installation quality, and a digital archive containing data from the entire process is generated.

Images