Safety monitoring system of high transverse supporting system for cable-stayed bridge man-shaped tower column construction

Abstract

The invention discloses a safety monitoring system of a high transverse support system for cable-stayed bridge man-shaped tower column construction, and relates to the technical field of bridge construction monitoring, the system comprises a creeping formwork integrated sensing module used for generating a point cloud model by using a mechanical arm integrated on a hydraulic creeping formwork platform and a laser radar scanning tower column curved surface, combining with a preset building information model coordinate, adaptively adjusting the mounting posture of the sensor, and outputting the mounting coordinate position of the sensor; the optical fiber sensing network module is used for collecting original strain and temperature data by using a distributed optical fiber sensor deployed along a main stress path of the support truss; an inertial navigation fusion positioning module; a digital twinning early warning module; and an edge calculation relay module. According to the invention, a full-process monitoring chain from data acquisition to risk early warning is constructed through cooperation of multiple modules, full-period and multi-dimensional dynamic control of construction of the human-shaped tower column high transverse support system is realized, and timely perception and overall control of potential risks are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bridge construction monitoring, in particular to a safety monitoring system for a high transverse support system of a cable-stayed bridge human-shaped tower column construction. BACKGROUND

[0002] The cable-stayed bridge human-shaped tower column has a unique shape with double limbs inclined, and has structural mechanics rationality and landscape, and is widely used in long-span bridges. Its construction process needs to go through high-altitude cantilever pouring, and under the action of tower column self weight, construction load and wind load, the double limbs are easy to produce transverse deformation, which poses a significant challenge to construction safety. The high transverse support system, as a key temporary structure in the construction stage, connects the double limb tower column through cross bracing, support and other components, bears the transverse restraint action, balances the deformation stress, and is the core facility to ensure the stability of the tower column construction.

[0003] Safety monitoring is crucial for the high transverse support system, and its significance not only lies in real-time capturing of the coordinated deformation of the tower column and the support, but also in early warning of potential failure paths. For example, local thermal expansion of the tower column caused by solar radiation temperature difference may cause additional stress on the support rod; and if the stress redistribution of the support system is not discovered in time during the climbing process of the construction segment, it may cause critical instability.

[0004] Safety monitoring of the high transverse support system is an important means to prevent structural instability and avoid collapse accidents. During construction, the support system needs to withstand dynamic load, temperature gradient and material aging, etc., and real-time grasp of its stress state and deformation trend is directly related to the safety of construction personnel and the progress of the project.

[0005] In the prior art, traditional monitoring often uses strain gauges, total stations and other equipment to collect data by manually laying sensors. However, there are obvious limitations in practice, such as temperature changes easily coupled with strain signals, leading to distorted data, and virtual strain caused by solar temperature variation often masks the real mechanical response, leading to misjudgment; insufficient multi-source data fusion, monitoring data scattered in different devices, lack of real-time fusion and dynamic mapping, difficult to fully reflect the overall state of the support system, lack of intelligent prediction mechanism; early warning is mostly based on threshold judgment, and the prediction ability for potential risks is insufficient, making it difficult to avoid gradual failure in time.

[0006] These defects make it difficult for existing monitoring methods to meet the high-precision, real-time safety control needs of the human-shaped tower column high transverse support system construction, and it is necessary to build a more intelligent and integrated monitoring system. SUMMARY

[0007] Therefore, it is necessary to provide a safety monitoring system for a high transverse support system of a cable-stayed bridge human-shaped tower column construction in view of the above technical problems.

[0008] The application provides a safety monitoring system of a high transverse support system for a cable-stayed bridge human-shaped tower column construction. The climbing form integrated sensing module is used for generating a point cloud model by using a mechanical arm integrated in a hydraulic climbing form platform and a laser radar scanning tower column curved surface, combining a preset building information model coordinate, adaptively adjusting a sensor installation posture, and outputting an installation coordinate position of the sensor; The optical fiber sensing network module is used for collecting original strain and temperature data by using distributed optical fiber sensors arranged along a main stress path of the support truss, dynamically correcting the original strain and temperature data by using an intelligent compensation algorithm, and outputting strain field distribution data after temperature compensation; The inertial navigation fusion positioning module is used for monitoring tower column deformation in real time based on satellite signals and gyroscope angular velocities collected by inertial sensors, and generating three-dimensional deformation data; The digital twin early warning module is used for fusing strain field distribution data and three-dimensional deformation data, using a digital twin model to map stress and displacement fields of the support system in real time, dynamically rendering deformation trends, predicting structural failure risks of the tower column, and outputting risk prediction results; The edge computing relay module is used for building a hybrid transmission network, acquiring original sensing signals and early warning instructions collected by the sensors, and outputting denoised structured data and real-time early warning signals.

[0009] Further, the climbing form integrated sensing module comprises: The mechanical arm control module is used for calculating a motion trajectory and a control signal of an end effector of the mechanical arm in real time based on preset mechanical arm joint parameters and target movement coordinates, so as to realize mechanical arm movement; The laser scanning module is used for acquiring a three-dimensional appearance of a tower column surface in real time by using a laser radar, extracting effective point cloud data, and constructing a point cloud model of the tower column by point cloud registration and reconstruction; The coordinate matching module is used for aligning the point cloud model actually scanned and built with a preset digital model, planning a target installation path of the mechanical arm according to target installation positions of a plurality of sensors, deploying the sensors at stress points of the tower column in sequence, and acquiring actual installation coordinate positions of the sensors; The posture adjustment module is used for adaptively calibrating the installation posture of the sensor based on a curved surface method, keeping the sensor adhering to a normal line direction, and calculating an installation inclination deviation value of the sensor; The installation verification module is used for triggering secondary scanning after the sensor is installed, comparing distance deviations of the installation coordinate position of the sensor and the target installation coordinate and installation inclination deviation values and a preset inclination deviation threshold value, and verifying installation precision of the sensor.

[0010] Further, the optical fiber sensing network module comprises: The photoelectric conversion conditioning module is configured to convert the backscattering light signal into an electrical signal, and perform noise reduction and pre-processing, and output raw strain data, raw temperature data, and a signal-to-noise ratio index. The temperature-strain decoupling module is configured to separate the coupling effect of temperature and strain in the optical fiber measurement based on the raw strain data and the raw temperature data by using a pre-constructed transfer learning decoupling model, and obtain independent strain data and temperature data. The dynamic temperature compensation module is configured to correct the strain data in real time based on an environmental temperature gradient, eliminate a virtual increase error caused by a sunshine temperature difference, and obtain a temperature-compensated strain field distribution. The quality evaluation output module is configured to package the compensated strain field distribution data into a standard format, and transmit the compensated strain field distribution data to a digital twin model through a hybrid network.

[0011] Further, the temperature-strain decoupling module is configured to separate the coupling effect of temperature and strain in the optical fiber measurement based on the raw strain data and the raw temperature data by using a pre-constructed transfer learning decoupling model, and obtain independent strain data and temperature data. The raw strain data and the raw temperature data are subjected to time-space alignment and feature enhancement to generate standardized input suitable for transfer learning, and a multi-dimensional feature vector is constructed as model input. A transfer learning decoupling model based on a bidirectional long short-term memory network is obtained in a source domain, a bottom feature extraction layer is frozen, and a top regression layer is replaced to adapt to the output dimension of a target domain, a multi-dimensional feature vector is input, and decoupled strain data and temperature data are output. A maximum mean difference loss is added after the feature extraction layer by using known labeled data in the target domain, the distribution difference between the source domain and the target domain is minimized, and the decoupling residual error of the strain data and the temperature data is calculated, and if the decoupling residual error is greater than a preset threshold, the model is recalibrated. For unlabeled data in the target domain, a standard confidence pseudo-label is generated by using the trained transfer learning decoupling model, and the model capability is iteratively optimized through self-supervised learning.

[0012] Further, the dynamic temperature compensation module is configured to correct the strain data in real time based on an environmental temperature gradient, eliminate a virtual increase error caused by a sunshine temperature difference, and obtain a temperature-compensated strain field distribution. Based on the decoupled strain data and the temperature data, a time sequence feature vector including a temperature gradient, a strain change rate, and a historical temperature trajectory is constructed, and an input time sequence is generated. A dynamic compensation model based on a time re-attention long short-term memory network is constructed by learning the time lag mapping relationship between the temperature field and the strain field through an encoder-decoder framework, and a strain prediction value is output by combining the current temperature; a thermal hysteresis compensation amount is calculated by using the deviation between the strain prediction value and the actual independent strain data, and a compensated strain field is dynamically generated by combining the temperature gradient. Based on the compensated strain residual and temperature gradient, a spatially distributed compensation coefficient matrix is dynamically generated, and combined with confidence evaluation, the dynamic compensation model is online learned and incrementally optimized.

[0013] Further, the time delay mapping relationship between the temperature field and the strain field is learned through the encoder-decoder framework, and a dynamic compensation model based on the time re-attention long short-term memory network is constructed, including: The historical temperature sequence is encoded into a hidden state vector using the encoder layer, the time attention module focuses on the historical labeled time, and the decoder layer predicts the strain compensation amount based on the current temperature; An entropy-based gating unit is added in the loss layer to compare the deviation between the predicted strain value and the actual strain data, and the attention weight of the decoder is corrected in reverse.

[0014] Further, the digital twin early warning module includes: A multi-source heterogeneous integration module is used to integrate strain field distribution data, three-dimensional deformation data, environmental temperature and humidity data, and load data, to realize spatio-temporal alignment and quality verification of multi-source heterogeneous data; A digital twin modeling module is used to construct a multi-scale digital twin model of the support system, establish a dynamic mapping between the physical entity and the virtual space, and generate a stress field-displacement field simulation calculation environment; A physical double-drive analysis module is used to couple the physical mechanism model and the deep learning algorithm, identify stress concentration areas and displacement mutation points, quantify the structural failure risk probability of the tower column, and locate the risk position; A dynamic risk deduction module is used to simulate the structural deformation trend in the tower column based on real-time twin data, predict the critical failure time and damage path, and generate a visual warning scenario; An early warning decision generation module is used to generate graded warning instructions based on the deduction results, drive human-machine collaborative response strategies, and feedback optimized digital twin model parameters.

[0015] Further, the coupling of the physical mechanism model and the deep learning algorithm, the identification of the stress concentration area and the displacement mutation point, the quantification of the structural failure risk probability of the tower column, and the positioning of the risk position include: The real-time stress field and displacement field output by the digital twin model are obtained, the risk candidate area is divided, and an independent risk positioning agent is initialized for each candidate area, and the initial resource weight is allocated; Based on the material fatigue characteristics and historical failure data, the game rules and reward mechanism between agents are set, a dynamic strategy space is constructed, and the agent strategy network is trained using a hierarchical architecture, and physical constraints consistent with the physical laws of material failure are introduced; Real-time monitoring of resource competition behavior between agents, dynamically adjusting agent resource allocation weight, eliminating local deadlock, and combining local risk prediction results of each agent to generate a cooperative positioning matrix and output the risk position coordinates of the tower column with abnormality; Based on the multi-agent game result, the structural failure probability of the tower column is quantified, and the positioning result is matched with the historical case to dynamically update the risk knowledge base.

[0016] Further, the dynamic strategy space includes: abnormal risk area agents occupying computing resources during the game process; non-abnormal risk area agents cooperating with each other during the game process; and a reward function that combines stress mutation gradient and displacement offset for quantifying game benefits; The hierarchical architecture adopts a federated learning framework, the local layer uses a Q-value algorithm to learn the target action strategy, and the federated layer aggregates global policy parameters and generates a shared knowledge base.

[0017] Further, based on real-time twin data, the structural deformation trend in the tower column is simulated, the critical failure time and damage path are predicted, and a visual warning scenario is generated, including: Based on the structural failure probability and risk position of the tower column, a virtual deduction environment is constructed based on the digital twin model, taking the risk unit as the agent, the agent state space and action space are set, and the game rule base of each agent is initialized; The controller coordinates the behavior of multiple agents, receives the state of all agents in real time, resolves action conflicts, generates a globally known damage path topology graph, and outputs a corrected deformation field; Based on the action sequence of the agent and the time-varying strength change rule of the material, the critical failure time is calculated, and a three-dimensional engine is integrated to convert the deduction process into a three-dimensional dynamic scene.

[0018] The beneficial effects of the present application are: 1. A full-process monitoring chain from data acquisition to risk warning is constructed through multi-module cooperation, realizing full-cycle, multi-dimensional dynamic control of the high transverse support system construction of the human-shaped tower column; including accurate deployment of sensors to dynamic correction of environmental interference, and high-precision acquisition of deformation data to intelligent deduction of structural risk, each monitoring link is organically linked, breaking the fragmentation limitations of traditional monitoring, and fully covering the stress state, deformation trend and environmental impact of the support system, providing systematic technical support for construction safety, ensuring timely perception and overall control of potential risks.

[0019] 2、The application effectively solves the monitoring interference problem in complex environment, especially in the processing of temperature and strain coupling effect, realizes dynamic adaptation and accurate correction of environmental gradient change, ensures the reliability of strain data, avoids monitoring distortion caused by environmental interference, provides scientific data support for accurately identifying stress concentration of support system and evaluating structure health state, and significantly improves the credibility and application value of monitoring results.

[0020] 3、The application relies on multi-source data fusion and intelligent deduction technology, realizes advanced identification and visual warning of structure risk, accurately locates risk area through coupling physical mechanism and intelligent algorithm, simulates deformation trend and predicts critical state, provides intuitive basis for construction decision, helps to take intervention measures in advance, reduces structure failure risk from the source, and effectively ensures the safety and stability of high transverse support system in construction process. BRIEF DESCRIPTION OF DRAWINGS

[0021] The drawings described herein are used to provide further understanding of the application, constitute a part of the application, the illustrative embodiments of the application and the description thereof are used to explain the application, and do not constitute improper limitation on the application. In the drawings: Figure 1 It is a system principle block diagram of a safety monitoring system of a high transverse support system of a cable-stayed bridge human-shaped tower column construction according to an embodiment of the application; Figure 2 It is a local structure schematic view of a hydraulic climbing formwork platform according to an embodiment of the application; Figure 3 It is a junction section support of a human-shaped tower column and a support system according to an embodiment of the application; Figure 4 It is a cable tower construction step schematic view of a human-shaped tower column according to an embodiment of the application; Figure 5 It is one of human-shaped tower column construction schematic views according to an embodiment of the application; Figure 6 It is the second human-shaped tower column construction schematic view according to an embodiment of the application; Figure 7 It is the third human-shaped tower column construction schematic view according to an embodiment of the application; Figure 8 It is the fourth human-shaped tower column construction schematic view according to an embodiment of the application; Figure 9 It is the fifth human-shaped tower column construction schematic view according to an embodiment of the application; Figure 10 It is the sixth human-shaped tower column construction schematic view according to an embodiment of the application.

[0022] Reference signs: 1, climbing form integrated sensing module; 2, optical fiber sensing network module; 3, inertial navigation fusion positioning module; 4, digital twin early warning module; 5, edge computing relay module. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0024] Please refer to Figure 1 , a safety monitoring system for a high lateral support system of a cable-stayed bridge human-shaped tower column construction is provided, and the system comprises: The climbing form integrated sensing module 1 is used to generate a point cloud model by using a mechanical arm integrated in a hydraulic climbing form platform and a laser radar scanning tower column curved surface, and to self-adaptively adjust a sensor installation posture in combination with a preset building information model coordinate, and to output an installation coordinate position of the sensor.

[0025] The hydraulic climbing form platform (also referred to as a climbing form system or a climbing form platform) is a core equipment in the construction of the cable-stayed bridge human-shaped tower column, and the structural design thereof directly serves the high-altitude pouring and shape control of the tower column. As shown in Figure 2 , a local schematic view of the hydraulic climbing form platform, the climbing form platform is closely attached to the outside of the tower column through a multi-layer nested structure: the main body of the outer layer frame is divided into-4 layers to 0 layers of functional platforms (including a steel bar operation layer, a formwork operation layer, a bearing layer and a curing operation layer), which are connected to the embedded nodes through triangular support frames and climbing poles to form a stable frame support system. The hydraulic climbing form platform is climbed along the tower column curved surface layer by layer through a hydraulic jacking mechanism, which adapts to the inclination angle and cross-section shrinkage of the human-shaped tower column. When the tower column transitions at the bifurcation, the east-west climbing frame dynamically adjusts the inclination angle through a steel wedge block to ensure that the multi-layer formwork accurately fits the concrete curved surface and synchronously supports the load transfer of the layered pouring.

[0026] The hydraulic system of the hydraulic climbing form platform alternately jacks the guide rail and the frame through the oil cylinder, so that the platform is climbed layer by layer along the embedded climbing cone, and continuous construction can be realized without external hoisting equipment.

[0027] In the construction of the cable-stayed bridge human-shaped tower column, the hydraulic climbing form platform needs to adapt to the inclined shape and spatial bifurcation structure of the tower column. The human-shaped tower column usually presents an "A" shape or a "Y" shape with the two columns inclined and folded, the inclination angle can be more than 70°, and the cross-sectional size changes with the height. The climbing form platform deals with such complexity through modular design: for example, at the bifurcation of the tower column, the east-west and upstream and downstream climbing frames are combined into a whole frame body through temporary connection, and the inclination angle is gradually adjusted through a steel wedge block during the climbing process, and finally converted to a vertical state to enter the upper tower column construction.

[0028] In the description of the present application, the climbing form integrated sensing module 1 comprises: a mechanical arm control module, a laser scanning module, a coordinate matching module, a posture adjustment module and an installation verification module.

[0029] The mechanical arm control module is used for real-time calculation of the motion trajectory and control signal of the end effector of the mechanical arm based on the preset mechanical arm joint parameters and the target movement coordinates, so as to realize the motion of the mechanical arm.

[0030] Specifically, the mechanical arm control module drives the motion of the mechanical arm on the hydraulic climbing form platform through an embedded real-time control system (such as STM32 or PLC). The system presets the mechanical arm joint parameters and the target installation coordinates, the mechanical arm joint parameters including the rotation range, speed curve and torque limit of each joint; the motion trajectory of the end effector is solved in real time based on the inverse kinematics algorithm, and a pulse width modulation (PWM) signal is generated to control the action of the servo motor.

[0031] During the motion process, the joint angle is fed back in real time through the rotary encoder, and the trajectory deviation is dynamically corrected by combining the PID controller to ensure the positioning accuracy. At the same time, a wireless communication module (such as 4G / WiFi) is integrated to receive remote instructions, supporting manual intervention or emergency braking.

[0032] The laser scanning module is used for real-time acquisition of the three-dimensional topography of the tower column surface by using a laser radar, extraction of effective point cloud data, and construction of a point cloud model of the tower column through point cloud registration and reconstruction.

[0033] Specifically, the laser scanning module uses a phase laser radar to perform 360° ring scanning on the tower column curved surface. The laser beam is emitted at a frequency of ≥100Hz, the distance is calculated by the time difference of the received reflection signal, and the original point cloud data is generated by combining the scanning angle (horizontal angle and vertical angle) recorded by the built-in gyroscope. The outlier elimination algorithm based on statistical filtering is used to remove noise points, and the multi-view point cloud is registered as a complete model through the iterative closest point (ICP) algorithm, and finally a high-precision three-dimensional point cloud of the tower column surface with a density of ≥500 points / ㎡ is output.

[0034] The coordinate matching module is used for aligning the point cloud model actually scanned and built with the preset digital model, planning the target installation path of the mechanical arm according to the target installation positions of a plurality of sensors, deploying the sensors on the stress points of the tower column in turn, and obtaining the actual installation coordinate positions of the sensors.

[0035] Specifically, the coordinate matching module aligns the scanned point cloud with the preset BIM (Building Information Model) model through a rigid transformation matrix (rotation matrix R and translation vector T). First, the feature corner points in the point cloud and the corresponding point set of the BIM model are extracted, and the optimal transformation parameters are solved by using the least squares method to realize the registration with a model space error of ≤3mm.

[0036] Subsequently, based on the sensor target position marked in the BIM model (such as the stress concentration area of the tower column), an A* path planning algorithm is used to calculate the obstacle avoidance moving path of the mechanical arm, and a motion sequence containing the path key node coordinates and joint angles is output to drive the mechanical arm to reach the target position in turn.

[0037] The posture adjustment module is used for adaptively calibrating the sensor installation posture based on the curved surface method, keeping the sensor adhering to the normal direction, and calculating the installation inclination deviation value of the sensor.

[0038] Specifically, the posture adjustment module dynamically calibrates the sensor installation posture according to the geometric characteristics of the tower column curved surface. By calculating the local curved surface normal vector of the sensor target position in the point cloud model, the Euler angle (pitch angle a, yaw angle β) that needs to be adjusted by the end effector of the mechanical arm is solved and converted into a rudder control instruction. The six-axis force sensor carried by the end effector monitors the contact pressure between the sensor and the tower column surface in real time, and triggers the posture fine adjustment when the pressure distribution is uneven, ensuring that the deviation between the sensor axis and the curved surface normal direction is maintained at a certain level. After adjustment, the actual installation inclination is recorded and used as a calibration reference.

[0039] The installation verification module is used for triggering a secondary scan after the sensor is installed, comparing the distance deviation between the actual installation coordinate position of the sensor and the target installation coordinate, and the installation inclination deviation value and the preset inclination deviation threshold, and verifying the installation accuracy of the sensor.

[0040] Specifically, the installation verification module starts a secondary laser scan after the sensor is installed, and extracts the actual contour point cloud of the sensor. By using a point cloud registration algorithm, the actual contour is aligned with the theoretical contour of the sensor in the BIM model, and the spatial distance deviation Δd and the axis angle deviation Δθ of the two mass centers are calculated.

[0041] If Δd or Δθ is out of limit, an alarm is triggered and a deviation correction vector (including translation direction and rotation angle) is output to guide the mechanical arm to readjust the installation position; if the verification is qualified, the actual installation coordinate and inclination are written into the digital twin model as the spatial reference for subsequent monitoring data.

[0042] In addition, it should be noted that the climbing form integrated sensing module 1 adaptively adjusts the sensor installation posture by combining the preset building information model coordinates, and outputs the installation coordinate position of the sensor. The sensor includes the distributed optical fiber sensor involved in the optical fiber sensing network module 2 and the inertial sensor involved in the inertial navigation fusion positioning module 3, as well as other monitoring sensors that can be installed and deployed according to actual scenes, which realize the all-around monitoring of the humanoid tower column and its supporting system through cooperation.

[0043] Wherein, the distributed optical fiber sensor (such as fiber Bragg grating FBG) is laid along the main load path of the support truss, such as the stringer of the Bailey truss, the node plate. The strain and temperature change are sensed by the reflection light center wavelength shift (1525nm~1565nm), the range is 300pm~20mm, and the accuracy is ±0.5%FS. Its armored cable packaging can withstand a temperature difference of-30~80℃ and chemical corrosion, and combined with the double gate structure to isolate the temperature-strain coupling effect, to provide the original strain field data for the optical fiber sensing network module.

[0044] The inertial sensor includes a high-precision MEMS inclinometer and a fiber-optic gyroscope, which is the core of the inertial navigation fusion positioning module combined with the GNSS receiver. The gyroscope measures angular velocity based on the Sagnac effect, and the accelerometer detects three-axis linear displacement. Through Kalman filtering, satellite positioning signals are fused to suppress cumulative errors, and real-time output of tower column torsion, inclination and other three-dimensional deformation data is realized.

[0045] The optical fiber sensing network module 2 is used to collect original strain and temperature data by deploying distributed optical fiber sensors along the main load path of the support truss, and to output temperature-compensated strain field distribution data through intelligent compensation algorithm.

[0046] Specifically, as shown in Figure 3 The distributed optical fiber sensor is deployed along the main load path of the support truss, and its installation path design is closely combined with the mechanical properties of the truss structure: first, the optical fiber grating chain is longitudinally laid at the cross section AB of the vertical rod of the bearing layer (-1 layer platform), because this area bears the peak value of vertical construction load (such as concrete wet weight and equipment pressure), and needs to monitor the micro-strain change caused by rod axial force; second, it is spirally wound and laid along the key diagonal brace BC segment (connecting the formwork operation layer and the bearing layer), which resists horizontal wind load, and the spiral arrangement can capture multi-directional strain gradient; the top horizontal cross bar CD (located between -2 layer and -3 layer) is laid along the lower flange because it bears bending stress, and the fiber is laid along the lower flange, focusing on covering the node D (the maximum bending moment point). The armored optical fiber is pre-penetrated inside the truss pre-buried pipe in the area to be poured concrete, to avoid damage caused by vibrating operation. For the steam curing area, temperature-compensated double-gate optical fiber is pasted on the surface of the heat source covered rod to separate the thermal expansion fictitious strain in real time.

[0047] All optical fibers are pre-reserved with redundant loops at the bifurcated nodes to adapt to the structural deformation during climbing. After the sensor is deployed, the laser radar is scanned again through the mechanical arm to verify the path coordinates, to ensure that the installation position meets the requirements, and the cable is fixed along the guide beam groove, finally forming a continuous monitoring network covering the main load-bearing components of the truss, providing high-credibility original data for the temperature compensation algorithm.

[0048] In the description of the present application, the optical fiber sensing network module 2 comprises: an optical-electric conversion conditioning module, a temperature-strain decoupling module, a dynamic temperature compensation module, and a quality evaluation output module.

[0049] The optical-electric conversion conditioning module is used for converting the backscattering light signal into an electric signal, and performing noise reduction and enhancement preprocessing, to output original strain data, original temperature data, and a signal-to-noise ratio index.

[0050] Specifically, the backscattering light signal is received by a high-sensitivity photoelectric detector (such as an InGaAs avalanche diode), and the photon energy is converted into a micro-current signal by using the photoelectric effect. The current signal is converted into a voltage signal by a transimpedance amplifier (TIA), and then the amplitude is dynamically adjusted by a programmable gain amplifier (PGA) to ensure signal linearity.

[0051] The noise reduction processing adopts two-stage filtering: the hardware layer suppresses power frequency interference and high-frequency noise by using a band-pass filter; and the software layer separates the noise spectrum by using a wavelet threshold noise reduction algorithm, and combines a sliding window average method to smooth the baseline drift.

[0052] The signal quality evaluation is based on real-time calculation of the signal-to-noise ratio, and uses a power spectral density analysis method to compare the signal band energy and the noise floor energy, to output the original strain data, the temperature data, and the SNR index.

[0053] The temperature-strain decoupling module is used for separating the coupling effect of temperature and strain in the optical fiber measurement based on the original strain data and the original temperature data, by using a pre-constructed transfer learning decoupling model, to obtain independent strain data and temperature data.

[0054] The temperature-strain decoupling module uses a transfer learning mechanism to solve the cross-sensitivity problem of temperature and strain in optical fiber sensing. Since the measurement signal of the optical fiber sensor is affected by both strain and temperature changes, the traditional method cannot distinguish the contributions of the two, resulting in the strain data containing virtual increase errors caused by temperature.

[0055] The temperature-strain decoupling module uses a pre-trained bidirectional long short-term memory network (Bi-LSTM) combined with domain adaptive technology to transfer the source domain knowledge calibrated in the laboratory to the actual engineering environment (target domain), to realize physical separation of strain and temperature.

[0056] In the cable-stayed bridge tower construction monitoring system, the temperature strain decoupling module plays a key role in eliminating environmental interference and restoring the true mechanical state. The herringbone tower column is easily affected by non-uniform sunlight due to its curved structure, and the actual strain of the support system truss under construction load is often masked by temperature noise. For example, local temperature rise during concrete pouring may cause the fiber optic signal to show false high strain, which may be misjudged as structural overload if not decoupled. The temperature strain decoupling module separates the temperature and strain contributions dynamically and outputs independent physical quantity data, providing a true and reliable strain field input for the subsequent digital twin model, and avoiding false alarms or missed reports caused by distorted data.

[0057] In the description of the present application, based on the original strain data and the original temperature data, the pre-constructed transfer learning decoupling model is used to separate the coupling effect of temperature and strain in fiber measurement, and independent strain data and temperature data are obtained, including: S11, the original strain data and the original temperature data are time-space aligned and feature enhanced to generate standardized input suitable for transfer learning, and a multi-dimensional feature vector is constructed as model input.

[0058] Specifically, the original strain and temperature data collected by the fiber sensing system may have sampling time or spatial position deviation. The dynamic time warping (DTW) algorithm is used to align the time stamp, and the strain and temperature data points at the same physical position are matched by combining the fiber spatial coordinate information, so as to eliminate the time-space misalignment caused by sensor layout difference.

[0059] The aligned data is subjected to multi-dimensional enhancement operation. For strain data, calculate the statistical features (mean, variance, peak factor) in the sliding window; for temperature data, extract time gradient features and spatial distribution features. Through wavelet transform, high-frequency noise is filtered out, and Z-score standardization is used to uniformly scale each feature to zero mean and unit variance, generating a standardized feature matrix with dimensions of 128xN, which is used as the input of transfer learning.

[0060] S12, obtain a transfer learning decoupling model based on bidirectional long short-term memory network (Bi-LSTM) pre-trained in the source domain, freeze the bottom feature extraction layer, replace the top regression layer to adapt to the output dimension of the target domain, input the multi-dimensional feature vector and output the decoupled strain data and temperature data.

[0061] Specifically, the pre-trained bidirectional long short-term memory network (Bi-LSTM) is loaded from the source domain (such as the bridge monitoring historical database). The model has learned the coupling rule of temperature-strain in the source domain, and the bottom layer contains 4 Bi-LSTM units for extracting deep features across time steps. Freeze the bottom Bi-LSTM weights to retain their general feature extraction capability and avoid overfitting caused by small sample data in the target domain (current monitoring scene).

[0062] The top layer of the full connection layer of the source model is removed and replaced with a double-branch regression structure. Each branch contains two layers of full connection networks, respectively outputting independent strain values and temperature values. The input layer dimension of the branch matches the output of the Bi-LSTM hidden layer, and the output layer uses a linear activation function to adapt to the regression task. This design ensures that the model outputs decoupled physical quantities simultaneously, meeting the requirements of the target domain.

[0063] S13, using known labeled data in the target domain, adding a maximum mean difference loss after the feature extraction layer, minimizing the distribution difference between the source domain and the target domain, and calculating the decoupling residual of the strain data and the temperature data. If the decoupling residual is greater than the preset threshold, the model recalibration is triggered.

[0064] Specifically, when training on target domain labeled data, a maximum mean difference loss is introduced after the feature extraction layer (Bi-LSTM output). The MMD distance of the feature vectors of the source domain and the target domain is calculated. By adjusting the model parameters through back propagation, the feature distributions of the two domains are aligned, and the decoupling deviation caused by environmental differences is reduced.

[0065] After the model is decoupled, the residual error of the decoupled strain value and the real strain value is compared. If the root mean square of the residual error of the last 10 time steps exceeds the threshold, the recalibration mechanism is triggered: the weights of the bottom Bi-LSTM are unfrozen, the model is fine-tuned with target domain labeled data, and the MMD loss weight is dynamically adjusted until the residual error returns to the normal range.

[0066] S14, for unlabeled data in the target domain, use the trained transfer learning decoupling model to generate standard confidence pseudo labels, and iteratively optimize the model capability through self-supervised learning.

[0067] Specifically, for unlabeled data in the target domain, use the trained model to output decoupling results and calculate the confidence score. Select data with a confidence score greater than 0.95 as high-confidence pseudo labels (i.e. standard confidence pseudo labels) to generate a pseudo label dataset. Add the pseudo label dataset to the training pool and enhance the model's feature discrimination ability through a contrastive learning strategy: construct positive and negative sample pairs, minimize the feature distance of positive samples and maximize the distance of negative samples. Update the pseudo labels after each iteration and repeat the optimization for 3-5 rounds to gradually improve the decoupling accuracy of the model in complex environments in the target domain.

[0068] The dynamic temperature compensation module is used to correct the strain data in real time based on the environmental temperature gradient, eliminate the virtual error caused by the sunshine temperature difference, and obtain the strain field distribution after temperature compensation.

[0069] The dynamic temperature compensation module captures the nonlinear time-delay mapping relationship between the temperature field and the strain field through a deep learning model, achieving physical separation of the sunshine temperature variation interference.

[0070] The dynamic temperature compensation module mainly aims at the unique non-uniform temperature field challenge in the construction of a cable-stayed bridge. The herringbone tower column has a significant temperature difference between the sun shadow area and the exposed area due to the curved surface structure, and the traditional static compensation model cannot adapt to rapid temperature changes. The encoder-decoder framework can accurately separate the coupling effects of concrete thermal expansion (reversible) and load strain (irreversible) through time delay mapping learning, and avoid misjudging thermal deformation as structural deformation. At the same time, the design of the spatial compensation coefficient matrix solves the problem of thermal inertia difference in different areas of the curved surface of the tower column, for example, the sun-facing concrete needs higher compensation weight due to fast heat absorption, and the shaded area needs to suppress overcompensation.

[0071] In the monitoring of the high lateral support system of a cable-stayed bridge, the dynamic temperature compensation module plays a key role in eliminating environmental noise and restoring the true mechanical state. The support truss is prone to stress redistribution under asymmetric loads, but the temperature difference caused by sunlight may mask the true strain signal. The dynamic temperature compensation module controls the strain measurement error within ±5με through dynamic compensation, providing reliable input for the digital twin model and directly supporting the accurate positioning of stress concentration areas. For example, during the concrete pouring stage, the module can identify the virtual high strain caused by local temperature rise of the support frame, avoiding false triggering of overload alarms and resulting in downtime.

[0072] In the description of the present application, the strain data is corrected in real time based on the environmental temperature gradient, the virtual increase error caused by the temperature difference of sunlight is eliminated, and the strain field distribution after temperature compensation includes: S21, based on the decoupled strain data and temperature data, a time sequence feature vector containing temperature gradient, strain change rate and historical temperature trajectory is constructed, and an input time sequence is generated.

[0073] Specifically, the time sequence features are extracted through a sliding window mechanism: for each monitoring point, the spatial temperature gradient is calculated to capture the temperature field changes caused by non-uniform sunlight; the strain change rate is calculated simultaneously to reflect the real-time dynamics of the structure response; and the historical temperature trajectory is introduced to generate a feature vector using a time encoder to represent the thermal cumulative effect.

[0074] S22, learn the time delay mapping relationship between the temperature field and the strain field through the encoder-decoder framework, construct a dynamic compensation model based on the time re-attention long short-term memory network, and predict the strain value output combined with the current temperature. The deviation between the strain prediction value and the actual independent strain data is used to calculate the thermal hysteresis compensation amount, and the compensation strain field is dynamically generated combined with the temperature gradient.

[0075] In the description of the present application, the time delay mapping relationship between the temperature field and the strain field is learned through the encoder-decoder framework, and the dynamic compensation model based on the time re-attention long short-term memory network includes: S221, encode the historical temperature sequence into a hidden state vector using an encoder layer, and the time attention module focuses on the historical labeled time, and the decoder layer combines the current temperature to predict the strain compensation amount.

[0076] Specifically, the encoder layer uses a bidirectional LSTM network to process the historical temperature sequence (time step ≥ 30), outputting a hidden state vector; the time attention module focuses on the key historical time through entropy weight distribution, enhancing the sensitivity to temperature mutation period; the decoder layer combines the current temperature value and the attention weighted hidden state to output the theoretical strain prediction value.

[0077] S222, add an entropy-based gating unit in the loss layer to compare the deviation of the predicted strain and the actual strain data, and reverse correct the attention weight of the decoder.

[0078] Specifically, embed a KL divergence-based gating unit in the loss function to compare the deviation of the predicted strain and the actual decoupled strain data, and when the deviation exceeds the threshold, trigger backpropagation to dynamically reduce the attention weight of the abnormal temperature fluctuation time, and calculate the thermal hysteresis compensation amount accordingly; Finally, combine the real-time temperature gradient to generate a spatially differentiated compensation strain field, solving the problem of virtual strain increase caused by solar temperature difference.

[0079] S23, based on the compensated strain residual and temperature gradient, dynamically generate a spatially distributed compensation coefficient matrix, and combine the confidence evaluation to perform online learning and incremental optimization of the dynamic compensation model.

[0080] Specifically, based on the compensated strain residual (the difference between the actual value and the compensation value) and the temperature gradient distribution, a spatial compensation coefficient matrix is dynamically generated: the residual-temperature gradient relationship is fitted by Gaussian process regression, higher coefficients are assigned to high temperature gradient areas, and lower coefficients are assigned to shadow areas; The confidence evaluation module calculates the residual variance and the correlation with the temperature gradient, and when the variance is below the threshold and the correlation coefficient is > 0.9, it is determined to be high confidence, triggering the incremental learning mechanism: using online gradient descent algorithm to update the LSTM network weight, and at the same time optimizing the compensation coefficient matrix parameters, adapting to the change of material thermal properties with the aging of the season.

[0081] The quality evaluation output module is used to package the compensated strain field distribution data into a standard format and transmit it to the digital twin model through a hybrid network.

[0082] Specifically, the compensated strain field is packaged: using HDF5 format hierarchical storage, metadata includes spatial coordinates (GPS positioning + fiber meter marker), time stamp (Beidou time synchronization), temperature compensation coefficient and SNR index.

[0083] The quality check includes integrity check and rationality check. The hybrid transmission network adopts a hierarchical strategy: high-frequency data are uploaded in batches after being locally cached through ZigBee protocol; and early warning signals are pushed in real time to the digital twin platform through a 5G private network.

[0084] The transmission process embeds a retransmission mechanism: if the packet loss rate is >5% or the delay is >200ms, the data packet is retransmitted through the LoRa backup channel to ensure that the data packet arrival rate is ≥99.9%.

[0085] The inertial navigation fusion positioning module 3 is configured to monitor the tower column deformation in real time based on the satellite signals and the gyroscope angular velocity collected by the inertial sensor, and generate three-dimensional deformation data.

[0086] Specifically, the inertial navigation fusion positioning module realizes millimeter-level deformation monitoring of the tower column through multi-source sensor data collaborative processing, and the core process includes four stages of data acquisition, time-space synchronization, fusion calculation and deformation reconstruction. The specific implementation mode is as follows: 1. The data acquisition layer uses a dual-frequency GNSS receiver to obtain satellite raw observation values at a frequency of ≥10Hz, and simultaneously collects three-axis angular velocity data of the key section of the tower column through a MEMS gyroscope array at a high frequency of 200Hz; the satellite positioning data is used to calculate the three-dimensional coordinates of the tower monitoring points through real-time kinematic (RTK) technology, and the gyroscope data is used to output the local attitude change of the tower column in real time through quaternion integration to form an initial positioning and attitude data set.

[0087] 2. The time-space synchronization stage relies on the Beidou / GPS joint timing module to provide a unified time scale for all sensors, and aligns the time series of GNSS (low frequency) and gyroscope (high frequency) data through interpolation algorithm: first, the gyroscope angular velocity data is integrated into a sequence of attitude quaternions, and then the spherical linear interpolation (Slerp) algorithm is used to complete the attitude data at intermediate time based on the GNSS data time, to eliminate the time asynchronous error of heterogeneous signals and ensure the time-space consistency of the data stream.

[0088] 3. The fusion calculation engine constructs a state space model based on the extended Kalman filter (EKF): The state vector is defined as a 12-dimensional vector, including position offset, velocity deviation, Euler angle deviation and gyroscope zero offset; the state prediction uses the gyroscope angular velocity and the attitude quaternion at the previous time to update the current state, and linearizes the system equation through the Jacobian matrix; The observation update takes the GNSS coordinate difference as an observation, designs an observation matrix to map a state vector to a measurement space, fuses a prediction value and an observation value through a Kalman gain, and updates an error covariance matrix in real time; in order to suppress divergence caused by satellite signal loss of lock, an adaptive noise adjustment mechanism is introduced: when the satellite signal-to-noise ratio is lower than 35 dB-Hz, the process noise covariance is dynamically increased to reduce the GNSS weight, and the gyroscope short-time high-precision attitude data is preferentially relied on.

[0089] 4. Convert the fused displacement field into the tower deformation characteristics: 4.1, rigid body displacement separation: calculate the overall translation component based on the fixed reference point at the bottom of the tower (such as the reference point of the pile cap), and deduct it from the fused displacement; 4.2, curvature calculation: calculate the local curvature by using the attitude angle difference of adjacent monitoring sections, and calculate the bending moment distribution combined with the inertia moment of the tower section; 4.3, three-dimensional deformation field generation: expand the displacement and curvature data of the discrete monitoring points to the full tower surface through the thin plate spline interpolation algorithm, output the three-dimensional deformation field, and mark the stress concentration area.

[0090] The digital twin early warning module 4 is used for fusing strain field distribution data and three-dimensional deformation data, using a digital twin model to map the stress and displacement field of the support system in real time, dynamically rendering the deformation trend, and predicting the structural failure risk of the tower, and outputting the risk prediction result.

[0091] In the description of the application, the digital twin early warning module 4 includes: a multi-source heterogeneous integration module, a digital twin modeling module, a physical double-drive analysis module, a dynamic risk deduction module, and an early warning decision generation module.

[0092] The multi-source heterogeneous integration module is used for integrating strain field distribution data, three-dimensional deformation data, environmental temperature and humidity, and load data, realizing spatio-temporal alignment and quality checking of multi-source heterogeneous data.

[0093] The digital twin modeling module is used for constructing a multi-scale digital twin model of the support system, establishing a dynamic mapping between the physical entity and the virtual space, and generating a stress field-displacement field simulation calculation environment.

[0094] The physical double-drive analysis module is used for coupling a physical mechanism model and a deep learning algorithm, identifying a stress concentration area and a displacement mutation point, quantifying a structural failure risk probability of the tower, and locating a risk position.

[0095] In the description of the application, coupling a physical mechanism model and a deep learning algorithm, identifying a stress concentration area and a displacement mutation point, quantifying a structural failure risk probability of the tower, and locating a risk position include: S31, obtain the real-time stress field and displacement field output by the digital twin model, divide the risk candidate area, and initialize an independent risk positioning agent for each candidate area and assign an initial resource weight.

[0096] Specifically, based on the real-time stress field and displacement field data output by the digital twin model, the stress concentration area and displacement mutation point are identified through a gradient threshold segmentation algorithm.

[0097] The area meeting the threshold is marked as a risk candidate area, and an independent risk positioning agent is initialized for each candidate area. Each agent is assigned an initial resource weight, which is dynamically set according to the risk level of the area: a high stress gradient area is assigned an initial weight of 0.6, and a low risk area is assigned an initial weight of 0.1, ensuring that resources are tilted towards key areas.

[0098] S32, based on the material fatigue characteristics and historical failure data, set the game rules and reward mechanism between the agents, construct a dynamic strategy space, and train the agent strategy network using a hierarchical architecture, and introduce physical constraints consistent with the material failure physical law.

[0099] In the description of the present application, the dynamic strategy space includes: the abnormal risk area agent occupies the computing resources in the game process. The non-abnormal risk area agent cooperates with each other in the game process. The reward function fuses the stress mutation gradient and displacement offset, which is used to quantify the game income.

[0100] Specifically, the game rules: the abnormal risk area agent (such as the stress mutation point) adopts a preemptive strategy and competes for additional computing resources through a bidding mechanism; the non-abnormal area agent adopts a cooperative strategy and shares historical failure mode data (such as crack propagation path, fatigue accumulation data), avoiding repeated analysis. The reward function fuses the physical law constraint: the stress mutation gradient contributes 60% of the weight, the displacement offset contributes 30%, and the material fatigue characteristics (such as the slope of the S-N curve) contribute 10%, ensuring that the reward value quantifies the game income while conforming to the material failure mechanism.

[0101] The hierarchical architecture adopts a federated learning framework, the local layer adopts a Q-value algorithm to learn the target action strategy, and the federated layer aggregates global strategy parameters and generates a shared knowledge base.

[0102] The local layer uses a distributed Q-learning algorithm, and each agent learns the optimal action strategy (such as applying resources, releasing resources, and sharing data) based on the local state (stress value, displacement history, resource occupancy rate); the federated layer aggregates the strategy network weights of each agent through a parameter server to generate a shared knowledge base. During training, physical constraints are introduced: the risk probability output by the agent is limited to satisfy Hooke's law and the energy conservation equation, avoiding predictions that violate the principles of mechanics.

[0103] S33, real-time monitoring of resource competition behavior between agents, dynamically adjusting the resource allocation weight of the agent, eliminating local deadlock, and fusing the local risk prediction results of each agent to generate a cooperative positioning matrix and output the risk position coordinates of the tower column abnormality.

[0104] Specifically, when multiple agents continuously compete for the same resource, a deadlock elimination protocol is triggered, which dynamically degrades the resource weight of the conflict party based on historical cooperation records and increases the weight of the cooperative party.

[0105] Each agent outputs a local risk prediction result (such as coordinate position, failure probability), which is integrated into a cooperative positioning matrix through a Bayesian fusion algorithm. The matrix elements represent the risk probability of the tower column surface grid, and the abnormal position coordinates are extracted through probability density clustering.

[0106] S34, based on the multi-agent game result, the structural failure probability of the tower column is quantified, and the positioning result is matched with the historical case to dynamically update the risk knowledge base.

[0107] Specifically, the maximum risk value of the cooperative positioning matrix is compared with the material strength threshold value, if P max >0.95 is determined as a high-risk state, triggering a level one alarm.

[0108] The probability calculation fuses the Nash equilibrium solution of the multi-agent game: when the high-risk area agent continuously occupies the resource and the reward value is saturated, the system determines that the region is in an unstable state, and the failure probability increases exponentially.

[0109] The positioning result is matched with the historical case library, and the risk knowledge base is dynamically updated: the feature vector (stress gradient pattern, displacement time history curve) of the new case is stored in the shared knowledge base, and the weight distribution of the federal layer strategy network is fine-tuned, and the subsequent game strategy is optimized.

[0110] The dynamic risk deduction module is used to simulate the structural deformation trend in the tower column based on real-time twin data, predict the critical failure time and damage path, and generate a visual warning scene.

[0111] In the description of the present application, simulating the structural deformation trend in the tower column based on real-time twin data, predicting the critical failure time and damage path, and generating a visual warning scene includes: S41, based on the structural failure probability and risk position of the tower column, a virtual deduction environment with risk units as agents is constructed based on the digital twin model, the agent state space and action space are set, and the game rule library of each agent is initialized.

[0112] Specifically, based on the structural failure probability and risk position coordinates of the tower column, a virtual deduction environment with risk units as intelligent agents is constructed in the digital twin model. Each risk unit agent is assigned an independent state space, including position coordinates, current stress value, material degradation rate, environmental load parameters, and other dynamic variables; the action space is defined as crack propagation direction selection, stress redistribution request, and collaborative early warning trigger.

[0113] When initializing the game rule library, reference historical failure cases and material mechanics properties to set the competition rules and collaboration rules between agents, and pre-load typical failure scenarios as initial training data through Monte Carlo sampling.

[0114] S42, build a controller to coordinate the behavior of multiple agents, use the controller to receive the state of all agents in real time, and eliminate action conflicts to generate a globally known damage path topology graph and output a corrected deformation field.

[0115] Specifically, the central controller coordinates the behavior of multiple agents: the controller receives the state vector of all agents in real time, and uses a conflict detection graph-based resolution algorithm to identify action conflicts. For conflicting behaviors, the controller dynamically adjusts the priority through Nash equilibrium calculation - for example, agents at the critical point of material brittle change are given higher weights, and their actions are forced to be executed first. After resolving conflicts, the controller integrates the local damage paths of each agent to generate a global damage path topology graph.

[0116] Then call the physics engine to simulate stress redistribution under the path topology, output the corrected deformation field, and update the digital twin model through real-time data synchronization mechanism.

[0117] S43, based on the action sequence of the agent and the time-varying strength variation law of the material, calculate the critical failure time, and integrate the three-dimensional engine to convert the deduction process into a three-dimensional dynamic scene.

[0118] Specifically, based on the action sequence of the agent and the time-varying strength law of the material, calculate the critical failure time: analyze the action sequence of each agent (such as crack propagation length, frequency), combine the time-varying strength model of the material, and calculate the unit failure time through cumulative damage integral. For example, for high-risk weld areas, calculate the critical failure time based on the current stress amplitude and crack length.

[0119] At the same time, integrate a three-dimensional game engine (such as Unity or Unreal Engine) to visualize the deduction process: map the damage path topology graph to the progressive coloring of the three-dimensional grid, dynamically render the structure deformation process, such as the frame-by-frame animation of the tower column bending and twisting deformation, and superimpose real-time sensor data to generate a warning hotspot map. Finally, output a three-dimensional dynamic scene containing a countdown label, supporting VR device interactive operation.

[0120] An early warning decision generation module is configured to generate hierarchical early warning instructions in combination with the inference result, drive a man-machine collaborative response strategy, and feedback optimization of the digital twin model parameters.

[0121] An edge computing relay module 5 is configured to build a hybrid transmission network, acquire original sensing signals and early warning instructions collected by sensors, and output denoised structured data and real-time early warning signals.

[0122] Specifically, the hybrid transmission network builds a double-channel redundancy mechanism: a high-priority channel directly transmits early warning signals through a 5G URLLC slice, adopts a double-time-slot ring redundancy protocol to ensure a transmission delay of <20 ms, and internally embeds a forward error correction coding to compensate for a 10% packet loss rate; a low-priority channel transmits structured data to an edge gateway in batches through a LoRa ad hoc network, the gateway aggregates data packets through a time-sensitive network scheduler, and encapsulates the data packets into an MQTT protocol for optical fiber backhaul to the cloud.

[0123] A three-stage pipeline architecture is adopted: first, the original signals are standardized, robust principal component analysis is applied to separate noise and effective data, and baseline drift caused by environmental disturbance is eliminated; second, a lightweight convolutional network is used to extract joint time-frequency domain features to generate a <=128-dimensional feature vector; and finally, a pre-trained random forest-LSTM hybrid model is used to output structured data and 0-4 level early warning signals, and when strain mutation or temperature gradient anomaly is detected, a three-level early warning is triggered.

[0124] The following describes the construction scene and construction process of the safety monitoring system of the high transverse support system of the cable-stayed bridge human-shaped tower column construction designed by the application in combination with specific embodiments.

[0125] As shown in Figure 4 The whole process of the construction of the cable-stayed bridge human-shaped tower column can be summarized as three stages, and the specific stages and their contents include: 1. Foundation and initial construction The process starts with site and resource preparation, then installs the tower crane foundation to ensure the support of heavy equipment, and completes the first 2 section construction of the tower column (usually using the turnover molding process). The elevator passage is added in the early stage to provide a safe entrance for high-altitude work.

[0126] 2. Standardized cycle and key node control The main body adopts the climbing formwork process to cycle the construction of the standard section of the tower column (3-6 sections).

[0127] Key control points include: 2.1, active cross brace installation: add transverse support to resist inclined load in sections 6 and 15 to constrain tower column linear deviation; 2.2, corbel and intersection section construction: set up corbel support to cast key nodes (herringbone bifurcation point) in section 7, and control complex curved surface (sections 16-18) through mold turning process; 2.3, process conversion: remove the climbing form after section 15, and erect special support for intersection section to ensure stability of the transition area.

[0128] 3, finishing and resource release After the completion of section 27, the system removes the climbing form and auxiliary equipment, and the process design highlights three core features: 3.1, process dynamic adaptation: climbing form efficiently covers standard sections, and mold turning precisely handles corbel and intersection section special-shaped structures; 3.2, safety coordination enhancement: active cross brace and scaffold intervene in layers to match wind resistance requirements during tower column height growth; 3.3, resource recycling logic: tower crane / elevator is early put into use to support personnel and material transportation throughout the whole process, and is removed at the end to release the site.

[0129] The following will detail the construction steps in combination with construction engineering drawings. Further, the tower column can be divided into six stages according to construction tasks, and the specific steps are shown as follows.

[0130] As shown in Figure 5 , step one includes: 1, before pouring the concrete of the pile cap, embed the tower base reinforcement and tower column reinforcement; 2, before pouring the concrete of the pile cap, embed 1 tower crane foundation section, corbel support steel column embedded parts, etc.; 3, simultaneously construct the 1st tower column and the tower base.

[0131] Strain gauges and inclination sensors are embedded in the tower crane foundation section during the pouring of the pile cap concrete, laser radar scans the tower base reinforcement cage to generate an initial point cloud model, and establishes a subsequent monitoring space reference.

[0132] As shown in Figure 6 , step two includes: 1, install the 2nd tower column reinforcement and pour the concrete; 2, cycle construction to section 4 and timely install the climbing form construction platform.

[0133] During the installation process of the climbing form platform, the mechanical arm distributes distributed optical fiber sensors along the curved surface of sections 2-4, the main path covers the template operation layer diagonal brace, and simultaneously calibrates the verticality in real time through the inertial navigation module.

[0134] As shown in Figure 7 , step three includes: 1, cycle construction to section 8; 2, install the corbel support and bottom form; 3, install the corbel reinforcement, embedded parts, etc. and side form; 4, pour the corbel concrete.

[0135] The fiber grating chain is pre-installed inside the corbel support steel column embedded part, and the temperature gradient is monitored in real time during the pouring of the corbel concrete; inertial sensors need to be deployed at the active cross brace node to capture abnormal vibration during the installation stage.

[0136] like Figure 8 As shown, step four includes: 1. Cyclic construction up to the 15th segment, and timely active horizontal bracing construction and tower crane lifting construction; 2. Removal of the inner part of the formwork platform.

[0137] like Figure 9 As shown, step five includes: 1. Removing the inner climbing formwork; 2. Installing the bottom formwork support for the intersection section, and installing the reinforcing steel and embedded parts for the 16th segment; 3. Installing the outer flip formwork for the 16th segment and pouring concrete for the 16th segment; 4. Removing the climbing formwork and repeating the construction until the 18th segment.

[0138] Pressure sensors were pre-installed on the bottom formwork support of the intersection section. When pouring the 16th to 18th sections, deformation data was matched to verify the settlement of the support. The process of dismantling the climbing formwork triggered a second scan by the lidar to correct the deformation field model.

[0139] like Figure 10 As shown, step six includes: 1. Reinstalling the climbing formwork, cyclically constructing to the top of the 27th segment, and gradually lifting the tower crane; 2. Using the tower crane to dismantle the formwork; 3. Dismantling elevators, pump pipes, and tower cranes that are not related to subsequent construction.

[0140] In summary, this invention constructs a full-process monitoring chain from data acquisition to risk warning through multi-module collaboration, realizing full-cycle, multi-dimensional dynamic control over the construction of the human-shaped tower column high lateral support system. This includes precise sensor deployment, dynamic correction of environmental interference, high-precision acquisition of deformation data, and intelligent inference of structural risks. The various monitoring links are organically connected, breaking the fragmented limitations of traditional monitoring, comprehensively covering the stress state, deformation trend, and environmental impact of the support system, providing systematic technical support for construction safety, and ensuring timely perception and overall control of potential risks.

[0141] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

Claims

1. A safety monitoring system for a high lateral support system in the construction of human-shaped tower columns for cable-stayed bridges, characterized in that, The system includes: The climbing formwork integrated sensing module is used to generate a point cloud model by scanning the curved surface of the tower column with a robotic arm integrated in the hydraulic climbing formwork platform and a lidar. Combined with the preset building information model coordinates, it adaptively adjusts the sensor installation posture and outputs the sensor installation coordinate position. The fiber optic sensing network module is used to collect raw strain and temperature data using distributed fiber optic sensors deployed along the main force path of the supporting truss, and to dynamically correct the data through an intelligent compensation algorithm, outputting strain field distribution data after temperature compensation. The inertial navigation fusion positioning module is used to monitor the tower deformation in real time based on satellite signals collected by inertial sensors and gyroscope angular velocity, and generate three-dimensional deformation data. The digital twin early warning module is used to integrate strain field distribution data and three-dimensional deformation data, use the digital twin model to map the stress and displacement field of the support system in real time, dynamically render the deformation trend, predict the structural failure risk of the tower column, and output the risk prediction results. Edge computing relay modules are used to build hybrid transmission networks, acquire raw sensor signals and early warning commands collected by sensors, and output noise-reduced structured data and real-time early warning signals.

2. The safety monitoring system for a high lateral support system for the construction of a human-shaped tower column of a cable-stayed bridge according to claim 1, characterized in that, The climbing formwork integrated sensing module includes: The robotic arm control module is used to calculate the motion trajectory and control signals of the robotic arm end effector in real time based on the preset robotic arm joint parameters and target movement coordinates, so as to realize the movement of the robotic arm. The laser scanning module is used to acquire the three-dimensional topography of the tower surface in real time using lidar, extract effective point cloud data, and construct the point cloud model of the tower through point cloud registration and reconstruction. The coordinate matching module is used to align the actual scanned point cloud model with the preset digital model. Based on the target installation positions of several sensors, it plans the target installation path of the robotic arm, deploys the sensors sequentially at the stress points of the tower column, and obtains the actual installation coordinate positions of the sensors. The attitude adjustment module is used to adaptively calibrate the sensor mounting attitude based on the surface method, maintain the sensor's alignment with the law direction, and calculate the sensor's mounting tilt angle deviation. The installation verification module is used to trigger a secondary scan after the sensor is installed, comparing the distance deviation between the sensor's installation coordinates and the target installation coordinates, as well as the installation tilt angle deviation value with the preset tilt angle deviation threshold, to verify the sensor's installation accuracy.

3. The safety monitoring system for a high lateral support system for the construction of a human-shaped tower column of a cable-stayed bridge according to claim 1, characterized in that, The fiber optic sensing network module includes: The photoelectric conversion conditioning module is used to convert backscattered light signals into electrical signals, perform noise reduction and enhancement preprocessing, and output raw strain data, raw temperature data, and signal-to-noise ratio indicators. The temperature-strain decoupling module is used to separate the coupling effect of temperature and strain in fiber optic measurement based on the original strain data and original temperature data, using a pre-built transfer learning decoupling model, and to obtain independent strain and temperature data. The dynamic temperature compensation module is used to correct strain data in real time based on the ambient temperature gradient, eliminate the artificial error caused by the solar temperature difference, and obtain the strain field distribution after temperature compensation. The quality assessment output module encapsulates the compensated strain field distribution data into a standard format and transmits it to the digital twin model via a hybrid network.

4. The safety monitoring system for a high lateral support system for the construction of a human-shaped tower column of a cable-stayed bridge according to claim 3, characterized in that, The process of separating the coupling effect of temperature and strain in fiber optic measurement using a pre-built transfer learning decoupling model based on raw strain and temperature data, and obtaining independent strain and temperature data, includes: Spatiotemporal alignment and feature enhancement are performed on the original strain data and original temperature data to generate standardized inputs suitable for transfer learning, and multidimensional feature vectors are constructed as model inputs. Obtain a transfer learning decoupling model based on a bidirectional long short-term memory network pre-trained in the source domain, freeze the bottom feature extraction layer, replace the top regression layer to adapt to the output dimension of the target domain, input multidimensional feature vectors and output decoupled strain data and temperature data; Using known labeled data in the target domain, a maximum mean difference loss is added after the feature extraction layer to minimize the difference in adjustment distribution between the source and target domains, and the decoupling residuals of strain and temperature data are calculated. If the decoupling residuals are greater than a preset threshold, model recalibration is triggered. For unlabeled data within the target domain, standard confidence pseudo-labels are generated using a trained transfer learning decoupling model, and the model's capabilities are iteratively optimized through self-supervised learning.

5. The safety monitoring system for a high lateral support system for the construction of a human-shaped tower column of a cable-stayed bridge according to claim 3, characterized in that, The method of real-time correction of strain data based on ambient temperature gradient to eliminate artificially inflated errors caused by solar radiation temperature difference, and obtaining the temperature-compensated strain field distribution includes: Based on the decoupled strain and temperature data, a time-series feature vector containing temperature gradient, strain change rate and historical temperature trajectory is constructed to generate the input time-series sequence. The time-delay mapping relationship between the temperature field and the strain field is learned through an encoder-decoder framework. A dynamic compensation model based on a time re-attention long short-term memory network is constructed, and the strain prediction value is output in combination with the current temperature prediction. The thermal hysteresis compensation amount is calculated by using the deviation between the strain prediction value and the actual independent strain data, and the compensation strain field is dynamically generated in combination with the temperature gradient. Based on the compensated strain residuals and temperature gradient, a spatially distributed compensation coefficient matrix is ​​dynamically generated, and combined with confidence assessment, the dynamic compensation model is subjected to online learning and incremental optimization.

6. The safety monitoring system for a high lateral support system for the construction of a human-shaped tower column of a cable-stayed bridge according to claim 5, characterized in that, The process of learning the time-delay mapping relationship between the temperature field and the strain field through an encoder-decoder framework and constructing a dynamic compensation model based on a time-re-attention long short-term memory network includes: The encoder layer encodes the historical temperature sequence into a hidden state vector, the time attention module focuses on the historical labeled moments, and the decoder layer combines the current temperature to predict the strain compensation amount. An entropy-based gating unit is added to the loss layer to compare the deviation between the strain prediction value and the actual strain data, and to correct the attention weights of the decoder in reverse.

7. The safety monitoring system for a high lateral support system for the construction of a human-shaped tower column of a cable-stayed bridge according to claim 1, characterized in that, The digital twin early warning module includes: The multi-source heterogeneous integration module is used to integrate strain field distribution data, three-dimensional deformation data, environmental temperature and humidity and load data to achieve spatiotemporal alignment and quality verification of multi-source heterogeneous data. The digital twin modeling module is used to construct a multi-scale digital twin model of the support system and establish a dynamic mapping between physical entities and virtual space to generate a simulation calculation environment for stress field-displacement field. The physical dual-drive analysis module is used to couple physical mechanism models with deep learning algorithms to identify stress concentration areas and displacement abrupt change points, quantify the probability of structural failure risk of tower columns, and locate the risk location. The dynamic risk simulation module is used to simulate the deformation trend of the internal structure of the tower column based on real-time twin data, predict the critical failure time and failure path, and generate a visual early warning scenario. The early warning decision generation module is used to generate tiered early warning instructions based on the simulation results, drive human-machine collaborative response strategies, and provide feedback to optimize the parameters of the digital twin model.

8. The safety monitoring system for a high lateral support system for the construction of a human-shaped tower column of a cable-stayed bridge according to claim 7, characterized in that, The coupled physical mechanism model and deep learning algorithm identify stress concentration zones and displacement abrupt change points, quantify the structural failure risk probability of the tower column, and locate the risk location, including: The system acquires the real-time stress and displacement fields output by the digital twin model, divides the candidate risk regions, initializes an independent risk localization agent for each candidate region, and assigns initial resource weights. Based on material fatigue characteristics and historical failure data, game rules and reward mechanisms are set between agents, a dynamic policy space is constructed, and a hierarchical architecture is used to train the agent policy network, introducing physical constraints that conform to the physical laws of material failure. Real-time monitoring of resource competition behavior among agents, dynamic adjustment of resource allocation weights for agents, elimination of local deadlock, and fusion of local risk prediction results of each agent to generate a cooperative localization matrix and output the coordinates of the risky locations where the tower has anomalies. The structural failure probability of the tower column is quantified based on the results of multi-agent game, and the location results are matched with historical cases to dynamically update the risk knowledge base.

9. The safety monitoring system for a high lateral support system for the construction of a human-shaped tower column of a cable-stayed bridge according to claim 8, characterized in that, The dynamic strategy space includes: agents in abnormal risk areas competing for computing resources during the game; agents in non-abnormal risk areas cooperating with each other during the game; and a reward function that integrates stress mutation gradient and displacement offset to quantify game payoffs. The hierarchical architecture adopts a federated learning framework. The local layer uses the Q-value algorithm to learn the target action policy, and the federated layer aggregates global policy parameters and generates a shared knowledge base.

10. The safety monitoring system for a high lateral support system for the construction of a human-shaped tower column of a cable-stayed bridge according to claim 7, characterized in that, The method of simulating the deformation trend of the internal structure of the tower column based on real-time twin data, predicting the critical failure time and failure path, and generating a visualized early warning scenario includes: Based on the structural failure probability and risk location of the tower column, a virtual simulation environment with risk units as intelligent agents is constructed based on the digital twin model. The state space and action space of the intelligent agents are set, and the game rule base of each intelligent agent is initialized. A controller is built to coordinate the behavior of multiple agents. The controller receives the state of all agents in real time, resolves action conflicts, generates a globally known topology of destruction paths, and outputs a corrected deformation field. Based on the action sequence of the intelligent agent and the time-varying intensity change law of the material, the critical failure time is calculated, and a three-dimensional engine is integrated to transform the simulation process into a three-dimensional dynamic scene.

Images