Anti-overturning dynamic adjusting method and system for steel bridge construction hanging basket

Abstract

The invention discloses a steel bridge construction hanging basket multi-process collaborative optimization scheduling method and system. The method comprises the following steps: collecting and preprocessing multi-source real-time data, and establishing a digital twinborn body; forward simulation prediction is carried out through the digital twinborn body, and an overturning risk probability sequence is generated; inputting the sequence into a reinforcement learning strategy network to output an adjustment instruction; and executing the instruction and feeding back optimization. The system comprises a data acquisition module, a digital twinborn management module, a risk prediction module, a reinforcement learning decision module, an instruction execution module and a collaborative optimization module. Through deep fusion of digital twinning and reinforcement learning, intelligent monitoring, risk advanced early warning and adaptive adjustment of hanging basket construction are realized, the problems of early warning lagging and high false alarm rate of a traditional method are effectively solved, and the construction safety and efficiency are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of steel bridge construction safety monitoring technology, specifically to a method and system for dynamic adjustment of anti-overturning of hanging baskets used in steel bridge construction. Background Technology

[0002] As a key piece of equipment in the construction of long-span bridges, the anti-overturning safety control of hanging baskets for steel bridge construction has always been a major issue in the engineering field. Traditional hanging basket monitoring mainly relied on manual inspections and simple sensor combinations. In recent years, it has gradually developed into a system based on fixed threshold early warning using inclinometers, stress sensors, and other similar devices. With the advancement of technology, digital methods such as finite element analysis and BIM modeling have been introduced for structural analysis, and some studies have attempted to conduct safety assessments by establishing static mechanical models. Meanwhile, the advantages of machine learning algorithms in pattern recognition have led to their application in structural anomaly detection. These technological advancements have, to a certain extent, improved the monitoring capabilities of hanging basket construction, providing a fundamental guarantee for construction safety.

[0003] However, existing technologies still have significant limitations. First, at the data sensing level, traditional methods mostly use independent sensors to collect single types of data, lacking an effective multi-source heterogeneous data fusion mechanism. This results in data quality being severely affected by noise interference, and the timing of the sensors is not synchronized, making it difficult to accurately reflect the true state of the structure. Second, in terms of risk warning, existing systems mostly rely on fixed thresholds or simplified static models for judgment, failing to make advanced and quantitative predictions of overturning risks under time-varying environmental loads. The problem of warning lag is prominent, essentially remaining a passive response mode. Third, in terms of control strategies, existing methods mostly rely on preset rules or PID control, lacking the ability to adapt to sudden working conditions and dynamic environments. The adjustment effect is limited in complex scenarios such as sudden changes in wind speed and load. Most importantly, the functional modules of existing systems are relatively independent, with monitoring, prediction, and control links disconnected, forming information silos. Furthermore, they lack self-optimization capabilities based on practical feedback, failing to continuously improve performance through system operation. These shortcomings collectively make it difficult for traditional methods to achieve accurate, forward-looking, and adaptive control of hanging baskets against overturning. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing hanging basket safety control methods have significant early warning lag, high false alarm rate, and insufficient adaptability of control strategies, as well as the problem of how to achieve anti-tipping dynamic adjustment and closed-loop optimization based on digital twins and reinforcement learning.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for dynamic adjustment of anti-overturning features of a formwork for steel bridge construction, comprising: collecting multi-source real-time data of the formwork body and the environment and preprocessing it to obtain preprocessed multi-source real-time data; establishing an initial digital twin based on the preprocessed multi-source real-time data; performing forward simulation prediction through the digital twin to generate a multi-step overturning risk probability sequence; inputting the multi-step overturning risk probability sequence into a reinforcement learning policy network trained based on the digital twin, and outputting anti-overturning dynamic adjustment instructions; executing the anti-overturning dynamic adjustment instructions, and feeding back the execution effect to the digital twin and the reinforcement learning policy network for collaborative optimization.

[0007] As a preferred embodiment of the anti-overturning dynamic adjustment method for the hanging basket used in steel bridge construction according to the present invention, the multi-source real-time data includes structural state data from the hanging basket body and environmental load data from the external environment; wherein, the structural state data includes tilt angle data, stress data and displacement data, and the environmental load data includes wind speed, wind direction, temperature and humidity data.

[0008] As a preferred embodiment of the anti-overturning dynamic adjustment method for the hanging basket used in steel bridge construction according to the present invention, the preprocessing includes adaptive filtering and noise reduction, time alignment and format standardization of multi-source real-time data.

[0009] As a preferred embodiment of the anti-overturning dynamic adjustment method for the hanging basket in steel bridge construction described in this invention, the method includes: after establishing the initial digital twin, continuously updating the digital twin in real time using preprocessed multi-source real-time data; the real-time update includes the following processes: data synchronization process: injecting the preprocessed multi-source real-time data into the digital twin as its input boundary conditions; model parameter correction process: dynamically adjusting the internal physical parameters of the mechanical simulation model based on the deviation between the output of the mechanical simulation model inside the digital twin and the measured values ​​of the physical entity; state synchronization process: resetting the current operating state of the digital twin based on the quality assessment of each data source and data with confidence scores higher than a preset threshold.

[0010] As a preferred embodiment of the anti-overturning dynamic adjustment method for the hanging basket in steel bridge construction described in this invention, the forward simulation prediction includes: based on the current state of the digital twin, parallel deducing various discretized future environmental conditions, and calculating the overturning evaluation index under each condition through multi-threaded dynamic solution.

[0011] As a preferred embodiment of the anti-overturning dynamic adjustment method for the hanging basket used in steel bridge construction according to the present invention, the overturning assessment indicators of each working condition are integrated to generate a multi-step overturning risk probability sequence; the sequence is an array of probability values ​​arranged in chronological order, wherein each probability value is obtained by weighted fusion of the extrapolation results of the multiple working conditions and environmental uncertainty factors.

[0012] As a preferred embodiment of the anti-overturning dynamic adjustment method for the formwork used in steel bridge construction according to the present invention, the reinforcement learning strategy network is a deep neural network obtained by offline training in the simulation environment constructed by the digital twin through interactive trial and error with the environment, with the goal of minimizing long-term overturning risk; the reinforcement learning strategy network takes the multi-step overturning risk probability sequence as the state input and outputs specific anti-overturning dynamic adjustment commands.

[0013] As a preferred embodiment of the anti-overturning dynamic adjustment method for steel bridge construction formwork described in this invention, the collaborative optimization includes the following processes: a digital twin correction process, which uses the actual state effect data after the execution of the anti-overturning dynamic adjustment command to perform reverse correction on the digital twin; and a reinforcement learning strategy network optimization process, which converts the execution effect of the anti-overturning dynamic adjustment command into a reward signal for reinforcement learning and performs online fine-tuning on the reinforcement learning strategy network.

[0014] Another objective of this invention is to provide an anti-overturning dynamic adjustment system for hanging baskets used in steel bridge construction. By constructing an intelligent closed-loop control system of "perception-prediction-decision-execution-optimization", it deeply integrates digital twins and reinforcement learning, solving the problem of passive response control caused by inaccurate data perception, delayed risk prediction, and rigid control strategies in traditional hanging basket construction safety control, and realizing a fundamental shift from passive emergency response to proactive prevention and control.

[0015] A preferred embodiment of the anti-overturning dynamic adjustment system for a steel bridge construction formwork, as described in this invention, includes a multi-source data acquisition and preprocessing module, a digital twin management module, an overturning risk prediction module, a reinforcement learning decision-making module, and an instruction execution and collaborative optimization module. The data acquisition and preprocessing module collects multi-source real-time data from the formwork and its environment, and preprocesses the data. The digital twin management module establishes and maintains a digital twin synchronized with the physical formwork based on the preprocessed multi-source real-time data. The overturning risk prediction module performs forward simulation prediction using the digital twin to generate a multi-step overturning risk probability sequence. The reinforcement learning decision-making module inputs the multi-step overturning risk probability sequence into a pre-trained reinforcement learning policy network and outputs anti-overturning dynamic adjustment instructions. The instruction execution and collaborative optimization module executes the anti-overturning dynamic adjustment instructions and feeds back the actual state effect data after instruction execution to the digital twin management module and the reinforcement learning decision-making module for collaborative optimization.

[0016] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the aforementioned method for dynamic adjustment of anti-overturning formwork for steel bridge construction.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned method for dynamic adjustment of anti-overturning formwork for steel bridge construction.

[0018] The beneficial effects of this invention are: This invention establishes a complete intelligent anti-overturning control system through five interconnected and synergistic steps. Data acquisition and preprocessing provide reliable perception for the system; digital twin construction enables high-fidelity simulation; forward simulation prediction provides risk foresight; reinforcement learning decision-making generates optimization instructions; and execution feedback and collaborative optimization ensure system evolution. This architectural design enables the system to not only possess real-time monitoring and early warning capabilities but also self-learning and adaptive intelligent features. It effectively solves key technical problems in traditional hanging basket anti-overturning control, such as delayed early warning, rigid decision-making, and lack of adaptability, thereby comprehensively improving construction safety and providing all-round technical protection for the safety of steel bridge construction. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 The first embodiment of the present invention provides an overall flowchart of a dynamic adjustment method and system for anti-overturning of a hanging basket used in steel bridge construction. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figure 1 As one embodiment of the present invention, a method for dynamic adjustment of anti-overturning of a hanging basket used in steel bridge construction is provided, comprising: S1: Collect multi-source real-time data of the hanging basket body and the environment and preprocess it to obtain preprocessed multi-source real-time data.

[0023] Furthermore, in practical implementation, 24 tilt sensors (range ±10°, accuracy ±0.02°), 16 strain gauges (range ±1500με, accuracy ±1με), and 8 displacement sensors (range 200mm, accuracy ±0.1mm) are installed at key sections of the main beam of the hanging basket, major load-bearing members, and connection nodes. Simultaneously, 4 three-dimensional ultrasonic anemometers (range 0-30m / s, accuracy ±0.1m / s) and 2 sets of temperature and humidity sensors are installed at the top and bottom of the hanging basket. These sensors collect data synchronously at a sampling frequency of 20Hz, generating approximately 104 million raw data records daily.

[0024] Furthermore, the preprocessing of multi-source real-time data includes a three-level processing architecture: adaptive filtering denoising dynamically adjusts the process noise covariance matrix Q and the observation noise covariance matrix R through an improved Kalman filter algorithm, when the mean of the residual sequence... And variance At the same time, the diagonal elements of the Q matrix are automatically increased by a factor of 1.5 and the diagonal elements of the R matrix are decreased by a factor of 0.8. This dynamic adjustment mechanism improves the signal-to-noise ratio from the original 15dB to over 28dB. Time alignment uses the GPS clock as a unified reference and employs a cubic spline interpolation algorithm to resample the data stream with millisecond-level delays, controlling the time synchronization error of multi-source data to within 10ms. Format standardization converts the raw signals output by each sensor into physical quantity data with standard units and dimensions according to their respective calibration parameters, and encapsulates them into standard data frames containing timestamps, data source identifiers, and checksums.

[0025] It should be noted that the adaptive filtering mechanism effectively suppressed noise caused by electromagnetic interference and mechanical vibration at the construction site, reducing the fluctuation range of tilt angle data from ±0.15° to ±0.03°, providing stable and reliable data input for the digital twin. Precise time alignment ensured the temporal consistency of cross-physical domain data such as stress changes and wind speed fluctuations, laying a temporal foundation for subsequent analysis of load-response correlation. The standardized data frame structure enabled the digital twin to uniformly parse heterogeneous sensor data, overcoming the technical obstacles to multi-source heterogeneous data fusion. The synergistic application of these conventional data processing techniques in this invention provides a high-quality data foundation for subsequent digital twin construction and reinforcement learning decision-making. The improved data quality directly led to an approximately 15% increase in model prediction accuracy, demonstrating the supporting value of fundamental technologies in high-level applications.

[0026] S2: Based on preprocessed multi-source real-time data, establish an initial digital twin.

[0027] Furthermore, after establishing the initial digital twin, the digital twin is continuously updated in real time using preprocessed multi-source real-time data; the real-time update includes the following processes: Data synchronization process: Preprocessed multi-source real-time data is injected into the digital twin as its input boundary conditions; Model parameter correction process: Based on the deviation between the output of the mechanical simulation model inside the digital twin and the measured values ​​of the physical entity, the internal physical parameters of the mechanical simulation model are dynamically adjusted; State synchronization process: The current operating state of the digital twin is reset according to the data with a confidence evaluation value higher than a preset threshold.

[0028] This step aims to establish a high-fidelity digital twin that evolves in sync with the physical hanging basket, providing an accurate simulation environment for subsequent risk prediction and intelligent decision-making.

[0029] It should be noted that the digital twin is constructed based on preprocessed multi-source real-time data and the design parameters of the hanging basket, establishing an initial digital twin. This twin adopts a parametric finite element model, containing approximately 12,800 mesh elements (using a hybrid model of Solid186 solid elements and Beam188 beam elements). The material constitutive relation is defined as a bilinear hardening model, and its initial parameters are set according to the design drawings (e.g., elastic modulus 210 GPa, Poisson's ratio 0.3). Boundary conditions are accurately simulated through coupled equations to simulate the actual connection constraints between the hanging basket and the bridge structure. The initial loads include the structural self-weight and the design and construction loads.

[0030] It should also be noted that, to ensure the digital twin always accurately reflects the true state of the physical entity, the system performs the following three real-time updates: Data synchronization process: At a frequency of 20Hz, pre-processed multi-source real-time data (including tilt angle, stress, displacement, wind speed, wind direction, temperature and humidity) are used as time-varying boundary conditions and injected into the digital twin.

[0031] Model parameter calibration process: When the deviation norm between the mechanical simulation output of the digital twin and the measured value of the physical entity exceeds a threshold (0.1), parameter calibration is triggered. This process is based on a sensitivity analysis algorithm.

[0032] in, For parameter correction vector, The sensitivity matrix was obtained through finite element perturbation analysis. Sensitivity matrix The transpose of the matrix, This is a weight matrix, a diagonal matrix, where each element represents the confidence weight for each measurement point. is the regularization coefficient, set to 0.01 to prevent ill-posed matrix inversion. It is the identity matrix. This is the deviation vector between the measured values ​​and the model predictions.

[0033] State synchronization process: Real-time data is filtered based on a confidence evaluation function. When the confidence evaluation value of the data... Immediately reset the current operating state of the digital twin; for data with medium confidence ( Kalman filtering is used for progressive adjustment. The confidence evaluation function is defined as:

[0034] in, For the first The confidence assessment value of real-time monitoring data from each sensor (the value ranges from 0 to 1, with a larger value indicating higher data reliability). Accuracy class (0-1 normalized value, reflecting the sensor's measurement accuracy); The signal quality index (0-1 normalized value, calculated based on the signal-to-noise ratio). Historical reliability (0-1 normalized value, based on historical data consistency assessment).

[0035] Through the aforementioned real-time update mechanism, this step enables the digital twin to maintain a high degree of consistency with the physical entity, resulting in significantly higher model prediction accuracy compared to traditional static models. This dynamically evolving digital twin provides a reliable simulation foundation for forward simulation prediction, and its continuous optimization characteristics provide crucial assurance for improving the overall system performance.

[0036] S3: Forward simulation prediction using digital twins generates a multi-step overturning risk probability sequence.

[0037] Furthermore, the forward simulation prediction includes: based on the current state of the digital twin, parallel extrapolation of various discretized future environmental conditions, and calculation of overturning assessment indicators under each condition through multi-threaded dynamics solution.

[0038] It should be noted that forward simulation prediction is performed based on the updated digital twin. The system adopts an MPI+OpenMP hybrid parallel computing architecture to simultaneously simulate eight discretized future environmental conditions (including typical risk scenarios such as linear wind speed increase, sudden wind direction change, and step load change). Through multi-threaded dynamics solving, it calculates the overturning assessment indicators (including safety factor, center of gravity offset, and support reaction force distribution) for the next 60-120 seconds under each condition. The center of gravity offset and support reaction force distribution are used to assist in verifying the reliability of the safety factor. The core indicator is the safety factor. To transform this physical quantity into a probability value suitable for risk assessment and decision-making, this invention uses the Sigmoid function as a probability mapping tool to convert the safety factor... Overturning probability mapped to the 0-1 interval The calculation formula is as follows:

[0039] in, for The probability of the basket tipping over at any given moment. For shape parameters, a value of 3.0 is used to control the rate of probability change. The critical safety factor is set at 2.0 (based on the requirements of bridge construction safety specifications). The actual safety factor of the hanging basket is calculated in real time. is a natural constant with a value of approximately 2.71828. Here, it serves as the base of the Sigmoid function and is an inherent part of this standard mathematical function.

[0040] For data loss due to transmission delay, an autoregressive (AR) model is used for short-term data imputation, and its formula is as follows:

[0041] in, for Fill in the missing moments with data values. For the first The autoregressive coefficients were obtained by training on historical monitoring data using the least squares method. The model order is set to 3-5 (determined based on the Akaike Information Criterion, AIC). It is a white noise sequence (mean is 0, variance is constant).

[0042] It should also be noted that in the steel bridge construction scenario, the structural status data of the hanging basket has strong temporal correlation on a short time scale. Using an AR model with low computational complexity to fill in the missing data at the millisecond to second level can maintain the temporal continuity of the digital twin input data while ensuring real-time performance, thus providing a stable data foundation for forward simulation and prediction.

[0043] Furthermore, the overturning assessment indicators of various working conditions are integrated to generate a multi-step overturning risk probability sequence; the sequence is an array of probability values ​​arranged in chronological order, where each probability value is obtained by weighted fusion of the extrapolation results of multiple working conditions and environmental uncertainty factors.

[0044] It should be noted that by integrating the simulation results of all operating conditions and incorporating environmental uncertainties (such as wind speed prediction errors), a weighting factor is assigned to each operating condition. (Calculated as: the ratio of the historical occurrence frequency of this working condition to the sum of the total occurrence frequencies of all working conditions, determined based on historical statistical likelihood), and finally, a multi-step overturning risk probability sequence arranged in chronological order is generated through weighted fusion. .

[0045] This step achieves advanced quantitative early warning of capsizing risk through parallel simulation of multiple operating scenarios and probabilistic risk assessment. The Sigmoid mapping function smoothly transforms continuous safety coefficients into probabilities, solving the inherent mutability and insensitivity of traditional binary threshold judgments (safe / dangerous), and providing richer and more stable gradient information for subsequent reinforcement learning decisions. The weighted fusion mechanism effectively reduces the interference of environmental prediction uncertainty on the overall risk assessment by quantifying the probability of different future scenarios, making the prediction results more robust.

[0046] S4: Input the multi-step overturning risk probability sequence into the reinforcement learning policy network generated based on digital twin training, and output the anti-overturning dynamic adjustment command.

[0047] Furthermore, the reinforcement learning policy network is a deep neural network trained offline in a simulation environment constructed by a digital twin, through trial and error with interaction with the environment, with the goal of minimizing long-term overturning risk. The reinforcement learning policy network takes a multi-step overturning risk probability sequence as the state input and outputs specific anti-overturning dynamic adjustment commands.

[0048] It should be noted that the reinforcement learning policy network was trained offline through 5 million steps of interactive trial and error in the simulation environment constructed by the digital twin. The Actor-Critic framework was adopted: the input layer contains 12 nodes (corresponding to multi-step risk probability sequences), the hidden layer has 3 layers (256 nodes in each layer), and the output layer contains 3 nodes (corresponding to three anti-overturning adjustment commands: counterweight adjustment, cable force adjustment, and position correction). The output value is constrained to the range of [-1, 1] by the Tanh activation function.

[0049] During training, the reward function adopts a multi-objective weighted form to guide the policy network to learn a safe, stable, and efficient regulation strategy.

[0050] in, for Total reward at each moment; The weighting coefficients are 0.7, 0.2, and 0.1 respectively (determined based on the priority order of safety, stability, and energy efficiency in hanging basket construction and engineering experience). As a safety reward, it is calculated as follows: ( The overturning probability at time t, defined in step S3, is used to penalize high-risk behaviors. As a stability reward, it is calculated as follows: ( for Adjustment command vector at time, The square of the Euclidean distance between adjacent commands is used to penalize drastic adjustments and ensure system stability. For energy efficiency bonuses, the calculation is as follows: To execute Time adjustment command The required energy consumption is used to incentivize low-energy regulation.

[0051] The strategy network takes a risk probability sequence as state input and outputs specific anti-overturning dynamic adjustment commands through forward propagation of a deep neural network. A safety protection mechanism is set in the output layer: when the command exceeds the physical limits of the actuator (such as the counterweight adjustment speed > 50 mm / s), the projection gradient method (correcting the command parameters through gradient descent until the physical constraints of the actuator are met) is used to map the command into the feasible region, ensuring the safety and executability of the adjustment action.

[0052] This step utilizes a reinforcement learning policy network to intelligently generate anti-overturning control commands, addressing the poor adaptability of traditional rule-based control strategies under complex operating conditions. A multi-objective reward function ensures a balance between safety, stability, and energy efficiency, while a safety protection mechanism prevents commands from exceeding limits, providing reliable and adaptive decision support for hanging basket construction.

[0053] S5: Execute anti-overturning dynamic adjustment commands and feed back the execution results to the digital twin and reinforcement learning policy network for collaborative optimization.

[0054] Furthermore, collaborative optimization includes the following processes: a digital twin calibration process, which uses the actual state effect data after the execution of anti-overturning dynamic adjustment commands to perform reverse calibration on the digital twin; and a reinforcement learning policy network optimization process, which transforms the execution effect of anti-overturning dynamic adjustment commands into strong... The reward signals for learning are used to fine-tune the reinforcement learning strategy network online.

[0055] It should be noted that the anti-overturning dynamic adjustment command output by step S4 is executed to control the hydraulic system and electric hoist to complete corresponding actions such as counterweight adjustment, cable force adjustment, and position correction; the actual state effect data after the command is executed (such as stress redistribution, attitude change, support reaction force adjustment results, etc.) are collected simultaneously, and the effect data is fed back to the digital twin and reinforcement learning policy network to achieve collaborative optimization between the two.

[0056] Collaborative optimization includes the following two processes: Digital twin calibration process: The digital twin is reverse-calibrated using actual state effect data, and a dual-trigger mechanism is employed to ensure model accuracy. Immediate correction: When the measured values ​​of key state parameters (such as main beam inclination angle, main member stress, and front displacement) deviate from the twin simulation values... (When the units are consistent with the corresponding parameters, such as tilt deviation in ° and displacement deviation in mm), a forced correction is triggered immediately.

[0057] Periodic calibration: A comprehensive parameter optimization is performed every 30 minutes, with the objective function being:

[0058] in, For the first The weighting coefficients for each measurement point (assigned based on the importance of the measurement point); For the first Actual measured values ​​of key state parameters (such as stress, displacement, and tilt angle); For the first Digital twin simulation values ​​of key state parameters; This is the regularization factor, set to 0.01 (to prevent excessive parameter adjustment). Adjust the parameter vector (as defined in step one, including physical parameters such as connection stiffness and damping coefficient).

[0059] Policy network optimization process: The effect of instruction execution is converted into a reward signal for reinforcement learning (calculated based on the reward function defined in step S4, where the safety reward is calculated). of The updated probability is the actual overturning probability after instruction execution. Online fine-tuning of the reinforcement learning policy network is then performed. Constrained policy optimization ensures update stability, requiring the KL divergence between the old and new policies to satisfy:

[0060] in, for Kullback-Leibler divergence is used to measure the difference in probability distributions between the old and new strategies. The old policy network before optimization; For the optimized new policy network; for The divergence threshold is set to 0.01 (to control the magnitude of policy updates).

[0061] The learning rate during the fine-tuning phase is set to 1 / 10 of that during the main training phase to ensure policy stability; simultaneously, a priority experience replay buffer is established, based on the temporal difference error. Assign priorities to each experience sample, giving priority to reusing samples with high learning value. Temporal difference error. The calculation formula is:

[0062] in, For instant rewards; Discount factor; The value function of the current state; The value function for the next state.

[0063] This step, through feedback and collaborative optimization, achieves continuous improvement in the accuracy of the digital twin model and the decision-making ability of the reinforcement learning policy network, forming a self-evolution mechanism for the system. Dual-triggered correction ensures the model's real-time performance and accuracy, while constrained policy optimization avoids the risk of policy mutation, effectively solving the problems of model rigidity and policy incompatibility with dynamic environments in traditional methods.

[0064] Example 2, an embodiment of the present invention, is a dynamic adjustment method for anti-overturning of a hanging basket for steel bridge construction. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0065] In a cantilever construction project of the main girder of a cross-river bridge, a 30-day field comparative test was conducted to verify the effectiveness of the anti-overturning dynamic adjustment method proposed in this invention. Two identical hanging basket systems were selected for the test, employing the traditional PID control method (control group) and the intelligent dynamic adjustment method proposed in this invention (experimental group), respectively. The rated load-bearing capacity of the experimental hanging basket was 225 tons, symmetrically arranged on both sides of the bridge pier, with a construction segment length of 6 meters. The data acquisition system was precisely arranged according to the invention requirements, with 24 tilt sensors (range ±10°, accuracy ±0.02°), 16 strain gauges (range ±1500με, accuracy ±1με), and 8 displacement sensors (range 200mm, accuracy ±0.1mm) placed at key sections of the main girder, main load-bearing members, and connection nodes of the hanging basket. Simultaneously, four three-dimensional ultrasonic anemometers (range 0-30m / s, accuracy ±0.1m / s) and two sets of temperature and humidity sensors were placed at the top and bottom of the hanging basket. All sensors collect data synchronously at a sampling frequency of 20Hz, generating approximately 104 million raw data records per day.

[0066] The digital twin was constructed using a parametric finite element model containing 12,800 mesh elements, employing a hybrid modeling approach using Solid186 solid elements and Beam188 beam elements. During the 30-day trial period, the digital twin management module continuously performed real-time updates, synchronizing data every 50 milliseconds and injecting preprocessed multi-source real-time data as boundary conditions into the digital twin. The model parameter calibration process dynamically adjusted the internal physical parameters of the mechanical simulation model based on the deviation between the output of the mechanical simulation model within the digital twin and the measured values ​​of the physical entity. The state synchronization process reset the digital twin's operating state based on the confidence scores of each data source, resetting data exceeding a preset threshold of 0.85.

[0067] The overturning risk prediction module, based on the current state of the digital twin, simulates eight discretized future environmental conditions in parallel, including typical risk scenarios such as linear wind speed increase, sudden wind direction change, and step load change. It calculates overturning assessment indicators for each condition through multi-threaded dynamics. The reinforcement learning decision-making module employs a deep neural network trained offline for 5 million steps. Using a multi-step overturning risk probability sequence as state input, it outputs three anti-overturning dynamic adjustment commands: counterweight adjustment, cable tension adjustment, and position correction. The command execution and collaborative optimization module converts the adjustment commands into control signals for the actuators. After command execution, it collects actual state effect data, performs reverse correction on the digital twin, and converts the execution effect into a reward signal for reinforcement learning, fine-tuning the policy network online.

[0068] The system performance was recorded under various conditions during the test, including normal construction, sudden strong winds, and load changes. To quantitatively evaluate system performance, seven key performance indicators were selected for comparative analysis. Specific test data are recorded below:

[0069] Based on comparative analysis of experimental data, the anti-overturning dynamic adjustment method proposed in this invention significantly outperforms the traditional PID control method in all performance indicators. Regarding the maximum overturning risk probability, the method of this invention controls the risk probability within the range of 23.1%-34.3% under normal operating conditions, a reduction of approximately 41.8% compared to the traditional method's 39.2%-62.9%. Particularly under extreme operating conditions, the method of this invention reduces the risk probability from 77.5% of the traditional method to 45.2%, a reduction of 41.7%. This significant improvement is mainly attributed to the well-designed sensor network layout and the forward simulation and prediction capabilities of the digital twin. The dense arrangement of 24 tilt sensors and 16 strain gauges provides the system with abundant state perception data, enabling the system to identify risk trends in advance and take preventative adjustment measures.

[0070] In terms of system response performance, the average adjustment response time of the method in this invention is only 2.0-2.6 seconds, which is about 49.5% faster than the 4.3-5.1 seconds of the traditional method. This significant improvement in response speed stems from the parallel computing architecture of the reinforcement learning policy network and its rapid processing capability for multi-source data. The system can process 104 million data records generated daily in real time and generate optimal adjustment instructions based on multi-step risk probability sequences. Simultaneously, the adjustment energy consumption of the method in this invention is significantly reduced, reaching 7.6-9.1 kWh under the same operating conditions, a reduction of approximately 31.2% compared to the 11.7-14.6 kWh of the traditional method, demonstrating the advantages of the multi-objective reward function in balancing safety and economy.

[0071] In terms of control accuracy, the method of this invention exhibits superior performance. The maximum tilt angle is controlled within the range of 1.13°-1.59°, an improvement of approximately 46.3% compared to the traditional method's 1.92°-3.22°; the stress fluctuation coefficient is reduced from 16.8%-21.9% in the traditional method to 9.4%-13.1%, a reduction of 41.2%; and the command overshoot is reduced from 19.9%-27.3% to 10.1%-13.8%, an improvement of approximately 46.5%. These improvements demonstrate the effectiveness of the dense sensor deployment and the real-time update mechanism of the digital twin. Through the synergistic effect of data synchronization, model parameter correction, and state synchronization, the consistency between the virtual model and the physical entity is ensured. The system stability index is improved from 72.4-78.6 in the traditional method to 87.4-91.8, reflecting the robustness and adaptability of the method of this invention under complex time-varying conditions. Comprehensive test data shows that this invention, through the deep integration of a sophisticated sensor network, digital twins, and reinforcement learning, has achieved an intelligent upgrade of the hanging basket anti-overturning control, providing a reliable technical guarantee for the safety of steel bridge construction.

[0072] Example 3, one embodiment of the present invention, provides a system for a dynamic adjustment method for anti-overturning of a hanging basket in steel bridge construction, including a multi-source data acquisition and preprocessing module, a digital twin management module, an overturning risk prediction module, a reinforcement learning decision-making module, and an instruction execution and collaborative optimization module.

[0073] The system comprises the following modules: a data acquisition and preprocessing module for collecting multi-source real-time data from the hanging basket and its environment, and preprocessing the data; a digital twin management module for establishing and maintaining a digital twin synchronized with the physical hanging basket based on the preprocessed multi-source real-time data; a tipping risk prediction module for performing forward simulation prediction using the digital twin to generate a multi-step tipping risk probability sequence; a reinforcement learning decision module for inputting the multi-step tipping risk probability sequence into a pre-trained reinforcement learning policy network and outputting anti-tipping dynamic adjustment commands; and a command execution and collaborative optimization module for executing anti-tipping dynamic adjustment commands and feeding back the actual state effect data after command execution to the digital twin management module and the reinforcement learning decision module for collaborative optimization.

[0074] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0076] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0077] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc. It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for dynamic adjustment of anti-overturning properties of a formwork for steel bridge construction, characterized in that, include: Collect multi-source real-time data of the hanging basket body and the environment and preprocess it to obtain preprocessed multi-source real-time data; An initial digital twin is established based on preprocessed multi-source real-time data; The digital twin is used for forward simulation prediction to generate a multi-step overturning risk probability sequence; The multi-step overturning risk probability sequence is input into a strong algorithm generated based on the digital twin training. The learning strategy network outputs dynamic adjustment commands to resist overturning. Execute the aforementioned anti-overturning dynamic adjustment command and feed back the execution effect to the digital twin and the strong Learn the learning strategy network for collaborative optimization.

2. The method for dynamic adjustment of anti-overturning of hanging basket for steel bridge construction according to claim 1, characterized in that: The multi-source real-time data includes structural state data from the hanging basket itself and environmental load data from the external environment; wherein, the structural state data includes tilt angle data, stress data and displacement data, and the environmental load data includes wind speed, wind direction, temperature and humidity data.

3. The method for dynamic adjustment of anti-overturning of hanging basket for steel bridge construction according to claim 2, characterized in that: The preprocessing includes adaptive filtering and noise reduction, time alignment, and format standardization of multi-source real-time data.

4. The method for dynamic adjustment of anti-overturning of hanging basket for steel bridge construction according to claim 3, characterized in that: After establishing the initial digital twin, the digital twin is continuously updated in real time using preprocessed multi-source real-time data; the real-time update includes the following process: Data synchronization process: The preprocessed multi-source real-time data is injected into the digital twin as its input boundary conditions; Model parameter calibration process: Based on the deviation between the output of the mechanical simulation model inside the digital twin and the measured values ​​of the physical entity, the internal physical parameters of the mechanical simulation model are dynamically adjusted; State synchronization process: Based on the quality assessment of each data source, the current operating state of the digital twin is reset according to the data whose confidence evaluation value is higher than the preset threshold.

5. The method for dynamic adjustment of anti-overturning of hanging basket for steel bridge construction according to claim 4, characterized in that: The forward simulation prediction includes: based on the current state of the digital twin, extrapolating multiple discretized future environmental conditions in parallel, and calculating the overturning assessment index under each condition through multi-threaded dynamics solution.

6. The method for dynamic adjustment of anti-overturning of hanging basket for steel bridge construction according to claim 5, wherein... The feature is that: the overturning assessment indicators of each working condition are integrated to generate a multi-step overturning risk probability sequence; the sequence is an array of probability values ​​arranged in chronological order, wherein each probability value is obtained by weighted fusion of the extrapolation results of the multiple working conditions and environmental uncertainty factors.

7. The method for dynamic adjustment of anti-overturning of hanging basket for steel bridge construction according to claim 6, characterized in that: The reinforcement learning policy network is a deep neural network trained offline in the simulation environment constructed by the digital twin, through trial and error with interaction with the environment, with the goal of minimizing long-term overturning risk; the reinforcement learning policy network takes the multi-step overturning risk probability sequence as the state input and outputs specific anti-overturning dynamic adjustment commands.

8. The method for dynamic adjustment of anti-overturning of hanging basket for steel bridge construction according to claim 7, characterized in that: The collaborative optimization Includes the following processes: The digital twin calibration process utilizes the actual state effect data after the execution of the anti-overturning dynamic adjustment command to perform reverse calibration on the digital twin; The reinforcement learning strategy network optimization process transforms the execution effect of anti-overturning dynamic adjustment commands into strong... The reward signals from the learning process are used to fine-tune the reinforcement learning strategy network online.

9. A system employing the anti-overturning dynamic adjustment method for a formwork used in steel bridge construction as described in any one of claims 1 to 8, characterized in that: It includes a multi-source data acquisition and preprocessing module, a digital twin management module, a tipping risk prediction module, a reinforcement learning decision-making module, and an instruction execution and collaborative optimization module; The data acquisition and preprocessing module is used to acquire multi-source real-time data of the hanging basket body and the environment, and to preprocess the data. The digital twin management module establishes and maintains a digital twin synchronized with the physical hanging basket based on preprocessed multi-source real-time data. The overturning risk prediction module is used to perform forward simulation prediction using the digital twin to generate a multi-step overturning risk probability sequence. The reinforcement learning decision module is used to input the multi-step overturning risk probability sequence into the pre-trained reinforcement learning policy network and output anti-overturning dynamic adjustment instructions. The instruction execution and collaborative optimization module is used to execute the anti-overturning dynamic adjustment instruction and feed back the actual state effect data after instruction execution to the digital twin management module and the reinforcement learning decision module for collaborative optimization.

Images