Bridge rotation real-time intelligent control method based on machine learning

Abstract

The invention discloses a bridge swivel real-time intelligent control method based on machine learning, and the method comprises the steps: constructing a multi-source sensing data fusion system, and deploying a global state sensing array; receiving a multi-source state data stream output by the deployed sensor array in real time; the time domain change rate and the frequency domain energy characteristics are subjected to parallel calculation, and intelligent mapping from physical signals to information-intensive characteristics is carried out; constructing a hybrid machine learning control model, and designing a multi-target reward function; establishing an online optimization and learning mechanism, and developing a Bayesian optimization-online learning cooperation system; creating a virtual-real interaction verification platform, and constructing a multi-class typical working condition test set; edge calculation and stability guarantee are deployed, an FPGA and GPU collaborative heterogeneous calculation architecture is designed, a real-time control period is achieved, the robustness of the system is evaluated through Monte Carlo simulation, and the contribution degree of each parameter to the control stability is analyzed; the method is suitable for multi-source disturbance suppression and millimeter-level precise control in swivel construction of the swivel bridge under complex construction conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent bridge construction and industrial control technology, and in particular to a real-time intelligent control method for bridge rotation based on machine learning. Background Technology

[0002] The real-time control accuracy and safety reliability of bridge rotation construction are influenced by multiple dynamic factors, including environmental disturbances (wind load, temperature), equipment status (hydraulic pressure, sensor drift), and structural response (stress concentration, beam deflection), exhibiting strong nonlinear and time-varying characteristics. Traditional manual control relies on empirical threshold adjustments and offline finite element simulation, which suffers from severe response lag, insufficient coverage of working conditions, and high verification costs, failing to meet the real-time risk management requirements of modern construction.

[0003] Existing improvement schemes, such as fuzzy PID control or offline reinforcement learning (e.g., the DDPG algorithm), improve some adaptability, but still have fundamental bottlenecks such as long computational delays in model predictive control, the inability of neural network feedforward control to evolve strategies online, and the lack of direct connection between traditional digital twin verification and edge controllers.

[0004] This patented technology utilizes a real-time control architecture that integrates machine learning, edge computing, and deep reinforcement learning to overcome the problems of traditional bridge rotation construction relying on manual labor and struggling to suppress dynamic disturbances. By constructing an LSTM-TD3 hybrid decision-making framework and dynamic feature engineering, and integrating a high-concurrency hardware acceleration platform (FPGA+GPU), it achieves 10ms-level generation and online evolution of hydraulic system commands. The deep integration of machine learning and digital twin technology reveals the sensitivity of key variables such as control delay, wind disturbance energy, and hydraulic fluctuations. This represents a cutting-edge exploration in the field of intelligent bridge construction, propelling bridge construction from "manual intervention" to "autonomous evolution," and possesses significant engineering application value. Summary of the Invention

[0005] To overcome the shortcomings in the aforementioned related technical research, and addressing the issues of lag and insufficient precision in traditional bridge rotation construction methods, the purpose of this invention is to propose a real-time intelligent control method for bridge rotation based on machine learning. This invention is applicable to the suppression of multi-source disturbances and millimeter-level precise control in the rotation construction of bridges under complex construction conditions.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a real-time intelligent control method for bridge rotation based on machine learning, comprising the following steps:

[0007] Step 1: Construct a multi-source sensor data fusion system, deploy a global state awareness array, establish a real-time data communication architecture based on the OPC UA protocol, and realize multi-source data timestamp synchronization;

[0008] Step 2: Develop a dynamic feature engineering module to receive multi-source state data streams from the sensor array deployed in Step 1 in real time at the input layer. Through a hybrid feature processing engine, the time-domain rate of change and frequency-domain energy features are calculated in parallel, and a third-order feature fusion architecture is used to achieve intelligent mapping from physical signals to information-intensive features.

[0009] Step 3: Construct a hybrid machine learning control model, build a hybrid control framework based on LSTM-TD3, and design a multi-objective reward function;

[0010] Step 4: Establish an online optimization and learning mechanism, develop a Bayesian optimization-online learning collaborative system, use the GP-UCB algorithm to globally optimize the hyperparameters of the control strategy, and embed an incremental online learning module: online data cache pool, elastic weight solidification (EWC) to prevent catastrophic forgetting, and priority experience replay (PER) to accelerate convergence;

[0011] Step 5: Create a virtual-real interaction verification platform, develop a digital twin system based on the Unity3D engine, and build a test set of 20 typical working conditions;

[0012] Step 6: Deploy edge computing and ensure stability. Design a heterogeneous computing architecture (FPGA+GPU collaboration) to achieve a 10ms real-time control cycle. Evaluate the system robustness through 100,000 Monte Carlo simulations (failure rate < 0.01%). Use the Sobol exponential method to analyze the contribution of each parameter to control stability.

[0013] Preferably, in step 1, key physical quantities of the bridge rotation system are collected in real time at the input layer, including rotation angle deviation (…). ), structural stress extreme values ​​( Real-time wind speed ), hydraulic pressure ( ) and other multi-dimensional parameters. A global state awareness network is constructed using a high-precision sensor array, with each node achieving clock synchronization error less than 0.5 mm based on the IEEE 1588 v2 precision time protocol. .

[0014] The general form of clock synchronization error is:

[0015]

[0016] in, The master clock timestamp This is the sensor's local timestamp.

[0017] Preferably, in step 1, deterministic transmission technology (TSN+OPC UA protocol stack) is used to ensure end-to-end latency <5ms. The final output time-aligned raw data stream (100Hz sampling rate) is directly transmitted to the feature extraction module in step 2.

[0018] Preferably, in step 2, the multi-source state data stream output from the sensor array deployed in step 1 is received in real time at the input layer, including the rotation angle deviation ( ), structural stress extreme values ​​( Real-time wind speed ), hydraulic pressure ( (and other original sensing vectors.)

[0019] The general form of the rate of change of tilt angle is:

[0020]

[0021] in, For the current time ( The actual tilt angle of the rotating beam, for The historical tilt from seconds ago.

[0022] The general form of the stress safety factor is:

[0023]

[0024] in, For triaxial stress components, This refers to the dynamic yield strength of steel.

[0025] As a preferred option, in step 2, a hybrid feature processing engine is used to calculate the time-domain rate of change and frequency-domain energy features in parallel, and a third-order feature fusion architecture is used to realize intelligent mapping from physical signals to information-intensive features.

[0026] The general form of effective energy from wind disturbance is:

[0027]

[0028] in, This represents the wind speed power spectral density.

[0029] The general form of the temperature-pressure cross-correlation factor is:

[0030]

[0031] in, Let i be the hydraulic oil temperature. Let i be the hydraulic pressure at time i. The number of samples in the sliding window. This is the arithmetic mean of temperature / pressure within the window.

[0032] Preferably, in step 3, the 15-dimensional real-time feature vector generated in step 2 is received at the input layer, including the tilt angle change rate (…). ), stress safety factor ( ), wind disturbance effective energy ( Key physical indicators such as [missing information]. An LSTM encoder processes the most recent 100 historical state sequences (with a 1-second time window), outputting a 128-dimensional spatiotemporal feature vector. The dynamic hysteresis characteristics of the rotation process are captured. The action value (Q value) is evaluated in parallel through a dual Critic network with a fully connected layer structure (256→128→64) to suppress Q value overestimation.

[0033] Preferably, in step 3, continuous control actions are generated based on spatiotemporal features. This is mapped to the traction machine power adjustment command (±1000kW), and the dynamic adjustment coefficient of the PID parameters is output simultaneously. Design a multi-objective reward function and define a dynamic weighting function.

[0034] The general form of a multi-objective reward function is:

[0035]

[0036] in, For a moment Even the reward value, This represents the absolute deviation between the actual tilt angle and the target value. It is the ratio of maximum stress to allowable stress. This represents the absolute deviation between the instantaneous oil pressure and the average oil pressure. for The absolute deviation of the real-time value from the reference. The control period (value is 0.01s) is used. For the safety status indication function (in) (Time is 1).

[0037] As a preferred option, in step 4, a probability mapping function from the hyperparameter space to the control value is constructed using the Gaussian process surrogate model (GP-UCB algorithm) to achieve Bayesian global optimization of the hyperparameters of the TD3 network.

[0038] The general form of Gaussian process modeling is:

[0039]

[0040] Among them, the mean function (Standardized processing) This is the kernel function.

[0041] The general form of the function for obtaining the upper confidence bound is:

[0042]

[0043] Among them, the balance coefficient (corresponding to the 95% confidence interval) The mean, The standard deviation is denoted as .

[0044] The general form of the optimization objective function is:

[0045]

[0046] in, The set of TD3 network parameters to be optimized ( Constrained by the stability threshold of the hydraulic system discovered in the step response experiment , (This represents the hydraulic response attenuation coefficient corresponding to the current β). The average performance score of the strategy on dataset D (D contains: 50 sets of historical engineering condition data and 10 sets of real-time wind disturbance data) is given. This represents the standard deviation of the bonus value under different working conditions.

[0047] Preferably, in step 4, the transfer learning initialization pre-training stage involves generating rotational data in a finite element simulation environment and training the LSTM encoder until convergence (loss function). During the fine-tuning phase, the underlying parameters of the encoder are frozen, and only the Actor network is optimized. Elastic Weight Fixation (EWC) is used to prevent catastrophic forgetting.

[0048] The general form of Elastic Weight Fixed (EWC) is as follows:

[0049]

[0050] in, For the total loss function, The loss function for the current task. Regularization coefficient (value is...) ), The diagonal elements of the Fisher information matrix. These are the parameters to be optimized during fine-tuning. These are the parameters for the pre-trained model.

[0051] Preferably, in step 4, Priority Experience Replay (PRE) uses Temporal Differential Error (TD Error) to calculate the initial priority for each new experience. After each training iteration, the priority of the corresponding experience is updated based on the latest TD Error. A base priority (initial priority) is assigned when storing new experiences, and the priority of the corresponding experience is updated after each sampling training iteration. A power-law distribution is used to adjust the sampling weights, and importance sampling and normalization are used to mitigate the bias introduced by priority sampling, ensuring convergence.

[0052] The general form of timing differential error (TD Error) is:

[0053]

[0054] in, For instant rewards, The current state action value. The next state action value predicted by the target network. This is the discount factor.

[0055] The general form of the initial priority is:

[0056]

[0057] in, =0.01 to prevent 0 priority.

[0058] The general form of the power-law distribution is:

[0059]

[0060] in, Priority intensity control factor ( , At that time, it degenerates into uniform sampling; (At that time, sampling is performed entirely according to priority)

[0061] The general form for calculating importance sampling weights is:

[0062]

[0063] Where N is the capacity of the experience pool. The deviation correction intensity control factor is set (initially 0.4, gradually increasing linearly to 1.0).

[0064] Preferably, in step 5, the multi-source state data stream output from the sensor array deployed in step 2 is received in real time at the physical interface layer, including the rotation angle deviation. Structural stress Wind speed Iso-sensing vectors. By integrating a finite element dynamics model with a fluid coupling solver using Unity3D's high-fidelity physics engine, a fully digital twin of the bridge-hydraulic system is constructed.

[0065] Preferably, in step 5, a working condition generation engine is designed. Based on the Latin hypercube sampling and fault injection protocol, it automatically creates 20 typical working condition test sets (including scenarios such as extreme wind load, sensor drift, and hydraulic leakage), and achieves millisecond-level data synchronization with the edge control system through the ROS# communication protocol. Driven by GPU cluster acceleration, it achieves real-time working condition rendering at 2000 frames per second, outputting quantitative verification reports such as trajectory tracking error and stress distribution cloud map, providing closed-loop certification with a working condition coverage of ≥98% for the control system.

[0066] Preferably, in step 6, a heterogeneous computing architecture based on FPGA and GPU is constructed. The data processing layer receives the optimization control strategy from step 4 and the output parameters of the verification platform from step 5, and allocates computing tasks through the Time-Sensitive Network (TSN) scheduling algorithm.

[0067] The general form of the Time-Sensitive Network (TSN) scheduling algorithm is:

[0068]

[0069] in, This is the closed-loop control cycle of the system. Calculate the time taken for the i-th subtask. To increase the number of cores for parallel computing, This refers to the FPGA-GPU communication latency.

[0070] Preferably, in step 6, the FPGA processing unit processes the real-time control flow, including state estimation and control command generation, while the GPU coprocessor performs Monte Carlo simulation and Sobol sensitivity analysis. Through 100,000 Monte Carlo perturbation tests, the system failure rate is less than 0.01%. Simultaneously, based on the Sobol exponent, the weights of multiple parameters on stability are analyzed, and the contribution of key factors such as hydraulic delay is output for real-time optimization of the control strategy.

[0071] The general form of the failure rate determination formula is:

[0072]

[0073] in, Number of failure simulations. (Failure condition is either excessive angular acceleration or excessive displacement deviation);

[0074] The general form of the Sobol sensitivity analysis formula is:

[0075]

[0076] in, Let i be the global sensitivity index for parameter i. For output indicators Total variance To fix other parameters ( )hour The expected condition.

[0077] This invention forms a closed-loop control architecture of "perception-decision-verification-execution" through high-precision sensor array deployment, dynamic feature engineering construction, LSTM-TD3 hybrid decision framework, Bayesian optimization incremental learning, edge heterogeneous acceleration platform, and stability verification system. The specific implementation steps are as follows: deploying multi-source sensors to achieve 10μs-level time synchronization of multi-source data based on the IEEE 1588 v2 protocol; constructing a dynamic feature vector with time-frequency fusion; using a dual-layer LSTM to encode historical state sequences and output 128-dimensional spatiotemporal features, which are then used to generate hydraulic power compensation commands and PID parameter dynamic tuning coefficients via a dual Critic-TD3 network; optimizing network hyperparameters through the GP-UCB algorithm, and combining Elastic Weight Consolidation (EWC) and Priority Experience Playback (PER) to achieve continuous evolution of the control strategy; achieving a 10ms control cycle based on FPGA+GPU heterogeneous computing, with a failure rate of <0.01% verified by 100,000 Monte Carlo simulations; and using Sobol sensitivity analysis to analyze key parameters such as hydraulic delay and drive hardware redundancy optimization. This invention solves the problems of low precision and high risk in traditional manual control, reduces rotation angle error, reduces energy consumption, and significantly improves the safety and efficiency of bridge rotation in complex environments.

[0078] The beneficial effects of this invention are:

[0079] This invention, through the collaborative innovation of a machine learning edge intelligent control architecture and a digital twin verification platform, constructs a full-link intelligent bridge rotation construction system encompassing "perception-decision-verification-control." It overcomes three major challenges in traditional manual bridge rotation construction: large trajectory deviation, slow response, and weak robustness, achieving real-time control optimization of the rotation construction process. Based on an LSTM-TD3 hybrid control framework, it tracks rotation angle errors and dynamically adjusts control strategy weights using a Bayesian optimization-online learning mechanism. Furthermore, it utilizes a digital twin platform for proactive defense against 20 types of extreme working conditions. Finally, a 10ms real-time control cycle is achieved through FPGA+GPU edge nodes, and the system failure rate is verified through 100,000 Monte Carlo simulations. This method simultaneously improves construction safety and energy efficiency, providing a reusable standardized technical paradigm for intelligent rotation of large bridges. Attached Figure Description

[0080] Figure 1 This is a flowchart of an embodiment of the present invention;

[0081] Figure 2 This is a comparison chart of angle tracking accuracy during the rotation process in an embodiment of the present invention;

[0082] Figure 3 This is a comparison chart of energy consumption for different control methods in embodiments of the present invention;

[0083] Figure 4 This is a comparison chart of control stability under different wind speed conditions in the embodiments of the present invention;

[0084] Figure 5 This is a schematic diagram illustrating the change in the reward function value during the training process in an embodiment of the present invention. Detailed Implementation

[0085] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0086] Example

[0087] like Figure 1 As shown, a real-time intelligent control method for bridge rotation based on machine learning specifically includes the following steps:

[0088] Step 1: Construct a multi-source sensor data fusion system, deploy a global state perception array, establish a real-time data communication architecture based on the OPC UA protocol, and realize multi-source data timestamp synchronization.

[0089] Step 2: Develop a dynamic feature engineering module to receive multi-source state data streams from the sensor array deployed in Step 1 in real time at the input layer. Through a hybrid feature processing engine, time-domain change rate and frequency-domain energy features are calculated in parallel, and a third-order feature fusion architecture is used to achieve intelligent mapping from physical signals to information-intensive features.

[0090] Step 3: Construct a hybrid machine learning control model, build a hybrid control framework based on LSTM-TD3, and design a multi-objective reward function.

[0091] Step 4: Establish an online optimization and learning mechanism, develop a Bayesian optimization-online learning collaborative system, use the GP-UCB algorithm to globally optimize the hyperparameters of the control strategy, and embed an incremental online learning module: online data cache pool, elastic weight solidification (EWC) to prevent catastrophic forgetting, and priority experience replay (PER) to accelerate convergence.

[0092] Step 5: Create a virtual-real interaction verification platform, develop a digital twin system based on the Unity3D engine, and build a test set of 20 typical working conditions.

[0093] Step 6: Deploy edge computing and ensure stability. Design a heterogeneous computing architecture (FPGA+GPU collaboration) to achieve a 10ms real-time control cycle. Evaluate the system robustness through 100,000 Monte Carlo simulations (failure rate < 0.01%). Use the Sobol exponential method to analyze the contribution of each parameter to control stability.

[0094] The working principle of this embodiment will be explained in detail below:

[0095] For step 1, key physical quantities of the bridge rotation system are collected in real time at the input layer, including rotation angle deviation ( ), structural stress extreme values ​​( Real-time wind speed ), hydraulic pressure ( ) and other multi-dimensional parameters. A global state awareness network is constructed using a high-precision sensor array, with each node achieving clock synchronization error less than 0.5 mm based on the IEEE 1588 v2 precision time protocol. .

[0096] Furthermore, the general form of clock synchronization error is:

[0097]

[0098] in, The master clock timestamp This is the sensor's local timestamp.

[0099] For step 1, deterministic transmission technology (TSN+OPC UA protocol stack) is used to ensure end-to-end latency <5ms. The final output is a time-aligned raw data stream (100Hz sampling rate). The performance data of the sensing and communication system in this embodiment are shown in Table 1.

[0100] Table 1. Performance data of sensing and communication systems

[0101]

[0102] For step 2, the multi-source state data stream output from the sensor array deployed in step 1 is received in real time at the input layer, including the rotation angle deviation ( ), structural stress extreme values ​​( Real-time wind speed ), hydraulic pressure ( (and other original sensing vectors.)

[0103] Furthermore, the general form of the rate of change of inclination angle is:

[0104]

[0105] in, For the current time ( The actual tilt angle of the rotating beam, for The historical tilt from seconds ago.

[0106] Furthermore, the general form of the stress safety factor is:

[0107]

[0108] in, For triaxial stress components, This refers to the dynamic yield strength of steel.

[0109] For step 2, a hybrid feature processing engine is used to calculate the time-domain rate of change and frequency-domain energy features in parallel, and a third-order feature fusion architecture is used to realize intelligent mapping from physical signals to information-intensive features.

[0110] Furthermore, the general form of effective energy from wind disturbance is:

[0111]

[0112] in, This represents the wind speed power spectral density.

[0113] Furthermore, the general form of the temperature-pressure cross-correlation factor is:

[0114]

[0115] in, Let i be the hydraulic oil temperature. Let i be the hydraulic pressure at time i. The number of samples in the sliding window. This represents the arithmetic mean of temperature / pressure within the window. The output range of the dynamic feature engineering in this embodiment is shown in Table 2.

[0116] Table 2 Output Range of Dynamic Feature Engineering

[0117]

[0118] For step 3, the input layer receives a 15-dimensional real-time feature vector generated in step 2, including the tilt rate of change ( ), stress safety factor ( ), wind disturbance effective energy ( Key physical indicators such as [missing information]. An LSTM encoder processes the most recent 100 historical state sequences (with a 1-second time window), outputting a 128-dimensional spatiotemporal feature vector. The dynamic hysteresis characteristics of the rotation process are captured. The action value (Q value) is evaluated in parallel through a dual Critic network with a fully connected layer structure (256→128→64) to suppress Q value overestimation.

[0119] For step 3, continuous control actions are generated based on spatiotemporal features. This is mapped to the traction machine power adjustment command (±1000kW), and the dynamic adjustment coefficient of the PID parameters is output simultaneously. Design a multi-objective reward function and define a dynamic weighting function.

[0120] Furthermore, the general form of the multi-objective reward function is:

[0121]

[0122] in, For a moment Even the reward value, This represents the absolute deviation between the actual tilt angle and the target value. It is the ratio of maximum stress to allowable stress. This represents the absolute deviation between the instantaneous oil pressure and the average oil pressure. for The absolute deviation of the real-time value from the reference. The control period (value is 0.01s) is used. For the safety status indication function (in) (Time is 1). The training performance metrics of the control model in this embodiment are shown in Table 3.

[0123] Table 3. Control Model Training Performance

[0124]

[0125] For step 4, a probability mapping function from the hyperparameter space to the control value is constructed using the Gaussian process surrogate model (GP-UCB algorithm), thereby realizing the Bayesian global optimization of the hyperparameters of the TD3 network.

[0126] Furthermore, the general form of Gaussian process modeling is:

[0127]

[0128] Among them, the mean function (Standardized processing) This is the kernel function.

[0129] Furthermore, the general form of the function for obtaining the upper confidence bound is:

[0130]

[0131] Among them, the balance coefficient (corresponding to the 95% confidence interval) The mean, The standard deviation is denoted as .

[0132] The general form of the optimization objective function is:

[0133]

[0134] in, The set of TD3 network parameters to be optimized ( Constrained by the stability threshold of the hydraulic system discovered in the step response experiment , (This represents the hydraulic response attenuation coefficient corresponding to the current β). The average performance score of the strategy on dataset D (D contains: 50 sets of historical engineering condition data and 10 sets of real-time wind disturbance data) is given. The standard deviation of the reward value under different operating conditions is shown in Table 4. The performance metrics of the optimization learning system in this embodiment are also shown in Table 4.

[0135] Table 4 Optimizing Learning System Performance

[0136]

[0137] For step 4, the transfer learning initialization pre-training stage, rotational data is generated in the finite element simulation environment, and the LSTM encoder is trained until convergence (loss function). During the fine-tuning phase, the underlying parameters of the encoder are frozen, and only the Actor network is optimized. Elastic Weight Fixation (EWC) is used to prevent catastrophic forgetting.

[0138] Furthermore, the general form of Elastic Weight Fixed (EWC) is as follows:

[0139]

[0140] in, For the total loss function, The loss function for the current task. Regularization coefficient (value is...) ), The diagonal elements of the Fisher information matrix. These are the parameters to be optimized during fine-tuning. These are the parameters for the pre-trained model.

[0141] For step 4, Priority Experience Replay (PRE), the initial priority is calculated for each new experience using Temporal Difference Error (TD Error). After each training iteration, the priority of the corresponding experience is updated based on the latest TD error. New experiences are assigned a base priority (initial priority) when stored, and the priority of the corresponding experience is updated after each sampling training iteration. A power-law distribution is used to adjust the sampling weights, and importance sampling and normalization are used to mitigate the bias introduced by priority sampling, ensuring convergence.

[0142] Furthermore, the general form of timing difference error (TD Error) is:

[0143]

[0144] in, For instant rewards, The current state action value. The next state action value predicted by the target network. This is the discount factor.

[0145] Furthermore, the general form of the initial priority is:

[0146]

[0147] in, =0.01 to prevent 0 priority.

[0148] Furthermore, the general form of the power-law distribution is:

[0149]

[0150] in, Priority intensity control factor ( , At that time, it degenerates into uniform sampling; (At that time, sampling is performed entirely according to priority).

[0151] Furthermore, the general form for calculating importance sampling weights is:

[0152]

[0153] Where N is the capacity of the experience pool. The deviation correction intensity control factor is set (initially 0.4, gradually increasing linearly to 1.0).

[0154] For step 5, the multi-source state data stream output from the sensor array deployed in step 2 is received in real time at the physical interface layer, including the rotation angle deviation. Structural stress Wind speed Iso-sensing vectors. By integrating a finite element dynamics model with a fluid coupling solver using Unity3D's high-fidelity physics engine, a fully digital twin of the bridge-hydraulic system is constructed.

[0155] For step 5, a working condition generation engine is designed. Based on the Latin hypercube sampling and fault injection protocol, it automatically creates 20 typical working condition test sets (including extreme wind load, sensor drift, hydraulic leakage, and other scenarios), and achieves millisecond-level data synchronization with the edge control system via the ROS# communication protocol. Driven by GPU cluster acceleration, it achieves real-time working condition rendering at 2000 frames per second, outputting quantitative verification reports such as trajectory tracking error and stress distribution cloud maps, providing closed-loop certification with a working condition coverage of ≥98% for the control system. The performance indicators of the digital twin verification platform in this embodiment are shown in Table 5.

[0156] Table 5 Performance of the Digital Twin Verification Platform

[0157]

[0158] For step 6, a heterogeneous computing architecture based on FPGA and GPU is constructed. The data processing layer receives the optimization control strategy from step 4 and the output parameters of the verification platform from step 5, and allocates computing tasks through the Time-Sensitive Network (TSN) scheduling algorithm.

[0159] Furthermore, the general form of the Time-Sensitive Network (TSN) scheduling algorithm is:

[0160]

[0161] in, This is the closed-loop control cycle of the system. Calculate the time taken for the i-th subtask. To increase the number of cores for parallel computing, This refers to the FPGA-GPU communication latency.

[0162] In step 6, the FPGA processing unit handles the real-time control flow, including state estimation and control command generation, while the GPU coprocessor performs Monte Carlo simulation and Sobol sensitivity analysis. Through 100,000 Monte Carlo perturbation tests, the system failure rate is less than 0.01%. Simultaneously, based on the Sobol exponent analysis, the weights of multiple parameters on stability are analyzed, and the contribution of key factors such as hydraulic delay is output for real-time optimization of the control strategy. The performance parameters of the edge computing system in this embodiment are shown in Table 6.

[0163] Table 6 Performance of Edge Computing Systems

[0164] Furthermore, the general form of the failure rate determination formula is:

[0165]

[0166] in, This represents the number of simulation attempts for failure. (The failure condition is either excessive angular acceleration or excessive displacement deviation.)

[0167] Furthermore, the general form of the Sobol sensitivity analysis formula is:

[0168]

[0169] in, Let i be the global sensitivity index for parameter i. For output indicators Total variance To fix other parameters ( )hour The expected conditions are as follows. The final rotation construction results of this embodiment are shown in Table 7.

[0170] Table 7 Final Rotation Construction Results

[0171] In the above embodiments, this invention, through edge intelligent control decision-making and multi-dimensional tolerance dynamic optimization, systematically overcomes the millimeter-level real-time control challenge under multi-source disturbances during bridge rotation construction, revealing the control performance boundaries of three core parameters: control delay sensitivity, wind load disturbance bandwidth, and hydraulic response attenuation coefficient. Combining Monte Carlo adaptive strategy verification and Pareto safety front constraints, a full-chain tolerance control system of "control command-actuator-structural feedback" is constructed. This method overcomes the limitations of traditional control relying on human experience and static thresholds, significantly improving the safety and economy of bridge rotation construction, and has significant promotional value for intelligent control and dynamic prediction of complex bridge construction.

[0172] From the appendix Figure 2 - Appendix Figure 5 The data presented shows that;

[0173] Regarding angle tracking accuracy, the intelligent control method proposed in this embodiment can track the target angle trajectory more accurately compared with traditional PID control and fuzzy control. During the rotation start and braking stages, the system responds quickly and the overshoot is significantly reduced. During the uniform rotation stage, the angle fluctuation range is controlled within ±0.1°, meeting the requirements of high-precision engineering.

[0174] In terms of energy consumption, this embodiment reduces energy consumption by approximately 25-30% compared to traditional PID control. It avoids frequent power peaks found in traditional methods, which is beneficial for extending equipment lifespan; and it achieves effective recovery of some energy during the braking phase.

[0175] Regarding wind stability, this embodiment maintains good control stability even under wind speeds of 3-5. The system can automatically adjust control parameters according to wind speed changes to maintain smooth rotation, and even in the event of sudden gusts, the system's safety margin still meets design requirements.

[0176] Regarding the optimization effect during the learning process, in the early stage of training, the reward value increases rapidly, indicating that the algorithm has high learning efficiency. During the training process, the reward function shows a stable upward trend without drastic fluctuations. As training progresses, the control strategy is continuously optimized, eventually reaching a high performance level.

[0177] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A real-time intelligent control method for bridge rotation based on machine learning, characterized in that, Includes the following steps: Step 1: Construct a multi-source sensor data fusion system, deploy a global state awareness array, establish a real-time data communication architecture based on the OPC UA protocol, and realize multi-source data timestamp synchronization; Step 2: Develop a dynamic feature engineering module to receive multi-source state data streams from the sensor array deployed in Step 1 in real time at the input layer; through a hybrid feature processing engine, calculate the time-domain rate of change and frequency-domain energy features in parallel, and use a third-order feature fusion architecture to realize intelligent mapping from physical signals to information-intensive features. Step 3: Construct a hybrid machine learning control model, build a hybrid control framework based on LSTM-TD3, and design a multi-objective reward function; Step 4: Establish an online optimization and learning mechanism, develop a Bayesian optimization-online learning collaborative system, use the GP-UCB algorithm to globally optimize the hyperparameters of the control strategy, and embed an incremental online learning module: online data cache pool, elastic weight solidification (EWC) to prevent catastrophic forgetting, and priority experience replay (PER) to accelerate convergence; Step 5: Create a virtual-real interaction verification platform, develop a digital twin system based on the Unity3D engine, and build test sets for multiple typical working conditions; Step 6: Deploy edge computing and ensure stability. Design a heterogeneous computing architecture that integrates FPGA and GPU to achieve real-time control cycle. Evaluate the system robustness through Monte Carlo simulation and use the Sobol exponential method to analyze the contribution of each parameter to control stability.

2. The real-time intelligent control method for bridge rotation based on machine learning according to claim 1, characterized in that, In step 1, key physical quantities of the bridge rotation system, including rotation angle deviation, are collected in real time at the input layer. Structural stress extreme values Real-time wind speed Hydraulic pressure Multi-dimensional parameters; a global state awareness network is constructed through a sensor array, and each node achieves clock synchronization error of less than a preset threshold based on the IEEE 1588 v2 precision time protocol; deterministic transmission technology using TSN and OPC UA protocol stacks ensures that end-to-end latency is less than a preset threshold.

3. The real-time intelligent control method for bridge rotation based on machine learning according to claim 1, characterized in that, In step 2, multi-source heterogeneous sensor nodes are arranged at the bridge ball joint and other key parts to perform real-time detection of the rotation attitude angle, detection of displacement synchronization error, detection of stress concentration area, early warning of abnormal bridge vibration, judgment of temperature rise fault, and real-time feedback of wind load.

4. The real-time intelligent control method for bridge rotation based on machine learning according to claim 1, characterized in that, In step 3, an LSTM encoder processes multiple recent historical state sequences and outputs a 128-dimensional spatiotemporal feature vector to capture the dynamic hysteresis characteristics of the rotation process. A dual Critic network is used to evaluate the action value, i.e., the Q value, in parallel. The network structure is a fully connected layer to suppress Q value overestimation. Continuous control actions are generated based on the spatiotemporal features and mapped to the traction machine power adjustment command. At the same time, the dynamic adjustment coefficients of the PID parameters are output. A multi-objective reward function is designed, and a dynamic weight establishment function is defined.

5. The real-time intelligent control method for bridge rotation based on machine learning according to claim 4, characterized in that, In step 4, the probability mapping function from the hyperparameter space to the control value is constructed using the Gaussian process surrogate model GP-UCB algorithm, thereby realizing the Bayesian global optimization of the hyperparameters of the TD3 network. In the initial pre-training stage of transfer learning, rotational data is generated in the finite element simulation environment, and the LSTM encoder is trained until convergence. In the fine-tuning stage, the underlying parameters of the encoder are frozen, and only the Actor network is optimized. Elastic weights are used to solidify the EWC to prevent catastrophic forgetting.

6. The real-time intelligent control method for bridge rotation based on machine learning according to claim 1, characterized in that, In step 4, the Priority Experience Replay PRE uses the Temporal Differential Error (TD Error) to calculate the initial priority for each new experience. After each training session, the priority of the corresponding experience is updated based on the latest TD error; When storing new experiences, a basic priority is assigned, i.e., the initial priority. The priority of the corresponding experience is updated after each sampling training. A power-law distribution is used to adjust the sampling weights, and the bias introduced by priority sampling is mitigated through importance sampling and normalization, thus ensuring convergence.

7. The real-time intelligent control method for bridge rotation based on machine learning according to claim 1, characterized in that, In step 5, a fully digital twin of the bridge-hydraulic system is constructed by integrating the finite element dynamics model with the fluid coupling solver through the Unity3D high-fidelity physics engine.

8. The real-time intelligent control method for bridge rotation based on machine learning according to claim 1, characterized in that, In step 5, a working condition generation engine is designed. Based on the Latin hypercube sampling and fault injection protocol, it automatically creates a variety of typical working condition test sets, including extreme wind load, sensor drift, and hydraulic leakage scenarios. It also achieves millisecond-level data synchronization with the edge control system through the ROS# communication protocol. Driven by GPU cluster acceleration, it realizes real-time working condition rendering and outputs trajectory tracking error and stress distribution cloud map quantitative verification reports, providing closed-loop certification with a working condition coverage of ≥98% for the control system.

9. The real-time intelligent control method for bridge rotation based on machine learning according to claim 1, characterized in that, In step 6, a heterogeneous computing architecture based on FPGA and GPU is constructed. The data processing layer receives the optimization control strategy and the output parameters of the verification platform, and the computing tasks are allocated through the Time-Sensitive Network (TSN) scheduling algorithm.

10. The real-time intelligent control method for bridge rotation based on machine learning according to claim 1, characterized in that, In step 6, the FPGA processing unit processes the real-time control flow, including state estimation and control command generation, while the GPU coprocessor performs Monte Carlo simulation and Sobol sensitivity analysis. Through Monte Carlo perturbation testing, the system failure rate is made lower than a preset threshold. At the same time, based on the Sobol exponent, the weights of multiple parameters on stability are analyzed, and the contribution of key hydraulic delay factors is output for real-time optimization of the control strategy.

Images