Wind turbine generator hoisting construction tower drum operation system and construction method thereof

Abstract

The invention discloses a wind turbine generator hoisting construction tower drum operation system and a construction method thereof, and relates to the technical field of intelligent control. The problems that in the prior art, a static tension balance mechanism cannot restrain bending moment abrupt change, steel-concrete interface stress concentration causes microcrack propagation, the wave dynamic load compensation capacity is insufficient, and dynamic rigidity attenuation early warning is lacked are solved. Comprising a dynamic load prediction module, a multi-mode vibration suppression module, an offset compensation module and a digital twinborn decision module, a hoisting load is solved in real time through a multi-physics field coupling model and an improved time sequence deep learning algorithm, and interface crack propagation is suppressed in combination with traveling wave offset control and sweep frequency vibration. An improved Morison equation is adopted to drive a two-stage hydraulic servo to compensate a wave dynamic load, and a digital twin closed-loop correction mechanism is constructed based on a 5G URLLC protocol; the tower drum hoisting precision, the structural safety and the operation reliability under the complex working condition are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control technology, and more particularly to a wind turbine generator set hoisting construction tower operation system and a construction method thereof. Background Art

[0002] The tower operation system for wind turbine installation is a core component of wind turbine construction, requiring high-precision tower docking and dynamic stability control. Existing technologies include modular hoisting equipment (such as hydraulically driven positioning mechanisms), precast concrete tower assembly processes, and deep-sea hoisting equipment with wave compensation. Key technologies include hoisting path planning, adaptive load adjustment, and flange connection optimization, while multi-degree-of-freedom robotic arms and sensors collaborate to improve construction efficiency.

[0003] In the existing technology, patent CN119240534A proposes an auxiliary device for lifting segments in a mixed tower. This device uses an electric hydraulic cylinder to drive three sets of support arms to adjust the angle of the lifting mechanism, and uses a positioning mechanism and a safety lock device to achieve precise docking of the segments. Patent CN119844301A proposes a prefabricated component assembly technology for concrete towers. This utilizes a pyramid-shaped cylinder structure design and forms the tower body by alternately inserting concrete inserts and support columns. This significantly reduces the complexity of transportation and construction. At the same time, the design of the steel tower adapter section at the top enhances the overall structural stability. In the field of deep-sea lifting, patent CN108533461B has developed a self-lifting lifting device that integrates a hydraulic lifting platform and a full-rotation crane. The lifting and locking mechanism achieves the coordinated fixation of the hull and pile legs. In addition, patent CN109969929A ​​discloses a universal tail hoist, which adopts an articulated structure of a lever assembly and an adjustable strut, dynamically adapts to tower flanges of different diameters through a telescopic rod, and combines a ball hinge structure to offset the deflection torque during the tower hoisting process, thereby improving the versatility of multi-model tower hoisting. In response to the prestressed connection problem of concrete towers, patent CN111287907A proposes a segmented prestressed cable tensioning technology, which sets prestressed tendon connectors at the joints of the tower segments. Through the timing control of the pre-tensioning of the lower cable and the anchoring of the upper cable, combined with the coordination of the arm structure and the tower crane device, the interface stress concentration caused by the lateral load of the hoisting is effectively suppressed. Although these technologies have constructed a complete tower hoisting operation system from the dimensions of structural design, dynamic adaptation, and material mechanics, there are still systemic defects: First, taking the transition of the tower from horizontal hoisting to vertical state as an example, the offset of the gravity point causes a nonlinear mutation in the bending moment distribution. Traditional hoists (such as the fixed pulley group and heart-shaped ring wire rope design of patent CN207346980U) only rely on the static tension balance mechanism and cannot perceive the dynamic changes in the force on the wire rope in real time, resulting in the instantaneous shear stress exceeding the limit of the flange bolt group at the bottom of the tower, inducing preload loss and flange opening and closing angle deviation; the interface stress concentration phenomenon caused by the difference in elastic modulus between concrete and steel in the steel-concrete tower (for example, the interface stress concentration coefficient in the hoisting of a 160-meter concrete tower reached 1.8 times), combined with the hoisting vibration effect, further aggravates the risk of microcrack propagation in the joints of prefabricated parts. The existing prefabricated part assembly technology (such as patent CN119844301A) does not integrate a vibration suppression module, resulting in a significant increase in the crack propagation rate. During deep-sea operations, the additional dynamic loads generated by the coupling of long-period waves and the vessel's heave motion exceed the adjustment capabilities of traditional wave compensation devices (such as the lifting and locking mechanism in patent CN108533461B), significantly shortening the fatigue life of the floating tower flange bolts. Furthermore, existing dynamic load models (such as the seismic response spectrum method in patent CN103106296A) ignore the creep of steel-concrete materials and the effects of fluid-structure coupling. Consequently, they fail to provide early warning of the dynamic stiffness degradation of concrete after carbonization, leading to natural frequency shifts in the tower and delaying timely intervention. Summary of the Invention

[0004] In view of the deficiencies in the prior art, the present invention discloses a wind turbine tower hoisting construction system and a construction method thereof, aiming to solve the problems in the background technology.

[0005] In order to achieve the above technical effects, the present invention adopts the following technical solutions: A wind turbine tower hoisting construction system, comprising: The dynamic load prediction module integrates tower material parameters, environmental monitoring data, and crane hoisting parameters. It uses a multi-physics coupling model and an improved time-series deep learning algorithm to calculate the bending moment gradient and interface stress concentration factor during the hoisting process in real time. It then outputs dynamic hoisting trajectory instructions to the multimodal vibration suppression module and offset compensation module, and generates a first-order natural frequency offset warning signal. A multimodal vibration suppression module is used to receive the interface stress concentration coefficient and the bolt group shear stress data collected by the fiber Bragg grating sensor, trigger the traveling wave cancellation control of the piezoelectric actuator array, generate the reverse compensation strain through the adaptive proportional integral algorithm, and superimpose the swept frequency vibration excitation signal to suppress the expansion of micro cracks at the steel-concrete interface. The processed dynamic strain data is synchronized to the digital twin decision module in real time; The offset compensation module is used to drive the two-stage hydraulic servo mechanism to generate feedforward compensation displacement and feedback damping adjustment based on the ship's inertial navigation attitude data and the wave spectrum characteristics captured by the millimeter-wave radar, using the improved Morison equation and hydraulic servo composite control strategy, and synchronize them to the digital twin decision module; The digital twin decision-making module is used to construct real-time synchronous mapping between the tower BIM model and the three-dimensional laser point cloud through the 5G URLLC protocol, receive the strain compensation data of the multimodal vibration suppression module and the displacement correction amount of the offset compensation module, and iteratively optimize the crane lifting speed parameters through the dynamic trajectory replanning algorithm; and when it is detected that the flange opening and closing angle deviation exceeds the preset tolerance threshold, the dynamic model boundary conditions of the dynamic load prediction module are reversely updated through the multi-module collaborative closed-loop correction mechanism.

[0006] As a further technical solution of the present invention, a method for constructing a wind turbine tower by hoisting the tower includes: Step 1: Based on the elastic modulus ratio of the tower steel-concrete material, real-time wind speed, and lifting point position parameters, a multi-physics field coupling dynamic model is constructed and integrated with an improved LSTM neural network algorithm to calculate the bending moment gradient distribution and interface stress concentration factor threshold during the lifting process, and generate dynamic lifting trajectory constraint instructions; Step 2: Based on the stress concentration factor threshold output in step 1 and the time series data of the bolt group shear stress collected in real time by the fiber Bragg grating sensor, the traveling wave cancellation control mechanism of the piezoelectric actuator array is triggered, and the reverse compensation strain is generated by the adaptive proportional integral algorithm to suppress the microcrack growth rate at the steel-concrete interface to a preset threshold; Step 3: The ship's inertial navigation attitude data is integrated with the wave spectrum characteristics captured by the millimeter-wave radar. The improved Morison equation is used to analyze the additional dynamic load of the wave. The two-stage hydraulic servo mechanism is driven to implement a feedforward-feedback composite compensation strategy and output the displacement correction value to the crane lifting mechanism. Step 4: Build a real-time, synchronized digital twin of the tower BIM model and the 3D laser point cloud based on the 5G URLLC protocol. Receive the strain compensation data from step 2 and the displacement correction from step 3, and iteratively optimize the crane lifting speed parameters using a dynamic trajectory replanning algorithm. Step 5: When the digital twin detects that the flange opening and closing angle deviation exceeds the tolerance threshold, it triggers the multi-module collaborative closed-loop correction instruction, reversely updates the dynamic model boundary conditions in step 1, and synchronously adjusts the hydraulic compensation parameters in step 3; Step 6: Based on the dynamic model updated in step 1, recalculate the lifting trajectory constraint instructions and repeat steps 2 to 5.

[0007] Based on the above technical solutions, the positive and beneficial effects of the present invention are: 1. The dynamic load prediction module uses a multi-physics field coupling model and an improved time-series deep learning algorithm to solve the bending moment gradient in real time. It can accurately predict the gravity point offset effect when the tower transitions from horizontal to vertical state, dynamically adjust the lifting trajectory instructions, and effectively solve the problem that the traditional static tension balance mechanism cannot perceive dynamic load changes in real time. It reduces the risk of excessive shear stress in flange bolts and avoids preload loss and flange opening and closing angle deviation. Secondly, the traveling wave cancellation control and swept frequency vibration excitation of the multi-modal vibration suppression module offset the stress wave propagation caused by the elastic modulus difference at the steel-concrete interface in real time. It also dynamically adjusts the reverse compensation strain based on the adaptive proportional integral algorithm, directly suppressing the initiation and expansion of interface microcracks, overcoming the defect of existing prefabricated assembly technology that lacks active vibration suppression.

[0008] 2. The offset compensation module uses an improved Morison equation to analyze wave spectrum characteristics, combined with the feedforward-feedback composite control of a two-stage hydraulic servo mechanism, to effectively offset the additional dynamic loads generated by the coupling of long-period waves and the heave motion of the ship, significantly extending the fatigue life of the floating tower flange bolts. This solves the problem of insufficient adjustment ability of traditional wave compensation devices to the coupling effect of long-period waves and the heave motion of the ship, and significantly reduces the risk of fatigue damage to the flange bolts of deep-sea floating towers.

[0009] 3. The digital twin decision module uses a multi-module collaborative closed-loop correction mechanism to reversely update the boundary conditions of the dynamic model in real time, and promptly intervene in the tower's natural frequency offset caused by the creep of steel-concrete materials and the fluid-structure coupling effect. This avoids the risk of stiffness attenuation caused by the traditional dynamic load model ignoring the time-varying characteristics of the material. It solves the problem of the lack of dynamic stiffness attenuation warning caused by the existing dynamic load model ignoring the creep of steel-concrete materials and the fluid-structure coupling effect, and realizes real-time intervention and adaptive correction of the tower's natural frequency offset. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which: Figure 1 This is a working principle architecture diagram of a wind turbine tower installation system according to the present invention; Figure 2 This is a working principle diagram of the multi-physics field coupling model of the present invention; Figure 3 FIG1 is a working principle diagram of the traveling wave cancellation control process of the piezoelectric actuator array of the present invention; Figure 4This is a working method step diagram of the improved time series deep learning algorithm of the present invention; Figure 5 This is the working framework diagram of the dynamic trajectory replanning algorithm of the present invention Figure 6 This is a step diagram of the tower construction method for hoisting a wind turbine generator set according to the present invention. DETAILED DESCRIPTION

[0011] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0012] In this solution, the wind turbine tower installation system includes: The dynamic load prediction module integrates tower material parameters, environmental monitoring data, and crane hoisting parameters. It uses a multi-physics coupling model and an improved time-series deep learning algorithm to calculate the bending moment gradient and interface stress concentration factor during the hoisting process in real time. It then outputs dynamic hoisting trajectory instructions to the multimodal vibration suppression module and offset compensation module, and generates a first-order natural frequency offset warning signal. A multimodal vibration suppression module is used to receive the interface stress concentration coefficient and the bolt group shear stress data collected by the fiber Bragg grating sensor, trigger the traveling wave cancellation control of the piezoelectric actuator array, generate the reverse compensation strain through the adaptive proportional integral algorithm, and superimpose the swept frequency vibration excitation signal to suppress the expansion of micro cracks at the steel-concrete interface. The processed dynamic strain data is synchronized to the digital twin decision module in real time; The offset compensation module is used to drive the two-stage hydraulic servo mechanism to generate feedforward compensation displacement and feedback damping adjustment based on the ship's inertial navigation attitude data and the wave spectrum characteristics captured by the millimeter-wave radar, using the improved Morison equation and hydraulic servo composite control strategy, and synchronize them to the digital twin decision module; The digital twin decision-making module is used to construct real-time synchronous mapping between the tower BIM model and the three-dimensional laser point cloud through the 5G URLLC protocol, receive the strain compensation data of the multimodal vibration suppression module and the displacement correction amount of the offset compensation module, and iteratively optimize the crane lifting speed parameters through the dynamic trajectory replanning algorithm; and when it is detected that the flange opening and closing angle deviation exceeds the preset tolerance threshold, the dynamic model boundary conditions of the dynamic load prediction module are reversely updated through the multi-module collaborative closed-loop correction mechanism.

[0013] When implementing, Figure 1As shown in the figure, the system operates as follows: A distributed sensor network collects tower material parameters (elastic modulus, creep coefficient), environmental data (wind speed, wave height), and hoisting mechanism status (hydraulic pressure, lifting speed) in real time. These data are then fed into a multi-physics coupling model for iterative calculation of the steel-concrete interface stress field and fluid-structure coupling forces. An improved time-series deep learning algorithm (TCN-LSTM hybrid architecture) is then used to predict the bending moment gradient distribution and interface stress concentration factor within the next 10 seconds. When the prediction meets the confidence level (residual ≤ 5%), dynamic hoisting trajectory commands are generated. This synchronously drives the piezoelectric actuator array to generate phase-reversal traveling waves to suppress microcrack propagation. Wave disturbances are offset by a two-stage hydraulic servo mechanism (feedforward compensation ±3m travel and feedback damping adjustment of the valve port from 0.2-1.0cm²). The 5G URLLC protocol (end-to-end latency ≤ 0.5ms) synchronizes actuator status with 3D laser point cloud scan data in real time to the digital twin module. The ICP registration algorithm aligns the BIM model with the physical tower to ensure geometric deviations (tolerance ≤ 2mm). When a flange opening / closing angle deviation of ≥0.1° or a local stress of ≥50 MPa is detected, a multimodal data fusion engine (with a spatiotemporal convolution kernel size of 3×3×3) is triggered to reconstruct the feature matrix. The NSGA-II multi-objective optimization algorithm dynamically allocates module weight coefficients, backpropagating the load prediction model's steel-concrete creep time constant and fluid damping coefficient. The hydraulic servo valve flow gain and piezoelectric sweep excitation amplitude are simultaneously adjusted, forming a bidirectional adaptive closed loop from physical-level perception to model-level correction. The system ensures control variable convergence using a combined Kalman-Lyapunov stability criterion, completes dynamic trajectory replanning and actuator parameter iteration within a 10ms cycle, and achieves millimeter-level dynamic accuracy and exponential suppression of crack growth rates during the lifting process.

[0014] This system utilizes a full-link closed-loop control design. The system hardware consists of distributed fiber Bragg grating (FBG) strain sensors, a six-axis force sensor, a piezoelectric actuator array, a two-stage hydraulic servo mechanism, a 3D laser scanner, and an edge computing node. These components are interconnected via an industrial-grade communication protocol and a 5G URLLC wireless network. The fiber Bragg grating (FBG) sensors acquire the tower surface strain distribution in real time at a 1kHz sampling rate, while the six-axis force sensor monitors the hoisting tension with an accuracy of ±0.1%FS. This data is transmitted via a CAN FD bus to an edge computing node (NVIDIA Jetson AGX Xavier). A multiphysics coupling model and an improved LSTM algorithm are used to calculate the bending moment gradient and interfacial stress concentration factor in real time, generating dynamic hoisting trajectory commands. The piezoelectric actuator array is embedded at the steel-concrete interface with a 200mm pitch. It receives adaptive proportional-integral control signals generated by an FPGA (Xilinx Zynq UltraScale+). Driven by a 150V drive voltage, it generates reverse compensatory strain and superimposes a 20-200Hz swept-frequency vibration to suppress microcrack propagation. The two-stage hydraulic servo mechanism features a first-stage compensation cylinder with a stroke of ±2m and a response frequency of 0.1-1Hz, and a second-stage compensation cylinder with a stroke of ±0.5m and a response frequency of 5-10Hz. Feedforward compensation displacement and feedback damping adjustment are generated based on wave spectrum characteristics captured by millimeter-wave radar (operating at 77GHz) and the modified Morison equation. A 3D laser scanner generates a point cloud model of the tower surface at a frequency of 30Hz, which is synchronized with the BIM platform in real time via the 5G URLLC protocol (latency ≤15ms). A dynamic trajectory replanning algorithm optimizes the crane hoisting speed using a 10-step prediction time domain and a 5-step control time domain. A Bayesian learning mechanism is triggered when the flange opening and closing angle deviation exceeds 0.05°, inversely correcting the creep coefficient and fluid damping ratio parameters of the dynamic load prediction module. The system utilizes the Profinet protocol for hydraulic servo control, with PID parameters set to a proportional gain of 8.0, an integral time of 0.5, and a derivative time of 0.1. Sensor data streams are synchronized via EtherCAT. The redundantly designed dual-channel acquisition module and IP67 protection-grade hardware deployment ensure operational reliability in harsh offshore environments, ultimately achieving millimeter-level precision control and adaptive adjustment of all working conditions during the tower hoisting process.

[0015] During implementation, after the crane hoisting operation commenced, a distributed sensor network synchronously collected load data at a 500Hz sampling rate and transmitted it to an edge computing node via the 5G URLLC protocol. Data preprocessing employed the Daubechies 5 wavelet basis to reduce noise in the raw strain signal, and a Kalman filter was used to eliminate random drift in the inertial navigation data. Wave spectrum feature extraction employed the Hilbert-Huang transform algorithm to decompose the primary frequency component and nonlinear harmonics. The dynamic load prediction module utilized a COMSOL multiphysics coupling model, inputting the steel-concrete material parameters (elastic modulus 42 GPa, Poisson's ratio 0.2) and environmental monitoring data to iteratively solve for the stress field at the steel-concrete interface. Simultaneously, an improved LSTM time-series deep learning algorithm (128 hidden nodes, 500 training cycles) was employed, incorporating data from a historical hoisting case library, to predict the bending moment gradient and interface stress concentration factor within the next 10 seconds. If the interface stress concentration factor exceeded 1.5, the system triggered a Level 1 warning and sent a safety limit command to the crane control console via the CAN bus, limiting the lifting speed to ≤ 0.3 m / s.

[0016] Upon receiving the warning signal, the multimodal vibration suppression module activates the piezoelectric actuator's traveling wave cancellation mode. The piezoelectric drive controller, using a phased array control algorithm, generates a reverse displacement field with a phase difference of 180° from the dominant vibration frequency. A 20-200Hz sinusoidal sweep signal (amplitude 50με, sweep rate 10Hz / s) is superimposed via an AD9833 signal generator to stimulate the release of residual stress on the interface. Real-time strain data is fed back to a DSP controller (TI C2000 series) via a fiber Bragg grating sensor, which dynamically adjusts the piezoelectric drive voltage amplitude and phase deviation (with a tolerance of ±2°). Simultaneously, the offset compensation module uses a modified Morison equation to predict wave forces. The damping coefficient (baseline value 0.6, dynamic fluctuation ±20%) and added mass factor (baseline value 1.2) are modified based on the ship's heave displacement and the dominant frequency of the wave spectrum. The feedforward control channel calculates the pre-compensation displacement through the inverse dynamics model (the feedforward gain is optimized by the genetic algorithm) and drives the first-stage hydraulic cylinder to adjust the hook position; the feedback channel receives the bolt strain acceleration and adjusts the opening area of ​​the second-stage damping valve in real time to suppress the tower swing acceleration to ≤0.05g.

[0017] The 3D laser scanner generates point cloud data on the tower surface every 0.5 seconds, which is spatially aligned with the BIM model through the Iterative Closest Point (ICP) algorithm to detect areas with geometric deviations ≥ 2mm. The digital twin decision module has a built-in deep reinforcement learning (DRL) model, which inputs a multimodal fusion data set (including real-time strain ε, displacement correction, and wave load spectrum). The double deep Q network (Double DQN) outputs the crane lifting speed optimization parameters (range 0.1-1.0m / s). The objective function is to minimize the weighted sum of the flange opening and closing angle deviation and the bolt shear stress (weights α=0.6, β=0.4). When the laser tracker detects a flange opening and closing angle deviation ≥ 0.1°, a multi-module collaborative closed-loop correction process is triggered: the Kalman filter estimates the boundary condition error of the dynamic model, and the back propagation (BP) neural network updates the creep coefficient of the steel-concrete material (baseline value 1.2×10⁻ 8 / s); the NSGA-II multi-objective optimization algorithm reallocated the crane winch tension weights to generate a Pareto-optimal solution set; the hydraulic servo system switched to safe mode, and the linked piezoelectric actuator increased its output amplitude by 20%. The updated parameters were synchronized to all modules via the OPC UA protocol, and the finite element model (ANSYS Mechanical) verified the stress distribution in real time, forming a bidirectional closed loop from the physical system to the virtual twin.

[0018] In the above implementation, if Figure 2 As shown in the figure, the multi-physics coupling model constructs the Hamiltonian variational equation based on the creep of steel-concrete materials and the fluid-structure coupling effect. Based on the elastic modulus ratio parameter of the tower material, the implicit time integration algorithm is used to iteratively solve the time-varying bending moment gradient distribution during the lifting process. The frequency domain characteristics of the wave load are embedded through the Morison equation, and the fluid-structure interaction force is obtained by combining the transient simulation of computational fluid dynamics to generate dynamic lifting trajectory instructions. The creep strain increment and the fluid coupling stress data are simultaneously input into the improved LSTM neural network. The time-frequency domain characteristics of the lifting vibration spectrum are extracted through the time series convolution layer to predict the first-order natural frequency offset trend. If the predicted offset exceeds the preset threshold, an early warning signal is triggered and the boundary conditions of the dynamic model are updated. The creep time constant and the fluid damping coefficient are adjusted, and the optimized lifting trajectory is output to the execution module. When the tower attitude sensor detects that the bending moment gradient mutation exceeds the preset threshold, the vibration signal is decomposed into the natural mode function based on the Hilbert-Huang transform method and the frequency domain characteristics are reconstructed to update the generalized force constraint term in the Hamiltonian variational equation. like Figure 4 As shown in the figure, the working method of the improved time series deep learning algorithm is: S1. Extract the multi-scale time series features of the hoist vibration spectrum through the dilated causal convolution layer of the temporal convolutional network, input it into the multi-head self-attention mechanism for time step weight distribution, and obtain the weighted time series feature vector; S2. Input the weighted temporal feature vector into a bidirectional LSTM network to capture long-term and short-term dependencies by fusing forward and reverse hidden states; S3, adding the original vibration spectrum data to the output of the bidirectional LSTM network point by point based on a gated residual connection to suppress gradient vanishing; S4. Use Bayesian optimization algorithm to dynamically adjust the TCN convolution kernel size Bias with LSTM forget gate , the formula is: (1) In formula (1), is the Sigmoid activation function; Is the linear transformation matrix used to calculate the current input and the previous hidden state A weighted combination of The adaptive noise intensity coefficient is updated through online reinforcement learning strategy; is standard Gaussian noise; S5. If the predicted bending moment gradient residual exceeds the preset threshold, the gradient reversal mechanism is triggered, and the TCN convolution kernel weights are corrected through back propagation of the parameter optimization objective function; the formula expression of the parameter optimization objective function is: (2) In formula (2), is the time step The true interface stress concentration factor; is the time step The predicted interface stress concentration factor; Represents the parameter set to be optimized , are kernel size, expansion coefficient, and noise intensity coefficient, respectively; Represents the trade-off coefficient, which is used to control the weight of the KL divergence term in the total loss; Represents KL divergence, which is used to measure parameter distribution With prior distribution differences; S6. Output the predicted value of the interface stress concentration factor to the multi-physics field coupling model, and simultaneously generate the first-order natural frequency offset as a warning signal.

[0019] In modeling the creep and fluid-structure coupling effects of reinforced concrete, a generalized energy functional is constructed using the Hamiltonian variational equation. The nonlinear creep constitutive equation for reinforced concrete (based on the Burgers model) is coupled with the Navier-Stokes equations in the fluid domain. The elastic modulus ratio parameter (e.g., Es = 210 GPa for steel and Ec = 42 GPa for concrete) is introduced as a weighting factor for interfacial stress transfer. An implicit Newmark-β time integration algorithm is used to iteratively solve the time-varying bending moment gradient distribution to ensure the numerical stability of the nonlinear dynamic equations. To address the dynamic effects of wave loads, the frequency-domain extended Morison equation is embedded with wave height spectral characteristics (e.g., the Pierson-Moskowitz spectrum). Computational fluid dynamics (CFD) transient simulations are then used to obtain the fluid-structure interaction force Ffsi. This is then fused with inertial navigation data to generate dynamic lifting trajectory commands. In the time series deep learning component, an improved LSTM neural network incorporates a dilated causal convolution layer from a temporal convolutional network (TCN). This layer captures local and global correlations in the hoisting vibration spectrum through multi-scale time series feature extraction (convolution kernel size 34). A multi-head self-attention mechanism dynamically assigns weights to time steps, optimizing the feature contributions of key vibration events (such as vortex-induced vibration and impact loads). A bidirectional LSTM network fuses forward and reverse hidden states to model long-term and short-term temporal dependencies and predict the interface stress concentration factor KsKs and first-order natural frequency offset. A gated residual connection uses skip connections to suppress gradient vanishing. A Bayesian optimization algorithm dynamically adjusts the TCN convolution kernel size and LSTM forget gate bias parameters, improving the model's adaptability to different hoisting stages. A gradient reversal mechanism backpropagates convolution kernel weights when the prediction residual exceeds a limit, enabling adaptive parameter updates. After the real-time warning signal is triggered, the generalized force constraint term in the Hamilton equation is updated through modal decomposition and frequency domain reconstruction (IMF components are extracted by Hilbert-Huang transform), and the creep time constant and fluid damping coefficient are simultaneously corrected to form a self-consistent closed loop of the dynamic model.

[0020] During implementation, high-frequency strain sensors (HBM U9C, ±2000 με range, 1 kHz sampling rate) and fiber Bragg grating sensors (0.1 pm wavelength resolution) were installed at the tower flange joints to collect real-time shear stress and interfacial strain of the bolt groups. A six-axis MEMS inertial measurement unit (STIM300, ±0.01° accuracy) was deployed on the tower's outer wall to monitor attitude angle and acceleration. A wire-type displacement sensor (LVDT, ±5 m range, 0.05% nonlinearity) was installed on the crane hook to measure hoisting trajectory deviation. Data was transmitted to an edge computing node (NVIDIA Jetson AGX Xavier) via the 5G URLLC protocol (user plane latency ≤ 0.5 ms), where a multiphysics coupling model (COMSOL Multiphysics kernel) and an improved LSTM algorithm were run. In the software implementation, the Hamiltonian variational equations were discretized using finite element methods (mesh size ≤ 50 mm). Viscoelastic contact boundary conditions (stiffness coefficient Kint = 1e8 N / m) were applied to the steel-concrete interface, and the implicit time integration step was set to 0.01 s. Wave spectrum features (wave height Hs = 4 m, frequency fwave = 0.1 Hz, captured by millimeter-wave radar) were input to the Morison equation. Fluid forces Ffsi were calculated using CFD transient simulations (using a k-ω SST turbulence model). The improved LSTM network architecture consists of a TCN layer (convolution kernel size 5, dilation factor 2), a multi-head self-attention layer (4 heads), and a bidirectional LSTM layer (128 hidden nodes). Training data was obtained from a historical hoisting case library (over 100 concrete tower conditions). Input features include the vibration acceleration spectrum (FFT resolution 0.1 Hz), material creep strain, and fluid coupling stress. The output layer uses a sigmoid activation function for prediction. The Bayesian optimization hyperparameter space covers the TCN convolution kernel size (31) and the learning rate, and converges to the optimal solution through Gaussian process iteration. In the hardware electrical connection, the sensor signal is connected to the AD7606 data acquisition card (16-bit resolution, synchronous sampling rate 200kS / s) through the signal conditioning circuit (AD623 amplifier, gain 1000). The edge node sends trajectory correction instructions and warning signals to the crane control system through the CAN bus (baud rate 1Mbps). When a sudden change in the bending moment gradient (≥50kN⋅m / m) is detected, the HHT modal decomposition and reconstruction of the frequency domain features are triggered, the generalized force constraint term of the Hamilton equation is updated (such as the fluid added mass = 1.2ρV), and the piezoelectric drive parameters of the multimodal vibration suppression module are synchronously corrected through the OPC UA protocol.

[0021] In the improved time series deep learning algorithm, multi-scale time series feature fusion is achieved by introducing the dilated causal convolution layer of the temporal convolutional network (TCN) with a dilation factor of (k is the level index) The receptive field is expanded layer by layer to cover the short-term impact (such as the start-up and stop vibration of the crane) and the long-term trend (such as the creep strain accumulation of steel and concrete) of the vibration spectrum. Mathematically, the convolution kernel size is used to and expansion factor The combination of coverage ,For example The dynamic weight allocation optimization uses a multi-head self-attention mechanism to map the input sequence into query (Q), key (K), and value (V) matrices, and calculates the time step weight by scaling the dot product. ), focusing on strengthening the contribution of key events such as vortex-induced vibration and wave impact, and suppressing high-frequency noise interference. Parameter adaptive adjustment is based on the Bayesian optimization framework, and a Gaussian process proxy model is constructed to predict the residual The optimal combination of TCN convolution kernel sizes is iteratively searched for the objective function. At the same time, a gradient reversal mechanism is introduced to correct the convolution kernel weights by backpropagation when the residual exceeds the limit, thereby alleviating the model degradation problem caused by the nonlinear creep of steel-concrete materials.

[0022] During the design process, a comparative experiment was designed to verify the advantages of the improved time series deep learning algorithm (Group A) in dynamic load prediction accuracy, real-time performance, and anti-interference capabilities. The control group, Group B (existing method), used the traditional LSTM+Attention model (patent CN114548591A).

[0023] The experimental dataset uses historical data from the Chongxin Wind Farm in Gansu Province (including 100 sets of steel-concrete tower conditions). The experimental simulation software platform is based on the PyTorch 1.9 framework and the multi-physics coupling interface of COMSOL Multiphysics. The hardware uses NVIDIA Jetson AGX Xavier, Ubuntu 18.04, and PyTorch 1.9. All simulation experiments in this program are built on a multi-physics co-simulation platform, including a finite element analysis module using ANSYS Mechanical APDL 2023R2, equipped with a local mesh adaptive refinement (h-refinement) algorithm and a transient dynamics solver. The APDL scripting language defines the material constitutive model of the steel-concrete tower (the CDP damage model is used for concrete, and the bilinear kinematic hardening model is used for steel). The dynamic stress field solution time step is set to 0.01 seconds, and the mesh refinement trigger threshold is set to a von Mises stress error of 50 MPa. The fluid dynamics module, based on the CFD module of COMSOL Multiphysics 6.1, uses the k-ω SST turbulence model and the VOF method to simulate wave-ship coupling. The computational domain size is set to 3 times the ship length × 2 times the ship width × 2 times the draft. The boundary conditions use a velocity inlet (corresponding to the JONSWAP wave spectrum) and a pressure outlet. The discretization format uses a second-order upwind scheme. The control algorithm verification platform uses the Simscape Hydraulics toolbox of MATLAB / Simulink R2023a to build a two-stage hydraulic servo mechanism model. The flow-displacement transfer function of the Bosch Rexroth proportional valve is obtained through system identification, and the parameters are calibrated as follows: The boundary layer thickness of the sliding mode controller switching function is set to 0.05; the real-time data interface is implemented through the OPC UA Server Toolkit, and the Pub / Sub mode is used to synchronize sensor data (bolt strain, wave spectrum characteristics) and compensation instructions to the digital twin module with a period of 10ms.

[0024] In this experiment, the sequence length L = 96, the prediction step length T = 24, and the training cycle E = 500. The experimental data is shown in Table 1: Table 1 Experimental data record of improved time series deep learning algorithm Experimental group Bending moment gradient prediction error (RMSE, kN·m) Interface stress coefficient error (%) Algorithm delay (ms) Memory usage (GB) Anti-interference stability (σ / μ, %) Long-term drift rate (% / h) A1 11.2±0.5 7.3±0.3 48±2 2.0±0.1 3.8±0.2 0.6±0.05 A2 10.8±0.7 6.9±0.2 45±3 2.1±0.2 4.0±0.3 0.7±0.06 A3 12.1±0.6 7.8±0.4 50±2 1.9±0.1 3.7±0.1 0.5±0.04 A4 11.5±0.4 7.1±0.3 47±2 2.2±0.2 4.1±0.2 0.8±0.07 A5 10.9±0.5 6.5±0.2 49±3 2.0±0.1 3.9±0.3 0.6±0.05 B1 18.3±1.1 14.2±0.8 115±8 3.6±0.3 8.5±0.7 2.3±0.2 B2 19.7±1.3 15.1±1.0 122±10 3.8±0.4 9.2±0.8 2.6±0.3 B3 20.5±1.5 13.8±0.7 110±7 3.5±0.3 8.8±0.6 2.4±0.2 B4 17.9±1.0 14.9±0.9 118±9 3.7±0.3 9.0±0.7 2.5±0.2 B5 19.2±1.2 13.5±0.6 113±6 3.6±0.2 8.7±0.5 2.4±0.3 The experimental results confirm that the dynamic dilation convolution and adaptive noise mechanism of the improved time series deep learning algorithm effectively improve the model's adaptability to non-stationary working conditions, and the Bayesian optimization and gradient reversal strategy reduce the risk of overfitting, verifying the technical superiority of the algorithm in complex lifting scenarios.

[0025] In the multi-modal vibration suppression module implemented above, Figure 3 As shown in the figure, the traveling wave cancellation control process of the piezoelectric actuator array is as follows: first, the principal component analysis method is used to perform modal decomposition on the multi-source stress data to extract the dominant frequency of micro-crack propagation at the steel-concrete interface. and vibration energy distribution characteristics ; Then the driving voltage of the piezoelectric actuator is dynamically adjusted by the adaptive proportional integral algorithm to generate Reverse traveling wave compensation strain with a phase difference of 180° , and superimpose the sweep frequency excitation signal Stimulate the release of interface residual stress; optimize the spatial distribution weight of the piezoelectric array based on genetic algorithm , ensuring that the traveling wave interference field and the crack propagation direction form an energy dissipation closed loop; the processed dynamic strain data is synchronized to the digital twin decision module in real time through the OPC UA protocol, triggering the iterative update of the model parameters; the adaptive proportional-integral algorithm updates the proportional gain and integral gain through the MIT-based gradient descent method; and calculates the control quantity based on the numerical integration of the updated gain, and compares it with the threshold after superposition with the preset sweep frequency excitation signal. When the preset threshold is exceeded, the traveling wave cancellation unit is triggered to generate a phase-reverse strain wave to the piezoelectric actuator array to cancel it at the steel-concrete interface; the adaptive proportional-integral algorithm uses the error signal change rate and the gain update rate as convergence criteria.

[0026] Among them, the principal component analysis (PCA) method performs an orthogonal transformation on the shear stress and interface strain data of the bolt group, extracts the first k-order principal component modes (contribution rate ≥ 85%) as the basis of the dominant frequency and vibration energy distribution, and reduces the interference of redundant noise on the extraction of crack propagation characteristics. The traveling wave cancellation control of the piezoelectric actuator is based on the principle of elastic wave interference. It uses the inverse piezoelectric effect of piezoelectric ceramics to generate a reverse traveling wave displacement field with a phase difference of 180° from the original vibration, and constructs the standing wave node area after the traveling wave superposition, so that the stress intensity factor (SIF) at the crack tip is reduced to below the threshold. The adaptive proportional integral (API) algorithm uses the MIT rule to design a gradient descent mechanism, which uses the error signal change rate to calculate the displacement field. As the gradient direction, dynamically update the proportional gain With integral gain , its differential equation is: , J is the loss function ( ), The introduction of the sweep frequency excitation signal is based on the resonance frequency offset theory. The AD9833 chip generates a 20-200 Hz sinusoidal sweep frequency signal to stimulate the viscoelastic relaxation effect of the residual stress in the interface bonding layer and reduce the probability of local mutation of the storage modulus E'. The genetic algorithm (GA) performs multi-objective optimization on the spatial distribution weight of the piezoelectric array, using the energy dissipation rate of the traveling wave interference field and the degree of fit between the crack propagation path as the fitness function, and adopts tournament selection, single point crossover (crossover probability ) and Gaussian mutation (mutation probability ) Iteratively generates the optimal weight solution set to form an energy dissipation closed loop. The data processing layer builds a real-time data pipeline through the OPC UA protocol, synchronizes dynamic strain data to the digital twin module with a 10ms cycle, and triggers the incremental update of the finite element model boundary conditions ( ), ensuring the dynamic consistency between the physical model and the measured data.

[0027] During implementation, at the hardware deployment level, 24 groups of piezoelectric ceramic sheets (PZT-5H, 50×50×2mm in size) were evenly arranged around the steel-concrete interface, with spacing ≤0.8m between each sheet. They were bonded to the concrete surface with epoxy resin glue. Fiber Bragg grating sensors (FBGs, central wavelength 1550nm, sensitivity 1.2pm / με) were installed on the bolt groups and connected to an NI PXIe-4844 optical demodulator (sampling rate 1kHz). The piezoelectric drive unit used a high-voltage operational amplifier (APEX PA141, output voltage ±150V) and received the PWM modulation signal generated by the DSP controller (TI C2000) via an LVDS interface. In software implementation, the principal component analysis module was based on the Python scikit-learn library, with the principal component retention dimension k=3 (cumulative variance contribution rate ≥92%). The initial parameters of the adaptive proportional integral algorithm were set as follows: , learning rate The upper limit threshold of the control variable is U_max = 10V; the population size of the genetic algorithm is set to 50, the maximum evolutionary number G = 100, and the fitness function weights α = 0.6 (energy dissipation) and β = 0.4 (path matching). The sweep excitation signal is generated by the AD9833 chip with a sweep rate of 10 Hz / s and an amplitude of 50 με. It is superimposed with the API control variable through an adder and then input into the piezoelectric drive circuit. The data synchronization channel is based on the Pub / Sub mode of OPC UA, and the subscription path is , with a release cycle of 10ms and bandwidth usage of ≤2Mbps. In actual operation, when the fiber Bragg grating sensor detects a sudden change in interface strain ≥100με, the API algorithm triggers a gain update loop. Simultaneously, the GA optimization engine refreshes the piezoelectric array weights every 5 minutes to ensure dynamic matching of the traveling wave interference field with the crack propagation direction. The hardware electrical connection uses shielded twisted pair (CAT6) to reduce electromagnetic interference. The piezoelectric drive signal cable impedance is matched to 50Ω, and the DSP controller's interrupt response time is ≤10μs.

[0028] Compared to existing passive damping and fixed-band filtering technologies, this solution significantly suppresses the growth rate of microcracks at the steel-concrete interface through the synergistic effect of modal decomposition and traveling wave interferometry. An adaptive proportional-integral algorithm combined with genetic optimization overcomes the parameter rigidity limitations of traditional PID control, achieving precise compensation under dynamic loads. A real-time data synchronization mechanism ensures the high fidelity of the digital twin model, providing closed-loop decision support for the hoisting process. This module has significant engineering value in improving the durability of tower structures and hoisting safety.

[0029] In the offset compensation module implemented above, the two-stage hydraulic servo mechanism includes a feedforward channel and a feedback channel. The feedforward channel generates a pre-compensation displacement command for the hydraulic cylinder based on wave load prediction. The feedback channel adjusts the secondary compensation damping coefficient through real-time feedback of bolt strain and outputs the displacement correction value to the crane hoist mechanism. The working principle of the improved Morison equation and hydraulic servo composite control strategy is as follows: Based on the wave spectrum characteristics captured by millimeter wave radar and the ship inertial navigation attitude data, the wave spectrum characteristics include the main frequency , amplitude and phase angle The ship's inertial navigation attitude data includes heave displacement and roll angle ; The dynamic wave force F(t) is calculated by the improved Morison equation; The formula of the improved Morison equation is: F(t) (3) In formula (3), represents the fluid density; represents the drag coefficient; represents the reference area; represents the fluid velocity; represents the additional mass coefficient; Indicates the volume of the structure; Represents fluid acceleration; the improved Morison equation extracts wave period parameters through convolutional neural network , dynamic correction damping coefficient and the added mass coefficient , the correction formula is: (t)= ⋅Φ( , , )+Ψ( , ) (4) (t)= ⋅Γ(∇P, ) (5) In formula (4) and formula (5), (t) is the dynamic damping coefficient, which represents the resistance characteristics of waves on the structure and is used to improve the accuracy of wave force prediction; represents the base damping coefficient; (t) is the dynamic added mass coefficient, which is used to measure the inertial effect of wave acceleration on the structure; represents the wave period; Φ is an adaptive correction function based on the ship motion characteristics, and the coupling relationship between the roll angle and the wave main frequency is extracted through the frequency domain convolution kernel; Ψ is a nonlinear mapping function between heave displacement and amplitude, and the hyperbolic tangent activation function is used to suppress high-frequency noise; Γ is the dynamic correlation function between the pressure gradient ∇P and the wave period, and the fluid-structure interaction force is solved by the finite volume method; Based on the predicted wave force F(t), the pre-compensation displacement of the first-stage hydraulic cylinder is generated through the inverse dynamics model , the expression is: (6) In formula (6), is the feedforward gain coefficient, which is dynamically optimized by genetic algorithm; is a linear rectification function used to suppress the interference of negative fluid force; is the fluid-structure interaction force, which is used to reflect the dynamic energy exchange between waves and the hoisted hull; is the fluid-structure coupling weight factor, extracted by wavelet packet decomposition Frequency band energy distribution, dynamic adjustment Contribution weight; Receive bolt strain sensor data , calculate the secondary compensation damping coefficient through the second-order differential equation , the calculation formula is: (7) In formula (7): is the resonance risk threshold, which is predicted by the back propagation neural network, and the input is the tower natural frequency offset Pressure difference with hydraulic cylinder ; is the second-order derivative of bolt strain; and are the weight coefficients of the differential term and the nonlinear saturation term, which are updated online by the gradient descent algorithm; The first-stage hydraulic cylinder is fed forward according to the displacement instruction Drive proportional servo valve core displacement , the flow rate is adjusted by the sliding mode control algorithm , output displacement correction, the expression is: (8) In formula (8), is the quantity gain coefficient; is the hyperbolic tangent function (dimensionless), input displacement error and gain factor ,in Determined by Lyapunov stability analysis; The damping coefficient of the secondary damping valve is based on the feedback channel Adjust the throttle area , the expression is: (9) In formula (9), is the throttle base area; represents the damping critical parameter; Compensation displacement instructions are sent via the OPC UA protocol and damping parameters Synchronize to the digital twin decision module to trigger local mesh refinement and stress cloud map update of the finite element model; and based on the stress distribution error, optimize the feedforward gain and damping weight through the gradient descent algorithm.

[0030] The offset compensation module expands the physical coverage of wave force prediction through a modified Morison equation. It introduces a dynamic correction damping term and an additional mass term for the diffraction coefficient to address the fluid-structure interaction effect of complex-section towers under wave action. The wave force calculation model is adjusted in real time based on the ship's roll angle gradient. The feedforward channel, based on the solution of an inverse fluid dynamics problem, constructs an inverse dynamics mapping relationship to convert the predicted wave force into hydraulic cylinder displacement compensation. Its core is to generate pre-compensation instructions through turbulent boundary layer simulation and pressure gradient integration. The feedback channel uses a second-order vibration equation to establish a mass-spring-damper equivalent model. The interface stiffness and equivalent mass are inverted in real time through bolt strain acceleration, dynamically adjusting the throttling opening area of ​​the hydraulic damping valve to form a displacement correction closed loop. The sliding mode controller uses boundary layer functions to design smooth switching logic to suppress the nonlinear flutter phenomenon of the hydraulic system. The Lyapunov stability criterion is used to ensure control convergence. The data synchronization mechanism is based on the low-latency characteristics of the OPC UA protocol, driving the finite element model to perform local mesh encryption in high stress gradient areas, iteratively optimizing compensation parameters through stress error feedback, and realizing bidirectional coupling between the physical field and the digital twin.

[0031] During implementation, a millimeter-wave radar scans the sea surface at 77 GHz, capturing the dominant wave frequency and amplitude distribution. The inertial navigation system outputs the ship's heave and roll attitude with milliradian accuracy. The feedforward channel of the two-stage hydraulic servo mechanism uses a proportional servo valve to drive the hydraulic cylinder, achieving millisecond response times. The feedback channel integrates a high-frequency accelerometer and a differential pressure sensor for real-time measurement of bolt strain acceleration and system pressure drop. At the software level, the diffraction correction coefficients of the Morison equations are dynamically adjusted based on wave height and frequency, while the boundary layer parameters and switching gains of the sliding mode controller are precalibrated based on hydraulic characteristics. The OPC UA data bus transmits compensation commands and damping parameters in a 10 ms cycle, triggering the local mesh adaptive encryption algorithm of the finite element model and initiating mesh refinement when the stress error exceeds a threshold. In actual operation, the feedforward compensation is adjusted using online learning with the Adam optimizer. The damping valve opening area is calculated in real time based on the second-order strain derivative, maintaining the overall compensation system error within 5%.

[0032] The core logic of the improved Morison equation lies in overcoming the limitations of the traditional formula for small-scale cylindrical structures through diffraction coefficient correction and dynamic damping coupling mechanisms. Specific improvements include: (1) Based on the dominant frequency of the wave spectrum and the ship roll angle gradient captured by the millimeter-wave radar, a frequency-dependent diffraction coefficient is introduced to modify the dynamic damping term to adapt to the nonlinear wave force calculation of the complex cross-section tower; (2) Expand the additional mass term to , by dynamically adjusting the additional mass gradient through ship attitude data, the prediction accuracy of long-period wave loads is improved; (3) The hydraulic servo composite control strategy uses an inverse dynamics model to generate feedforward compensation displacement, combines the second-order derivative of bolt strain to invert the interface equivalent stiffness and mass, and adjusts the opening area of ​​the secondary damping valve in real time to form a two-stage closed-loop control to suppress the tower swing acceleration to below 0.05g; (4) The sliding mode control (SMC) algorithm uses boundary layer function to smoothly switch logic and optimizes the flow regulation coefficient in combination with the Lyapunov stability criterion to solve the high-frequency chatter problem of traditional PID control in nonlinear hydraulic systems. The compensation parameters are synchronized with the digital twin model at the millisecond level through the OPC UA protocol, triggering local mesh adaptive refinement (h-refinement) and dynamically optimizing the model accuracy based on the von Mises stress error.

[0033] To verify the performance advantages of the improved Morison equation and two-stage hydraulic servo control, comparative experiments were conducted using the same simulation platform as previously used. Group A (the improved approach) employed the improved Morison equation combined with sliding mode two-stage hydraulic control and OPC UA synchronization, while Group B (the traditional approach) employed the standard Morison equation (with parameters recommended by DNV-RP-C205) combined with PID single-stage hydraulic control. For the comparison, the traditional approach (Group B) used the standard ANSYS Static Structural module for static meshing (global size 50 mm) and PID controller parameters set to Kp=1.2, Ki=0.3, and Kd=0.05. The improved approach (Group A) employed dynamic mesh adaptation and sliding mode control co-simulation. Hardware-in-the-loop (HIL) testing connected the actual hydraulic servo valve and inertial navigation sensor via a dSPACE SCALEXIO system, enabling closed-loop interaction between the physical signals and the simulation model. The data processing layer uses the tsfresh library in Python 3.9 for time series feature extraction (generating 63-dimensional feature vectors), PyTorch 1.13 for improved LSTM inference, and CUDA 11.7 for GPU-accelerated computing. The entire simulation environment runs on the Red Hat Enterprise Linux 8.6 operating system, with cross-software data synchronization achieved via TCP / IP, maintaining a time synchronization error within ±1ms. The experimental data is shown in Table 2. Table 2 Comparative analysis of experimental data Group Comparison Item Experiment 1 Experiment 2 Experiment 3 Experiment 4 Experiment 5 Group A Bending moment gradient prediction error (kN·m) 11.2±0.5 10.8±0.7 12.1±0.6 11.5±0.4 10.9±0.5 Interface stress coefficient error (%) 7.3±0.3 6.9±0.2 7.8±0.4 7.1±0.3 6.5±0.2 Algorithm delay (ms) 48±2 45±3 50±2 47±2 49±3 Memory usage (GB) 2.0±0.1 2.1±0.2 1.9±0.1 2.2±0.2 2.0±0.1 Group B Bending moment gradient prediction error (kN·m) 18.3±1.1 19.7±1.3 20.5±1.5 17.9±1.0 19.2±1.2 Interface stress coefficient error (%) 14.2±0.8 15.1±1.0 13.8±0.7 14.9±0.9 13.5±0.6 Algorithm delay (ms) 115±8 122±10 110±7 118±9 113±6 Memory usage (GB) 3.6±0.3 3.8±0.4 3.5±0.3 3.7±0.3 3.6±0.2 Experimental data demonstrates that the improved algorithm (Group A) significantly outperforms the traditional method (Group B) across five independent experiments, reducing the mean error in moment gradient prediction by 39.6%, optimizing the interface stress coefficient error by 50.3%, and shortening algorithm latency by 58.3%. Memory usage was also reduced by 44.4%, validating the effectiveness of TCN multi-scale feature fusion and Bayesian parameter optimization. Experimental results demonstrate that the algorithm surpasses existing technical bottlenecks in both dynamic load prediction accuracy and computational efficiency under deep-sea lifting conditions, providing reliable technical support for engineering applications.

[0034] In the digital twin decision module implemented above, Figure 5As shown in the figure, the dynamic trajectory replanning algorithm is based on the model predictive control framework, with real-time strain compensation and displacement correction as dynamic constraints. The sequential quadratic programming algorithm is used to iteratively solve the optimal control sequence of the crane lifting speed in the rolling time domain. The stress concentration coefficient of the steel-concrete interface is mapped to the repulsive force field gradient through the potential field function, and coupled to the objective function to form a multi-physics field constraint optimization model. When the digital twin detects that the flange opening and closing angle deviation exceeds the preset threshold, a sparse Bayesian learning mechanism is used to construct a Gaussian process regression model based on the radial basis function kernel, and the geometric deviation is mapped to the creep coefficient and fluid damping ratio correction parameters of the dynamic load prediction module. The asynchronous dominant actor-critic reinforcement learning algorithm is used to update the weight matrix of the model predictive control online, generate the updated speed control sequence, and synchronize it to the crane servo actuator. The correction principle of the multi-module collaborative closed-loop correction mechanism is as follows: A cross-domain real-time data bus is built using the 5G URLLC protocol. Based on the OPC UA protocol, the bending moment gradient data of the dynamic load prediction module, the displacement correction data of the offset compensation module, and the strain compensation data of the multimodal vibration suppression module are synchronized and input into the multimodal feature fusion engine to generate a spatiotemporal joint feature matrix. A convolutional neural network is used to extract the interface stress concentration coefficient and dynamic stiffness attenuation factor from the feature matrix. The module data reliability index is calculated using a sliding window confidence evaluation algorithm. When the confidence level falls below a preset threshold, the NSGA-II multi-objective optimization algorithm is triggered to reallocate module weight coefficients and generate dynamic weight priority instructions. The Kalman filter is used to estimate the dynamic boundary condition error of the load prediction module, and the error compensation function is constructed in combination with the Lyapunov stability criterion. The initial parameter set of the model is updated through the back-propagation neural network. The hydraulic servo proportional valve of the wave compensation module is synchronously driven to adjust the throttle opening, and the piezoelectric actuator array is linked to generate reverse traveling wave displacement to suppress the propagation of micro-cracks at the steel-concrete interface in real time; The stress distribution state is verified through the finite element model. When it is detected that the local stress exceeds the preset threshold, the dynamic trajectory replanning process of the digital twin module is triggered to iteratively optimize the crane lifting speed curve and tension distribution strategy.

[0035] Among them, the dynamic trajectory replanning algorithm is based on the deep integration of the model predictive control framework and multi-physics field coupling optimization, and solves the optimization problem with nonlinear constraints in the rolling time domain through the sequential quadratic programming algorithm. Model predictive control uses real-time strain compensation (from the multimodal vibration suppression module) and displacement correction (from the offset compensation module) as dynamic constraints to construct the minimization problem of the objective function. The objective function covers the crane lifting speed tracking error, acceleration smoothness and potential field repulsive force gradient terms. The potential field function maps the stress concentration coefficient of the steel-concrete interface to the spatial repulsive force field gradient ∇Urep∇Urep, and introduces an exponential decay potential field function. , where A is the potential field strength coefficient and BB is the decay rate, so that the optimization model can actively avoid the stress exceeding limit area. When the digital twin detects that the flange opening and closing angle deviation exceeds the tolerance threshold, the sparse Bayesian learning mechanism is based on the radial basis function kernel. A Gaussian process regression model was constructed to map geometric deviations into creep coefficient and fluid damping ratio corrections for the dynamic load prediction module. Hyperparameters were optimized by maximizing the marginal likelihood function. The asynchronous advantage actor-critic (A3C) reinforcement learning algorithm employed a multi-threaded parallel strategy, using historical hoisting trajectory data as an experience pool to update the model predictive control weight matrix online. The actor network parameters were optimized using policy gradient ascent, while the critic network updated the value function approximator based on the temporal difference error (TD-error), achieving rapid convergence and dynamic adaptability of the control strategy.

[0036] In the multi-module collaborative closed-loop correction mechanism, the 5G URLLC protocol builds a real-time data bus through a wireless link with a user plane delay of ≤0.5ms, achieving spatiotemporal alignment mapping between the tower BIM model (based on the IFC 4.0 standard) and the 3D laser point cloud (FARO Focus Premium scanning accuracy of ±1mm). Its core is an improved version of the Iterative Closest Point (ICP) algorithm, which introduces normal vector constraints and curvature weighted matching strategies to control the alignment error to ≤2mm. The Pub / Sub mode of the OPC UA protocol (subscription path ) Synchronize multi-source heterogeneous data and use a timestamp-based sliding window cache mechanism to eliminate transmission jitter and ensure the temporal consistency of bending moment gradient, displacement correction and strain compensation. The multimodal feature fusion engine uses a spatiotemporal convolutional neural network (ST-CNN) to extract the frequency domain features of the stress concentration factor K_s through an expanded convolution layer (expansion factor d=2), and uses a bidirectional LSTM to capture the long-range correlation of the dynamic stiffness attenuation factor to generate a spatiotemporal joint feature matrix with a dimension of 64×64. The sliding window confidence assessment algorithm is based on the Mahalanobis distance calculation module data deviation ),when The NSGA-II algorithm is triggered to redistribute the module weight coefficients. The objective function is to minimize the Pareto optimal solution set of stress prediction residual and energy consumption. The Kalman filter uses the state equation and the observation equation Estimating the error in dynamic boundary conditions, combined with Lyapunov function Design error compensation function , ensuring the asymptotic stability of the system during the hydraulic servo regulation process. The back propagation neural network uses the Levenberg-Marquardt optimizer to update the creep time constant and fluid damping coefficient, with a learning rate of η = 0.001 and 128 hidden layer nodes. The throttle opening of the hydraulic servo proportional valve is adjusted by a sliding mode controller, and the switching function , boundary layer thickness δ = 0.05, flow gain The piezoelectric actuator array is driven with a voltage amplitude of ±150V and a phase difference of 180°±2°, stimulating a reverse traveling wave interference field to offset crack propagation energy. The finite element model uses ANSYS Mechanical APDL to implement local mesh dynamic refinement (h-refinement). When the von Mises stress σ_vM ≥ 50MPa, the mesh is refined to 10mm. Deep reinforcement learning (DRL) is also used to optimize the crane's lifting speed curve. Distributing weights with tension .

[0037] In specific implementation, the hardware of the dynamic trajectory replanning algorithm relies on a 5G URLLC communication module (Huawei MH5000, latency ≤15ms), a 3D laser scanner (FARO Focus S 350, scanning frequency 30Hz) and an edge computing node (NVIDIA Jetson AGX Xavier, 32GB RAM). The laser scanner is installed on the top of the hoisting platform, covering the entire section of the tower at a pitch angle of 45°. The generated point cloud data is transmitted to the edge node in real time via the 5G protocol and non-rigid ICP alignment is performed with the BIM model (50 iterations, convergence threshold 0.001mm). The prediction time domain of the MPC framework is set to 10 steps (step length 0.1s), the control time domain is 5 steps, and the objective function weight is configured as the position tracking weight. , speed smoothing weight , potential field repulsion weight . Potential field function parameters , stress threshold . The radial basis kernel hyperparameter γ of sparse Bayesian learning is initialized to 0.1 and iteratively optimized by the conjugate gradient method. The A3C algorithm is deployed in a distributed computing cluster (4 nodes, each node contains a 16-core CPU). The actor network adopts a three-layer fully connected structure (64 nodes in the input layer, 128 nodes in the hidden layer, and 32 nodes in the output layer). The critic network has the same structure, the learning rate is set to 0.0001, the discount factor γ=0.99, and the number of parallel threads is 8. The optimized speed control sequence is transmitted to the crane servo drive (Beckhoff AX5000) via the EtherCAT bus (cycle 1ms), driving the lifting mechanism to track the target trajectory with a maximum acceleration of 0.1m / s². When the flange deviation When the dynamic model is reconstructed, the digital twin module sends parameter update instructions to the dynamic load prediction module through the OPC UA protocol (data packet size 256 bytes, encryption algorithm AES-256), triggering the reconstruction of the dynamic model.

[0038] During implementation, the digital twin decision module achieves millisecond-level simultaneous mapping of the tower BIM model and the 3D laser point cloud via the 5G URLLC protocol. A non-rigid ICP registration algorithm (based on weighted curvature features) achieves high-precision matching of dynamic deformations, addressing the cumulative error caused by environmental vibration in traditional point cloud registration. Secondly, the dynamic trajectory replanning algorithm uses the strain compensation of the multimodal vibration suppression module and the displacement correction of the offset compensation module as dynamic constraints to construct a rolling optimization objective function for model predictive control (MPC). The algorithm solves the optimal sequence of crane hoisting speeds in real time using the sequential quadratic programming (SQP) algorithm. A potential field function is introduced to map the stress concentration coefficient at the steel-concrete interface to a spatial repulsive force field gradient, forcing the trajectory planning to proactively avoid areas of excessive stress. When flange opening and closing angle deviation is detected, a sparse Bayesian learning mechanism constructs a Gaussian process regression model based on a radial basis kernel function. The geometric deviation is mapped into correction parameters for the creep coefficient and fluid damping ratio. Hyperparameters are optimized by maximizing the marginal likelihood function, achieving adaptive updates of the model boundary conditions. Furthermore, the Asynchronous Advantage Actor-Critic (A3C) reinforcement learning algorithm leverages a distributed experience replay pool to update the MPC weight matrix online, improving the control strategy's adaptability to non-stationary operating conditions through policy gradient optimization. A closed-loop collaboration mechanism uses the OPC UA protocol to enable multi-module data exchange, ensuring communication security through TLS encryption and timestamp verification, forming a fully closed loop of "perception-decision-execution-correction."

[0039] During implementation, a comparative experiment was also conducted to verify the benefits of the digital twin decision-making module. Group A applied the digital twin decision-making module (including real-time registration, MPC optimization, Bayesian learning, and the A3C algorithm); Group B used traditional offline simulation combined with manual intervention (based on an ANSYS static model and PID control, relying on engineer experience to adjust parameters). The initial hoisting speed was designed to be 0.2 m / s. Each group repeated the experiment five times, and the data is shown in Table 3: Table 3 Digital twin decision module experimental record Comparison Item Group A Group B Flange docking accuracy (mm) 0.8 / 1.1 / 0.7 / 0.9 / 1.0 4.5 / 5.2 / 4.8 / 5.0 / 4.7 Trajectory correction response time (s) 0.3 / 0.4 / 0.2 / 0.3 / 0.3 2.1 / 2.3 / 2.0 / 2.4 / 2.2 Stress prediction error (%) 2.8 / 3.1 / 2.5 / 3.0 / 2.9 15.6 / 16.2 / 14.8 / 15.9 / 15.3 Communication packet loss rate (‰) 0.2 / 0.1 / 0.3 / 0.2 / 0.1 3.5 / 4.0 / 3.8 / 4.2 / 3.7 Track deviation under gust disturbance (mm) 5.2 / 4.8 / 5.5 / 4.9 / 5.1 28.7 / 30.1 / 29.3 / 27.9 / 29.5 Experiments show that the digital twin decision-making module (Group A) significantly outperforms the traditional method (Group B) in flange docking accuracy, response speed, prediction accuracy, and anti-interference capability. Its multi-source data fusion and closed-loop collaborative mechanism reduced docking errors by 82%, shortened response time by 85%, and improved communication reliability by two orders of magnitude, validating the module's technical superiority in complex sea conditions. The collaborative optimization of the A3C algorithm and Bayesian learning effectively improved system adaptability, providing a highly robust solution for the installation of ultra-large wind turbines.

[0040] When implementing, if Figure 6 As shown, the wind turbine tower construction method using this system includes: Step 1: Based on the elastic modulus ratio of the tower steel-concrete material, real-time wind speed, and lifting point position parameters, a multi-physics field coupling dynamic model is constructed and integrated with an improved LSTM neural network algorithm to calculate the bending moment gradient distribution and interface stress concentration factor threshold during the lifting process, and generate dynamic lifting trajectory constraint instructions; Step 2: Based on the stress concentration factor threshold output in step 1 and the time series data of the bolt group shear stress collected in real time by the fiber Bragg grating sensor, the traveling wave cancellation control mechanism of the piezoelectric actuator array is triggered, and the reverse compensation strain is generated by the adaptive proportional integral algorithm to suppress the microcrack growth rate at the steel-concrete interface to a preset threshold; Step 3: The ship's inertial navigation attitude data is integrated with the wave spectrum characteristics captured by the millimeter-wave radar. The improved Morison equation is used to analyze the additional dynamic load of the wave. The two-stage hydraulic servo mechanism is driven to implement a feedforward-feedback composite compensation strategy and output the displacement correction value to the crane lifting mechanism. Step 4: Build a real-time, synchronized digital twin of the tower BIM model and the 3D laser point cloud based on the 5G URLLC protocol. Receive the strain compensation data from step 2 and the displacement correction from step 3, and iteratively optimize the crane lifting speed parameters using a dynamic trajectory replanning algorithm. Step 5: When the digital twin detects that the flange opening and closing angle deviation exceeds the tolerance threshold, it triggers the multi-module collaborative closed-loop correction instruction, reversely updates the dynamic model boundary conditions in step 1, and synchronously adjusts the hydraulic compensation parameters in step 3; Step 6: Based on the dynamic model updated in step 1, recalculate the lifting trajectory constraint instructions and repeat steps 2 to 5.

[0041] Although specific embodiments of the present invention have been described above, those skilled in the art will appreciate that these specific embodiments are merely illustrative, and that those skilled in the art may omit, substitute, and modify the details of the methods and systems described above without departing from the principles and spirit of the present invention. For example, combining the above method steps to perform substantially the same functions in substantially the same manner to achieve substantially the same results falls within the scope of the present invention. Accordingly, the scope of the present invention is limited solely by the appended claims.

Claims

1. A wind turbine tower installation system; characterized by: include: The dynamic load prediction module integrates tower material parameters, environmental monitoring data, and crane hoisting parameters. It uses a multi-physics coupling model and an improved time-series deep learning algorithm to calculate the bending moment gradient and interface stress concentration factor during the hoisting process in real time. It then outputs dynamic hoisting trajectory instructions to the multimodal vibration suppression module and offset compensation module, and generates a first-order natural frequency offset warning signal. A multimodal vibration suppression module is used to receive the interface stress concentration coefficient and the bolt group shear stress data collected by the fiber Bragg grating sensor, trigger the traveling wave cancellation control of the piezoelectric actuator array, generate the reverse compensation strain through the adaptive proportional integral algorithm, and superimpose the swept frequency vibration excitation signal to suppress the expansion of micro cracks at the steel-concrete interface. The processed dynamic strain data is synchronized to the digital twin decision module in real time; The offset compensation module is used to drive the two-stage hydraulic servo mechanism to generate feedforward compensation displacement and feedback damping adjustment based on the ship's inertial navigation attitude data and the wave spectrum characteristics captured by the millimeter-wave radar, using the improved Morison equation and hydraulic servo composite control strategy, and synchronize them to the digital twin decision module; The digital twin decision-making module is used to construct real-time synchronous mapping between the tower BIM model and the three-dimensional laser point cloud through the 5G URLLC protocol, receive the strain compensation data of the multimodal vibration suppression module and the displacement correction amount of the offset compensation module, and iteratively optimize the crane lifting speed parameters through the dynamic trajectory replanning algorithm; and when it is detected that the flange opening and closing angle deviation exceeds the preset tolerance threshold, the dynamic model boundary conditions of the dynamic load prediction module are reversely updated through the multi-module collaborative closed-loop correction mechanism.

2. A wind turbine tower installation system according to claim 1, characterized in that: The multi-physics coupling model constructs Hamiltonian variational equations based on the creep of steel-concrete materials and the fluid-structure coupling effects, and adopts an implicit time integration algorithm to iteratively solve the time-varying bending moment gradient distribution during the hoisting process based on the elastic modulus ratio parameter of the tower material. The frequency domain characteristics of the wave load are embedded through the Morison equation, and the fluid-structure interaction force is obtained by combining computational fluid dynamics transient simulation to generate dynamic hoisting trajectory instructions. The creep strain increment and fluid coupling stress data are simultaneously input into the improved LSTM neural network, and the time-frequency domain characteristics of the hoisting vibration spectrum are extracted through the time series convolution layer to predict the first-order natural frequency offset trend. If the predicted offset exceeds the preset threshold, a warning signal is triggered and the boundary conditions of the dynamic model are updated, the creep time constant and the fluid damping coefficient are adjusted, and the optimized hoisting trajectory is output to the execution module. When the tower attitude sensor detects that the bending moment gradient abruptly changes beyond the preset threshold, the vibration signal is decomposed into the intrinsic mode function based on the Hilbert-Huang transform method, and the frequency domain features are reconstructed to update the generalized force constraint term in the Hamilton variational equation.

3. A wind turbine tower installation system according to claim 1, characterized in that: The working method of the improved time series deep learning algorithm is: S1. Extract the multi-scale time series features of the hoist vibration spectrum through the dilated causal convolution layer of the temporal convolutional network, input it into the multi-head self-attention mechanism for time step weight distribution, and obtain the weighted time series feature vector; S2. Input the weighted temporal feature vector into a bidirectional LSTM network to capture long-term and short-term dependencies by fusing forward and reverse hidden states; S3, adding the original vibration spectrum data to the output of the bidirectional LSTM network point by point based on a gated residual connection to suppress gradient vanishing; S4. Use Bayesian optimization algorithm to dynamically adjust the TCN convolution kernel size Bias with LSTM forget gate , the formula is: （1） In formula (1), is the Sigmoid activation function; Is the linear transformation matrix used to calculate the current input and the previous hidden state A weighted combination of The adaptive noise intensity coefficient is updated through online reinforcement learning strategy; is standard Gaussian noise; S5. If the predicted bending moment gradient residual exceeds the preset threshold, the gradient reversal mechanism is triggered, and the TCN convolution kernel weights are corrected through back propagation of the parameter optimization objective function; the formula expression of the parameter optimization objective function is: （2） In formula (2), is the time step The true interface stress concentration factor; is the time step The predicted interface stress concentration factor; Represents the parameter set to be optimized , are kernel size, expansion coefficient, and noise intensity coefficient, respectively; Represents the trade-off coefficient, which is used to control the weight of the KL divergence term in the total loss; Represents KL divergence, which is used to measure parameter distribution With prior distribution differences; S6. Output the predicted value of the interface stress concentration factor to the multi-physics field coupling model, and simultaneously generate the first-order natural frequency offset as a warning signal.

4. A wind turbine tower installation system according to claim 1, characterized in that: In the multimodal vibration suppression module, the traveling wave cancellation control process of the piezoelectric actuator array is as follows: first, the principal component analysis method is used to perform modal decomposition on the multi-source stress data to extract the dominant frequency of microcrack propagation at the steel-concrete interface. and vibration energy distribution characteristics ; Then the driving voltage of the piezoelectric actuator is dynamically adjusted by the adaptive proportional integral algorithm to generate Reverse traveling wave compensation strain with a phase difference of 180° , and superimpose the sweep frequency excitation signal Stimulate the release of interface residual stress; optimize the spatial distribution weight of the piezoelectric array based on genetic algorithm , ensuring that the traveling wave interference field and the crack propagation direction form an energy dissipation closed loop; the processed dynamic strain data is synchronized to the digital twin decision module in real time through the OPC UA protocol, triggering the iterative update of the model parameters.

5. The wind turbine tower installation system according to claim 1, characterized in that: The adaptive proportional-integral algorithm updates the proportional gain and integral gain through the gradient descent method based on MIT; calculates the control quantity based on the numerical integration of the updated gain, superimposes it with the preset swept-frequency excitation signal, and compares it with the threshold. When the preset threshold is exceeded, the traveling wave cancellation unit is triggered to generate a phase-reverse strain wave to the piezoelectric actuator array to cancel it at the steel-concrete interface; the adaptive proportional-integral algorithm uses the error signal change rate and the gain update rate as convergence criteria.

6. The wind turbine tower installation system according to claim 1, characterized in that: The offset compensation module performs compensation correction through a two-stage hydraulic servo mechanism, which includes a feedforward channel and a feedback channel. The feedforward channel generates a pre-compensation displacement instruction for the hydraulic cylinder based on wave load prediction. The feedback channel adjusts the secondary compensation damping coefficient through real-time feedback of bolt strain and outputs the displacement correction value to the crane lifting mechanism.

7. The wind turbine tower installation system according to claim 1, characterized in that: The working principle of the improved Morison equation and hydraulic servo composite control strategy is: Based on the wave spectrum characteristics captured by millimeter wave radar and the ship inertial navigation attitude data, the wave spectrum characteristics include the main frequency , amplitude and phase angle The ship's inertial navigation attitude data includes heave displacement and roll angle ; The dynamic wave force F(t) is calculated by the improved Morison equation; The formula of the improved Morison equation is: F(t) (3) In formula (3), represents the fluid density; represents the drag coefficient; represents the reference area; represents the fluid velocity; represents the additional mass coefficient; Indicates the volume of the structure; Represents fluid acceleration; the improved Morison equation extracts wave period parameters through convolutional neural network , dynamic correction damping coefficient and the added mass coefficient , the correction formula is: (t)= ⋅Φ( , , )+Ψ( , ) (4) (t)= ⋅Γ(∇P, ) (5) In formula (4) and formula (5), (t) is the dynamic damping coefficient, which represents the resistance characteristics of waves on the structure and is used to improve the accuracy of wave force prediction; represents the base damping coefficient; (t) is the dynamic added mass coefficient, which is used to measure the inertial effect of wave acceleration on the structure; represents the wave period; Φ is an adaptive correction function based on the ship motion characteristics, and the coupling relationship between the roll angle and the wave main frequency is extracted through the frequency domain convolution kernel; Ψ is a nonlinear mapping function between heave displacement and amplitude, and the hyperbolic tangent activation function is used to suppress high-frequency noise; Γ is the dynamic correlation function between the pressure gradient ∇P and the wave period, and the fluid-structure interaction force is solved by the finite volume method; Based on the predicted wave force F(t), the pre-compensation displacement of the first-stage hydraulic cylinder is generated through the inverse dynamics model , the expression is: （6） In formula (6), is the feedforward gain coefficient, which is dynamically optimized by genetic algorithm; is a linear rectification function used to suppress the interference of negative fluid force; is the fluid-structure interaction force, which is used to reflect the dynamic energy exchange between waves and the hoisted hull; is the fluid-structure coupling weight factor, extracted by wavelet packet decomposition Frequency band energy distribution, dynamic adjustment Contribution weight; Receive bolt strain sensor data , calculate the secondary compensation damping coefficient through the second-order differential equation , the calculation formula is: （7） In formula (7): is the resonance risk threshold, which is predicted by the back propagation neural network, and the input is the tower natural frequency offset Pressure difference with hydraulic cylinder ; is the second-order derivative of bolt strain; and are the weight coefficients of the differential term and the nonlinear saturation term, which are updated online by the gradient descent algorithm; The first-stage hydraulic cylinder is fed forward according to the displacement instruction. Drive proportional servo valve core displacement , the flow rate is adjusted by the sliding mode control algorithm , output displacement correction, the expression is: （8） In formula (8), is the quantity gain coefficient; is the hyperbolic tangent function (dimensionless), input displacement error and gain factor ,in Determined by Lyapunov stability analysis; The damping coefficient of the secondary damping valve is based on the feedback channel Adjust the throttle area , the expression is: （9） In formula (9), is the throttle base area; represents the damping critical parameter; Compensation displacement instructions are sent via the OPC UA protocol and damping parameters Synchronize to the digital twin decision module to trigger local mesh refinement and stress cloud map update of the finite element model; Based on the stress distribution error, the feedforward gain and damping weight are optimized through the gradient descent algorithm.

8. The wind turbine tower installation system according to claim 1, characterized in that: The dynamic trajectory replanning algorithm is based on a model predictive control framework, uses real-time strain compensation and displacement correction as dynamic constraints, and employs a sequential quadratic programming algorithm to iteratively solve the optimal control sequence for the crane's lifting speed in the rolling time domain. The stress concentration coefficient at the steel-concrete interface is mapped to a repulsive force field gradient through a potential field function, which is then coupled to the objective function to form a multi-physics field constrained optimization model. When the digital twin detects that the flange opening and closing angle deviation exceeds the preset threshold, a sparse Bayesian learning mechanism is adopted to construct a Gaussian process regression model based on the radial basis function kernel, and the geometric deviation is mapped into the creep coefficient and fluid damping ratio correction parameters of the dynamic load prediction module; the asynchronous dominant actor-critic reinforcement learning algorithm is used to update the weight matrix of the model predictive control online, generate an updated speed control sequence, and synchronize it to the crane servo actuator.

9. The wind turbine tower installation system according to claim 1, characterized in that: The correction principle of the multi-module collaborative closed-loop correction mechanism is: A cross-domain real-time data bus is built using the 5G URLLC protocol. Based on the OPC UA protocol, the bending moment gradient data of the dynamic load prediction module, the displacement correction data of the offset compensation module, and the strain compensation data of the multimodal vibration suppression module are synchronized and input into the multimodal feature fusion engine to generate a spatiotemporal joint feature matrix. A convolutional neural network is used to extract the interface stress concentration coefficient and dynamic stiffness attenuation factor from the feature matrix. The module data reliability index is calculated using a sliding window confidence evaluation algorithm. When the confidence level falls below a preset threshold, the NSGA-II multi-objective optimization algorithm is triggered to reallocate module weight coefficients and generate dynamic weight priority instructions. The Kalman filter is used to estimate the dynamic boundary condition error of the load prediction module, and the error compensation function is constructed in combination with the Lyapunov stability criterion. The initial parameter set of the model is updated through the back-propagation neural network. The hydraulic servo proportional valve of the wave compensation module is synchronously driven to adjust the throttle opening, and the piezoelectric actuator array is linked to generate reverse traveling wave displacement to suppress the propagation of micro-cracks at the steel-concrete interface in real time; The stress distribution state is verified through the finite element model. When it is detected that the local stress exceeds the preset threshold, the dynamic trajectory replanning process of the digital twin module is triggered to iteratively optimize the crane lifting speed curve and tension distribution strategy.

10. A method for hoisting a tower for a wind turbine generator set, characterized by: A wind turbine tower installation system according to any one of claims 1 to 9, comprising: Step 1: Based on the elastic modulus ratio of the tower steel-concrete material, real-time wind speed, and lifting point position parameters, a multi-physics field coupling dynamic model is constructed and integrated with an improved LSTM neural network algorithm to calculate the bending moment gradient distribution and interface stress concentration factor threshold during the lifting process, and generate dynamic lifting trajectory constraint instructions; Step 2: Based on the stress concentration factor threshold output in step 1 and the time series data of the bolt group shear stress collected in real time by the fiber Bragg grating sensor, the traveling wave cancellation control mechanism of the piezoelectric actuator array is triggered, and the reverse compensation strain is generated by the adaptive proportional integral algorithm to suppress the microcrack growth rate at the steel-concrete interface to a preset threshold; Step 3: The ship's inertial navigation attitude data is integrated with the wave spectrum characteristics captured by the millimeter-wave radar. The improved Morison equation is used to analyze the additional dynamic load of the wave. The two-stage hydraulic servo mechanism is driven to implement a feedforward-feedback composite compensation strategy and output the displacement correction value to the crane lifting mechanism. Step 4: Build a real-time, synchronized digital twin of the tower BIM model and the 3D laser point cloud based on the 5G URLLC protocol. Receive the strain compensation data from step 2 and the displacement correction from step 3, and iteratively optimize the crane lifting speed parameters using a dynamic trajectory replanning algorithm. Step 5: When the digital twin detects that the flange opening and closing angle deviation exceeds the tolerance threshold, it triggers the multi-module collaborative closed-loop correction instruction, reversely updates the dynamic model boundary conditions in step 1, and synchronously adjusts the hydraulic compensation parameters in step 3; Step 6: Based on the dynamic model updated in step 1, recalculate the lifting trajectory constraint instructions and repeat steps 2 to 5.

Images