Scaffold intelligent leveling method and system based on digital twinning and space-time prediction

Abstract

The invention relates to the technical field of building construction safety monitoring and automatic control, and discloses an intelligent scaffold leveling method and system based on digital twinning and space-time prediction. The method comprises the following steps: constructing a scaffold digital twinborn body and a graph structure; acquiring sensor data and outputting a deformation prediction sequence through the space-time diagram convolutional network; injecting the prediction sequence into the digital twinborn body to simulate and evaluate a plurality of leveling strategies, and selecting an optimal strategy based on the leveling precision and the structural stability; and leveling is executed through an electromechanical execution mechanism, and model parameters are updated and a prediction network is trained by using leveled data. According to the method, through deep fusion of digital twinning and space-time prediction, technical spanning from passive response to active prediction is achieved, the problems of response lag and control oscillation existing in a traditional leveling method are effectively solved, the leveling precision, efficiency and structural safety are remarkably improved, and the adaptive optimization capability is achieved.

Description

Technical Field

[0001] This invention relates to the field of construction safety and automation control technology, specifically to a method and system for intelligent scaffolding leveling based on digital twins and spatiotemporal prediction. Background Technology

[0002] In recent years, with the deepening of the smart construction site concept, scaffolding safety monitoring technology has gradually evolved from mechanization to automation and intelligence. Existing technologies mainly rely on deploying sensor networks for tilt angles, stress, etc., at key nodes of the scaffolding, and achieving real-time monitoring of the structural status through data acquisition and transmission systems. Some advanced solutions have further introduced automated leveling actuators, such as electric push rods or hydraulic jacking systems, which can perform closed-loop control based on real-time attitude information fed back by sensors, reducing the labor intensity and safety risks of manual leveling to a certain extent. Meanwhile, digital twin technology, as a key means to achieve interactive integration between physical entities and virtual spaces, has been initially explored in the field of health monitoring of some complex engineering structures, providing a new paradigm for the digitalization of structural safety management.

[0003] However, existing intelligent scaffolding leveling methods still face several technical bottlenecks that urgently need to be addressed. First, their control logic is essentially "lagging response," meaning the system only initiates the correction procedure after sensors detect deformation exceeding a threshold. This feedback control, based on current or historical states, cannot anticipate and address cumulative deformations caused by dynamic factors such as continuous loads, wind loads, or concrete pouring. This results in significant delays in the leveling process and is prone to repeated adjustments and "oscillations," affecting leveling efficiency and the instantaneous stability of the structure. Second, existing methods largely rely on single-point or local sensor information, lacking a deep understanding of the overall spatial topology and deformation transmission patterns of the scaffolding structure. Their control strategies are often isolated, making it difficult to achieve coordinated optimization of multiple leveling actuators at the global level. This can lead to secondary stresses in adjacent members due to local over-adjustment, or even structural damage. Furthermore, existing systems generally lack self-learning and adaptive evolution capabilities. The control models and parameters of scaffolding are usually static after deployment, making it impossible to self-correct and iteratively optimize based on actual leveling effects and structural response data. This makes it difficult to adapt to the time-varying characteristics of scaffolding structures (such as loose connections and material fatigue) and complex and ever-changing on-site conditions, leading to a gradual decline in reliability and accuracy over long-term use. How to achieve the leap from passive response to active prediction, from local control to global coordination, and from static models to dynamic self-learning is a key technical challenge in this field. Summary of the Invention

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

[0005] Therefore, the technical problem solved by this invention is that existing intelligent scaffolding leveling methods have response lag and are prone to control oscillations, lack spatial collaborative control capabilities for overall structural deformation, and have fixed system models that lack self-learning and adaptive optimization capabilities. The problem is how to achieve a forward-looking, collaborative, and self-evolving intelligent leveling method based on deformation prediction and digital simulation.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for intelligent scaffolding leveling based on digital twins and spatiotemporal prediction, comprising: constructing a digital twin based on the geometric structure and material properties of the scaffolding, and constructing a graph structure of the scaffolding based on the structural connection information in the digital twin; collecting real-time sensor data of key nodes of the scaffolding, preprocessing the collected sensor data, inputting the preprocessed real-time sensor data and the graph structure into a spatiotemporal graph convolutional network, extracting spatiotemporal features through the spatiotemporal graph convolutional network, and calculating and outputting the deformation prediction sequence of the key nodes; The deformation prediction sequence is injected into the digital twin, and various leveling strategies are simulated using the deformation prediction sequence as initial conditions. The optimal leveling strategy is selected based on the leveling accuracy and structural stability evaluation. The optimal leveling strategy is sent to the electromechanical actuators corresponding to each key node for leveling, and sensor data of the key nodes after leveling are collected to update the model parameters of the digital twin and iteratively train the spatiotemporal graph convolutional network.

[0007] In a preferred embodiment of the method described in this invention, the construction of the digital twin and the graph structure of the scaffolding includes: constructing an initial digital twin based on geometric structural data and material properties obtained from a building information model or design drawings; the material properties include at least the elastic modulus, Poisson's ratio, and material density; defining the pipe connectors or load-bearing nodes of the scaffolding as vertices of the graph structure based on the structural connection information in the digital twin, and defining the members connecting adjacent vertices as edges of the graph structure; and establishing an adjacency matrix of the graph structure according to the connection relationship between the vertices and edges, wherein when there are directly connected members between two vertices, the element value at the corresponding position in the adjacency matrix is ​​1, otherwise it is 0.

[0008] As a preferred embodiment of the method described in this invention, the preprocessing of the collected sensor data includes: collecting real-time data through multiple types of sensors deployed at key nodes; performing integrity verification and missing value imputation on the collected raw data; sequentially performing Kalman filtering dynamic smoothing and wavelet transform threshold denoising on the imputed data; and performing timestamp alignment and spatial registration on the denoised multi-source sensor data.

[0009] As a preferred embodiment of the method described in this invention, the step of extracting spatiotemporal features through a spatiotemporal graph convolutional network to calculate and output the deformation prediction sequence of key nodes includes: inputting preprocessed sensor data and the adjacency matrix of the graph structure into the spatiotemporal graph convolutional network; extracting the spatial topological features of the scaffolding structure through the graph convolutional layers of the network; extracting the temporal series features of the sensor data through the one-dimensional convolutional layers of the network; fusing spatial and temporal features and performing feature enhancement through an attention mechanism; and outputting the deformation prediction sequence of key nodes for multiple future time steps.

[0010] In a preferred embodiment of the method described in this invention, the step of simulating multiple leveling strategies and selecting the optimal leveling strategy includes: loading the deformation prediction sequence as an initial condition into a digital twin to generate multiple leveling strategies; performing dynamic simulation on each leveling strategy to calculate the leveling accuracy index and structural stability index under each strategy; comprehensively considering the leveling accuracy and structural stability through an evaluation function to calculate the comprehensive score of each strategy; and selecting the leveling strategy with the highest comprehensive score as the optimal leveling strategy.

[0011] In a preferred embodiment of the method described in this invention, the step of sending the optimal leveling strategy to the electromechanical actuators for leveling includes: parsing the optimal leveling strategy into independent control commands for each key node; distributing the control commands to the corresponding electromechanical actuators via an industrial bus network; each electromechanical actuator performing precise displacement adjustment according to the received control commands; and monitoring the leveling process in real time, pausing the leveling and re-evaluating when an abnormal deviation is detected.

[0012] As a preferred embodiment of the method described in this invention, the step of updating the digital twin model parameters and iteratively training the spatiotemporal graph convolutional network includes: collecting sensor data of key nodes as a verification dataset after leveling; comparing and analyzing the verification data with the simulation prediction data of the digital twin; updating the model parameters of the digital twin based on the comparison results; adding the time-series data of the leveling process to the training set and iteratively training the spatiotemporal graph convolutional network.

[0013] Another objective of this invention is to provide an intelligent scaffolding leveling system based on digital twins and spatiotemporal prediction. Through the coordinated operation of four core modules, it constructs a complete technical closed loop of "perception-prediction-decision-execution-evolution", effectively solving the technical problems of slow response, oscillation control, lack of coordination and adaptive capabilities in the prior art.

[0014] As a preferred embodiment of the intelligent scaffolding leveling system based on digital twins and spatiotemporal prediction described in this invention, the system includes: a digital twin modeling module, a perception and prediction module, a decision simulation module, and an execution evolution module; the digital twin modeling module is used to construct a digital twin based on the geometric structure and material properties of the scaffolding, and to construct the graph structure of the scaffolding based on structural connection information; the perception and prediction module is used to collect real-time sensor data of key nodes of the scaffolding and perform preprocessing, input the processed data and graph structure into a spatiotemporal graph convolutional network, calculate and output the deformation prediction sequence of the key nodes; the decision simulation module is used to inject the deformation prediction sequence into the digital twin, simulate various leveling strategies using the sequence as initial conditions, and select the optimal leveling strategy based on the evaluation of leveling accuracy and structural stability; the execution evolution module is used to send the optimal leveling strategy to the electromechanical actuators corresponding to each key node for leveling, collect the leveled sensor data, update the digital twin model parameters, and iteratively train the spatiotemporal graph convolutional network.

[0015] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for intelligent scaffolding leveling based on digital twins and spatiotemporal prediction.

[0016] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor of steps in a smart scaffolding leveling method based on digital twins and spatiotemporal prediction.

[0017] The beneficial effects of this invention: The intelligent scaffolding leveling method based on digital twins and spatiotemporal prediction provided by this invention achieves a technological leap in scaffolding safety control from passive response to active prediction, and from local adjustment to global optimization by constructing an intelligent leveling system that combines digital twins and spatiotemporal prediction. Traditional leveling methods rely on sensor lag data and can only perform passive correction after deformation occurs. This invention, however, uses a spatiotemporal graph convolutional network to deeply fuse and analyze multi-source sensor data and scaffolding topology, enabling accurate prediction of structural deformation trends in future periods. This gives the leveling system a forward-looking intervention capability, effectively solving the technical problems of control lag and repeated oscillations. Based on this, the system injects the predicted sequence into a digital twin environment for multi-strategy simulation verification. By comprehensively evaluating the leveling accuracy and structural stability under each strategy in virtual space, optimal decision-making based on quantitative indicators is achieved, avoiding blindness and safety risks in actual leveling. This predictive simulation mechanism not only ensures the scientific validity and reliability of the leveling process but also significantly improves adaptability under complex working conditions. More importantly, this invention constructs a complete self-evolutionary closed loop. By collecting the discrepancies between actual leveling data and predicted results, the system continuously updates the parameters of the digital twin model and optimizes the weights of the prediction network. This allows the system to adapt to the time-varying characteristics of the scaffolding structure and environmental changes over time, achieving a fundamental improvement from fixed-parameter control to adaptive learning. This continuous evolution capability ensures the long-term reliability and accuracy of the system, solving the pain point of traditional systems where performance degrades with time and changing working conditions. Overall, this invention, through three core innovations—predictive control, virtual simulation optimization, and adaptive learning—achieves a synergistic and enhanced technical effect. It not only significantly improves leveling accuracy, response speed, and safety but also achieves breakthroughs in long-term system stability and self-learning capabilities, providing a completely new technological paradigm for scaffolding safety monitoring. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The first embodiment of the present invention provides an overall flowchart of a smart scaffolding leveling method based on digital twins and spatiotemporal prediction. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Example 1, referring to Figure 1 As one embodiment of the present invention, a method for intelligent scaffolding leveling based on digital twins and spatiotemporal prediction is provided, comprising: S1: Construct a digital twin based on the geometric structure and material properties of the scaffold, and construct the graph structure of the scaffold based on the structural connection information in the digital twin.

[0022] Furthermore, based on the geometric data and material properties obtained from building information models or design drawings, an initial digital twin is constructed, with material properties including at least the elastic modulus, Poisson's ratio, and material density.

[0023] First, export the scaffolding's geometric data from a pre-built Building Information Model (BIM). If the BIM model is unavailable, obtain key coordinate data by scanning the scaffolding design drawings. The geometric data should include: the bottom and top spatial coordinates (x, y, z) of all uprights, the coordinates of the connection endpoints of horizontal members, the coordinates of the fixing points of diagonal braces, and the spatial location and type information of connectors.

[0024] Then, the above geometric data is imported into finite element analysis software (such as ABAQUS or ANSYS) to establish a preliminary geometric model of the scaffold. Subsequently, according to the scaffold design drawings, corresponding material properties are assigned to each member. Among these, the material property assignment includes: elastic modulus. Typical values ​​for Q235 steel Poisson's ratio The standard value for steel is 0.3, and the material density is... It is 7850 kg / m³.

[0025] Finally, after assigning material properties, the scaffolding geometric model was discretized into a finite element mesh using the mesh generation function of the finite element software, and a digital twin computational model containing the mass matrix and stiffness matrix was established.

[0026] Furthermore, based on the structural connection information in the digital twin, the pipe connectors or load-bearing nodes of the scaffolding are defined as vertices of the graph structure, and the members connecting adjacent vertices are defined as edges of the graph structure. An adjacency matrix of the graph structure is established based on the connection relationships between vertices and edges. In this matrix, the element at the corresponding position is 1 when there is a directly connected member between two vertices, and 0 otherwise.

[0027] First, define the vertices by defining each pipe connector or main load-bearing node in the scaffolding as a vertex in the graph structure. Each vertex contains attribute information such as spatial coordinates, node type, and carrying capacity level.

[0028] Then, we define edges, defining the members connecting adjacent vertices as edges in the graph structure. Each edge contains attribute information such as the connected vertex pairs, member type, member length, and cross-sectional properties.

[0029] Finally, the adjacency matrix is ​​constructed based on the connection relationships between vertices and edges, establishing the adjacency matrix of the graph structure. The matrix is Symmetric matrix ( (where the total number of vertices is 1), and the assignment of its elements follows these rules: When the vertex... When there are directly connected rods, matrix elements ;otherwise And diagonal elements .

[0030] It should be noted that in the actual software implementation, a graph computing library is used to construct the graph structure described above. The specific process includes: first, creating an undirected graph data structure; then, adding all vertices and setting their attributes; next, adding edges based on the link connections and setting their attributes; and finally, generating an adjacency matrix representation based on the graph's connectivity.

[0031] It should also be noted that the output of this implementation process includes a digital twin model file containing complete physical properties, a set of vertices and edges of the graph structure, and an adjacency matrix representation of the scaffold topology.

[0032] The accuracy of the digital twin structure construction is verified by comparing the mechanical simulation results of the digital twin with the actual scaffolding measurement data. When the simulation error is less than a predetermined threshold, the digital twin structure construction is considered to meet the engineering accuracy requirements, and the process can proceed to the next step.

[0033] This step establishes a digital twin and graph structure that accurately reflects the physical characteristics and topological relationships of the scaffolding, providing a reliable digital foundation for subsequent deformation prediction and intelligent leveling.

[0034] S2: Collect real-time sensor data of key nodes of the scaffolding, preprocess the collected sensor data, input the preprocessed real-time sensor data and graph structure into the spatiotemporal graph convolutional network, extract spatiotemporal features through the spatiotemporal graph convolutional network, and calculate and output the deformation prediction sequence of key nodes.

[0035] Furthermore, real-time data is collected by multiple types of sensors deployed at key nodes; the collected raw data is verified for integrity and missing values ​​are filled in; the filled data is then subjected to Kalman filtering dynamic smoothing and wavelet transform threshold denoising; and the denoised multi-source sensor data is then time-stamp aligned and spatially registered.

[0036] It should be noted that the multi-source sensor data is collected by deploying various types of sensor arrays at key nodes of the scaffolding, including tilt sensors, stress sensors, and optional displacement sensors. The tilt sensors measure the three-dimensional attitude angle of the nodes at a sampling frequency of 10Hz, the stress sensors monitor the load distribution of the connection nodes, and the displacement sensors measure the relative displacement of the nodes. All sensor data is transmitted to the central processing unit in real time via a wireless transmission module.

[0037] It should also be noted that data integrity verification and missing value imputation employ a sliding time window mechanism (window size of 50 sampling points) to verify the integrity of the original sensor data. When data is missing, a linear interpolation algorithm is used for imputation. This method performs linear calculations based on the values ​​of the nearest valid data points before and after the missing point to ensure the continuity of the data sequence.

[0038] It should also be noted that the Kalman filter dynamic smoothing process applies the standard Kalman filter algorithm to dynamically smooth the imputed data. This algorithm recursively estimates the system state through the state equation and observation equation, effectively eliminating high-frequency measurement noise. Both process noise and observation noise are assumed to have a zero-mean Gaussian distribution, and the filtering effect is optimized by adjusting the covariance matrix parameters. Wavelet transform threshold denoising uses the sym4 wavelet basis to perform a three-level wavelet decomposition on the Kalman-filtered data. A soft threshold function is applied to the detail coefficients obtained from the decomposition for denoising. The threshold is adaptively calculated based on the noise standard deviation and signal length to separate the effective signal from the noise components. Timestamp alignment and spatial registration establish a unified time reference, resampling all sensor data to a 10Hz frequency. A spatial transformation matrix is ​​used to transform the sensor data from each node into a unified scaffold coordinate system, ensuring the consistency of multi-source data in the spatiotemporal dimensions.

[0039] Furthermore, the preprocessed sensor data and the adjacency matrix of the graph structure are input into a spatiotemporal graph convolutional network; the spatial topological features of the scaffolding structure are extracted through the graph convolutional layers of the network; the temporal series features of the sensor data are extracted through the one-dimensional convolutional layers of the network; the spatial and temporal features are fused and feature enhancement is performed through an attention mechanism; and the deformation prediction sequence of key nodes in the future multiple time steps is output.

[0040] First, the preprocessed sensor data and the graph adjacency matrix are input into a spatiotemporal graph convolutional network. The sensor data is represented as follows: The adjacency matrix is ​​represented as ,in, The tensor field is the real number field, indicating that the tensor elements are real numbers. The number of critical nodes in the scaffolding. This represents the length of the time series (number of historical time steps). The sensor features (such as tilt angle, stress, and other measured values) are used as the feature dimensions. Spatial topological features of the scaffolding structure are extracted using graph convolutional layers. The spectral graph convolution operation is implemented using a Chebyshev polynomial approximation, as shown in the following formula:

[0041] in, The result of the graph convolution operation ( For parameterized convolution kernels, As input features, (for graph convolution operations) For a trainable parameter vector, Let k be the Chebyshev polynomial of order k. The scaled Laplacian matrix is ​​calculated using the following formula:

[0042] , For the graph Laplace matrix, The largest eigenvalue of the Laplace matrix. It is an N-dimensional identity matrix. Let be the order of the Chebyshev polynomial, with a value of 3.

[0043] The spatial characteristics output by this operation The structural relationships between nodes were accurately captured, laying the foundation for subsequent time series analysis.

[0044] Then, spatial features The data is fed into a temporal convolutional layer, where a one-dimensional convolution operation is used to extract the dynamic changes over time. The formula is as follows:

[0045] in, For time-dynamic feature matrix, The activation function for the rectified linear unit is... For multi-scale temporal convolution kernels, This is the bias vector for the temporal convolutional layer.

[0046] Simultaneously capture short-term fluctuations and long-term trend features using convolution kernels of different widths.

[0047] Finally, spatial features With time characteristics By concatenating along the feature dimension and adaptively adjusting the feature weights through an attention mechanism, the ability to represent key information is enhanced, and the resulting fusion yields... (Spatiotemporal fusion feature matrix). Deformation prediction sequences are generated through mapping using a fully connected output layer, as shown in the following formula:

[0048] in, The output deformation prediction sequence is a three-dimensional tensor. The tensor field is the real number field, indicating that the tensor elements are real numbers. This is the mapping function for the output layer, typically a fully connected layer. The feature tensor after spatiotemporal feature fusion and attention enhancement. This represents the total number of critical nodes in the scaffolding. For the predicted number of future time steps, The deformation dimension predicted for each node at each future time step, typically representing three-dimensional spatial displacement (e.g., displacement along the X, Y, Z axes or a composite displacement vector).

[0049] Furthermore, the prediction accuracy is evaluated using the root mean square error (RMSE), calculated as follows:

[0050] in, To represent the scaffolding The key node at the 1st The actual deformation value at each time step To represent the scaffolding The key node at the 1st Predicted deformation values ​​at each time step. The total number of critical nodes in the scaffolding is compared with the structure shown in the previous diagram and the prediction formula. The definition remains consistent. The predicted future time steps are compared with those in the previous prediction formula. The definition remains consistent.

[0051] When the RMSE of the test set is lower than the preset threshold of 0.5mm, the model is considered to have met the engineering application standard and can proceed to the subsequent leveling decision process.

[0052] This step realizes a complete processing flow from raw sensor data to deformation prediction sequences, providing an accurate predictive foundation for intelligent leveling. The retained formulas are key technical features for achieving this invention; other prior art aspects have been described in text to ensure clarity and conciseness.

[0053] S3: Inject the deformation prediction sequence into the digital twin, simulate various leveling strategies using the deformation prediction sequence as the initial condition, and select the optimal leveling strategy based on the leveling accuracy and structural stability evaluation.

[0054] Furthermore, dynamic simulations were performed for each leveling strategy.

[0055] First, the deformation prediction sequence output by the spatiotemporal graph convolutional network is... The initial displacement boundary conditions are applied to the finite element model of the digital twin. The output is the deformation prediction sequence. The tensor field is the real number field, indicating that the tensor elements are real numbers. This represents the total number of critical nodes in the scaffolding. For the predicted number of future time steps, These are three-dimensional displacement components (corresponding to spatial X, Y, and Z axis displacements).

[0056] Then, based on the predicted deformation pattern, three types of leveling strategies are automatically generated in the digital twin: single-point independent leveling: priority leveling is given to the node with the largest deformation, and the leveling order is arranged in descending order of deformation magnitude; regional collaborative leveling: the region with the largest deformation gradient is identified, and all nodes in the region are leveled synchronously; global progressive leveling: leveling is carried out sequentially from the root region to the edge region according to the deformation propagation path.

[0057] Next, transient dynamic simulations were performed for each leveling strategy to simulate the structural response under leveling forces. The simulation time span covered the entire prediction period. (τ is the prediction time step), with the time step set to 0.1 seconds to ensure the capture of the dynamic leveling process. For each strategy... The height of the leveled nodes was calculated using finite element analysis. This result provides foundational data for subsequent multi-objective evaluation.

[0058] Furthermore, the leveling accuracy index and structural stability index under each strategy are calculated.

[0059] First, the root mean square error is used to measure the leveling accuracy:

[0060] in, For the first The leveling accuracy index of each strategy. For the first The target horizontal height of each node For the first After the first strategy was adjusted downwards, the second... The actual height of each node The number of key nodes participating in the evaluation.

[0061] At the same time, for each strategy Calculate the structural stability index:

[0062] in, For the first The structural stability index for each strategy ranges from [0,1]. A value closer to 1 indicates a higher structural safety margin. For the first During the simulation of each strategy, the maximum structural output... , The yield strength of the scaffolding material.

[0063] Furthermore, by comprehensively considering both leveling accuracy and structural stability through an evaluation function, a comprehensive score for each strategy is calculated, and the leveling strategy with the highest comprehensive score is selected as the optimal leveling strategy.

[0064] First, a weighted comprehensive evaluation function is established to balance the leveling accuracy and structural stability. The standardized formula is as follows:

[0065] in, For the first The overall strategy score for each strategy These are weighting coefficients (scalars), corresponding to the importance weights of leveling accuracy and structural stability, respectively, satisfying the constraint w1+w2=1; default values: w1=0.6 (leveling accuracy weight), w2=0.4 (structural stability weight). This is the maximum allowable leveling error threshold.

[0066] Then, select to make the overall score The strategy that maximizes the yield is the optimal balancing strategy, i.e.:

[0067] in, This is the optimal leveling strategy.

[0068] Furthermore, the effectiveness of the decision-making system is verified by comparing simulation results of different strategies. The optimal strategy must simultaneously satisfy: balancing accuracy: Structural safety: Overall score: .

[0069] This step, through the aforementioned predictive simulation and multi-objective decision-making process, enables the pre-verification and optimization of the leveling scheme in a virtual environment, ensuring the safety, efficiency, and accuracy of the actual leveling operation, and effectively avoiding the risks and inefficiencies associated with traditional trial-and-error leveling.

[0070] S4: Send the optimal leveling strategy to the electromechanical actuators corresponding to each key node for leveling, and collect sensor data of the key nodes after leveling to update the model parameters of the digital twin and iteratively train the spatiotemporal graph convolutional network.

[0071] Furthermore, the optimal leveling strategy is parsed into independent control commands for each key node; the control commands are distributed to the corresponding electromechanical actuators via an industrial bus network; each electromechanical actuator performs precise displacement adjustment according to the received control commands; the leveling process is monitored in real time, and leveling is paused and re-evaluated when abnormal deviations are detected.

[0072] First, the optimal leveling strategy is parsed into independent control command sets for each key node. Each command contains the target displacement. Maximum leveling speed: Acceleration limit: Control commands are distributed to the corresponding electromechanical actuators at each node via an industrial bus network (using the PROFIBUS-DP protocol with a transmission rate of 12Mbps).

[0073] Then, after each electromechanical actuator (consisting of a servo motor, reducer, and ball screw) receives the control command, it initiates closed-loop position control. The position controller employs an incremental PID algorithm.

[0074] in, For time index Increment of control quantity at any given time, For time The displacement deviation is calculated using the following formula: , For time The actual displacement, Set the proportional, integral, and derivative coefficients for the PID controller, and set the control cycle to 10ms to ensure real-time response.

[0075] Displacement deviation The actuator is driven to converge toward the target position. The output of the PID controller directly acts on the actuator drive unit, forming a closed-loop control loop of "command → execution → feedback → adjustment" to ensure displacement control accuracy.

[0076] Finally, a dual-criteria anomaly detection mechanism is established, with the standardized criteria as follows: Displacement deviation criterion: ,in , (for deviation threshold symbol); Duration criterion: The above displacement deviation criterion is satisfied for 3 consecutive control cycles (the default value of the control cycle is 10ms). When the above criteria are satisfied at the same time, all actuators are immediately suspended and a re-evaluation process is triggered to ensure the safety of system operation.

[0077] Furthermore, after the leveling is completed, sensor data from key nodes are collected as a validation dataset; the validation data is compared and analyzed with the simulation prediction data of the digital twin; the model parameters of the digital twin are updated based on the comparison results; and the time-series data of the leveling process is added to the training set to iteratively train the spatiotemporal graph convolutional network.

[0078] First, the validation dataset was collected. After the balancing process was completed, the system entered a 30-second rest period. Then, multiple rounds of sensor data were collected (sampling frequency 10Hz, lasting 5 seconds), and the average value of each sensor reading was calculated to form the validation dataset.

[0079] in, For the first Verification data for each node, For the first The node of the first Each sample value, The total number of sampling points is a fixed value. =50.

[0080] Then, the parameters of the digital twin model are updated, and the verification data is compared with the digital twin simulation prediction data to calculate the relative error of the key parameters, as shown in the following formula:

[0081] in, The relative error of the key parameters. For the simulation prediction parameter values ​​of the digital twin, To verify the actual parameter values ​​corresponding to the data.

[0082] When the relative error of the elastic modulus of the material At that time, a gradient descent-based parameter identification algorithm is used to update the material constitutive model parameters in the digital twin.

[0083] Finally, iterative training of the spatiotemporal graph convolutional network is performed. The complete temporal data of the leveling process (including deformation before leveling, the leveling process, and the stable state after leveling) is used to form new training samples, which are expanded to the original training set at a ratio of 1:5 (i.e., the number of new samples: the number of original samples = 1:5). The stochastic gradient descent algorithm with momentum is used for fine-tuning training, as shown in the following formula:

[0084] in, This is the set of weight parameters for a spatiotemporal graph convolutional network. For the first Time, Number Time, Number Network weight parameters at time ( (Index for training iteration steps) The learning rate is a fixed value. =0.001 (control parameter update step size) The momentum coefficient has a fixed value. =0.9 (Accelerates gradient descent convergence and suppresses oscillations). loss function For weight parameters The gradient (representing the direction of parameter update). θ is the network loss function (used to measure the deviation between predicted and actual values). During training, the network infrastructure is kept fixed, and only the weight parameters of the last two convolutional kernels are updated. The network parameters θ are updated through incremental training, allowing the prediction model to learn from actual leveling experience, continuously improving deformation prediction accuracy and achieving self-evolution of the model.

[0085] Furthermore, verification and performance evaluation were conducted: First, the leveling effect is verified. After the leveling operation is completed, the following three core performance indicators must be met: Leveling accuracy: (RMSE root mean square error); Maximum displacement deviation: ( For the target displacement, (actual displacement); structural stress safety: ( For the maximum von Mises stress, (This refers to the material's yield strength).

[0086] Then, the effect of the model update is evaluated. The optimization effect is verified by comparing the prediction accuracy of the model before and after the update: the simulation error of the digital twin should be reduced by ≥15% after the parameter update; the prediction RMSE of the spatiotemporal graph convolutional network should be reduced by ≥10%; and the loss function value of the model on the validation set should continue to decrease.

[0087] This step, through the above specific implementation process, establishes a complete "execution-verification-optimization" technical closed loop, which not only achieves precise leveling control, but more importantly, endows the system with the ability to continuously evolve, enabling it to adapt to changes in the field environment and time-varying structural characteristics, and maintain excellent performance over the long term.

[0088] Example 2, one embodiment of the present invention, provides an intelligent scaffolding leveling method based on digital twin and spatiotemporal prediction. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiment.

[0089] First, a test site simulating a high-rise building construction environment was set up, constructing a modular steel pipe scaffold with dimensions of 10m × 8m × 15m (length × width × height). This scaffold structure contained 128 standard nodes, with high-precision tilt sensors (range ±10°, accuracy ±0.05°) and miniature stress sensors (range 0-20kN, accuracy ±0.1%FS) deployed at 32 key nodes. All sensor data was collected via an industrial wireless IoT gateway and transmitted to a central processing unit configured with an Intel Xeon E5 processor and an NVIDIA Tesla V100 computing card. The implementation process strictly followed the steps in claims 1-7: First, based on the BIM design drawings of the scaffolding and the measured material properties of Q235 steel (elastic modulus 206 GPa, Poisson's ratio 0.29, density 7850 kg / m³), a high-fidelity digital twin was constructed in ANSYS software. Based on its structural connection relationships, 128 cup-lock connectors were defined as vertices of the graph structure, and 256 connecting rods were defined as edges, generating a nonlinear grid graph structure with 128 vertices and 256 edges, along with its corresponding 128×128 adjacency matrix. Subsequently, the system entered the real-time monitoring and leveling phase. During a 48-hour experiment, a rigorous preprocessing procedure was performed on the collected raw sensor data through the perception and prediction modules. This included linear interpolation of missing data based on time windows, dynamic smoothing using Kalman filtering, and 3-layer wavelet thresholding denoising using the sym4 wavelet basis, ensuring the quality of the input data. The preprocessed multi-source spatiotemporal data and the graph adjacency matrix were input into a pre-trained spatiotemporal graph convolutional network. This network captures the spatial topological relationships between nodes through its graph convolutional layers and learns the evolution of deformation over time through one-dimensional convolutional layers, ultimately outputting a deformation prediction sequence for 32 key nodes within the next 30 seconds. The decision simulation module injects this prediction sequence as initial conditions into the digital twin, simulating three strategies in parallel in virtual space: single-point priority leveling, regional collaborative leveling, and global progressive leveling. A comprehensive evaluation function (balancing leveling accuracy and maximum von Mises stress of the structure) calculates the comprehensive score for each strategy, ultimately selecting regional collaborative leveling as the optimal strategy. The execution evolution module parses this strategy into specific control commands, which are then sent to the servo electric cylinder actuators corresponding to the 32 nodes via the PROFINET industrial bus. The entire leveling process is monitored in real time. When the displacement deviation of node N-07 exceeds the 2.1mm threshold for three consecutive control cycles, the system automatically pauses and triggers a strategy re-evaluation, demonstrating the system's safety and fault tolerance capabilities. After stabilization, the system collected sensor data in a static state as a validation set to fine-tune the material parameters of the digital twin. The full-cycle time-series data from this stabilization was used as new samples to incrementally learn the spatiotemporal graph convolutional network, completing the first iterative update of the model.For a comprehensive comparison, the traditional threshold-triggered PID balancing method was also tested on the same platform with the same test cases.

[0090] Table 1: Performance Comparison Test Results of Intelligent Balancing Method and Traditional PID Method

[0091] Table 1 clearly demonstrates the significant advantages and innovative effects of the method of this invention compared to the traditional PID leveling method in many aspects. First, in terms of leveling accuracy, this invention achieves superior leveling control in all test cases, with a significantly lower root mean square error (RMSE) after leveling compared to the traditional method. For example, under normal load (Case_01), the RMSE of this invention is 1.58 mm, while that of the traditional method is 3.25 mm, representing an accuracy improvement of over 51%. Under the more complex off-center load condition (Case_02), the advantage is even more pronounced (2.01 mm vs 5.87 mm), which directly reflects the value of the spatiotemporal graph convolutional network for forward-looking deformation prediction: the system can anticipate the deformation trend caused by non-uniform loads, thereby formulating a more targeted collaborative leveling strategy, rather than performing delayed, isolated point corrections after deformation occurs as in the traditional method. Second, in terms of leveling efficiency, this invention, thanks to its prediction and simulation optimization capabilities, achieves shorter leveling times in all comparable cases. In Case_01, the present invention took 127 seconds, while the traditional method took 185 seconds, representing an efficiency improvement of approximately 31%. This is because the strategy simulation stage in the digital twin pre-eliminates inefficient or oscillating leveling schemes, making the actual execution path nearly optimal and avoiding repeated adjustments caused by feedback delay and overshoot in the traditional PID method. Third, regarding structural safety, the present invention explicitly constrains the maximum stress level during the leveling process through an evaluation function, as shown in the "Maximum Stress to Yield Strength Ratio" column in Table 1. Under all conditions, this value is lower than that of the traditional method, particularly evident in the wind load condition in Case_03 (71.3% vs 88.9%). This indicates that the present invention successfully minimizes the impact of the leveling process on structural safety, solving the problem that the traditional method might neglect the risk of instantaneous structural overload in pursuit of leveling speed, demonstrating its novelty in global optimization. Furthermore, Case_05 and Case_06 demonstrate the system's performance after one iterative update. Compared to the previous versions (Case_01, Case_02), both prediction error and leveling accuracy are further improved, and leveling time is slightly reduced. This strongly proves the effectiveness of the closed-loop evolution mechanism. The system can learn from actual leveling results and adaptively optimize its core model, thus possessing a long-term performance maintenance and improvement capability that traditional static control methods cannot achieve.

[0092] In summary, this embodiment verifies through detailed comparative data that the present invention achieves a leap from passive response to active prediction, from local control to global optimization, and from fixed strategy to adaptive learning by deeply integrating digital twins and spatiotemporal prediction, providing a creative solution to resolve the contradiction between accuracy, efficiency, and safety in scaffolding leveling.

[0093] Example 3, one embodiment of the present invention, provides a system for an intelligent scaffolding leveling method based on digital twins and spatiotemporal prediction, including a digital twin modeling module, a perception and prediction module, a decision simulation module, and an execution evolution module.

[0094] The digital twin modeling module is used to construct a digital twin based on the geometric structure and material properties of the scaffolding, and to construct the graph structure of the scaffolding based on the structural connection information. The perception and prediction module is used to collect real-time sensor data of key nodes of the scaffolding and perform preprocessing. The processed data and graph structure are input into the spatiotemporal graph convolutional network to calculate and output the deformation prediction sequence of key nodes. The decision simulation module is used to inject the deformation prediction sequence into the digital twin, simulate various leveling strategies with the sequence as the initial condition, and select the optimal leveling strategy based on the leveling accuracy and structural stability evaluation. The execution evolution module is used to send the optimal leveling strategy to the electromechanical actuators corresponding to each key node for leveling, collect the leveled sensor data, update the digital twin model parameters, and iteratively train the spatiotemporal graph convolutional network.

[0095] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0096] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0097] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0098] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc. It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0099] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent scaffolding leveling based on digital twins and spatiotemporal prediction, characterized in that, include: A digital twin is constructed based on the geometric structure and material properties of the scaffolding, and a graph structure of the scaffolding is constructed based on the structural connection information in the digital twin. Real-time sensor data of key nodes of scaffolding are collected and preprocessed. The preprocessed real-time sensor data and the graph structure are input into a spatiotemporal graph convolutional network. The spatiotemporal features are extracted through the spatiotemporal graph convolutional network to calculate and output the deformation prediction sequence of key nodes. The deformation prediction sequence is injected into the digital twin, and various leveling strategies are simulated using the deformation prediction sequence as the initial condition. The optimal leveling strategy is selected based on the leveling accuracy and structural stability evaluation. The optimal leveling strategy is sent to the electromechanical actuators corresponding to each key node for leveling, and sensor data of the key nodes after leveling are collected to update the model parameters of the digital twin and iteratively train the spatiotemporal graph convolutional network.

2. The method according to claim 1, characterized in that, The construction of the digital twin and the graph structure of the scaffolding include: Based on the geometric data and material properties obtained from the building information model or design drawings, an initial digital twin is constructed; the material properties include at least the elastic modulus, Poisson's ratio, and material density; based on the structural connection information in the digital twin, the pipe connectors or load-bearing nodes of the scaffolding are defined as vertices of the graph structure, and the rods connecting adjacent vertices are defined as edges of the graph structure; according to the connection relationship between the vertices and edges, an adjacency matrix of the graph structure is established, wherein when there are directly connected rods between two vertices, the element at the corresponding position in the adjacency matrix has a value of 1, otherwise it has a value of 0.

3. The method according to claim 2, characterized in that, The preprocessing of the collected sensor data includes: collecting real-time data through multiple types of sensors deployed at key nodes; performing integrity verification and missing value imputation on the collected raw data; performing Kalman filtering dynamic smoothing and wavelet transform threshold denoising on the imputed data in sequence; and performing timestamp alignment and spatial registration on the denoised multi-source sensor data.

4. The method according to claim 3, characterized in that, The step of extracting spatiotemporal features through a spatiotemporal graph convolutional network to calculate and output the deformation prediction sequence of key nodes includes: inputting preprocessed sensor data and the adjacency matrix of the graph structure into the spatiotemporal graph convolutional network; extracting the spatial topological features of the scaffolding structure through the graph convolutional layers of the network; extracting the temporal series features of the sensor data through the one-dimensional convolutional layers of the network; fusing spatial and temporal features and performing feature enhancement through an attention mechanism; and outputting the deformation prediction sequence of key nodes for multiple future time steps.

5. The method according to claim 4, characterized in that, The simulation of multiple leveling strategies and selection of the optimal leveling strategy includes: loading the deformation prediction sequence as an initial condition into the digital twin to generate multiple leveling strategies; performing dynamic simulation on each leveling strategy to calculate the leveling accuracy index and structural stability index under each strategy; comprehensively considering the leveling accuracy and structural stability through an evaluation function to calculate the comprehensive score of each strategy; and selecting the leveling strategy with the highest comprehensive score as the optimal leveling strategy.

6. The method according to claim 5, characterized in that, The step of sending the optimal leveling strategy to the electromechanical actuators for leveling includes: parsing the optimal leveling strategy into independent control commands for each key node; distributing the control commands to the corresponding electromechanical actuators through an industrial bus network; each electromechanical actuator making precise displacement adjustments according to the received control commands; monitoring the leveling process in real time, pausing the leveling and re-evaluating when an abnormal deviation is detected.

7. The method according to claim 6, characterized in that, The process of updating the digital twin model parameters and iteratively training the spatiotemporal graph convolutional network includes: collecting sensor data from key nodes as a verification dataset after leveling; comparing and analyzing the verification data with the simulation prediction data of the digital twin; updating the model parameters of the digital twin based on the comparison results; adding the time-series data of the leveling process to the training set and iteratively training the spatiotemporal graph convolutional network.

8. A scaffolding system based on digital twins and spatiotemporal prediction as described in any one of claims 1 to 7. The system of intelligent leveling method is characterized by: It includes a digital twin modeling module, a perception and prediction module, a decision simulation module, and an execution evolution module; The digital twin modeling module is used to construct a digital twin based on the geometric structure and material properties of the scaffolding, and to construct the graphical structure of the scaffolding based on the structural connection information. The perception and prediction module is used to collect real-time sensor data of key nodes of the scaffolding and preprocess it. The processed data and graph structure are input into the spatiotemporal graph convolutional network to calculate and output the deformation prediction sequence of the key nodes. The decision simulation module is used to inject the deformation prediction sequence into the digital twin, simulate various leveling strategies with the sequence as the initial condition, and select the optimal leveling strategy based on the leveling accuracy and structural stability evaluation. The execution evolution module is used to send the optimal leveling strategy to the electromechanical actuators corresponding to each key node for leveling, collect the leveled sensor data, update the digital twin model parameters, and iteratively train the spatiotemporal graph convolutional network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent scaffolding leveling method based on digital twins and spatiotemporal prediction as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent scaffolding leveling method based on digital twins and spatiotemporal prediction as described in any one of claims 1 to 7.

Images