High-speed railway continuous beam closure section bridge building machine posture control method based on digital twinning

Abstract

The application discloses a high-speed railway continuous beam closure section bridge builder posture control method based on digital twinning, belongs to the technical field of general control and regulation, and comprises the following steps: acquiring beam end joint coordinates, fulcrum coordinates, posture angles and real-time meteorological data, constructing a digital twinning model of the bridge builder and inversely time-sequentially generating a target pose sequence of each construction node; calculating a mechanical displacement vector and a meteorological response displacement vector based on the current construction node; constructing a joint input matrix; performing diffusion sampling and reverse denoising to generate a hydraulic oil cylinder displacement sequence; forwardly calculating a virtual running track; and outputting a servo drive electric signal according to a variance threshold value. The application adopts a control mode combining digital twinning, time sequence prediction and diffusion generation, and realizes posture control of the closure section bridge builder.

Description

Technical Field

[0001] This invention relates to the field of general control and regulation technology, and in particular to a method for attitude control of a bridge-building machine for the closure section of a continuous beam in a high-speed railway based on digital twins. Background Technology

[0002] During the construction of the closure section of the continuous beam of the high-speed railway, the bridge-building machine needs to complete the positioning, attitude adjustment and closure of the beam segment at different construction nodes. The attitude of the bridge-building machine during the construction process will affect the position of the beam end joint and the spatial position of the closure section. Therefore, attitude control during the construction stage is an important part of the construction of the closure section of the continuous beam.

[0003] In the prior art, Chinese invention patent with announcement number CN116430721A discloses a method for controlling the alignment of a cantilevered structure based on elevation prediction. This method selects control points during the assembly of the cantilevered beam, obtains the actual elevation values ​​of the segments, constructs predicted elevation values ​​of the control points, and guides the next stage of assembly construction based on the updated predicted elevation values ​​of the control points.

[0004] Regarding the aforementioned technologies, existing construction control methods mostly focus on predicting and adjusting segment elevations, alignments, or single-stage construction parameters. They give little consideration to the continuous control relationship between the closure endpoint state and the current construction node during construction. Furthermore, during the attitude adjustment of construction nodes, the degree of joint processing between structural stress state, construction weather changes, and bridge-building machine attitude control is limited, making it difficult to adapt to the application scenario of multi-node continuous control during the construction of continuous beam closure segments. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a posture control method for a bridge-building machine in the closure section of a high-speed railway continuous beam based on digital twins. The method employs a control approach that combines digital twins, time-series prediction, and diffusion generation to achieve posture control of the bridge-building machine in the closure section.

[0006] The above objectives can be achieved through the following approach:

[0007] A method for attitude control of a bridge-building machine in the closure section of a continuous beam in high-speed railway based on digital twins includes acquiring the coordinates of the beam end joints, support coordinates, attitude angles, and real-time meteorological data of the continuous beam closure section; constructing a digital twin model of the bridge-building machine; performing inverse time-series solving based on the joint coordinates and attitude angles at the final moment; generating target pose sequences for each construction node; the digital twin model of the bridge-building machine includes a rigid body dynamics module and a time-series prediction module; collecting sensor force data of the bridge-building machine based on the current construction node and inputting it in parallel to the rigid body dynamics module and the time-series prediction module; outputting mechanical displacement vectors and meteorological response displacement vectors for the current construction node; extracting the target coordinate vector of the current construction node using the target pose sequence; and calculating the mechanical displacement vector and the meteorological response displacement vector. The Euclidean distance and included angle value of the meteorological response displacement vector are concatenated with the target coordinate vector, the Euclidean distance, and the included angle value to generate a joint input matrix for the current construction node. Based on the joint input matrix, diffusion sampling and reverse denoising are performed to output the displacement sequence of each hydraulic cylinder of the bridge-building machine. The displacement sequence is extracted and input into the digital twin model of the bridge-building machine to perform forward calculation of the virtual running trajectory from the current construction node to the endpoint, and output the joint coordinate difference and attitude angle difference at the endpoint. The fluctuation variance of the real-time meteorological data is calculated to generate a variance threshold. The displacement sequence in which both the joint coordinate difference and the attitude angle difference are less than the variance threshold is extracted, and the displacement sequence is converted into a servo drive electrical signal and output to the hydraulic cylinder.

[0008] Optionally, generating the target pose sequence for each construction node includes: extracting the coordinates of the beam end joints and the coordinates of the support points, constructing the spatial structure topology, assembling the physical parameters of the structural components, and generating a rigid body dynamics module; coupling the time-series prediction module with the rigid body dynamics module to construct a digital twin model of the bridge-building machine; setting the joint coordinates and attitude angles at the end point as inverse time-series boundary constraints, and progressively performing inverse dynamics solutions along the time axis from the end point to the current construction node, outputting the target pose sequence.

[0009] Optionally, the time-series prediction module includes: acquiring historical meteorological monitoring data and historical structural displacement data of the same type of bridge construction phase, performing timestamp alignment and noise cleaning to obtain a historical training dataset; using the historical meteorological monitoring data as network input and the historical structural displacement data as network output, establishing and training a recurrent neural network using the historical training dataset, and updating the network weight parameters to obtain the time-series prediction module.

[0010] Optionally, the step of outputting the mechanical displacement vector and the meteorological response displacement vector for the current construction node includes: parsing the force data of the bridge-building machine sensor into a node load tensor, inputting it into the rigid body dynamics module, and mapping and outputting the mechanical displacement vector; extracting the force data of the bridge-building machine sensor and the real-time meteorological data, constructing multi-dimensional time-series input features, calling the time-series prediction module to perform forward propagation inference calculation, and outputting the meteorological response displacement vector.

[0011] Optionally, generating a joint input matrix for the current construction node includes: matching the timestamp identifier of the current construction node and extracting the target coordinate vector from the target pose sequence; performing vector subtraction and L2 norm calculation on the mechanical displacement vector and the meteorological response displacement vector to output the Euclidean distance, and calculating the quotient of the dot product and the product of the modulus, generating an angle value through an inverse cosine function mapping; converting the Euclidean distance and the angle value into feature scalars, and performing tensor concatenation operations along the feature dimension of the target coordinate vector to generate a joint input matrix.

[0012] Optionally, the output displacement sequence of each hydraulic cylinder of the bridge-building machine includes: performing dimensionality reduction feature extraction on the joint input matrix to generate latent space conditional coding, and extracting the random normal distribution noise features of each hydraulic cylinder, combining them to form a hybrid latent space state; guided by the latent space conditional coding, performing inverse feature denoising transformation on the hybrid latent space state, and performing dimensionality-upgrading reconstruction calculation to output the displacement sequence of each hydraulic cylinder of the bridge-building machine.

[0013] Optionally, the output of the joint coordinate difference and attitude angle difference at the endpoint includes: converting the displacement sequence into driving boundary conditions and inputting them into the digital twin model of the bridge-building machine; performing stepwise forward numerical integration calculations along the time axis from the current construction node to the endpoint to generate a virtual running trajectory; extracting the predicted joint coordinates and predicted attitude angles of the virtual running trajectory at the endpoint, and performing numerical subtraction operations with the joint coordinates and attitude angles at the endpoint respectively, and outputting the joint coordinate difference and attitude angle difference at the endpoint.

[0014] Optionally, converting the displacement sequence into a servo drive electrical signal for output to the hydraulic cylinder includes: extracting the observation sequence of the real-time meteorological data at the current construction node, calculating the fluctuation variance, and mapping the fluctuation variance to generate a variance threshold; comparing the joint coordinate difference, the attitude angle difference, and the variance threshold to extract the displacement sequence, and encoding the displacement sequence into a pulse width modulation signal through digital-to-analog conversion, which is then output to the hydraulic cylinder as a servo drive electrical signal.

[0015] Optionally, the step of mapping the fluctuation variance to generate a variance threshold includes: truncating a time window along the time axis, extracting meteorological data within the time window as an observation sequence, and solving for the second-order central moment to obtain the fluctuation variance; extracting the maximum allowable deformation of the rigid body dynamics module, and using the maximum allowable deformation of the structure as the upper limit of the value, performing extreme value normalization scaling on the fluctuation variance to map and generate a variance threshold.

[0016] Based on the same inventive concept, this invention also provides a digital twin-based attitude control system for a bridge-building machine in the closure section of a continuous beam for high-speed railway. The system includes: a data acquisition and inverse kinematics planning module, used to acquire the coordinates of the beam end joints, support coordinates, attitude angles, and real-time meteorological data of the continuous beam closure section, construct a digital twin model of the bridge-building machine, and perform inverse time-series solving based on the joint coordinates and attitude angles at the final moment to generate target pose sequences for each construction node. The digital twin model of the bridge-building machine includes a rigid body dynamics module and a time-series prediction module; a dual-parallel solution module, used to acquire sensor force data of the bridge-building machine based on the current construction node, and input it in parallel to the rigid body dynamics module and the time-series prediction module, outputting mechanical displacement vectors and meteorological response displacement vectors for the current construction node; and a deviation feature fusion module, used to extract the target coordinate vector of the current construction node using the target pose sequence. The system calculates the Euclidean distance and angle between the mechanical displacement vector and the meteorological response displacement vector, concatenates the target coordinate vector, the Euclidean distance, and the angle, and generates a joint input matrix for the current construction node. A diffusion sampling generation module performs diffusion sampling and reverse denoising based on the joint input matrix, outputting the displacement sequence of each hydraulic cylinder of the bridge-building machine. A forward simulation verification module extracts the displacement sequence and inputs it into the digital twin model of the bridge-building machine, forward-calculating the virtual trajectory from the current construction node to the endpoint, and outputting the joint coordinate difference and attitude angle difference at the endpoint. A meteorological screening and drive output module calculates the fluctuation variance of the real-time meteorological data, generates a variance threshold, extracts displacement sequences where both the joint coordinate difference and the attitude angle difference are less than the variance threshold, and converts the displacement sequence into a servo drive electrical signal output to the hydraulic cylinder.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] 1. A digital twin model of the bridge-building machine is constructed based on the coordinates of the beam end joints, the coordinates of the support points, the attitude angles, and real-time meteorological data. The target pose sequence of each construction node is generated in reverse time sequence from the joint coordinates and attitude angles at the end point. The end point state of the closure is moved forward to the control process of each construction node, reducing the problem of discontinuous control between nodes caused by adjusting only based on the current state.

[0019] 2. Mechanical displacement vectors and meteorological response displacement vectors are generated through the rigid body dynamics module and the time series prediction module. A joint input matrix is ​​constructed by combining the target pose vector, and then the hydraulic cylinder displacement sequence is generated. This allows structural force information, construction node pose information and meteorological change information to participate in the solution within the same control process, reducing the problem of insufficient response of a single model to complex construction disturbances.

[0020] 3. The hydraulic cylinder displacement sequence is input into the digital twin model for forward calculation. The difference in joint coordinates and attitude angle at the end point is extracted. The displacement sequence is then filtered by combining the variance of real-time meteorological data and output as a servo drive electrical signal. This forms a closed-loop control chain from displacement generation and trajectory verification to drive output, reducing the need for manual verification of the end point state and repeated adjustments on site.

[0021] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the attitude control method for a bridge-building machine in the closure section of a high-speed railway continuous beam based on digital twins, according to an embodiment of the present invention.

[0024] Figure 2 This is a data distribution evaluation diagram of the pose deviation of construction nodes under different control strategies in embodiments of the present invention.

[0025] Figure 3 This is a joint probability kernel density distribution map of meteorological fluctuations and spatial deviations in an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the attitude control system of the bridge-building machine for the closure section of a high-speed railway continuous beam based on digital twin, according to an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Reference Figure 1 One embodiment of the present invention proposes a posture control method for a bridge-building machine in the closure section of a high-speed railway continuous beam based on digital twins. The method adopts a control approach that combines digital twins, time-series prediction, and diffusion generation to achieve posture control of the bridge-building machine in the closure section.

[0029] The method described in this embodiment specifically includes:

[0030] The coordinates of the beam end joints, support coordinates, attitude angles and real-time meteorological data of the continuous beam closure segment are obtained, and a digital twin model of the bridge-building machine is constructed. Based on the joint coordinates and attitude angles at the end time, a reverse time sequence solution is performed to generate a target pose sequence for each construction node. The digital twin model of the bridge-building machine includes a rigid body dynamics module and a time sequence prediction module.

[0031] Based on the force data collected from the bridge-building machine's sensors at the current construction node, the data is input in parallel to the rigid body dynamics module and the time-series prediction module, and the mechanical displacement vector and meteorological response displacement vector are output to the current construction node respectively.

[0032] The target coordinate vector of the current construction node is extracted using the target pose sequence. The Euclidean distance and the included angle between the mechanical displacement vector and the meteorological response displacement vector are calculated. The target coordinate vector, the Euclidean distance, and the included angle are concatenated to generate a joint input matrix for the current construction node.

[0033] Based on the joint input matrix, diffusion sampling and inverse denoising are performed to output the displacement sequence of each hydraulic cylinder of the bridge-building machine;

[0034] The displacement sequence is extracted and input into the digital twin model of the bridge-building machine. The virtual running trajectory from the current construction node to the end time is calculated in a forward direction, and the joint coordinate difference and attitude angle difference at the end time are output.

[0035] The fluctuation variance of the real-time meteorological data is calculated, a variance threshold is generated, and displacement sequences whose joint coordinate difference and attitude angle difference are both less than the variance threshold are extracted. The displacement sequences are then converted into servo drive electrical signals and output to the hydraulic cylinder.

[0036] Optionally, generating the target pose sequence for each construction node includes:

[0037] Extract the coordinates of the beam end joints and the coordinates of the support points, construct the spatial structure topology, assemble the physical parameters of the structural components, and generate a rigid body dynamics module;

[0038] The design drawings and survey data of the continuous beam closure segment are obtained, from which the coordinates of the beam end joints and the coordinates of each temporary or permanent support are extracted as geometric references. These three-dimensional spatial coordinate data constitute the basic framework of the model. Based on the coordinate points, the spatial topology of the overall structure is constructed, clarifying the spatial connection relationships and motion degree-of-freedom constraints between each steel structural component. For example, the main beam is simplified into a continuous load-bearing beam component, and the hydraulic cylinder is defined as a variable-length mechanical drive component. Subsequently, the physical parameters of the structural components are assembled for this spatial structural topology. The physical parameters of the structural components include the modulus, Poisson's ratio, and density obtained through mechanical tests of building materials at the factory, as well as the bending moment of inertia and torsional stiffness calculated based on the cross-sectional shape marked on the drawings. By integrating the above topological associations and physical characteristic parameters, a rigid body dynamics module that can reflect the real mechanical transmission characteristics and deformation response characteristics of the physical bridge-building machine is generated.

[0039] The timing prediction module is coupled with the rigid body dynamics module to construct a digital twin model of the bridge-building machine;

[0040] The execution time-series prediction module and the rigid body dynamics module are deeply coupled at the data flow and interaction level. The coupling mechanism is specifically manifested in the bidirectional dynamic transfer of boundary conditions and equivalent forces. The time-series prediction module, based on meteorological data, derives environmental disturbance displacements, converts them into equivalent environmental compensation nodal loads, and applies them to the corresponding structural nodes of the rigid body dynamics module. Simultaneously, the mechanical strain state data output by the rigid body dynamics module is fed back to the time-series prediction module as auxiliary correction features. Through the fusion of physical and data field information across modules, a digital twin model of the bridge-building machine is ultimately constructed, capable of synchronously simulating mechanical stress deformation and dynamic environmental and meteorological disturbances.

[0041] For example, when facing extreme construction conditions of continuous strong winds and large temperature differences, the time-series prediction module first predicts the beam contraction displacement caused by a sudden drop in temperature and the lateral displacement caused by crosswinds within the next two hours. Then, the temperature-induced beam contraction displacement is converted into an equivalent axial tensile force acting on the main truss nodes, and the lateral displacement caused by crosswinds is converted into a uniformly distributed lateral thrust. After receiving the equivalent axial tensile force and the uniformly distributed lateral thrust, the rigid body dynamics module, combined with the bridge-building machine's own gravity load matrix, recalculates the overall three-dimensional deformation distribution. This data coupling method allows the bridge-building machine's digital twin model to not only accurately reflect the static deflection caused by the structure's self-weight but also to superimpose the nonlinear dynamic drift caused by sudden weather changes.

[0042] Set the joint coordinates and attitude angles at the end point as inverse time-series boundary constraints, and perform inverse dynamics solution step by step along the time axis from the end point to the current construction node to output the target pose sequence.

[0043] The ideal joint 3D coordinates and theoretical spatial attitude angles at the closure endpoint, as required by the construction drawings and specifications, are extracted and set as the absolute inverse temporal boundary constraints for the entire evolution process. Using this endpoint state as the starting point, inverse dynamics solutions are performed progressively towards the current actual construction node along the time axis, with fixed time steps. In each inverse deduction step, the pose state of the preceding construction node required to offset the self-weight deflection and meteorological disturbances within that time step, thus enabling the bridge-building machine to reach the pose of the subsequent construction node, is calculated. This inverse dynamics solution process follows a nonlinear state inverse deduction formula:

[0044] ,

[0045] Among them, in order to improve the calculation accuracy, the symbols The internal calculation logic involves multiplying the mass matrix by the environmental response acceleration matrix and then multiplying by a safety amplification factor. The calculation expression is as follows: ,symbol This represents the target pose state vector of the previously constructed node to be solved, which is composed of the target's three-dimensional translational coordinate components and spatial Euler angle components. (Symbol) This represents the pose state vector of a later construction node whose pose data is known, even though the node is later in time than the earlier one. (Symbol) This represents a fixed time step between two adjacent construction nodes. (Symbol) This represents a nonlinear inverse solver operator that encapsulates the physical characteristics of a digital twin model of a bridge-building machine. (Symbol) This represents the matrix of all generalized environmental forces acting on the bridge-building machine during the time interval between two construction nodes. (Symbols) This represents the structural equivalent mass matrix statically derived from the rigid body dynamics module. (Symbol) This represents the environmental response acceleration matrix resolved by the time-series prediction module. (Number) This represents the safety amplification factor obtained through wind tunnel testing and calibration. The physical meaning of the nonlinear inverse solver in the formula lies in the fact that, by inputting the desired future target state and the external force disturbance matrix during this period, iteratively calculates the initial deformation compensation attitude that must be present at the current moment. By iteratively calculating the state vector along the time axis, a set containing the precise three-dimensional coordinates and attitude angles of all intermediate construction nodes is obtained, thereby generating and outputting the target pose sequence.

[0046] For example, assume that the construction drawings specify the ideal joint coordinates at the closure endpoint as the absolute zero origin and the attitude angle as completely horizontal. A fixed time step of fifteen minutes is set, and a reverse calculation is performed. In the penultimate fifteen-minute calculation cycle, the pose state vector of the subsequent construction node is known to be the zero-error closure endpoint state. At this time, the timing prediction module predicts that downward gravitational deformation and lateral wind load displacement will occur within this fifteen-minute period. Using the nonlinear state inverse calculation formula, it is calculated that the target pose state vector of the preceding construction node must include a two-millimeter upward pre-camber and a one-millimeter lateral inverse offset. This process is repeated, continuously performing reverse calculations towards the current actual construction time node, ultimately generating a complete target pose sequence to guide the adjustment actions of the bridge-building machine's hydraulic cylinders.

[0047] Optionally, the time series prediction module includes:

[0048] Historical meteorological monitoring data and historical structural displacement data of bridges of the same type during the construction phase were obtained, and timestamp alignment and noise removal were performed to obtain a historical training dataset.

[0049] First, historical meteorological monitoring data and historical structural displacement data from the construction phase of similar bridges are retrieved from the engineering database. Historical meteorological monitoring data typically includes physical parameters such as wind speed, wind direction, temperature, and solar radiation intensity in the surrounding environment of the bridge site. Historical structural displacement data includes the three-dimensional coordinate changes of bridge control nodes collected by GNSS global navigation satellite systems or high-precision total stations. Due to differences in the data acquisition hardware and communication protocols of various sensors, the sampling frequencies of the data are inconsistent, necessitating timestamp alignment. This operation uses a cubic spline interpolation algorithm to resample time-series data of different frequencies onto a unified time resolution reference plane. Subsequently, noise cleaning is performed, using a Kalman filter algorithm to filter out random high-frequency noise generated by sensors under harsh operating conditions and baseline drift errors caused by temperature influences. After the above interpolation alignment and filtering cleaning processes, redundancy and abnormal interference are eliminated, resulting in a historical training dataset with perfectly matched spatiotemporal characteristics and numerically smoothed data.

[0050] For example, the original acquisition frequency of the anemometer at the engineering site is 10 Hz, while the displacement acquisition frequency of the high-precision total station used to monitor deformation is 1 Hz. By running a cubic spline interpolation algorithm, the historical meteorological monitoring data collected by the anemometer and the historical structural displacement data collected by the high-precision total station are uniformly aligned to a standard timestamp once per second. When the GNSS global navigation satellite system signal is temporarily blocked by large machinery, causing abrupt spikes in the historical structural displacement data, the Kalman filter algorithm is used to remove the abnormal spike data that exceeds the normal physical deformation threshold, and the abnormal point is replaced with a smoothed state inference value. Finally, a high-quality historical training dataset that can be used for efficient learning of the subsequent network is generated.

[0051] Using the historical meteorological monitoring data as network input and the historical structural displacement data as network output, a recurrent neural network is established and trained using the historical training dataset, and the network weight parameters are updated to obtain the time series prediction module.

[0052] A recurrent neural network consisting of multiple long short-term memory (LSTM) layers is constructed in memory. Historical meteorological monitoring data is used as the network input, and synchronous historical structural displacement data is used as the network output annotation. The supervised learning training process of the network is driven by the historical training dataset. During the forward propagation stage of the recurrent neural network, meteorological feature information is transformed into the internal hidden layer states and ultimately mapped to the predicted displacement response. The core computational logic of this hidden layer mapping process is described by the following state transition formula:

[0053] ,

[0054] Among them, symbols The hidden state tensor representing the current time step physically represents the cumulative influence of the current weather environment on the bridge's attitude, extracted and memorized by the recurrent neural network. (Symbol) This represents the historical meteorological monitoring data tensor input at the current time step, with its specific values ​​derived from the historical training dataset. (Symbol) This represents the hidden state tensor of the previous time step immediately preceding the current time step. (Symbol) Represents the weight parameter matrix from the input layer to the hidden layer, denoted by [symbol]. This represents the weight parameter matrix for state transitions from one hidden layer to another. (Symbol) Represents the hidden layer bias vector. (Symbol) The nonlinear activation function is represented by a hyperbolic tangent function to avoid gradient explosion. In the initial stage of model training, all weight parameter matrices and bias vectors are randomly initialized using a standard normal distribution. During training iterations, the mean square error between the predicted displacement and the actual historical structural displacement data is calculated as the model loss function, and an adaptive moment estimation optimization algorithm is used to perform backpropagation calculations. The values ​​of the aforementioned weight parameter matrices and bias vectors are continuously updated based on the calculated gradients. When the mean square error converges to the set lower error limit or reaches the maximum number of iterations, training stops and all updated network weight parameters are fixed, thus obtaining a time-series prediction module with meteorological bias prediction capabilities.

[0055] For example, historical meteorological monitoring data is structured into an input tensor containing five feature dimensions, specifically temperature, relative humidity, wind speed, wind direction, and solar radiation intensity. Correspondingly, historical structural displacement data is structured into an output tensor containing three dimensions: longitudinal offset, lateral offset, and vertical deflection. The initial learning rate of the adaptive moment estimation optimization algorithm is set to 0.001, the specific value of which is the optimal convergence hyperparameter determined through multiple grid search comparisons on the validation set data. The historical training dataset containing 100,000 independent samples is input into the recurrent neural network in batches. After two hundred rounds of backpropagation iterations, the model's mean square error decreases from the initial 5.3 and stabilizes at 0.0002. At this point, weight updates are stopped, and the currently converged network structure and parameter array are extracted, formally encapsulated as a time-series prediction module that can be deployed online in the bridge-building machine control system.

[0056] Optionally, the step of outputting the mechanical displacement vector and the meteorological response displacement vector for the current construction node includes:

[0057] The force data of the bridge-building machine sensor is parsed into nodal load tensors, input to the rigid body dynamics module, and mapped and output as a mechanical displacement vector.

[0058] The system acquires force data from force sensors installed at the bottom of the hydraulic cylinders and support legs of the bridge-building machine. Based on the geometric coordinate mapping of the sensors in the digital twin model of the bridge-building machine, the original electrical signals or pressure values ​​are converted into concentrated or distributed force matrices acting on the corresponding mesh nodes. Then, the force information of all nodes is assembled and analyzed into nodal load tensors. These nodal load tensors are then input into the rigid body dynamics module. The rigid body dynamics module internally solves the static equilibrium matrix equations and maps out the mechanical displacement vectors. This mapping calculation process follows the following elasticity equilibrium formula:

[0059] ,

[0060] Among them, symbols The mechanical displacement vector representing the output of the mapping represents the three-dimensional spatial positional change of each key node of the bridge-building machine under the action of purely mechanical external forces. (Symbol) Represents the global flexibility matrix, which is the inverse of the global stiffness matrix obtained by the rigid body dynamics module during initialization based on physical parameters such as the moment of inertia of the assembled sections. It is used to characterize the inherent physical properties of the structure in resisting mechanical deformation. (Symbol) This represents the nodal load tensor generated by the preceding steps. By performing this matrix multiplication operation, the deformation component dominated by physical mechanical loads is directly quantized.

[0061] For example, during the pouring of a main girder, the bridge-building machine's sensor data on the left front outrigger was collected in real time, showing a compression of 30 MPa. Based on the calibration conversion factor, this pressure was converted into a concentrated downward vertical load of 500 kN, and this was filled into the tensor element at the corresponding three-dimensional coordinates of the left front outrigger, thus completing the construction of the nodal load tensor. Subsequently, the rigid body dynamics module was invoked, and a matrix multiplication operation was performed between the global flexibility matrix containing the main girder's stiffness properties and the nodal load tensor. The calculation showed that this stress state would cause a downward mechanical deflection of 3 mm at the cantilever end, and this result was encapsulated as a mechanical displacement vector for output.

[0062] Extract the force data from the bridge-building machine sensor and the real-time meteorological data, construct multi-dimensional time-series input features, call the time-series prediction module to perform forward propagation inference calculation, and output the meteorological response displacement vector.

[0063] While acquiring mechanical deformation characteristics, real-time meteorological data containing wind speed and temperature change sequences, as well as bridge-building machine sensor force data characterizing the structural stress status, are simultaneously extracted. The two types of heterogeneous data are normalized in terms of dimensions and then concatenated along the time step axis to construct a multi-dimensional time-series input feature that integrates environmental disturbances and physical states. Next, the time-series prediction module, which is in a frozen weight state, is invoked to feed this multi-dimensional time-series input feature into the network input layer for forward propagation inference calculation. This forward propagation inference calculation process does not perform error backpropagation; it only utilizes the network parameter matrix fixed during the training phase for linear transformation and activation mapping, directly outputting the meteorological response displacement vector. The logic of this forward propagation inference calculation is expressed by the following formula:

[0064] ,

[0065] Among them, symbols This represents the meteorological response displacement vector output by the model inference, used to accurately quantify environmental variables such as thermal expansion and contraction or wind-induced displacement caused by meteorological elements. (Symbol) Representing a nonlinear activation function, a hyperbolic tangent function is used here to improve the smoothness and computational efficiency of the forward inference mapping. (Symbol) Represents the forward mapping weight matrix from the input layer to the hidden layer, denoted by [symbol]. Representing the bias vector, both of these values ​​originate from the final model parameters fixed after iterative updates using the historical training dataset in the previous training phase. (Symbol) This represents the multidimensional temporal input features constructed in the current step. Through this matrix multiplication operation and nonlinear activation mapping, the structural additional displacements driven by highly nonlinear environmental factors are quickly derived.

[0066] For example, the construction area where the bridge-building machine is located encountered a sudden strong crosswind and a sharp drop in temperature. Real-time meteorological data showing a wind speed increase to 15 meters per second and a temperature drop of 8 degrees Celsius within the last ten minutes are extracted. Simultaneously, force data from the bridge-building machine's sensors, which maintain stable pressure in the hydraulic circuit, are extracted. These two types of heterogeneous data are aligned according to time series and subjected to minimum-maximum normalization to construct a 10x5 numerical matrix as a multi-dimensional time-series input feature. This numerical matrix is ​​input into the time-series prediction module, multiplied by the module's pre-stored forward mapping weight matrix, and a bias vector is added. After activation by the hyperbolic tangent function, the model outputs a 2mm lateral offset and a 1mm longitudinal contraction. These pure environmental deformation offsets caused by drastic weather changes are packaged and output as a meteorological response displacement vector.

[0067] Optionally, generating the joint input matrix for the current construction node includes:

[0068] Match the timestamp of the current construction node and extract the target coordinate vector from the target pose sequence;

[0069] The real-time clock signal of the bridge-building machine's control motherboard is acquired and converted into a timestamp identifier for the current construction node in a standard time format. Subsequently, the target pose sequence, containing a full-cycle time series, generated by the preceding inverse dynamics solution steps, is retrieved from the memory database. By executing a timestamp comparison command, the data row absolutely corresponding to the current construction node is precisely matched, and a three-dimensional spatial coordinate feature, i.e., the target coordinate vector, is extracted from it. The physical significance of this target coordinate vector lies in defining the theoretical three-dimensional absolute position of the bridge-building machine's closure joint point at the current moment under absolutely ideal working conditions without any external interference, providing a basic targeting benchmark for deviation correction.

[0070] Perform vector subtraction and L2 norm calculation on the mechanical displacement vector and the meteorological response displacement vector to output the Euclidean distance, and calculate the quotient of the dot product and the product of the modulus, and generate the included angle value through the inverse cosine function mapping.

[0071] Through rigorous mathematical and geometric analytical calculations, the spatial conflict divergence between two independent displacement effects caused by mechanical load and external meteorological dynamic disturbances is quantified. First, a vector subtraction command is invoked to calculate the difference vector between the mechanical displacement vector and the meteorological response displacement vector. Then, the square root of the sum of squares of all elements of this difference vector is taken to complete the L2 norm calculation and output the Euclidean distance. This Euclidean distance characterizes the absolute divergence error magnitude of the two displacement components in three-dimensional physical space. Next, the dot product of the two displacement vectors is calculated, along with their modulus product. Dividing the dot product by the modulus product yields the spatial cosine ratio. Finally, the inverse cosine function is invoked to perform a nonlinear mapping operation on this spatial cosine ratio, generating an angle value. This angle value characterizes the synergistic or antagonistic relationship between the two displacement effects in spatial direction. The mathematical formula for the above multi-source displacement deviation quantification process is as follows:

[0072] ,

[0073] ,

[0074] Among them, symbols Represents the Euclidean distance, which physically represents the absolute value of the distance deviation between purely mechanical displacement and purely meteorological response displacement in three-dimensional space. (Symbol) Represents the mechanical displacement vector output by the rigid body dynamics module. Symbol This represents the meteorological response displacement vector output by the forward propagation inference of the time-series prediction module. (Symbol) This represents the standard mathematical operator for finding the L2 norm of an interior vector. (Symbol) This represents the included angle value, ranging from zero to pi in radians, and is used to indicate the degree of directional deviation of conflicting commands. (Symbol) This represents the inverse cosine mathematical mapping function. In the above formula, and The values ​​are all directly derived from the output of the preceding dual-path parallel solution step. By performing the above spatial geometric operations, the coupled conflict state of the multi-source physical fields is transformed into specific numerical indicators that can be recognized by the deep learning model.

[0075] For example, the output of the parallel calculation module yields a mechanical displacement vector of three millimeters vertically downwards, represented by spatial coordinates of 0, 0, and -3. Simultaneously, the meteorological response displacement vector caused by transient gusts is obtained as four millimeters horizontally, also represented by spatial coordinates of 0, 4, and 0. Substituting these values ​​into the Euclidean distance formula, the square root of the sum of the squares of three and four yields five millimeters, which is the output Euclidean distance. Furthermore, since one displacement is vertically downwards and the other horizontally, their dot product is zero. Substituting this into the inverse cosine formula, the calculated angle is half the radius of pi in radians. These two indicators depict a completely orthogonal antagonistic physical state between the current stress deformation and wind load deformation.

[0076] The Euclidean distance and the included angle value are converted into feature scalars, and tensor concatenation operations are performed along the feature dimensions of the target coordinate vector to generate a joint input matrix.

[0077] The calculated Euclidean distance and angle values ​​are converted from basic floating-point format to feature scalars compatible with deep learning tensor frameworks. Then, using the extracted target coordinate vector as the skeleton, tensor concatenation is performed on the two feature scalars representing spatial conflict states along the column direction of the target coordinate vector or the feature extension dimension. After tensor dimension concatenation, the target coordinate vector, originally containing only three-dimensional spatial location information, is successfully expanded into a multimodal fusion state matrix containing spatial constraint targets, deviation magnitudes, and conflict angles—that is, a joint input matrix. This joint input matrix achieves high-density data encoding of the control endpoint geometric constraints and the current multiphysics game conflict state. It serves as the standardized condition-guided input feature for the subsequent diffusion sampling denoising model, ensuring that the large model can perform generative computations based on complete physical and environmental conflict boundaries.

[0078] Optionally, the displacement sequence of each hydraulic cylinder of the output bridge-building machine includes:

[0079] Dimensionality reduction feature extraction is performed on the joint input matrix to generate latent space conditional coding, and random normal distribution noise features of each hydraulic cylinder are extracted and combined to form a hybrid latent space state.

[0080] After acquiring high-density data encoding containing conflict states of multiple physical fields, dimensionality reduction feature extraction is performed on the joint input matrix. The physical purpose of this dimensionality reduction feature extraction is to remove redundant dimensions from the high-dimensional input matrix, compressing and mapping mechanical deformation, meteorological response, and spatial target into a low-dimensional latent space feature vector with highly nonlinear expressive power, i.e., generating latent space conditional encoding. Subsequently, the number of active hydraulic cylinders and the expected control time step in the current bridge-building machine control system are analyzed to determine the tensor dimension of the output sequence. Based on the determined tensor dimension, a pseudo-random number generation algorithm is used to independently sample from the standard normal distribution, extracting the white noise matrix of each action channel, i.e., generating the random normal distribution noise features of each hydraulic cylinder. Finally, by introducing a cross-attention mechanism or a channel splicing algorithm, the random normal distribution noise features are fused with the aforementioned generated latent space conditional encoding to jointly constitute a hybrid latent space state. This hybrid latent space state serves as the absolutely disordered initial starting point for the subsequent large model generation process.

[0081] Guided by the latent space conditional encoding, the inverse feature denoising transformation is performed on the hybrid latent space state, and the dimension-up reconstruction calculation is performed to output the displacement sequence of each hydraulic cylinder of the bridge-building machine.

[0082] By progressively eliminating disordered noise components from the mixed latent space state, the optimal adjustment action that conforms to the current physical forces and meteorological boundary constraints is extracted. Starting from the set maximum discrete diffusion time step, the inverse feature denoising transformation is iteratively executed along the decreasing time step direction. In each discrete time step, a denoising neural network, pre-trained and converged with massive samples, is invoked. Guided by latent space conditional encoding, it predicts the noise component contained in the noisy tensor of the current time step and removes this noise component through subtraction. The iterative solution process of this inverse feature denoising transformation is described by the following nonlinear Markov evolution formula:

[0083] ,

[0084] Among them, symbols This represents the result after one denoising calculation. The denoised latent space state tensor of the time step. (Symbol) Representative at The noisy latent space state tensor of the time step, when When the maximum discrete diffusion time step is reached, This refers to the mixed latent space state generated by the preceding steps. (Symbol) This represents the latent space conditional encoding that remains numerically constant throughout the denoising cycle. Its physical meaning is to guide the denoising process to converge towards the correct pose that counteracts gravity and wind deflection. (Symbol) This represents the variable at the current backward time step. (Symbol) This represents the fixed network weight parameter matrix within the denoising neural network. This parameter matrix is ​​obtained by calculating the mean square error between the predicted noise and the actual added noise in a historical simulation dataset and updating it through backpropagation. (Symbol) This represents a nonlinear evolution operator that encapsulates the entire denoising prediction and feature subtraction calculation process. At time step... When the iteration decreases to zero, the initial pure noise tensor is completely transformed into a clear latent feature sequence. Subsequently, a dimensionality-upgrading reconstruction calculation is performed on this latent feature sequence to map it from the abstract latent space back to a three-dimensional mechanical motion space with actual physical meaning. Finally, a numerical matrix containing the specific actions of all hydraulic cylinders at each future time point is output, which is the displacement sequence of each hydraulic cylinder of the bridge-building machine.

[0085] For example, the maximum discrete diffusion time step is set to one thousand steps. In the initial stage of iteration, the mixed latent space state of the input network appears as completely chaotic white noise. Guided by the continuous physical constraints of latent space conditional encoding, a denoising neural network performs one thousand iterations of formula evolution calculations. In each calculation, the network gradually removes tiny random fluctuations according to the current step number, and finally outputs a smooth latent space motion trajectory that conforms to mechanical safety logic at the zeroth step. Then, a decoder network composed of deconvolutional layers performs up-dimensional reconstruction calculations to parse the latent space motion trajectory into mechanical commands that are readable by humans and servo motors. Finally, the displacement sequence of each hydraulic cylinder of the bridge-building machine is output, which clearly indicates that in the next ten minutes, hydraulic cylinder No. 1 needs to lift five millimeters at a constant speed, while hydraulic cylinder No. 2 needs to remain stationary for two minutes and then slowly retract two millimeters, thereby achieving closure control under complex working conditions.

[0086] Optionally, the seam coordinate difference and attitude angle difference at the output endpoint include:

[0087] The displacement sequence is converted into driving boundary conditions and input into the digital twin model of the bridge-building machine. A stepwise forward numerical integration calculation is performed along the time axis from the current construction node to the end time to generate a virtual running trajectory.

[0088] The displacement sequence output by the dimensionality reduction and denoising process is obtained and analyzed from the abstract numerical matrix into kinematic constraints with physical dimensions, i.e., transformed into driving boundary conditions. The physical significance of this transformation is that, in the digital twin model of the bridge-building machine, the mesh nodes representing the action points at the top of each hydraulic cylinder no longer participate in the force deformation solution as degrees of freedom, but are forcibly assigned time-varying displacement values ​​specified by the displacement sequence, thereby reproducing the execution process of the servo mechanism in virtual space. Subsequently, along the time axis, starting from the current construction node and ending at the set closure time, stepwise forward numerical integration calculations are performed. Within each discrete time step, the multibody dynamics differential equations are solved by integrating the structural self-weight load, nonlinear geometric large deformation characteristics, and forward environmental wind and temperature disturbance loads synchronously derived by the time-series prediction module, all contained in the digital twin model of the bridge-building machine. By accumulating and recording the three-dimensional spatial positions, velocities, and acceleration states of all key mesh nodes in the model throughout the entire prediction period, a high-fidelity time-series spatial state dataset, i.e., the virtual running trajectory, is finally generated.

[0089] Extract the predicted seam coordinates and predicted attitude angle of the virtual running trajectory at the end time, and perform numerical subtraction operations with the seam coordinates and attitude angle at the end time respectively, and output the difference in seam coordinates and attitude angle at the end time.

[0090] To quantitatively assess whether the displacement sequence generated by diffusion can meet the requirements of extremely high-precision closure, the final state is extracted and compared for verification. The virtual trajectory dataset is analyzed, locating the last data frame at the end of the time axis. From this, the three-dimensional spatial position vectors representing the closure joint feature points and the three-dimensional Euler rotation angles are extracted and defined as the predicted joint coordinates and predicted attitude angles, respectively. Next, the ideal closure control target set in the initial reverse engineering stage is retrieved, namely the joint coordinates and attitude angles at the final moment, and the internal arithmetic logic operation component is invoked to perform a numerical subtraction operation of the vectors. This numerical subtraction operation is strictly defined by the following two multi-dimensional vector space difference formulas:

[0091] ,

[0092] ,

[0093] Among them, symbols The difference in seam coordinates at the final output moment is represented by a three-dimensional feature vector. Its physical meaning is the absolute translational error distance remaining in three-dimensional space at the seam control point after the virtual execution of the control command. (Symbol) This represents the predicted joint coordinates extracted from the end of the virtual trajectory, derived from the final state reference calculated using forward numerical integration. (Symbol) The joint coordinates represent the endpoint specified in the design drawings, and their values ​​are derived from an ideally set reference. (Symbol) The attitude angle difference value at the final output moment is a three-dimensional feature vector containing three rotational degrees of freedom. Its physical meaning is the residual angular deviation of the seam surface in the rotational directions about the three spatial projection axes. (Symbol) Represents the predicted attitude angle extracted from the end of the virtual trajectory. (Symbol) This represents the attitude angle at the endpoint required by the design. By performing the absolute value subtraction operation of the above multi-dimensional features, the difference in seam coordinates and the difference in attitude angle at the endpoint are output, which serve as the core closed-loop feedback index for measuring the physical effectiveness of the control sequence. Figure 2 As shown, a violin diagram was used to compare the three-dimensional spatial comprehensive deviation distribution at the closure section joint using the traditional static control method and the digital twin diffusion algorithm. It can be seen that not only is the median error of the joint coordinates significantly reduced, but the dispersion of the data distribution is also greatly narrowed, verifying the extremely high robustness and control accuracy of the multimodal feature fusion and denoising iterative algorithm in dealing with complex meteorological disturbances.

[0094] For example, from the virtual trajectory data frame corresponding to 12 o'clock, the predicted joint coordinates of the cantilever end joint point are extracted as 1001 mm horizontally, 2000 mm vertically, and 502 mm vertically. Simultaneously, the predicted attitude angle is extracted as a pitch angle of ±0.01 radians. The engineering construction specification document is read, and the theoretical values ​​of the joint coordinates at the endpoint are determined to be 1000 mm horizontally, 2000 mm vertically, and 500 mm vertically, with the theoretical requirement that the attitude angle at the endpoint be an absolute horizontal zero radian. The corresponding coordinates and angle components are then substituted one by one into the spatial difference formula, and numerical subtraction is performed. After the calculation, the joint coordinate difference at the endpoint is output as ±1 mm horizontally, 0 mm vertically, and ±2 mm vertically, and the attitude angle difference at the endpoint is output as a pitch angle of ±0.01 radians. These spatial difference data, accurate to the millimeter and microradian level, will serve as the direct basis for determining whether the current displacement sequence meets the tolerance threshold release standard.

[0095] Optionally, converting the displacement sequence into a servo drive electrical signal and outputting it to the hydraulic cylinder includes:

[0096] Extract the observation sequence of the real-time meteorological data at the current construction node, calculate the fluctuation variance, and map the fluctuation variance to generate a variance threshold;

[0097] To establish the boundary of dynamic adaptive construction control accuracy, meteorological fluctuation characteristics are extracted and mathematically mapped. Using the current construction node as the time reference point, a set time window length is traced back along the historical timeline to extract high-frequency real-time meteorological data within that time period, forming a one-dimensional or multi-dimensional observation sequence. Next, the sum of squared deviations of the values ​​at each sampling point in the observation sequence from the arithmetic mean is calculated and divided by the total number of sampling points, thereby rigorously calculating the fluctuation variance reflecting the recent intensity of weather events. To address the dimensional mismatch between meteorological physical quantities and spatial geometric tolerances, a linear mapping function is introduced to map the fluctuation variance to generate a variance threshold. The physical and mathematical formulas for this mapping process are as follows:

[0098] ,

[0099] Among them, symbols The variance threshold representing the output, physically signifies the maximum three-dimensional spatial comprehensive deviation allowed by the control algorithm during bridge-building machine docking under the current meteorological disturbance level. (Symbol) This represents the variance of real-time meteorological data calculated using the statistical methods described above. (Symbol) The dimensional transformation sensitivity coefficient, whose value is derived from statistical regression analysis of historical construction phases of similar bridges, is used to scale the magnitude of meteorological variance to the magnitude of spatial displacement, achieving data bridging across physical fields. (Symbol) This represents the foundation construction tolerance constant under ideal, static conditions with absolutely no wind and no temperature difference. This value is usually obtained directly from national bridge construction and acceptance standards. By implementing this mapping formula, it is ensured that the control threshold is appropriately relaxed during severe weather to avoid high-frequency oscillations and deadlocks in the servo mechanism, while the threshold is tightened during stable weather to pursue the ultimate closure accuracy.

[0100] For example, given the frequent gusts of wind at the site, an observation sequence was constructed from 180 wind speed sampling points collected over the past 30 minutes. The arithmetic mean wind speed of this observation sequence was calculated to be 5 meters per second, and the variance of the wind speed fluctuation was calculated to be four. Parameters were read from the internal read-only memory, obtaining a dimension conversion sensitivity coefficient of 0.5 and a foundation construction tolerance constant of 2 millimeters as specified in the national bridge construction acceptance standards. Substituting these values ​​into the linear mapping formula, 0.5 multiplied by four and then added to two resulted in a dynamic variance threshold of four millimeters. This four-millimeter variance threshold serves as the baseline for determining whether the prediction error meets the standard in the current time period.

[0101] By comparing the seam coordinate difference, the attitude angle difference, and the variance threshold, the displacement sequence is extracted, and the displacement sequence is encoded into a pulse width modulation signal through digital-to-analog conversion, which is then output to the hydraulic cylinder as a servo drive electrical signal.

[0102] After obtaining the dynamically set accuracy benchmark, the final control sequence screening and underlying hardware driver conversion are performed. The spatial L2 norm or equivalent endpoint displacement of the joint coordinate difference and attitude angle difference output from the forward simulation are calculated respectively, and these are compared with a strict logical relationship threshold. If and only if the comprehensive quantitative indicators of the joint coordinate difference and attitude angle difference are both strictly less than or equal to the variance threshold, the candidate control trajectory is determined to meet the dual requirements of safety and accuracy under the current weather conditions, and a displacement sequence meeting the conditions is extracted from the memory cache queue. Subsequently, in order to convert the purely digital control matrix into an energy signal that hydraulic machinery in the physical world can recognize and execute, the analog-to-digital conversion interface of the underlying hardware abstraction layer is called. Based on the displacement and voltage response characteristic curves of the hydraulic system servo proportional valve, the discrete values ​​of the target displacement in the displacement sequence are interpolated and encoded into a pulse width modulation signal with a specific high and low level time ratio. The duty cycle percentage of this pulse width modulation signal directly maps to the valve core opening size of the hydraulic cylinder servo valve. Finally, the pulse width modulation signal is amplified by the drive circuit and used as a servo drive signal to be precisely output to the hydraulic cylinders of each support node of the bridge building machine through industrial fieldbus or hardwiring, driving the mechanical structure to perform physical posture adjustment actions, and completing the final closed loop of virtual and real integration.

[0103] Optionally, mapping the volatility variance to generate a variance threshold includes:

[0104] A time window is extracted along the time axis, and the meteorological data within the time window is used as an observation sequence. The second-order central moments are then calculated to obtain the fluctuation variance.

[0105] Using the current physical time as the origin, a fixed-length time window is extracted by retrospectively calculating along the historical time axis. The length of this time window is determined using the Nyquist sampling law, which analyzes the frequency of meteorological changes and the natural vibration period of the bridge-building machine structure. Subsequently, all continuous meteorological data, such as wind speed or temperature, within this time window range are extracted from the underlying sensor database and arranged chronologically to form a one-dimensional observation sequence. To objectively measure the dispersion and jumps in the data within this observation sequence, the underlying arithmetic logic module is invoked to solve for the second-order central moment of the observation sequence, thereby deriving the variance of the fluctuation. The calculation process of this variance is rigorously defined by the following statistical formula:

[0106] ,

[0107] Among them, symbols This represents the variance of the fluctuation obtained from the solution, and its physical meaning is an indicator of the degree of energy dispersion due to drastic weather changes within the time sliding window. (Symbol) This represents the total number of meteorological data sampling points included in the observation sequence. This value is calculated by multiplying the time span of the time window by the sampling frequency of the meteorological sensor. (Symbol) This represents the actual value of the i-th meteorological observation in the observation sequence. (Symbol) This represents the arithmetic mean obtained by summing all values ​​in the entire observation sequence and dividing by the total number of sampling points. (Symbol) This represents a standard mathematical operator that performs a cumulative summation over all sampling points within a time window. By applying this second-order central moment formula, the disordered original meteorological waveform is transformed into a single and precise discrete energy scalar.

[0108] For example, on the eve of the bridge's main span closure, a time window of thirty minutes is set. The anemometer samples once per minute, thus extracting thirty wind speed samples within this time window to form an observation sequence. The arithmetic mean of these thirty wind speed values ​​is calculated to be ten meters per second. Then, the arithmetic mean of ten meters per second is subtracted from each wind speed observation, and the square is calculated. For example, if the wind speed at a certain point is twelve meters per second, the squared deviation is four. These thirty squared deviations are summed, assuming the total sum is one hundred and twenty, and finally divided by the total number of sampling points (thirty), to calculate the variance, which is also four. This value of four is a quantitative representation of the degree of wind speed turbulence within the current thirty minutes.

[0109] Extract the maximum allowable deformation of the rigid body dynamics module, and use the maximum allowable deformation of the structure as the upper limit of the value to perform extreme value normalization scaling on the fluctuation variance, and map to generate a variance threshold.

[0110] The rigid body dynamics module is accessed via a memory interface, and the maximum allowable deformation of the structure is extracted from its material property configuration file. This maximum allowable deformation is an absolute geometric deformation limit derived from the yield strength limit of steel and the engineering safety factor, representing the physical limit of deformation of the continuous beam without permanent failure or misalignment. Next, the statistical dimension of meteorological fluctuation indicators is transformed into geometric tolerances of the control dimension. Using this maximum allowable deformation as the absolute numerical upper limit, extreme value normalization scaling is performed on the fluctuation variance, ultimately mapping and generating a variance threshold. The calculation logic for generating this variance threshold is described by the following extreme value normalization constraint formula:

[0111] ,

[0112] Among them, symbols This represents the variance threshold generated by the final mapping. Its dimension is consistent with spatial coordinates, for example, millimeters, and it is used directly as the tolerance judgment standard for issuing control commands. (Symbol) This represents the lower limit of the tolerance for foundation construction. This parameter is derived from the docking error requirements under ideal, windless conditions specified in the National High-Speed ​​Railway Bridge Construction and Acceptance Standard. (Symbol) Represents the variance of the fluctuation calculated in real time. (Symbol) Represents the lower bound of the historical stationary variance, sign Representing the upper limit of historical extreme fluctuation variance, both parameters are extreme value constants extracted from the historical meteorological dataset of bridges in the same region over the previous year's period, and are used to determine the denominator benchmark for normalization. (Symbols) This represents the maximum allowable deformation of the structure extracted from the rigid body dynamics module. This extreme value normalization scaling formula ensures that when weather fluctuations are extremely severe and approach historical extremes, the generated variance threshold can safely approach but never exceed the maximum allowable deformation of the structure; while when the weather is extremely stable, the variance threshold can be strictly shrunk to the lower limit of the foundation construction tolerance, thus achieving a dynamic adaptive safety mapping of control accuracy. Figure 3 As shown, the kernel density contour plot illustrates the joint probability distribution between the variance of real-time meteorological data fluctuations and the final output actual spatial comprehensive deviation under long-term high-frequency monitoring conditions. From the Gaussian kernel density contour distribution trend and the superimposed dynamic tolerance threshold boundary in the figure, it can be seen that in the high-density convergence region and under extreme divergence conditions, the actual pose deviation is always strictly limited below the safety envelope generated by the linear mapping equation, fully demonstrating the physical absolute security of the dynamic release mechanism and the robust convergence of the large model control algorithm.

[0113] Based on the same inventive concept, this invention also provides a posture control system for a bridge-building machine for the closure section of a high-speed railway continuous beam based on digital twins, such as... Figure 4 As shown, the system includes:

[0114] The data acquisition and inverse kinematics planning module is used to acquire the beam end joint coordinates, support coordinates, attitude angles and real-time meteorological data of the continuous beam closure segment, construct a digital twin model of the bridge building machine, perform inverse kinematics based on the joint coordinates and attitude angles at the end point, and generate target pose sequences for each construction node. The digital twin model of the bridge building machine includes a rigid body dynamics module and a time series prediction module.

[0115] The dual-path parallel calculation module is used to collect the force data of the bridge building machine sensor based on the current construction node, and input it in parallel to the rigid body dynamics module and the time series prediction module, and output the mechanical displacement vector and the meteorological response displacement vector for the current construction node respectively.

[0116] The deviation feature fusion module is used to extract the target coordinate vector of the current construction node using the target pose sequence, calculate the Euclidean distance and the included angle between the mechanical displacement vector and the meteorological response displacement vector, and concatenate the target coordinate vector, the Euclidean distance and the included angle to generate a joint input matrix for the current construction node.

[0117] The diffusion sampling generation module is used to perform diffusion sampling and inverse denoising based on the joint input matrix, and output the displacement sequence of each hydraulic cylinder of the bridge building machine;

[0118] The forward simulation verification module is used to extract the displacement sequence and input it into the digital twin model of the bridge building machine, forward calculate the virtual running trajectory from the current construction node to the end time, and output the joint coordinate difference and attitude angle difference at the end time.

[0119] The meteorological screening and drive output module is used to calculate the fluctuation variance of the real-time meteorological data, generate a variance threshold, extract displacement sequences where both the seam coordinate difference and the attitude angle difference are less than the variance threshold, and convert the displacement sequences into servo drive electrical signals and output them to the hydraulic cylinder.

[0120] It should be noted that the functional division and information interaction between the various modules described above are logical, but in terms of physical implementation, they can be integrated on the same software platform or deployed in a distributed manner. The connections between them represent data flow and control flow, aiming to collaboratively achieve the objectives of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of protection of this invention.

Claims

1. A method for attitude control of a bridge-building machine for the closure section of a continuous beam in a high-speed railway based on digital twins, characterized in that... The method includes: The coordinates of the beam end joints, support coordinates, attitude angles and real-time meteorological data of the continuous beam closure segment are obtained, and a digital twin model of the bridge-building machine is constructed. Based on the joint coordinates and attitude angles at the end time, a reverse time sequence solution is performed to generate a target pose sequence for each construction node. The digital twin model of the bridge-building machine includes a rigid body dynamics module and a time sequence prediction module. Based on the force data collected from the bridge-building machine's sensors at the current construction node, the data is input in parallel to the rigid body dynamics module and the time-series prediction module, and the mechanical displacement vector and meteorological response displacement vector are output to the current construction node respectively. The target coordinate vector of the current construction node is extracted using the target pose sequence. The Euclidean distance and the included angle between the mechanical displacement vector and the meteorological response displacement vector are calculated. The target coordinate vector, the Euclidean distance, and the included angle are concatenated to generate a joint input matrix for the current construction node. Based on the joint input matrix, diffusion sampling and inverse denoising are performed to output the displacement sequence of each hydraulic cylinder of the bridge-building machine; The displacement sequence is extracted and input into the digital twin model of the bridge-building machine. The virtual running trajectory from the current construction node to the end time is calculated in a forward direction, and the joint coordinate difference and attitude angle difference at the end time are output. The fluctuation variance of the real-time meteorological data is calculated, a variance threshold is generated, and displacement sequences whose joint coordinate difference and attitude angle difference are both less than the variance threshold are extracted. The displacement sequences are then converted into servo drive electrical signals and output to the hydraulic cylinder.

2. The attitude control method for bridge-building machine of high-speed railway continuous beam closure segment based on digital twin as described in claim 1, characterized in that, The generation of target pose sequences for each construction node includes: Extract the coordinates of the beam end joints and the coordinates of the support points, construct the spatial structure topology, assemble the physical parameters of the structural components, and generate a rigid body dynamics module; The timing prediction module is coupled with the rigid body dynamics module to construct a digital twin model of the bridge-building machine; Set the joint coordinates and attitude angles at the end point as inverse time-series boundary constraints, and perform inverse dynamics solution step by step along the time axis from the end point to the current construction node to output the target pose sequence.

3. The attitude control method for bridge-building machine of high-speed railway continuous beam closure segment based on digital twin as described in claim 1, characterized in that, The time series prediction module includes: Historical meteorological monitoring data and historical structural displacement data of bridges of the same type during the construction phase were obtained, and timestamp alignment and noise removal were performed to obtain a historical training dataset. Using the historical meteorological monitoring data as network input and the historical structural displacement data as network output, a recurrent neural network is established and trained using the historical training dataset, and the network weight parameters are updated to obtain the time series prediction module.

4. The attitude control method for bridge-building machine of high-speed railway continuous beam closure segment based on digital twin as described in claim 1, characterized in that, The step of outputting the mechanical displacement vector and the meteorological response displacement vector for the current construction node includes: The force data of the bridge-building machine sensor is parsed into nodal load tensors, input to the rigid body dynamics module, and mapped and output as a mechanical displacement vector. Extract the force data from the bridge-building machine sensor and the real-time meteorological data, construct multi-dimensional time-series input features, call the time-series prediction module to perform forward propagation inference calculation, and output the meteorological response displacement vector.

5. The attitude control method for bridge construction machine of high-speed railway continuous beam closure segment based on digital twin as described in claim 1, characterized in that, The process of generating a joint input matrix for the current construction node includes: Match the timestamp of the current construction node and extract the target coordinate vector from the target pose sequence; Perform vector subtraction and L2 norm calculation on the mechanical displacement vector and the meteorological response displacement vector to output the Euclidean distance, and calculate the quotient of the dot product and the product of the modulus, and generate the included angle value through the inverse cosine function mapping. The Euclidean distance and the included angle value are converted into feature scalars, and tensor concatenation operations are performed along the feature dimensions of the target coordinate vector to generate a joint input matrix.

6. The posture control method for bridge-building machine of high-speed railway continuous beam closure segment based on digital twin as described in claim 1, characterized in that, The displacement sequence of each hydraulic cylinder of the output bridge-building machine includes: Dimensionality reduction feature extraction is performed on the joint input matrix to generate latent space conditional coding, and random normal distribution noise features of each hydraulic cylinder are extracted and combined to form a hybrid latent space state. Guided by the latent space conditional encoding, the inverse feature denoising transformation is performed on the hybrid latent space state, and the dimension-up reconstruction calculation is performed to output the displacement sequence of each hydraulic cylinder of the bridge building machine.

7. The attitude control method for bridge construction machine of high-speed railway continuous beam closure segment based on digital twin as described in claim 1, characterized in that, The seam coordinate difference and attitude angle difference at the output endpoint include: The displacement sequence is converted into driving boundary conditions and input into the digital twin model of the bridge-building machine. A stepwise forward numerical integration calculation is performed along the time axis from the current construction node to the end time to generate a virtual running trajectory. Extract the predicted seam coordinates and predicted attitude angle of the virtual running trajectory at the end time, and perform numerical subtraction operations with the seam coordinates and attitude angle at the end time respectively, and output the difference in seam coordinates and attitude angle at the end time.

8. The posture control method for bridge construction machine of high-speed railway continuous beam closure segment based on digital twin as described in claim 1, characterized in that, The step of converting the displacement sequence into a servo drive electrical signal and outputting it to the hydraulic cylinder includes: Extract the observation sequence of the real-time meteorological data at the current construction node, calculate the fluctuation variance, and map the fluctuation variance to generate a variance threshold; By comparing the seam coordinate difference, the attitude angle difference, and the variance threshold, the displacement sequence is extracted, and the displacement sequence is encoded into a pulse width modulation signal through digital-to-analog conversion, which is then output to the hydraulic cylinder as a servo drive electrical signal.

9. The attitude control method for bridge-building machine of high-speed railway continuous beam closure segment based on digital twin as described in claim 8, characterized in that, The step of mapping the volatility variance to generate a variance threshold includes: A time window is extracted along the time axis, and the meteorological data within the time window is used as an observation sequence. The second-order central moments are then calculated to obtain the fluctuation variance. Extract the maximum allowable deformation of the rigid body dynamics module, and use the maximum allowable deformation of the structure as the upper limit of the value to perform extreme value normalization scaling on the fluctuation variance, and map to generate a variance threshold.

10. A digital twin-based attitude control system for a bridge-building machine for the closure section of a continuous beam in a high-speed railway, applied to the attitude control method for a bridge-building machine for the closure section of a continuous beam in a high-speed railway as described in any one of claims 1-9, characterized in that... The system includes: The data acquisition and inverse kinematics planning module is used to acquire the beam end joint coordinates, support coordinates, attitude angles and real-time meteorological data of the continuous beam closure segment, construct a digital twin model of the bridge building machine, perform inverse kinematics based on the joint coordinates and attitude angles at the end point, and generate target pose sequences for each construction node. The digital twin model of the bridge building machine includes a rigid body dynamics module and a time series prediction module. The dual-path parallel calculation module is used to collect the force data of the bridge building machine sensor based on the current construction node, and input it in parallel to the rigid body dynamics module and the time series prediction module, and output the mechanical displacement vector and the meteorological response displacement vector for the current construction node respectively. The deviation feature fusion module is used to extract the target coordinate vector of the current construction node using the target pose sequence, calculate the Euclidean distance and the included angle between the mechanical displacement vector and the meteorological response displacement vector, and concatenate the target coordinate vector, the Euclidean distance and the included angle to generate a joint input matrix for the current construction node. The diffusion sampling generation module is used to perform diffusion sampling and inverse denoising based on the joint input matrix, and output the displacement sequence of each hydraulic cylinder of the bridge building machine; The forward simulation verification module is used to extract the displacement sequence and input it into the digital twin model of the bridge building machine, forward calculate the virtual running trajectory from the current construction node to the end time, and output the joint coordinate difference and attitude angle difference at the end time. The meteorological screening and drive output module is used to calculate the fluctuation variance of the real-time meteorological data, generate a variance threshold, extract displacement sequences where both the seam coordinate difference and the attitude angle difference are less than the variance threshold, and convert the displacement sequences into servo drive electrical signals and output them to the hydraulic cylinder.

Images