Large-span bridge swivel in-place control method based on neural network and computer system

Abstract

The application provides a large-span bridge swivel in-place control method and computer system based on a neural network, which decouples time-space features by collecting swivel motion data sequences generated in the swivel construction process of a large-span bridge, obtains swivel posture evolution features and environmental interference coupling features, then calls a pre-trained swivel dynamics prediction network to jointly model the two, generates a state prediction sequence of the swivel process, then compares the state prediction sequence with a preset swivel target trajectory sequence, generates a swivel deviation risk evaluation matrix containing position deviation and posture deviation, finally drives a swivel control parameter optimization network to perform compensation calculation based on the swivel deviation risk evaluation matrix, generates a swivel in-place control instruction set containing a traction force adjustment coefficient and a swivel speed correction value, and transmits the swivel in-place control instruction set to a swivel driving system. Through the application, the accuracy and reliability of the large-span bridge swivel in-place control can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a large-span bridge swivel in-place control method based on a neural network and a computer system. BACKGROUND

[0002] With the rapid development of transportation infrastructure construction, the large-span bridge swivel construction technology is widely used in complex terrain scenes such as crossing valleys, rivers and existing traffic lines, and the swivel in-place control as a core link of the swivel construction directly affects the final butt joint accuracy and construction safety of the bridge structure, and the technical essence thereof lies in that the posture and position in the swivel process are real-time regulated and controlled, so that the bridge structure completes rotation according to a preset trajectory and realizes accurate butt joint. At present, the traditional large-span bridge swivel in-place control method is difficult to fully consider the complex coupling relationship between the posture change of the bridge itself and the external environmental interference in the swivel process, so that the swivel state prediction accuracy is limited, and the dynamic risk assessment ability of the swivel deviation is insufficient, so that the timeliness and accuracy of the control parameter adjustment are limited, and the accuracy and reliability of the swivel in-place are affected. SUMMARY

[0003] Therefore, the present application provides a large-span bridge swivel in-place control method based on a neural network and a computer system.

[0004] The technical scheme of the embodiment of the present application is implemented in this way: on the one hand, the present application provides a large-span bridge swivel in-place control method based on a neural network, which comprises the following steps: collecting a swivel motion data sequence generated in the large-span bridge swivel construction process, wherein the swivel motion data sequence comprises multiple groups of monitoring data units with time stamp marks collected synchronously at a preset sampling frequency; decoupling the space-time features of the swivel motion data sequence to obtain the swivel posture evolution features and the environmental interference coupling features in the monitoring data units; calling a pre-trained swivel dynamics prediction network to perform joint dynamics modeling on the swivel posture evolution features and the environmental interference coupling features to generate a state prediction sequence of the swivel process; comparing the state prediction sequence with a preset swivel target trajectory sequence to generate a swivel deviation risk assessment matrix comprising a position deviation amount and a posture deviation amount; driving a swivel control parameter optimization network to perform compensation calculation based on the swivel deviation risk assessment matrix to generate a swivel in-place control instruction set comprising a traction force adjustment coefficient and a swivel speed correction value, and transmitting the swivel in-place control instruction set to a swivel driving system.

[0005] On the other hand, the present application provides a computer system, which comprises a memory and a processor, the memory stores a computer program capable of running on the processor, and the processor implements the steps in the above method when executing the program.

[0006] The present invention provides a neural network-based control method for the rotation and positioning of long-span bridges. This method constructs a structured rotation motion data sequence by collecting multiple sets of time-stamped monitoring data units synchronously at a preset sampling frequency. By decoupling the spatiotemporal features of the rotation motion data sequence, the method obtains rotation attitude evolution features and environmental disturbance coupling features. This systematically breaks down the complex rotation motion data into features reflecting the bridge's own dynamic changes and features characterizing the coupling effect of external environmental factors, providing structured feature support for accurate modeling. A pre-trained rotation dynamics prediction network is then invoked to perform joint dynamic modeling of the rotation attitude evolution features and environmental disturbance coupling features. The nonlinear fitting capability of the neural network is used to fuse the two features for collaborative modeling, improving the rotation process. The accuracy of state prediction is improved by comparing the state prediction sequence with the preset rotation target trajectory sequence to generate a rotation deviation risk assessment matrix that includes position and attitude deviations. This upgrades the deviation information from single-point numerical values ​​to a structured matrix that integrates time, deviation type, and risk dimensions, enhancing the comprehensiveness and dynamism of deviation analysis. Based on the rotation deviation risk assessment matrix, a rotation control parameter optimization network is driven to perform compensation calculations. Through the neural network, multi-dimensional deviation risk information is nonlinearly mapped to dynamically generate traction adjustment coefficients and speed correction values. This overcomes the limitations of traditional linear adjustment modes in adapting to dynamic deviation risks, achieving intelligent dynamic optimization of rotation control parameters and thus improving the accuracy and reliability of rotation positioning control for long-span bridges. Attached Figure Description

[0007] Figure 1 A schematic diagram illustrating the implementation process of a neural network-based control method for the rotation and positioning of a long-span bridge, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the composition of a rotation and positioning control device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0008] This invention provides a neural network-based method for controlling the rotation and positioning of long-span bridges. This method can be executed by a processor in a computer system. The computer system can refer to devices with data processing capabilities, such as servers, laptops, tablets, and desktop computers.

[0009] Figure 1 This is a schematic diagram illustrating the implementation process of a neural network-based rotation and positioning control method for long-span bridges, as provided in an embodiment of the present invention. Figure 1As shown, the method comprises: step S100: collecting a rotation body motion data sequence generated in the rotation body construction process of a long-span bridge, the rotation body motion data sequence containing multiple groups of monitoring data units with timestamp labels collected synchronously at a preset sampling frequency. The rotation body motion data sequence is a data set generated during the rotation body construction of the long-span bridge, reflecting various motion information in the bridge rotation body process. The preset sampling frequency is a time interval for data collection set in advance, for example, it can be set to collect data once per second. The timestamp label is a time identifier given to each group of monitoring data units, used to clearly indicate the specific time point of data collection. The monitoring data unit is the basic data unit in the rotation body motion data sequence, containing various data related to the bridge rotation body. In the actual collection process, various sensors such as displacement sensors and angle sensors can be installed at the preset positions of the long-span bridge. Taking the displacement sensor as an example, it can measure the position change of the bridge during the rotation body process; the angle sensor can measure the rotation angle of the bridge. These sensors collect data synchronously at the preset sampling frequency and add timestamp labels to each group of collected data.

[0010] Step S200: decoupling the space-time features of the rotation body motion data sequence to obtain the rotation body posture evolution features and the environmental interference coupling features in the monitoring data units. The space-time feature decoupling is to separate and analyze the time and space related features contained in the rotation body motion data sequence to extract different types of key features. The rotation body posture evolution features reflect the changes of the posture of the long-span bridge with time during the rotation body process, such as the dynamic changes of the rotation angle, axis offset and other features of the bridge. The environmental interference coupling features reflect the influence features of environmental factors on the rotation body posture of the bridge, and there is a coupling relationship between these environmental factors and the rotation body posture changes.

[0011] Specifically, step S200 can specifically include the following steps S210-S250: step S210: performing time axis division processing on the rotation body motion data sequence, dividing the rotation body motion data sequence into multiple data segment groups with continuous time association according to a preset time window length, each data segment group containing a preset number of continuous monitoring data units. The time axis division processing is an operation of dividing the rotation body motion data sequence according to the time dimension. The preset time window length is a time interval length for dividing data set in advance, for example, it can be set to 10 seconds. The continuous time association means that the divided data segment groups are continuous in time, without time jump. The preset number is the number of continuous monitoring data units contained in each data segment group.

[0012] Step S220: Perform spatial coordinate system conversion processing on the monitoring data units in the data segment group, converting each monitoring data unit from the local sensor coordinate system to the bridge turn global coordinate system, and generating a spatial position data set under a unified coordinate reference. Spatial coordinate system conversion processing is the process of converting the coordinates of a monitoring data unit from one coordinate system to another. The local sensor coordinate system is the coordinate system adopted by the sensor itself, and the local coordinate systems of different sensors may be different, which will bring difficulties to the unified analysis of data. The bridge turn global coordinate system is a unified coordinate system suitable for the entire bridge turn process, and the bridge turn movement can be more conveniently analyzed based on this coordinate system. When performing spatial coordinate system conversion, the conversion relationship between the local sensor coordinate system and the bridge turn global coordinate system needs to be determined. This can be achieved by measuring the position and direction of the sensor in the global coordinate system. For example, the origin of the local sensor coordinate system of a displacement sensor installed on a long-span bridge may be located at the installation position of the sensor. By measuring the coordinates and direction of the sensor in the bridge turn global coordinate system, a conversion matrix is established. Then, the monitoring data units in the local sensor coordinate system are converted to the bridge turn global coordinate system using the conversion matrix. Finally, all the converted monitoring data units are combined into a spatial position data set under a unified coordinate reference.

[0013] Step S230: Extract the rotation angle change sequence, axis offset sequence and center of gravity displacement trajectory in the turn process from the spatial position data set as the basic constituent elements of the turn posture evolution feature. The rotation angle change sequence reflects the change of the rotation angle of the bridge with time during the turn process, and is an important feature for describing the rotation posture of the bridge. The axis offset sequence represents the change of the offset degree of the bridge axis with time during the turn process, and embodies the dynamic change of the position of the bridge axis. The center of gravity displacement trajectory is the moving path of the center of gravity of the bridge during the turn process, and reflects the position change of the whole bridge. These elements together constitute the turn posture evolution feature, which can comprehensively describe the posture change of the bridge during the turn process.

[0014] Specifically, step S230 can specifically include steps S231-S235: step S231: performing time series difference processing on the three-dimensional coordinate data in the spatial position data set, calculating the coordinate change amount between adjacent time stamps, and generating a sequence of instantaneous velocity vectors of the rotational motion. The time series difference processing is an operation of calculating the difference value between adjacent data points in the three-dimensional coordinate data in the spatial position data set in time sequence. The adjacent time stamps are two data acquisition time points adjacent in time sequence. The coordinate change amount is the difference value of the three-dimensional coordinates between adjacent time stamps, reflecting the change of the bridge position in the time period. The sequence of instantaneous velocity vectors is a sequence composed of instantaneous velocity vectors in each time interval, and the direction and size of the instantaneous velocity vector represent the motion direction and speed of the bridge in the time interval, respectively. In actual calculation, for each three-dimensional coordinate data point in the spatial position data set, the coordinate difference value between the adjacent two data points is calculated in time sequence. For example, for the three-dimensional coordinate data (x n ,y n ,z n ) of the nth time stamp and the three-dimensional coordinate data (x n+1 ,y n+1 ,z n+1 ) of the n+1th time stamp, the coordinate change amount is calculated as (x n+1 -x n ,y n+1 -y n ,z n+1 -z n ). Then, the coordinate change amount is divided by the time interval to obtain the instantaneous velocity vector in the time interval. Arranging the instantaneous velocity vector of each time interval in time sequence can generate a sequence of instantaneous velocity vectors of the rotational motion.

[0015] Step S232: based on the sequence of instantaneous velocity vectors, the angle calculation processing is performed to obtain the rotation angle change value in the horizontal plane and the inclination angle change value in the vertical plane, and a three-dimensional angle change matrix containing azimuth angle, pitch angle and roll angle is constructed.

[0016] The angle calculation process is an operation of calculating the rotation angle and the inclination angle change value according to the instantaneous velocity vector sequence. The rotation angle change value in the horizontal plane reflects the rotation of the bridge in the horizontal direction, and the inclination angle change value in the vertical plane reflects the inclination change of the bridge in the vertical direction. Azimuth angle, pitch angle and roll angle are three angle parameters used to describe the three-dimensional attitude of an object, and the three-dimensional angle change matrix is a matrix composed of the change values of the three angles. When calculating the angles, the direction information of the velocity in the instantaneous velocity vector sequence is used to calculate the rotation angle change value in the horizontal plane and the inclination angle change value in the vertical plane by mathematical methods such as trigonometric functions. For example, according to the projection direction of the instantaneous velocity vector in the horizontal plane, the change value of the azimuth angle is calculated; according to the projection direction of the instantaneous velocity vector in the vertical plane, the change values of the pitch angle and the roll angle are calculated. These angle change values are arranged in order to form a three-dimensional angle change matrix. For example, the first row of the matrix can represent the change value of the azimuth angle, the second row represents the change value of the pitch angle, and the third row represents the change value of the roll angle.

[0017] Step S233: filtering the three-dimensional angle change matrix to eliminate noise interference and generating a smooth rotation angle change sequence. The filtering process is an operation of processing the three-dimensional angle change matrix to remove noise interference therein. The noise interference can be random fluctuations in the angle change value caused by sensor errors, environmental interference and other factors. The smooth rotation angle change sequence is a sequence of rotation angle changes over time after filtering to remove noise interference. In practical applications, various filtering algorithms can be used to process the three-dimensional angle change matrix, such as Kalman filtering algorithm, moving average filtering algorithm, etc. Taking the Kalman filtering algorithm as an example, first, the state model and the observation model of the system need to be established, and the state prediction and update are performed according to the data in the three-dimensional angle change matrix. Through continuous iterative calculation, the true rotation angle value is estimated to eliminate noise interference. Finally, the filtered rotation angle values are arranged in time sequence to generate a smooth rotation angle change sequence.

[0018] Step S234: Calculate the distance deviation value of each monitoring point in the spatial position data set relative to the axis center of the rotating body, obtain the axis offset sequence through fitting processing, and the sampling frequency of the axis offset sequence is the same as that of the rotation angle change sequence. The distance deviation value is the difference between the actual distance of each monitoring point to the axis center of the rotating body and the theoretical distance. The fitting processing is the operation of curve fitting of the calculated distance deviation value. The axis offset sequence is a sequence describing the degree of offset of the bridge axis during the rotation process. The sampling frequency is the time interval frequency of data acquisition or generation. When calculating the distance deviation value, the position of the axis center of the rotating body is determined first, and then the actual distance of each monitoring point in the spatial position data set to the axis center of the rotating body is calculated. The actual distance is subtracted from the theoretical distance to obtain the distance deviation value. Through fitting processing, such as polynomial fitting, spline curve fitting, etc., the calculated distance deviation value is fitted into a continuous curve to obtain the axis offset sequence. In order to facilitate subsequent analysis and comparison, the sampling frequency of the axis offset sequence needs to be the same as that of the rotation angle change sequence.

[0019] Step S235: Perform curve fitting processing on the discrete barycenter coordinate data through an interpolation algorithm to generate a continuous and derivable barycenter displacement trajectory curve, and the time dimension of the barycenter displacement trajectory curve is aligned with that of the rotation angle change sequence. In actual operation, various interpolation algorithms can be used, such as Lagrange interpolation algorithm, Newton interpolation algorithm, cubic spline interpolation algorithm, etc. Taking the cubic spline interpolation algorithm as an example, first, according to the discrete barycenter coordinate data points, the entire time interval is divided into multiple subintervals. In each subinterval, a cubic polynomial is used to fit the data points. By satisfying the set boundary conditions and continuity conditions, the coefficients of these cubic polynomials are determined. Finally, the cubic polynomials of each subinterval are combined into a continuous and derivable curve, i.e. the barycenter displacement trajectory curve. In order to ensure that the barycenter displacement trajectory curve can accurately reflect the attitude change of the bridge during the rotation process, its time dimension needs to be aligned with that of the rotation angle change sequence. That is, the barycenter position corresponding to each time point on the barycenter displacement trajectory curve corresponds to the rotation angle corresponding to the same time point in the rotation angle change sequence.

[0020] Step S240: Collect environment perception data synchronized with the time stamp of the rotating motion data sequence, which includes wind speed and direction time series data, temperature field distribution data, and ground vibration conduction data.

[0021] The environmental perception data is data used to describe the surrounding environment during the bridge rotation process. The wind speed and direction time series data is a sequence data recording the change of wind speed and direction over time, reflecting the influence of atmospheric flow on the bridge rotation. The temperature field distribution data is data describing the temperature distribution around the bridge. Changes in temperature can cause thermal expansion and contraction of the bridge structure, affecting the rotation posture. The ground vibration conduction data is data recording the ground vibration situation. Ground vibrations can be transmitted to the bridge structure through the bridge foundation, causing interference to the rotation process. When collecting environmental perception data, corresponding sensor equipment needs to be used. For example, an anemometer is used to collect wind speed and direction time series data. The anemometer can measure wind speed and direction in real time and record their changes over time. Temperature sensors are used to collect temperature field distribution data. Multiple temperature sensors can be installed at different positions around the bridge to obtain more comprehensive temperature distribution information. Vibration sensors are used to collect ground vibration conduction data. The vibration sensor is installed near the bridge foundation to measure the ground vibration.

[0022] Step S250: Time stamp alignment processing is performed on the environmental perception data and the spatial position data set. A coupling correlation model between environmental factors and rotation posture changes is established through sliding window correlation analysis, and an environmental interference coupling feature vector is generated.

[0023] When performing time stamp alignment processing, first, the common range of time stamps in the environmental perception data and the spatial position data set needs to be found. Then, the data is filtered and sorted according to the time stamps, so that the data at the same time point can be corresponded. When performing sliding window correlation analysis, a fixed size sliding window is set. The window is synchronously slid on the environmental perception data and the spatial position data set. The correlation index between each environmental factor in the environmental perception data and the rotation posture change related data (such as rotation angle change sequence, axis offset sequence, etc.) in each window is calculated, such as Pearson correlation coefficient, etc. According to these correlation indexes, a coupling correlation model between environmental factors and rotation posture changes is established. Finally, based on the coupling correlation model, the influence features of environmental factors on the rotation posture are extracted. These features are combined into an environmental interference coupling feature vector. The dimension of the vector is matched with the dimension of the rotation posture evolution feature, so as to perform joint modeling subsequently.

[0024] Specifically, step S250 can specifically include the following steps S251-S255: step S251: time stamp calibration processing is performed on the environment perception data, and the wind speed and direction time series data, temperature field distribution data and ground vibration conduction data collected by different sensors are unified to the same time reference system as the rotation motion data sequence. The time stamp calibration processing is an operation of adjusting and correcting the time stamp of the environment perception data to ensure the consistency of the data collected by different sensors in time. Different sensors may have clock errors, different data collection frequencies, etc., resulting in deviations in the time stamps of the collected data. Unifying these data to the same time reference system as the rotation motion data sequence facilitates subsequent analysis and processing. In actual operation, the time reference system of the rotation motion data sequence is first determined. Then, the time stamp of the environment perception data collected by different sensors is calibrated. For example, for wind speed and direction time series data, temperature field distribution data and ground vibration conduction data, the difference between their time stamps and the time stamps of the rotation motion data sequence is checked. If the data time stamp of a certain sensor is overall early or late, a fixed time offset can be added or subtracted for calibration. In this way, the time stamps of all environment perception data are unified to the same time reference system as the rotation motion data sequence.

[0025] Step S252: resampling processing is performed on the calibrated environment perception data based on the time interval of the rotation angle change sequence to ensure that the sampling frequency of the environment perception data is consistent with the sampling frequency of the spatial position data set.

[0026] Resampling processing is an operation of resampling the calibrated environment perception data to change its sampling frequency. The time interval of the rotation angle change sequence is the time difference between adjacent two data points in the sequence, reflecting the sampling frequency of the rotation angle change. In order to ensure that the environment perception data can be effectively analyzed and compared with the spatial position data set, it is necessary to make its sampling frequency consistent with the sampling frequency of the spatial position data set. When resampling processing is performed, the calibrated environment perception data is interpolated or extracted according to the time interval of the rotation angle change sequence. If the sampling frequency of the environment perception data is higher than the sampling frequency of the rotation angle change sequence, the extraction method can be used to select a data point every set time interval to reduce its sampling frequency; if the sampling frequency of the environment perception data is lower than the sampling frequency of the rotation angle change sequence, the interpolation method can be used to estimate the value between adjacent data points to increase its sampling frequency. In this way, the sampling frequency of the environment perception data is consistent with the sampling frequency of the spatial position data set.

[0027] Step S253: Construct a sliding analysis window containing a time window size parameter, and input the resampled environment perception data and the rotation angle change sequence into the sliding analysis window for synchronous sliding processing.

[0028] The sliding analysis window is a data processing window for correlation analysis, with a set time window size parameter. The time window size parameter is the length of the time interval covered by the sliding analysis window. Synchronous sliding processing is to slide the resampled environment perception data and the rotation angle change sequence on the sliding analysis window at the same time, so as to perform correlation analysis in each time window. When constructing the sliding analysis window, the time window size parameter is set according to actual needs. For example, the time window size is set to 10 seconds. After arranging the resampled environment perception data and the rotation angle change sequence in chronological order, the sliding analysis window is slid from the starting position of the data sequence. At each sliding position, the window contains the environment perception data and the rotation angle change sequence data in the set time interval. Through synchronous sliding processing, correlation analysis can be performed on the environment perception data and the rotation angle change sequence in different time windows.

[0029] Step S254: In each sliding window, calculate the correlation index between each environmental factor in the environment perception data and the rotation angle change value, and generate a multi-dimensional correlation coefficient matrix.

[0030] The correlation index is a numerical value used to measure the degree of correlation between two variables. In each sliding window, the correlation between each environmental factor (such as wind speed, wind direction, temperature, ground vibration, etc.) in the environment perception data and the rotation angle change value needs to be calculated. The multi-dimensional correlation coefficient matrix is a matrix composed of these correlation indexes. Each row and each column of the matrix corresponds to an environmental factor and a rotation angle change value, respectively, and the elements in the matrix represent the correlation coefficients between them. When calculating the correlation index, various methods can be used, such as Pearson correlation coefficient, Spearman correlation coefficient, etc. Taking the Pearson correlation coefficient as an example, for the environmental factor data sequence and the rotation angle change value data sequence in each sliding window, first calculate their mean and standard deviation. Then, according to the calculation formula of the Pearson correlation coefficient, calculate the correlation coefficient between them. The correlation coefficients between each environmental factor and the rotation angle change value are combined into a matrix, which is the multi-dimensional correlation coefficient matrix.

[0031] Step S255: Based on the multi-dimensional correlation coefficient matrix, extract the main environmental interference factors affecting the rotation attitude change through principal component analysis, and perform weighted fusion of the time series data of the main environmental interference factors and the axis offset sequence to generate an environmental interference coupling feature vector. The dimension of the environmental interference coupling feature vector matches the dimension of the rotation attitude evolution feature.

[0032] In performing principal component analysis, the multidimensional correlation coefficient matrix is first standardized to have the same scale and range. Then, the covariance matrix of the matrix is calculated, and the principal components of the data are obtained by solving the eigenvalues and eigenvectors of the covariance matrix. The first N principal components with larger eigenvalues are selected as the main features, and the environmental factors corresponding to these principal components are the main environmental interference factors. When weighting and fusing the time series data of the main environmental interference factors and the axis offset sequence, appropriate weights need to be determined. For example, the weights can be determined according to the influence degree of each main environmental interference factor on the change of the rotating body attitude. The time series data of the main environmental interference factors are multiplied by the corresponding weights, and then added or combined with the axis offset sequence to generate an environmental interference coupling feature vector. In order to ensure that the environmental interference coupling feature vector can be effectively jointly modeled with the rotating body attitude evolution feature, the dimension of the environmental interference coupling feature vector is matched with the dimension of the rotating body attitude evolution feature.

[0033] Step S300: calling a pre-trained rotating body dynamics prediction network to perform joint dynamics modeling on the rotating body attitude evolution feature and the environmental interference coupling feature, and generating a state prediction sequence of the rotating body process. In actual application, the rotating body attitude evolution feature and the environmental interference coupling feature are input into the pre-trained rotating body dynamics prediction network as input data. The network processes and analyzes the input data according to the pre-learned knowledge and rules to establish a joint dynamics model. Through the model, the state parameters of the bridge at different future time points are predicted to generate a state prediction sequence of the rotating body process.

[0034] Specifically, step S300 can include the following steps S310-S360: step S310: inputting the rotating body attitude evolution feature and the environmental interference coupling feature into the feature fusion layer of the rotating body dynamics prediction network, merging the rotating body attitude evolution feature and the environmental interference coupling feature into a joint input feature vector through feature concatenation, and the channel number of the joint input feature vector is the sum of the channel number of the rotating body attitude evolution feature and the channel number of the environmental interference coupling feature. In actual operation, the rotating body attitude evolution feature and the environmental interference coupling feature are input into the feature fusion layer of the rotating body dynamics prediction network. Assuming that the channel number of the rotating body attitude evolution feature is m and the channel number of the environmental interference coupling feature is n. Through the feature concatenation operation, the rotating body attitude evolution feature and the environmental interference coupling feature are arranged in the channel dimension in turn to form a new feature vector, i.e., a joint input feature vector. The channel number of the joint input feature vector is m+n. For example, if the rotating body attitude evolution feature is represented as a vector of length m and the environmental interference coupling feature is represented as a vector of length n, then the joint input feature vector will be a vector of length m+n.

[0035] Step S320: The dynamic characteristic extraction processing of the spatiotemporal correlation modeling layer of the bridge rotation dynamics prediction network is performed on the joint input feature vector. The dynamic graph convolution module is used to capture the time-varying topological relationship of the bridge rotation structure. At the same time, the time sequence attention mechanism is used to adjust the feature response strength of different stages of the rotation process, and a spatiotemporal correlation feature map is generated.

[0036] The spatiotemporal correlation modeling layer is a layer in the bridge rotation dynamics prediction network for processing spatiotemporal correlation information. The dynamic characteristic extraction processing is an operation of extracting feature information reflecting the dynamic changes of the bridge rotation process from the joint input feature vector. The dynamic graph convolution module is a convolution module for processing graph structure data, which can capture the topological relationship changes of the bridge rotation structure at different time points. The time sequence attention mechanism is a mechanism for adjusting the importance of features at different time points, which can adjust the response strength of the features according to different stages of the rotation process. The spatiotemporal correlation feature map is a feature map obtained after processing by this layer, which can reflect the spatiotemporal correlation information of the bridge rotation process.

[0037] Specifically, step S320 can include the following steps S321-S326: Step S321: The joint input feature vector is reconstructed into a three-dimensional feature tensor, and the dimensions of the three-dimensional feature tensor correspond to the time dimension, the space dimension, and the feature channel dimension, respectively. The three-dimensional feature tensor is a data structure with three dimensions, which can better represent the spatiotemporal information in the joint input feature vector. The time dimension represents the change of data over time, the space dimension represents the spatial distribution information of the bridge rotation structure, and the feature channel dimension represents different types of feature components. In the reconstruction process, the joint input feature vector is converted into a three-dimensional feature tensor according to its length and data characteristics. For example, if the joint input feature vector is collected at multiple time points, each time point contains feature information of multiple spatial positions, and each spatial position has multiple feature components, it can be reconstructed into a three-dimensional feature tensor.

[0038] Step S322: Construct an initial topology adjacency matrix based on the design parameters of the bridge rotation structure, update the edge weights of the adjacency matrix in real time according to the rotation angle change, and generate a time-varying topology matrix that dynamically changes with the rotation process. The initial topology adjacency matrix is a matrix constructed according to the design parameters of the bridge rotation structure, which is used to represent the connection relationship between the nodes of the bridge structure. The rotation angle change is the real-time change of the rotation angle of the bridge during the rotation process. The edge weight is the numerical value in the adjacency matrix that represents the connection strength between nodes. The time-varying topology matrix is a topology adjacency matrix that dynamically changes with the rotation process, which can reflect the topology relationship change of the bridge rotation structure at different rotation angles. When constructing the initial topology adjacency matrix, according to the design drawings and parameters of the bridge rotation structure, the connection relationship between each node in the bridge structure is determined. If there is a connection between two nodes, the corresponding element in the adjacency matrix is assigned a value of 1; otherwise, it is assigned a value of 0. As the bridge rotation process proceeds, the rotation angle will change. According to the real-time change of the rotation angle, the weight of the edge in the adjacency matrix is adjusted. For example, when the bridge rotates to a certain angle, the connection strength between some nodes may change, and the value of the corresponding element in the adjacency matrix can be adjusted according to this change. By continuously updating the weight of the edge, a time-varying topology matrix that dynamically changes with the rotation process is generated.

[0039] Step S323: Input the three-dimensional feature tensor and the time-varying topology matrix into the dynamic graph convolution module, model the structural correlation in the spatial dimension through graph convolution operation, and generate a structural correlation feature map containing structural dynamic change features.

[0040] The dynamic graph convolution module is a convolution module specially designed for processing graph structure data, which can perform convolution operations on nodes and edges in the graph. Graph convolution operation is a convolution operation performed on graph structure, which can capture the structural correlation between nodes. The structural correlation feature map is a feature map obtained through graph convolution operation, which reflects the spatial correlation information and dynamic change features of the bridge rotation structure. When performing graph convolution operation, the three-dimensional feature tensor represents the feature information of each node, and the time-varying topology matrix represents the connection relationship between nodes. The dynamic graph convolution module will aggregate and update the node features in the three-dimensional feature tensor according to the time-varying topology matrix. Through multiple graph convolution operations, the long-distance dependence relationship between nodes and the structural dynamic change features are captured. Finally, a structural correlation feature map containing structural dynamic change features is generated.

[0041] Step S324: Construct a time sequence attention mechanism module, calculate the correlation weight between the current timestamp and the historical timestamp feature vector based on the time sequence constraint of the rotation process, generate a time sequence attention weight matrix, and ensure that the attention calculation does not depend on the feature information of future time.

[0042] The time sequence attention mechanism module is a module for implementing the time sequence attention mechanism. The time sequence constraint of the rotation process is that the time sequence needs to be followed when calculating the attention weight, and the feature information of the future time cannot be used. The correlation weight is a numerical value for measuring the degree of correlation between the feature vector of the current timestamp and the historical timestamp. The time sequence attention weight matrix is a matrix composed of these correlation weights. When constructing the time sequence attention mechanism module, a model architecture based on the attention mechanism can be used, such as the attention mechanism in the Transformer. Specifically, for the feature vector of the current timestamp, the similarity between it and all historical timestamp feature vectors is calculated, for example, the similarity score is obtained by dot product operation. Then, the similarity scores are normalized by softmax to obtain the correlation weights. Arranging these correlation weights in time sequence can generate the time sequence attention weight matrix.

[0043] Step S325: Apply the time sequence attention weight matrix to the structure correlation feature map to dynamically adjust the feature contribution degree of the rotation start stage, the uniform speed stage and the deceleration stage, and enhance the feature response intensity of the rotation start stage and the deceleration stage. Applying the time sequence attention weight matrix to the structure correlation feature map is a weighted multiplication operation of the time sequence attention weight matrix and the feature vector in the structure correlation feature map. The rotation start stage, the uniform speed stage and the deceleration stage are three different stages in the bridge rotation process, and in different stages, the dynamic characteristics and attitude change of the bridge are different. The feature contribution degree is the contribution degree of different features in the structure correlation feature map to the final prediction result. In actual operation, according to the weight value in the time sequence attention weight matrix, the feature vector of each time point in the structure correlation feature map is weighted. In the rotation start stage and the deceleration stage, due to the relatively violent attitude change of the bridge, the features of these stages have a greater impact on the prediction result. Therefore, the time sequence attention mechanism can enhance the response intensity of the features of these stages by adjusting the weight value, so that they have greater contribution degree in the final prediction. In the uniform speed stage, the attitude change of the bridge is relatively small, and the response intensity of the features is relatively weak.

[0044] Step S326: Integrate the original joint input feature vector and the attention weighted structure correlation feature map through residual connection, adjust the channel dimension through 1x1 convolution layer, and generate a spatiotemporal correlation feature map containing time-varying topological relationship and time sequence characteristics.

[0045] In the feature integration process, the original joint input feature vector is first element-wise added to the structure-related feature map weighted by attention to obtain a residual connected feature map. This addition operation can retain the information in the original joint input feature vector while fusing the time-varying topological relationship and time sequence characteristic information in the structure-related feature map weighted by attention. Then, the residual connected feature map is input into a 1x1 convolution layer. The 1x1 convolution layer adjusts the channel dimension of the feature map so that the number of channels meets the input requirements of the subsequent network layer. In this way, a spatiotemporal correlation feature map containing time-varying topological relationship and time sequence characteristics is generated.

[0046] Step S330: Perform multi-scale feature fusion processing on the spatiotemporal correlation feature map, cross-scale connect feature maps of different levels, balance the contribution of local detail features and global trend features through a feature weighting algorithm, and generate a fusion feature tensor. Multi-scale feature fusion processing is an operation of fusing feature information of different scales. Feature maps of different levels are obtained in different layers of the neural network, and these feature maps have different scales and resolutions and contain information of different levels. Cross-scale connection is an operation of connecting feature maps of different levels in the spatial dimension. The feature weighting algorithm is an algorithm for determining the weight of different features in the fusion process, which can balance the contribution of local detail features and global trend features. The fusion feature tensor is a tensor containing multi-scale feature information obtained after multi-scale feature fusion processing. In multi-scale feature fusion processing, first, the spatiotemporal correlation feature map is extracted at different levels to obtain feature maps of different scales. For example, local detail information can be obtained in the shallow layer of the neural network, and global trend information can be obtained in the deep layer. Then, the feature maps of different levels are cross-scale connected. The local detail features in the shallow layer and the global trend features in the deep layer can be connected by using a skip connection or the like, so that the feature information is transmitted and fused between different scales. Then, the feature weighting algorithm is used to determine the weight of different features in the fusion process. For example, the weight can be determined according to the variance, information entropy, or other indicators of the features, so that the local detail features and the global trend features can be reasonably reflected in the fusion feature tensor. Finally, the feature maps of different levels are weighted and summed according to the determined weight to generate a fusion feature tensor.

[0047] Step S340: Input the fusion feature tensor into the dynamic memory processing layer of the body rotation dynamics prediction network, construct a state cache unit and a trend memory unit through a hierarchical memory mechanism, store the historical state information in the body rotation process and update the current state representation, and generate a dynamic evolution feature sequence.

[0048] The dynamic memory processing layer is a layer in the rotor dynamics prediction network, used to process the fusion feature tensor and generate a dynamic evolution feature sequence. The hierarchical memory mechanism is a mechanism for storing and updating historical state information, including a state cache unit and a trend memory unit. The state cache unit is used to store the recent historical state information in the rotor process, and the trend memory unit is used to store the historical trend information of the rotor process. The current state representation is generated according to the historical state information and the current input fusion feature tensor, and can reflect the current state of the rotor process. The dynamic evolution feature sequence is obtained after processing by the dynamic memory processing layer, which contains the dynamic fluctuations and historical evolution trend information of the rotor process. In practical applications, the dynamic memory processing layer can adopt an architecture based on a recurrent neural network (RNN), such as a long short-term memory network (LSTM), a gated recurrent unit (GRU), etc. These networks have a memory function and can store and update historical state information. The state cache unit and the trend memory unit can be implemented through the hidden state and the cell state of the network.

[0049] Specifically, step S340 can include the following steps S341-S346: step S341: expanding the fusion feature tensor into a time series feature vector sequence in the time dimension, each time series feature vector corresponding to the fusion feature representation of a timestamp in the rotor process. The time dimension expansion is an operation of splitting the fusion feature tensor according to the time dimension. The time series feature vector sequence is a sequence composed of a series of time series feature vectors, each time series feature vector corresponding to the fusion feature at a timestamp in the rotor process. In actual operation, according to the dimension information of the fusion feature tensor, it is split according to the time dimension. For example, if the dimension of the fusion feature tensor is (T, S, C), where T represents the time dimension, S represents the spatial dimension, and C represents the feature channel dimension, then after expanding it in the time dimension, a time series feature vector sequence with a length of T is obtained. The dimension of each time series feature vector is (S, C), representing the fusion feature information at that timestamp. In this way, the fusion feature tensor is converted into a time series feature vector sequence for subsequent processing and analysis.

[0050] Step S342: initializing the state cache unit and the trend memory unit of the dynamic memory processing layer, setting the memory window parameter and the memory decay parameter, the capacity of the state cache unit being consistent with the dimension of the time series feature vector, and the capacity of the trend memory unit being a preset multiple of the capacity of the state cache unit.

[0051] The initialization is an initial setting operation of the state cache unit and the trend memory unit. The state cache unit is a unit for temporarily storing recent historical state information, and the trend memory unit is a unit for storing long-term historical trend information. The memory window parameter is a time range of the historical state information stored in the state cache unit, which determines the number of historical data that the state cache unit can store. The memory decay parameter is a parameter for controlling the decay speed of the historical information in the trend memory unit. In the initialization process, first, the capacities of the state cache unit and the trend memory unit are determined. The capacity of the state cache unit is consistent with the dimension of the time series feature vector, so as to ensure that the feature information of each timestamp can be accurately stored. The capacity of the trend memory unit is a preset multiple of the capacity of the state cache unit, for example, it can be set to 3 times, so that the trend memory unit can store more historical trend information. Then, the memory window parameter and the memory decay parameter are set. The memory window parameter can be adjusted according to actual needs, for example, it can be set to contain the historical state information of the last 10 timestamps. The memory decay parameter can be set according to the importance and timeliness of the historical information, for example, it can be set to a small positive number, so that the historical information gradually decays over time.

[0052] Step S343: The time series feature vectors are sequentially input into the state cache unit in chronological order, and the cache content is updated in real time through the sliding window mechanism. When a new feature vector is input, the earliest historical feature vector in the window is removed, and the latest feature vector in the window is retained.

[0053] In actual operation, the time series feature vectors are sequentially input into the state cache unit according to the chronological order of the timestamps. When the first feature vector is input, the state cache unit starts to store the feature vector. As new feature vectors are continuously input, the sliding window slides forward. When the number of feature vectors in the window reaches the number set by the memory window parameter, each time a new feature vector is input, the earliest historical feature vector in the window is removed, and the latest feature vector in the window is retained. In this way, the state cache unit can update the cache content in real time, and always store the latest historical state information.

[0054] Step S344: The output features of the state cache unit are processed by a trend extraction module to perform a moving average, and a trend feature vector is calculated by combining a memory decay parameter to update the storage content of the trend memory unit. The trend extraction module is a module for extracting historical trend information from the output features of the state cache unit. The moving average processing is an operation of averaging the output features of the state cache unit, which can smooth the data and reduce the influence of noise. The memory decay parameter is used to control the weight of historical information in trend calculation, so that the influence of older information on the trend gradually decreases. The trend feature vector is a vector that can reflect historical trend information calculated by the trend extraction module. In actual operation, the trend extraction module can use a moving average algorithm or an exponential moving average algorithm.

[0055] Step S345: The window-in feature vector output by the state cache unit and the historical trend feature vector output by the trend memory unit are weighted and fused based on a gating fusion mechanism to generate a comprehensive state representation vector for the current timestamp, and the weighting coefficient is dynamically adjusted according to the rotation phase feature, which is adjusted to be biased towards the window-in feature vector in the start-up phase and the deceleration phase, and is adjusted to be biased towards the historical trend feature vector in the uniform speed phase.

[0056] The gating fusion mechanism is a mechanism for weighted fusion of different feature vectors. Weighted fusion is an operation of merging the window-in feature vector output by the state cache unit and the historical trend feature vector output by the trend memory unit according to the set weight. The comprehensive state representation vector is obtained after weighted fusion, which can comprehensively reflect the current state and historical trend information of the rotation process. The weighting coefficient is a numerical value used to determine the weight of each feature vector, and the weighting coefficient is dynamically adjusted according to the rotation phase feature. In the start-up phase and the deceleration phase, since the rotation process changes greatly, more attention needs to be paid to the current state information, so the weighting coefficient is biased towards the window-in feature vector; in the uniform speed phase, the rotation process is relatively stable, and the historical trend information can better reflect the overall situation, so the weighting coefficient is biased towards the historical trend feature vector. In actual operation, the gating fusion mechanism can use a model architecture based on a gating unit, such as the gating mechanism in LSTM. The weighting coefficient is calculated by a gating unit, and then the window-in feature vector output by the state cache unit and the historical trend feature vector output by the trend memory unit are weighted and summed according to the weighting coefficient to obtain the comprehensive state representation vector.

[0057] Step S346: The comprehensive state representation vectors of each timestamp are arranged in chronological order to generate a dynamic evolution feature sequence containing the dynamic fluctuations and historical evolution trends in the rotation process window, and the length of the dynamic evolution feature sequence is consistent with the length of the input time series feature vector sequence.

[0058] Arranging the comprehensive state representation vectors of each timestamp in chronological order is an operation of arranging the comprehensive state representation vectors corresponding to different timestamps in chronological order. The dynamic evolution feature sequence is a feature sequence that can reflect the dynamic fluctuations and historical evolution trends in the bridge rotation process after being arranged. In actual operation, the comprehensive state representation vectors calculated at each timestamp are arranged in chronological order. Since each timestamp has a corresponding comprehensive state representation vector, the length of the dynamic evolution feature sequence is consistent with the length of the input time sequence feature vector sequence. In this way, the dynamic evolution feature sequence can completely record the state changes of the bridge during the rotation process, including dynamic fluctuation information and historical evolution trend information within the window.

[0059] Step S350: calling the sequence prediction layer of the rotation dynamics prediction network to perform future state inference processing on the dynamic evolution feature sequence, generating a rotation process state prediction sequence containing multiple future time points through a time series prediction algorithm, and each element of the state prediction sequence corresponding to a predicted pose parameter of a timestamp in the rotation process.

[0060] The sequence prediction layer is a level in the rotation dynamics prediction network for future state prediction. The future state inference processing is an operation of predicting the state of the bridge at future time points according to the current dynamic evolution feature sequence. The time series prediction algorithm is an algorithm for predicting future values of time series data, which can predict data values at future time points according to the laws and trends of historical data. The rotation process state prediction sequence is a sequence containing bridge state parameters at multiple future time points generated by the time series prediction algorithm, and each element corresponds to a predicted pose parameter such as a rotation angle and a position offset at a timestamp in the rotation process. When performing future state inference processing, the dynamic evolution feature sequence is input to the sequence prediction layer of the rotation dynamics prediction network. This layer can use time series prediction algorithms such as recurrent neural networks (RNN), long short-term memory networks (LSTM), and gated recurrent units (GRU) to process the dynamic evolution feature sequence. These algorithms learn the time-dependent relationships and laws in the dynamic evolution feature sequence and predict bridge state parameters at multiple future time points based on current state information. For example, when using an LSTM network for prediction, the LSTM unit updates its hidden state based on the input dynamic evolution feature sequence. Through continuous iteration, bridge state parameters at different timestamps in the future are predicted. Finally, these predicted parameters are arranged in chronological order to generate a rotation process state prediction sequence containing multiple future time points.

[0061] Step S360: performing confidence calibration processing on the state prediction sequence, assigning confidence weights to the predicted pose parameters of each timestamp based on the historical prediction error distribution, and generating a state prediction sequence with confidence labels.

[0062] The confidence calibration process is an operation of evaluating and adjusting the reliability of the prediction results in the state prediction sequence. The historical prediction error distribution is the error distribution between the prediction results and the actual results in the past prediction process. The confidence weight is a numerical value used to represent the reliability of each timestamped prediction attitude parameter, and the higher the weight, the more reliable the prediction result. The state prediction sequence with confidence labels is a sequence obtained by assigning confidence weights to each prediction attitude parameter in the state prediction sequence. When performing the confidence calibration process, the historical prediction error distribution is first collected and analyzed. The prediction error of each timestamp can be calculated by comparing the past prediction results with the actual results. The distribution of these errors is calculated, such as calculating the mean, variance, and other statistics of the errors. Then, according to the historical prediction error distribution, a confidence weight is assigned to each timestamped prediction attitude parameter in the state prediction sequence. For example, if the historical prediction error of a certain timestamp is small, it means that the prediction result of that timestamp is more reliable, and a higher confidence weight can be assigned; on the contrary, if the historical prediction error is large, a lower confidence weight is assigned. Finally, the confidence weight is associated with the prediction attitude parameter in the state prediction sequence to generate the state prediction sequence with confidence labels.

[0063] Step S400: Deviation comparison between the state prediction sequence and the preset rotation target trajectory sequence to generate a rotation deviation risk assessment matrix containing position deviation and attitude deviation.

[0064] The state prediction sequence is a sequence of state parameters of the bridge at different time points in the rotation process predicted by the rotation dynamics prediction network, containing predicted position, attitude, and other information. The preset rotation target trajectory sequence is a sequence of ideal state parameters that the bridge should reach during the rotation process, which specifies the target position and attitude of the bridge at different time points. Deviation comparison is an operation of comparing the state prediction sequence with the preset rotation target trajectory sequence to calculate the difference between the two. The position deviation is the deviation between the predicted position and the target position, and the attitude deviation is the deviation between the predicted attitude and the target attitude. The rotation deviation risk assessment matrix is a matrix generated by deviation comparison for assessing the deviation risk in the bridge rotation process, containing information such as position deviation and attitude deviation at different time points.

[0065] Specifically, step S400 can include the following steps S410-S460: step S410: time axis unification processing is performed on the state prediction sequence with confidence labels and the rotation target trajectory sequence, the timestamps of the two sequences are aligned to the same time reference system, and the prediction sequence and the target sequence form a point-by-point correspondence relationship in the time dimension. The time axis unification processing is an operation of adjusting the timestamps of the state prediction sequence with confidence labels and the rotation target trajectory sequence to the same time reference system. When performing the time axis unification processing, the time range and the time interval of the two sequences are first determined. Then, an optimal alignment path between the two sequences can be calculated using, for example, a dynamic time warping algorithm. The algorithm finds a mapping relationship between the timestamps of the two sequences, so that they can form a point-by-point correspondence relationship in the time dimension.

[0066] Step S420: from the state prediction sequence after time axis unification, the prediction position coordinates, the prediction attitude angles, and the prediction confidence vectors corresponding to each timestamp are extracted; from the rotation target trajectory sequence after time axis unification, the target position coordinates and the target attitude angles corresponding to each timestamp are extracted.

[0067] The prediction position coordinates are the three-dimensional coordinate values of the predicted position of the bridge corresponding to each timestamp in the state prediction sequence, reflecting the spatial position of the bridge in the predicted state. The prediction attitude angle is the angle value of the predicted attitude of the bridge corresponding to each timestamp in the state prediction sequence, such as the azimuth angle, the pitch angle, the roll angle, etc., describing the attitude direction of the bridge in the predicted state. The prediction confidence vector is the confidence weight vector of the prediction result corresponding to each timestamp in the state prediction sequence, indicating the reliability of the prediction result. The target position coordinates are the three-dimensional coordinate values of the target position of the bridge corresponding to each timestamp in the rotation target trajectory sequence, specifying the ideal position that the bridge should reach at the time point. The target attitude angle is the angle value of the target attitude of the bridge corresponding to each timestamp in the rotation target trajectory sequence, specifying the ideal attitude that the bridge should reach at the time point.

[0068] Step S430: calculate the three-dimensional deviation vector of the prediction position coordinates and the target position coordinates, the three-dimensional angle deviation vector of the prediction attitude angle and the target attitude angle at each timestamp, and simultaneously calculate the first time derivative of the three-dimensional deviation vector as the position deviation rate, and the first time derivative of the three-dimensional angle deviation vector as the attitude deviation rate.

[0069] The three-dimensional deviation vector is a vector describing the deviation of the predicted position coordinates from the target position coordinates in three-dimensional space, including deviation values in X-axis, Y-axis and Z-axis directions. The three-dimensional angle deviation vector is a vector describing the deviation of the predicted attitude angle from the target attitude angle in three-dimensional space, including deviation values of roll angle, pitch angle and yaw angle. The position deviation rate is the first-order time derivative of the three-dimensional deviation vector, reflecting the speed of change of the position deviation with time. The attitude deviation rate is the first-order time derivative of the three-dimensional angle deviation vector, reflecting the speed of change of the attitude deviation with time. In calculating the three-dimensional deviation vector, the differences between the predicted position coordinates and the target position coordinates in X-axis, Y-axis and Z-axis directions are calculated respectively, and these differences are combined into a three-dimensional vector. In calculating the three-dimensional angle deviation vector, the differences between the predicted attitude angle and the target attitude angle in roll angle, pitch angle and yaw angle directions are calculated respectively, and these differences are combined into a three-dimensional vector. For the position deviation rate and the attitude deviation rate, numerical differentiation methods can be used for calculation, such as finite difference method, which calculates the difference between the deviation vectors at adjacent time stamps, and then divides by the time interval to obtain the deviation rate.

[0070] In step S430, the three-dimensional deviation vector of the predicted position coordinates and the target position coordinates at each time stamp is calculated, which can specifically include the following steps S431-S436: step S431: extracting the X-axis coordinate predicted value, the Y-axis coordinate predicted value, the Z-axis coordinate predicted value and the corresponding coordinate confidence component from the predicted position coordinates of the state prediction sequence with confidence marking, to form a three-dimensional predicted coordinate vector and a coordinate confidence vector.

[0071] The predicted position coordinates are the coordinate information of the predicted position of the bridge corresponding to each time stamp in the state prediction sequence, including coordinate values in X-axis, Y-axis and Z-axis directions. The coordinate confidence component is the confidence evaluation value of the predicted coordinate on each coordinate axis, reflecting the reliability of the predicted value. The three-dimensional predicted coordinate vector is a vector combining the coordinate predicted values of X-axis, Y-axis and Z-axis, used to represent the spatial position of the bridge in the predicted state. The coordinate confidence vector is a vector composed of the coordinate confidence components corresponding to the three coordinate axes, reflecting the confidence level of each coordinate predicted value.

[0072] Step S432: extracting the X-axis coordinate target value, the Y-axis coordinate target value and the Z-axis coordinate target value from the target position coordinates of the swivel target trajectory sequence to form a three-dimensional target coordinate vector.

[0073] The swivel target trajectory sequence defines the ideal position that the bridge should reach at each timestamp during the swivel process, wherein the target position coordinates include the coordinate target values of the X-axis, the Y-axis and the Z-axis. The three-dimensional target coordinate vector is a vector composed of the target values of the three coordinate axes, used to represent the spatial position of the bridge in the ideal state. In the extraction process, for each timestamp in the swivel target trajectory sequence, the X-axis coordinate target value, the Y-axis coordinate target value and the Z-axis coordinate target value are taken from the target position coordinate data respectively.

[0074] Step S433: Obtain the position sensitivity parameter of the current stage of the swivel, which is dynamically determined according to the stiffness parameters of the swivel structure in different directions. The smaller the stiffness parameter, the higher the value of the position sensitivity parameter.

[0075] The stiffness of the swivel structure in different directions is different, and the stiffness reflects the ability of the structure to resist deformation. The position sensitivity parameter is used to measure the sensitivity of the swivel structure to position deviation in different directions. Since the direction with smaller stiffness is more prone to deformation under stress, a small change in position in these directions may have a greater impact on the entire swivel process, so the smaller the stiffness parameter, the higher the value of the position sensitivity parameter.

[0076] Step S434: Calculate the difference between the corresponding coordinate axis components of the three-dimensional predicted coordinate vector and the three-dimensional target coordinate vector to obtain the X-axis original deviation, the Y-axis original deviation and the Z-axis original deviation.

[0077] The three-dimensional predicted coordinate vector represents the predicted position of the bridge, and the three-dimensional target coordinate vector represents the ideal position of the bridge. By calculating the difference between the corresponding coordinate axis components of the two vectors, the original deviation value on each coordinate axis can be obtained, which reflects the direct difference between the predicted position and the target position in each direction. In the calculation process, the X-axis coordinate prediction value of the three-dimensional predicted coordinate vector is subtracted from the X-axis coordinate target value of the three-dimensional target coordinate vector to obtain the X-axis original deviation; the Y-axis coordinate prediction value is subtracted from the Y-axis coordinate target value to obtain the Y-axis original deviation; and the Z-axis coordinate prediction value is subtracted from the Z-axis coordinate target value to obtain the Z-axis original deviation.

[0078] Step S435: Multiply the X-axis original deviation, the Y-axis original deviation and the Z-axis original deviation respectively by the position sensitivity parameter and the coordinate confidence component of the corresponding coordinate axis to obtain the X-axis weighted deviation, the Y-axis weighted deviation and the Z-axis weighted deviation.

[0079] The position sensitivity parameter considers the sensitivity of the rotating body structure to the position deviation in different directions, and the coordinate confidence component reflects the reliability of each coordinate prediction value. By multiplying the original deviation with the position sensitivity parameter and the coordinate confidence component, the structure sensitivity and the prediction reliability can be considered comprehensively to obtain a more practically meaningful weighted deviation value. In the calculation, the X-axis original deviation is multiplied by the position sensitivity parameter and the coordinate confidence component of the X-axis to obtain the X-axis weighted deviation; similarly, the Y-axis weighted deviation and the Z-axis weighted deviation are calculated. For example, if the position sensitivity parameter of the X-axis is s x , the coordinate confidence component is c xp , and the X-axis original deviation is x p -x t , then the X-axis weighted deviation is (x p -x t ) × s x × c xp .

[0080] Step S436: combining the X-axis weighted deviation, the Y-axis weighted deviation and the Z-axis weighted deviation into a three-dimensional deviation vector, each component of the three-dimensional deviation vector comprehensively reflecting the size of the deviation in the direction, the structure sensitivity and the prediction reliability.

[0081] The three-dimensional deviation vector is a vector combining the weighted deviations of the three coordinate axes, which comprehensively considers the size of the deviation in each direction, the sensitivity of the rotating body structure in the direction and the reliability of the prediction value. Through this vector, the deviation between the predicted position of the bridge and the target position can be more comprehensively and accurately evaluated. The calculated X-axis weighted deviation, Y-axis weighted deviation and Z-axis weighted deviation are combined in order to form a three-dimensional deviation vector. For example, if the X-axis weighted deviation is d x , the Y-axis weighted deviation is d y , and the Z-axis weighted deviation is d n , then the three-dimensional deviation vector is (d x , d y , d n ). Each component of the vector contains information such as the size of the deviation in the direction, the structure sensitivity and the prediction reliability.

[0082] In step S430, the three-dimensional angle deviation vector of the predicted attitude angle and the target attitude angle at each timestamp is calculated, which can include the following steps S437-S4312: Step S437: extracting the roll angle prediction value, the pitch angle prediction value, the yaw angle prediction value and the corresponding attitude confidence component from the predicted attitude angle of the state prediction sequence with confidence marking to form a three-dimensional predicted attitude angle vector and an attitude confidence vector.

[0083] The predicted attitude angle is the angle information of the bridge predicted attitude corresponding to each timestamp in the state prediction sequence, including three angle values of roll angle, pitch angle and yaw angle. The attitude confidence component is the confidence evaluation value of each attitude angle prediction value, reflecting the reliability of the prediction value. The three-dimensional predicted attitude angle vector is a vector composed of the predicted values of roll angle, pitch angle and yaw angle, used to represent the attitude direction of the bridge in the predicted state. The attitude confidence vector is a vector composed of the confidence components corresponding to the three attitude angles, reflecting the confidence level of each attitude angle prediction value. In actual operation, for each timestamp in the state prediction sequence with confidence marking, the roll angle prediction value, pitch angle prediction value and yaw angle prediction value are extracted from the predicted attitude angle data respectively. At the same time, the attitude confidence components corresponding to the three attitude angles are obtained. For example, at a certain timestamp, the roll angle prediction value r p , the pitch angle prediction value p p , and the yaw angle prediction value y p are extracted, and the corresponding attitude confidence components are c rp , c pp , and c yp . The r p , p p , and y p are combined into a three-dimensional predicted attitude angle vector (r p , p p , y p ), and the c rp , c pp , and c yp are combined into an attitude confidence vector (c rp , c pp , c yp ).

[0084] Step S438: Extracting roll angle target value, pitch angle target value and yaw angle target value from target attitude angle in the swivel target trajectory sequence to form a three-dimensional target attitude angle vector.

[0085] The swivel target trajectory sequence defines the ideal attitude that the bridge should reach at each timestamp during the swivel process, and the target attitude angle therein contains the target values of roll angle, pitch angle and yaw angle. The three-dimensional target attitude angle vector is a vector composed of the target values of the three attitude angles, used to represent the attitude direction of the bridge in the ideal state. In the extraction process, for each timestamp in the swivel target trajectory sequence, the roll angle target value, pitch angle target value and yaw angle target value are taken out from the target attitude angle data respectively. For example, at the timestamp corresponding to step S437, the roll angle target value r t , the pitch angle target value p t , and the yaw angle target value y t are extracted, and they are combined into a three-dimensional target attitude angle vector (rt , p t , y t ).

[0086] Step S439: Construct a posture coupling matrix based on the rotational inertia parameters of the rotating body structure. The element values of the posture coupling matrix represent the dynamic coupling relationship between different posture angles. The coupling coefficients between the roll angle and the pitch angle, and between the pitch angle and the yaw angle are dynamically determined according to the inertia parameters of the rotating body main shaft.

[0087] The rotational inertia parameters of the rotating body structure reflect the inertia characteristics of the structure when rotating around the shaft. There is a dynamic coupling relationship between different posture angles, that is, the change of one posture angle may cause the change of other posture angles. The posture coupling matrix is a matrix used to describe this coupling relationship, and its element values reflect the degree of mutual influence between different posture angles. The coupling coefficients between the roll angle and the pitch angle, and between the pitch angle and the yaw angle are closely related to the inertia parameters of the rotating body main shaft. Different inertia parameters of the rotating body main shaft result in different degrees of coupling between different posture angles. In practical applications, detailed analysis of the rotational inertia parameters of the rotating body structure is required, especially the inertia parameters of the rotating body main shaft. According to these parameters, the values of the elements in the posture coupling matrix are dynamically determined, especially the coupling coefficients between the roll angle and the pitch angle, and between the pitch angle and the yaw angle. For example, when the inertia of the rotating body main shaft in a certain direction is large, the coupling coefficient between the corresponding posture angles may be small; on the contrary, the coupling coefficient may be large.

[0088] Step S4310: Calculate the difference between the corresponding angle components of the three-dimensional predicted posture angle vector and the three-dimensional target posture angle vector to obtain the roll angle original deviation, the pitch angle original deviation, and the yaw angle original deviation. The original deviations are coupled and corrected by the posture coupling matrix to generate the coupled and corrected roll angle deviation, the coupled and corrected pitch angle deviation, and the coupled and corrected yaw angle deviation. The three-dimensional predicted posture angle vector represents the predicted posture of the bridge, and the three-dimensional target posture angle vector represents the ideal posture of the bridge. By calculating the difference between the corresponding angle components of the two vectors, the original deviation values of each posture angle can be obtained, which reflect the direct differences between the predicted posture and the target posture in each angle. Since there is a coupling relationship between different posture angles, the original deviations need to be corrected using the posture coupling matrix. The element values in the posture coupling matrix reflect the degree of mutual influence between different posture angles. In the correction process, the roll angle original deviation, the pitch angle original deviation, and the yaw angle original deviation are substituted into the posture coupling matrix for calculation, considering the coupling effect between different posture angles, to obtain the coupled and corrected deviation values. For example, the original deviation of the roll angle may be affected by the pitch angle and the yaw angle. Through the operation of the posture coupling matrix, this influence can be considered to obtain more accurate coupled and corrected roll angle deviation.

[0089] Step S4311: Obtain the attitude stability parameter of the current stage of the rotating body. The attitude stability parameter is dynamically adjusted according to the angular acceleration of the attitude change in the rotating process. The greater the angular acceleration, the higher the value of the attitude stability parameter in this stage.

[0090] The attitude stability parameter is used to measure the sensitivity of the rotating structure to the attitude deviation in different stages. In the rotating process, the angular acceleration of the attitude change reflects the degree of the attitude change. When the angular acceleration is large, the attitude change is relatively violent, and the rotating structure is more sensitive to the attitude deviation in this stage, so the attitude stability parameter has a higher value. Conversely, when the angular acceleration is small, the attitude stability parameter has a lower value. In practical applications, the angular acceleration of the attitude change in the rotating process needs to be monitored in real time. According to the size of the angular acceleration, the attitude stability parameter is dynamically adjusted. For example, in the starting stage and the deceleration stage of the rotating process, the angular acceleration of the attitude change is usually large, and at this time the attitude stability parameter is set to a high value. In the uniform speed stage, the angular acceleration of the attitude change is relatively small, and the attitude stability parameter is relatively low.

[0091] Step S4312: Multiply the coupling corrected roll angle deviation, the coupling corrected pitch angle deviation, and the coupling corrected yaw angle deviation by the attitude stability parameter and the attitude confidence component of the corresponding attitude angle, respectively, to form a three-dimensional angle deviation vector. Each component of the three-dimensional angle deviation vector comprehensively reflects the coupling corrected angle deviation, the attitude stability requirement, and the prediction reliability.

[0092] The attitude stability parameter considers the sensitivity of the rotating structure to the attitude deviation in different stages, and the attitude confidence component reflects the reliability of the predicted value of each attitude angle. By multiplying the coupling corrected deviation by the attitude stability parameter and the attitude confidence component, the coupling corrected angle deviation, the attitude stability requirement, and the prediction reliability can be considered comprehensively to obtain a more practically meaningful deviation value. The coupling corrected roll angle deviation, the coupling corrected pitch angle deviation, and the coupling corrected yaw angle deviation are calculated and multiplied by the attitude stability parameter and the attitude confidence component of the corresponding attitude angle, respectively, and then combined in order to form a three-dimensional angle deviation vector. For example, if the coupling corrected roll angle deviation is δr, the corresponding attitude stability parameter is s r , the attitude confidence component is c rp , the coupling corrected pitch angle deviation is δp, the corresponding attitude stability parameter is s p , the attitude confidence component is c pp , the coupling corrected yaw angle deviation is δy, the corresponding attitude stability parameter is s y , and the attitude confidence component is c yp , then the three-dimensional angle deviation vector is (δr×s r ×crp , δp x s p x c pp , δy x s y x c yp Each component of this vector contains information about the angle deviation size after coupling correction, attitude stability requirements, and prediction reliability, providing an important basis for subsequent deviation risk assessment.

[0093] Step S440: Construct a deviation risk propagation model based on the dynamic characteristic parameters of the rotating structure, which include the stiffness distribution parameters and mass inertia parameters of the rotating structure. Input the prediction confidence vector, three-dimensional deviation vector, position deviation rate, three-dimensional angle deviation vector, and attitude deviation rate into the deviation risk propagation model, and output the deviation risk propagation coefficient of each timestamp, which represents the influence degree of the current deviation on the subsequent rotating process.

[0094] The dynamic characteristic parameters of the rotating structure are important parameters for describing the mechanical properties of the rotating structure. The stiffness distribution parameters reflect the ability of the rotating structure to resist deformation at different parts, and the mass inertia parameters represent the inertia characteristics of the rotating structure during motion. The deviation risk propagation model is a model for evaluating the influence of the current deviation on the subsequent rotating process, which is constructed based on the dynamic characteristic parameters of the rotating structure, and can consider the influence of the mechanical properties of the rotating structure on deviation propagation. The prediction confidence vector reflects the reliability of the prediction results, the three-dimensional deviation vector and the three-dimensional angle deviation vector represent the deviation of position and attitude respectively, and the position deviation rate and the attitude deviation rate reflect the change of deviation with time. By inputting these information into the deviation risk propagation model, the model analyzes and calculates based on the dynamic characteristic parameters of the rotating structure. For example, when the stiffness distribution of the rotating structure is uneven, the deviation of some parts may be more easily propagated and amplified; the motion state change of the rotating structure with larger mass inertia may be more complex when affected by deviation. Through the calculation of the model, the deviation risk propagation coefficient of each timestamp is output. The larger the coefficient, the greater the influence of the current deviation on the subsequent rotating process.

[0095] Step S450: Perform probabilistic reasoning on the deviation risk propagation coefficient, position deviation rate, and attitude deviation rate through the Bayesian network to generate a deviation risk transition probability matrix, and analyze the spatial correlation features of the covariance matrix of the three-dimensional deviation vector and the three-dimensional angle deviation vector.

[0096] Bayesian networks are graphical models based on probabilistic reasoning, which can infer the probability of unknown events according to known information and conditional probability relationships. The bias risk propagation coefficient, the position bias rate, and the attitude bias rate are important parameters that describe the bias situation and the bias change. By performing probabilistic reasoning on these parameters using Bayesian networks, the transition probabilities between different bias states can be analyzed. In practical applications, the bias risk propagation coefficient, the position bias rate, and the attitude bias rate are taken as input information, and the nodes and edges in the Bayesian network are used to represent the probability relationships between variables. By performing inference calculations on the network, a bias risk transition probability matrix is obtained. The elements in this matrix represent the probability of transitioning from one bias state to another, reflecting the dynamic changes of bias between different states. At the same time, the covariance matrix of the three-dimensional bias vector and the three-dimensional angle bias vector is analyzed. The covariance matrix can measure the linear correlation between two vectors, and by analyzing this matrix, the spatial correlation characteristics between the position bias and the attitude bias can be found. For example, the element values in the covariance matrix can reflect the strength of the correlation between the position bias in a certain direction and the attitude angle bias. These spatial correlation characteristics can provide important information for further understanding the propagation and impact of bias, and help to more accurately assess the bias risk during the rotation process.

[0097] Step S460: Integrate the three-dimensional bias vector, the three-dimensional angle bias vector, the bias risk propagation coefficient, the bias risk transition probability matrix, and the spatial correlation characteristics in the time dimension, the bias type dimension, and the risk propagation dimension to generate a rotation bias risk assessment matrix containing dynamic risk evolution characteristics.

[0098] The three-dimensional deviation vector and the three-dimensional angle deviation vector respectively reflect the deviation of position and attitude, the deviation risk propagation coefficient reflects the influence degree of the current deviation on the subsequent rotation process, the deviation risk transfer probability matrix describes the transfer probability of the deviation between different states, and the spatial correlation feature reveals the correlation between the position deviation and the attitude deviation. By integrating these information in the time dimension, the deviation type dimension and the risk propagation dimension, the deviation risk in the rotation process can be comprehensively and systematically evaluated. In the time dimension, the deviation and risk propagation at different time stamps are considered, reflecting the dynamic change of the deviation and risk over time. In the deviation type dimension, the position deviation and the attitude deviation are distinguished, and different types of deviation are analyzed and evaluated. In the risk propagation dimension, the propagation coefficient and the transfer probability of the deviation risk are considered, reflecting the influence of the deviation on the subsequent rotation process and the transfer of the deviation state. By integrating these information, a rotation deviation risk evaluation matrix is generated. Each element in the matrix contains the deviation risk information in the time stamp, the deviation type and the risk propagation, forming a matrix containing dynamic risk evolution characteristics. This matrix can provide important decision basis for subsequent rotation control, helping the operator to discover and handle the deviation risk in the rotation process in time.

[0099] Step S500: driving the rotation control parameter optimization network based on the rotation deviation risk evaluation matrix to perform compensation calculation, generating a rotation positioning control instruction set containing a traction force adjustment coefficient and a rotation speed correction value, and transmitting the rotation positioning control instruction set to a rotation driving system.

[0100] The rotation control parameter optimization network is a neural network for optimizing rotation control parameters according to the rotation deviation risk evaluation matrix. The compensation calculation refers to calculating the amount of adjustment needed for the rotation control parameters according to the information in the deviation risk evaluation matrix. The traction force adjustment coefficient is a coefficient for adjusting the size of the traction force in the rotation process, and the rotation speed correction value is a numerical value for correcting the rotation speed. The rotation positioning control instruction set is a set of instructions containing the traction force adjustment coefficient and the rotation speed correction value at each time stamp. The rotation driving system is a system responsible for performing the rotation operation, and the rotation positioning control instruction set is transmitted to the rotation driving system, so that the system can control the rotation process according to the instructions.

[0101] In practical applications, the swivel control parameter optimization network can adopt a neural network structure such as a multi-layer perceptron (MLP), the input is the swivel deviation risk assessment matrix, and the output is a vector containing the traction force adjustment coefficient and the rotation speed correction value. First, the swivel control parameter optimization network is trained. Using a large amount of historical swivel data, including the swivel deviation risk assessment matrix and the corresponding optimal control parameters (traction force adjustment coefficient and rotation speed correction value), the weights and biases of the network are adjusted through the back propagation algorithm, so that the network can accurately output appropriate control parameters according to the input deviation risk assessment matrix. When the current swivel deviation risk assessment matrix is obtained, it is input into the trained swivel control parameter optimization network, and the network performs forward propagation calculation to output a vector containing the traction force adjustment coefficient and the rotation speed correction value. The traction force adjustment coefficient and the rotation speed correction value of each timestamp are arranged in chronological order to form a swivel positioning control instruction set. Finally, the swivel positioning control instruction set is transmitted to the swivel driving system through the communication interface, and the driving system adjusts the traction force and rotation speed according to the instructions to realize accurate control of the swivel process, so that the bridge can be accurately positioned.

[0102] Specifically, step S500 can include steps S510-S560: step S510: input the swivel deviation risk assessment matrix into the deviation feature extraction layer of the swivel control parameter optimization network, and perform deep feature mining on the swivel deviation risk assessment matrix through the convolutional neural network to generate a high-dimensional deviation feature vector. The swivel deviation risk assessment matrix contains rich deviation risk information, but this information exists in the form of a matrix and needs to be further processed and analyzed to extract useful features. The deviation feature extraction layer is a level in the swivel control parameter optimization network, which uses a convolutional neural network to process the swivel deviation risk assessment matrix.

[0103] In the deviation feature extraction layer, the convolutional neural network performs multiple convolution and pooling operations on the swivel deviation risk assessment matrix. Convolution operation can extract local feature patterns in the matrix, such as deviation trend under different timestamps and deviation types; the pooling operation can reduce the dimension and compress the extracted features, reducing the number of features while retaining important feature information. Through multiple convolution and pooling operations, the convolutional neural network can perform deep feature mining on the swivel deviation risk assessment matrix, converting the information in the matrix into a high-dimensional deviation feature vector. This high-dimensional deviation feature vector contains key feature information in the swivel deviation risk assessment matrix, providing a basis for subsequent risk classification and parameter calculation. For example, some elements in the vector may reflect the trend of position deviation at the corresponding timestamp, and other elements may reflect the correlation between attitude deviation and risk propagation.

[0104] Step S520: input the high-dimensional deviation feature vector into the risk grading processing layer of the rotation control parameter optimization network, calculate the probability distribution of different deviation levels through the fully connected layer and activation processing, and determine the risk level label of the current rotation deviation.

[0105] The risk grading processing layer is a layer in the rotation control parameter optimization network for grading the deviation risk. The high-dimensional deviation feature vector contains key feature information of the rotation deviation risk. After inputting it into the risk grading processing layer, the fully connected layer will further process and combine these features.

[0106] Each neuron in the fully connected layer is connected to all elements in the input high-dimensional deviation feature vector, and performs linear combination and weighted summation on the input features. In this way, the fully connected layer can integrate and transform the information in the high-dimensional deviation feature vector to obtain different feature representations. Activation processing is applied to the output results of the fully connected layer using an activation function, which can introduce non-linear factors to enable the network to learn more complex patterns and relationships. In the risk grading processing layer, the probability distribution of different deviation levels is calculated through the fully connected layer and activation processing. The deviation level can be divided according to the actual situation during the rotation process, such as low risk, medium risk, and high risk levels. According to the calculated probability distribution, the risk level label of the current rotation deviation is determined. For example, if the probability of the high-risk level is the highest, the risk level label of the current rotation deviation is determined as high risk. This risk level label provides an important basis for subsequent parameter adjustment, and different risk levels may require different control strategies.

[0107] Step S530: according to the risk level label, call the corresponding parameter adjustment rule from the pre-set control strategy library, which contains the adjustment direction of the traction force adjustment coefficient and the calculation method of the rotation speed correction value. The pre-set control strategy library is a pre-established database containing control strategies for different risk levels. According to experience and theoretical analysis during the rotation process, parameter adjustment rules for different risk levels are developed. The risk level label determines the risk degree of the current rotation deviation, and according to this label, the appropriate parameter adjustment rule can be selected from the control strategy library. The parameter adjustment rule contains the adjustment direction of the traction force adjustment coefficient and the calculation method of the rotation speed correction value. For example, when the risk level label is low risk, the parameter adjustment rule may specify that the adjustment amplitude of the traction force adjustment coefficient is small, and the calculation method of the rotation speed correction value is relatively simple; when the risk level label is high risk, the parameter adjustment rule may require a large adjustment of the traction force adjustment coefficient, and a more complex calculation method is used to determine the rotation speed correction value. By calling the corresponding parameter adjustment rule according to the risk level label, appropriate control strategies can be adopted to adjust the rotation control parameters under different deviation risk conditions.

[0108] Step S540: input the high-dimensional deviation feature vector and the parameter adjustment rule into the compensation calculation layer of the rotation control parameter optimization network, perform nonlinear mapping processing through the nonlinear mapping layer, and generate a preliminary traction force adjustment coefficient and a preliminary rotation speed correction value.

[0109] The compensation calculation layer is a layer in the rotation control parameter optimization network for calculating the control parameter adjustment value. The high-dimensional deviation feature vector contains key feature information of the rotation deviation risk, and the parameter adjustment rule specifies the adjustment direction and calculation method of the traction force adjustment coefficient and the rotation speed correction value. The two are input into the nonlinear mapping layer of the compensation calculation layer. The nonlinear mapping layer performs complex nonlinear mapping on the input high-dimensional deviation feature vector and parameter adjustment rule. During the mapping process, the nonlinear mapping layer will calculate and process according to the requirements in the parameter adjustment rule and in combination with the feature information in the high-dimensional deviation feature vector. For example, according to the adjustment direction of the traction force adjustment coefficient in the parameter adjustment rule, the nonlinear mapping layer will calculate the specific adjustment amplitude according to the deviation reflected in the high-dimensional deviation feature vector; for the rotation speed correction value, it will calculate according to the calculation method in the parameter adjustment rule in combination with the related features in the high-dimensional deviation feature vector.

[0110] Specifically, step S540 can include the following steps S541-S545: step S541: feature splicing the high-dimensional deviation feature vector and the rule code vector in the parameter adjustment rule to generate a fusion input vector, and the rule code vector is a digital representation of the parameter adjustment rule, and its dimension is the same as that of the high-dimensional deviation feature vector.

[0111] The high-dimensional deviation feature vector contains key feature information of the rotation deviation risk, and the rule code vector in the parameter adjustment rule is a digital representation of the parameter adjustment rule. The rule code vector converts the information in the parameter adjustment rule into a vector form for easy processing in the neural network. In order to fully utilize the information of the high-dimensional deviation feature vector and the parameter adjustment rule, they are spliced. In the feature splicing process, the high-dimensional deviation feature vector and the rule code vector are combined together in a set order to form a fusion input vector. Since the dimension of the rule code vector is the same as that of the high-dimensional deviation feature vector, such splicing can ensure that the fusion input vector can contain complete deviation feature information and parameter adjustment rule information. For example, some elements in the high-dimensional deviation feature vector may reflect the characteristics of the position deviation, and the corresponding elements in the rule code vector may represent the adjustment rule for the position deviation.

[0112] Step S542: input the fusion input vector into the first hidden layer of the nonlinear mapping layer, perform nonlinear transformation processing on the fusion input vector through activation processing, generate a first hidden layer feature vector, and the dimension of the first hidden layer feature vector is a preset multiple of the dimension of the fusion input vector. The first hidden layer is the first level in the nonlinear mapping layer and receives the fusion input vector as input. In the first hidden layer, neurons perform weighted summation on the fusion input vector and then perform nonlinear transformation through an activation function. The activation function can introduce a nonlinear factor, enabling the network to learn more complex patterns and relationships. Through activation processing, the fusion input vector is nonlinearly transformed to generate a first hidden layer feature vector. The dimension of the first hidden layer feature vector is a preset multiple of the dimension of the fusion input vector, which can increase the number and expression ability of features, enabling the network to better capture complex features in the input information. For example, the preset multiple can be set according to actual conditions. If the network needs to learn more complex feature relationships, the preset multiple can be set larger; otherwise, it can be set smaller. Through the processing of the first hidden layer, the information in the fusion input vector is further converted and expanded, providing richer feature representations for subsequent processing.

[0113] Step S543: input the first hidden layer feature vector into the second hidden layer of the nonlinear mapping layer, perform further nonlinear transformation processing through activation processing, generate a second hidden layer feature vector, and the dimension of the second hidden layer feature vector is half of the dimension of the first hidden layer feature vector.

[0114] The second hidden layer is the second level in the nonlinear mapping layer and receives the first hidden layer feature vector as input. In the second hidden layer, the input feature vector is also subjected to weighted summation and activation processing. Through further nonlinear transformation, the information in the first hidden layer feature vector is further mined and converted. The dimension of the second hidden layer feature vector is half of the dimension of the first hidden layer feature vector, which is a dimension reduction operation. Through dimension reduction, the number of features can be reduced, removing some redundant feature information while retaining important feature information. This can improve the computational efficiency of the network and avoid overfitting problems. For example, some repeated or unimportant features may be learned in the first hidden layer, and through the dimension reduction processing of the second hidden layer, these features can be removed, making the second hidden layer feature vector more refined and effective.

[0115] Step S544: input the second hidden layer feature vector into the output layer of the nonlinear mapping layer, perform mapping processing through linear mapping processing, and generate a control parameter vector containing two components, the first component of the control parameter vector being a preliminary traction adjustment coefficient, and the second component being a preliminary speed correction value.

[0116] The output layer is the last layer of the nonlinear mapping layer, which receives the second hidden layer feature vector as input. In the output layer, a linear mapping process is performed, i.e., a linear combination and weighted summation of the input second hidden layer feature vector. Through this linear mapping, the feature information in the second hidden layer feature vector is converted into a control parameter vector. The control parameter vector contains two components, the first component is a preliminary traction adjustment coefficient, and the second component is a preliminary speed correction value. Through the linear mapping process of the output layer, according to the feature information in the second hidden layer feature vector, the preliminary traction adjustment coefficient and the preliminary speed correction value are calculated. For example, some elements in the second hidden layer feature vector may be related to the features of position deviation and attitude deviation, and the output layer will calculate the corresponding preliminary traction adjustment coefficient and preliminary speed correction value according to these feature information and the mapping relationship learned by the network.

[0117] Step S545: Scale transformation processing is performed on the control parameter vector to convert the preliminary traction adjustment coefficient and the preliminary speed correction value from the standardized space output by the network to the actual physical quantity space, generating preliminary control parameter values with actual engineering significance. The preliminary traction adjustment coefficient and the preliminary speed correction value output by the network are numerical values in the standardized space, and the standardized space is a data processing method for facilitating network training and calculation, whose numerical range and unit may be different from the actual physical quantity space. In order to make these parameters applicable to actual engineering, scale transformation processing is needed. Scale transformation processing is the process of converting the numerical values output by the network from the standardized space to the actual physical quantity space. In this process, the preliminary traction adjustment coefficient and the preliminary speed correction value need to be adjusted according to the physical quantity range and unit in actual engineering. For example, the preliminary traction adjustment coefficient output by the network may be a standardized numerical value between 0 and 1, while the traction adjustment coefficient in actual engineering may be a specific percentage value or an actual force adjustment amount.

[0118] Step S550: Boundary constraint processing is performed on the preliminary traction adjustment coefficient and the preliminary speed correction value to ensure that their numerical values fall within the pre-set safe value range, generating the final traction adjustment coefficient and the speed correction value.

[0119] The preliminary traction force adjustment coefficient and the preliminary rotation speed correction value are obtained through compensation calculation, but these values can exceed the safety range allowed in actual engineering. In order to ensure the safety and stability of the rotation process, boundary constraint processing is needed for these preliminary values. The preset safety value range is determined according to the mechanical properties of the rotation structure, the capacity of the driving system, and the actual requirements of the engineering, etc. In the boundary constraint processing, the preliminary traction force adjustment coefficient and the preliminary rotation speed correction value are compared with the preset safety value range. If the preliminary traction force adjustment coefficient is less than the lower limit value of the safety value range, it is adjusted to the lower limit value; if it is greater than the upper limit value, it is adjusted to the upper limit value; if it is within the safety value range, it remains unchanged. The same method is applied to the preliminary rotation speed correction value.

[0120] Through boundary constraint processing, it is ensured that the final traction force adjustment coefficient and the rotation speed correction value are within the safety value range. This can avoid safety problems in the rotation process caused by excessively large or small parameter values, and ensure the smooth progress of the rotation process. For example, if the traction force adjustment coefficient is too large, it can cause the rotation structure to bear too much force, thereby damaging the structure; if the rotation speed correction value is too large, it can cause the rotation speed to be too fast, affecting the accuracy and safety of the rotation.

[0121] Specifically, step S550 can include the following steps S551-S555: step S551: obtaining a preset traction force adjustment coefficient safety value range and a rotation speed correction value safety value range, the traction force adjustment coefficient safety value range being a positive number interval, and the rotation speed correction value safety value range being a real number interval.

[0122] The preset traction force adjustment coefficient safety value range and the rotation speed correction value safety value range are preset according to the actual situation and requirements of the rotation engineering. The traction force adjustment coefficient is usually a coefficient used to adjust the size of the traction force, and its value should be positive, because a negative traction force adjustment coefficient has no meaning in actual engineering. Therefore, the traction force adjustment coefficient safety value range is a positive number interval, and the lower limit and the upper limit of this interval are determined according to the bearing capacity of the rotation structure, the power of the driving system, etc. The rotation speed correction value is a value used to correct the rotation speed of the rotation, which can be positive or negative, positive indicating an increase in rotation speed and negative indicating a decrease in rotation speed. Therefore, the rotation speed correction value safety value range is a real number interval, and its lower limit and upper limit are also determined according to the stability of the rotation structure, the performance of the driving system, and the accuracy requirements of the rotation process, etc.

[0123] Step S552: Compare the preliminary traction force adjustment coefficient with the lower limit value and the upper limit value of the traction force adjustment coefficient safe value range. If the preliminary traction force adjustment coefficient is less than the lower limit value, adjust it to the lower limit value. If the preliminary traction force adjustment coefficient is greater than the upper limit value, adjust it to the upper limit value. Otherwise, keep the preliminary traction force adjustment coefficient unchanged to obtain the constrained traction force adjustment coefficient.

[0124] After obtaining the traction force adjustment coefficient safe value range, compare the preliminary traction force adjustment coefficient with the lower limit value and the upper limit value of the range. If the preliminary traction force adjustment coefficient is less than the lower limit value, it means that the coefficient is too small and may not provide enough traction to achieve the rotation control, so adjust it to the lower limit value to ensure that the traction can meet the basic requirements of the rotation. If the preliminary traction force adjustment coefficient is greater than the upper limit value, it means that the coefficient is too large and may cause the rotation structure to bear excessive force, leading to structural damage or other safety problems, so adjust it to the upper limit value to ensure the safety of the rotation process. If the preliminary traction force adjustment coefficient is within the safe value range, keep it unchanged because the value meets the actual engineering requirements. Through this comparison and adjustment, the constrained traction force adjustment coefficient is obtained, which is within the safe value range and can ensure that the rotation driving system works under reasonable traction.

[0125] Step S553: Compare the preliminary rotation speed correction value with the lower limit value and the upper limit value of the rotation speed correction value safe value range. If the preliminary rotation speed correction value is less than the lower limit value, adjust it to the lower limit value. If the preliminary rotation speed correction value is greater than the upper limit value, adjust it to the upper limit value. Otherwise, keep the preliminary rotation speed correction value unchanged to obtain the constrained rotation speed correction value.

[0126] Similar to the processing of the preliminary traction force adjustment coefficient, compare the preliminary rotation speed correction value with the lower limit value and the upper limit value of the rotation speed correction value safe value range. If the preliminary rotation speed correction value is less than the lower limit value, it means that the rotation speed is reduced too much, which may cause the rotation speed to be too slow, affecting the efficiency and accuracy of the rotation, so adjust it to the lower limit value to ensure that the rotation speed is within a reasonable range. If the preliminary rotation speed correction value is greater than the upper limit value, it means that the rotation speed is increased too much, which may cause the rotation speed to be too fast, leading to unstable rotation process and even safety accidents, so adjust it to the upper limit value to ensure the safety and stability of the rotation process. If the preliminary rotation speed correction value is within the safe value range, keep it unchanged because the value meets the requirements of the rotation process. Through this comparison and adjustment, the constrained rotation speed correction value is obtained, which is within the safe value range and can ensure that the rotation driving system operates at a reasonable rotation speed.

[0127] Step S554: Smooth the constrained traction force adjustment coefficient to eliminate the fluctuation component of the traction force adjustment coefficient and generate the final traction force adjustment coefficient.

[0128] The constrained traction force adjustment coefficient may have some fluctuation components within the safe value range. These fluctuations may be caused by calculation errors, sensor noise, or other factors. Fluctuations in the traction force adjustment coefficient may cause the traction force of the swivel drive system to be unstable, affecting the smoothness of the swivel process. Smoothing processing can use algorithms such as moving average method, exponential smoothing method, etc. Through smoothing processing, the fluctuation components in the constrained traction force adjustment coefficient are eliminated, making the traction force adjustment coefficient more stable. The final traction force adjustment coefficient generated in this way can ensure that the swivel drive system provides stable traction force, improving the smoothness and accuracy of the swivel process.

[0129] Step S555: Rate limiting processing is performed on the constrained speed correction value to ensure that the speed correction value change between adjacent time stamps does not exceed the preset maximum allowed change rate, generating the final speed correction value.

[0130] The constrained speed correction value may have a large change between adjacent time stamps. This rapid change may cause impact on the swivel drive system and swivel structure, affecting the safety and stability of the swivel process. In order to avoid this situation, rate limiting processing needs to be performed on the constrained speed correction value. The preset maximum allowed change rate is determined according to the mechanical properties of the swivel structure, the response ability of the drive system, and the accuracy requirements of the swivel process, etc. In the rate limiting processing process, the speed correction value change between adjacent time stamps is calculated. If the change exceeds the preset maximum allowed change rate, the speed correction value is adjusted to a value that makes the change equal to the maximum allowed change rate; if the change is within the maximum allowed change rate, the speed correction value remains unchanged.

[0131] Through rate limiting processing, the speed correction value change between adjacent time stamps is ensured not to exceed the preset maximum allowed change rate, making the speed correction value change more gradual. The final speed correction value generated in this way can ensure that the swivel drive system adjusts the speed at a stable rate, reduces the impact on the swivel structure, and improves the safety and stability of the swivel process.

[0132] Step S560: Arrange the traction force adjustment coefficient and the speed correction value in time stamp order to generate a swivel positioning control instruction set aligned with the swivel process time axis, each instruction unit of the swivel positioning control instruction set containing the traction force adjustment coefficient and the speed correction value of the corresponding time stamp.

[0133] After the foregoing steps, the final traction adjustment coefficient and the final rotation speed correction value have been obtained. In order to enable these parameters to be accurately applied to the control of the rotation process, it is necessary to arrange them in chronological order. The timestamp records information at each moment in the rotation process, and the traction adjustment coefficient and the rotation speed correction value are arranged in chronological order of the timestamp to generate a rotation in-place control instruction set. This instruction set is aligned with the time axis of the rotation process, and each instruction unit corresponds to a timestamp and contains the traction adjustment coefficient and the rotation speed correction value at the timestamp.

[0134] For example, at the 1st timestamp of the rotation process, the corresponding traction adjustment coefficient is a1, and the rotation speed correction value is b1; at the 2nd timestamp, the corresponding traction adjustment coefficient is a2, and the rotation speed correction value is b2; and so on. These information is combined in chronological order into a rotation in-place control instruction set. After the instruction set is transmitted to the rotation driving system, the rotation driving system can accurately adjust the traction force and the rotation speed according to the parameters of each timestamp in the instruction set, to ensure that the bridge can be accurately positioned according to the predetermined trajectory and requirements.

[0135] Figure 2 A composition structure diagram of a rotation in-place control device provided by an embodiment of the present application is shown in Figure 2 The rotation in-place control device 200 includes: a data acquisition module 210, configured to acquire a rotation movement data sequence generated in a rotation construction process of a long-span bridge, the rotation movement data sequence containing multiple groups of monitoring data units with timestamp markers acquired synchronously at a preset sampling frequency; a feature decoupling module 220, configured to perform space-time feature decoupling on the rotation movement data sequence to obtain rotation posture evolution features and environmental interference coupling features in the monitoring data units; a state prediction module 230, configured to call a pre-trained rotation dynamics prediction network to perform joint dynamics modeling on the rotation posture evolution features and the environmental interference coupling features, to generate a state prediction sequence of the rotation process; a deviation comparison module 240, configured to compare the state prediction sequence with a preset rotation target trajectory sequence to generate a rotation deviation risk evaluation matrix containing position deviation amounts and posture deviation amounts; a compensation intervention module 250, configured to drive a rotation control parameter optimization network to perform compensation calculation based on the rotation deviation risk evaluation matrix, to generate a rotation in-place control instruction set containing traction adjustment coefficients and rotation speed correction values, and to transmit the rotation in-place control instruction set to a rotation driving system.

[0136] The description of the above device embodiments is similar to the description of the above method embodiments, and has similar beneficial effects as the method embodiments. In some embodiments, the device provided by the embodiments of the present application has functions or includes modules that can be used to perform the methods described in the above method embodiments. For technical details not disclosed in the device embodiments of the present application, please refer to the description of the method embodiments of the present application for understanding.

[0137] It can be understood that, in the above introduction of the embodiments of the present application, various algorithms involved can be known from related contents in the prior art, and in order to save space, they will not be expanded too much in the embodiments of the present application. In addition, those skilled in the art can supplement details according to common knowledge in the art when implementing the scheme of the present application, for example, normalization can be used to eliminate the dimensional conflict before feature fusion according to common knowledge in the art. For example, in step S255, based on the multi-dimensional correlation coefficient matrix, the main environmental interference factors affecting the change of the body attitude are extracted by principal component analysis, the time series data of the main environmental interference factors and the axis offset sequence are weighted and fused to generate an environmental interference coupling feature vector. Specifically, based on the physical magnitude of each environmental factor and the correlation weight between the axis offset, the influence coefficient sequence of the environmental interference on the axis offset can be calculated, and then the time series data of the main environmental interference factors is normalized and combined with the influence coefficient sequence to generate the environmental interference coupling feature vector. Other similar cases will not be exemplified one by one.

[0138] In addition, those skilled in the art can also use interpolation algorithms to eliminate dimensional differences, set reasonable thresholds based on historical data, experience or business scenario requirements, train models based on general model training methods, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present application will not introduce redundant details of the implementation process.

[0139] Figure 3 A hardware entity schematic diagram of a computer system provided by the embodiments of the present application is shown in FIG. 10, which includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program executable on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program. Figure 3 The hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program executable on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

Claims

1. A method for construction of a large-span bridge based on a neural network, characterized in that, The method comprises: collecting a rotation body movement data sequence generated in the rotation body construction process of a long-span bridge, the rotation body movement data sequence comprising multiple groups of monitoring data units with timestamp markers collected synchronously at a preset sampling frequency; decoupling the rotation body movement data sequence in time and space to obtain rotation body posture evolution features and environmental interference coupling features in the monitoring data units; calling a pre-trained rotation body dynamics prediction network to jointly model the rotation body posture evolution features and the environmental interference coupling features to generate a state prediction sequence of the rotation body process; comparing the state prediction sequence with a preset rotation body target trajectory sequence to generate a rotation body deviation risk assessment matrix comprising position deviation amounts and posture deviation amounts; driving a rotation body control parameter optimization network based on the rotation body deviation risk assessment matrix to perform compensation calculation, generating a rotation body positioning control instruction set comprising a traction force adjustment coefficient and a rotation speed correction value, and transmitting the rotation body positioning control instruction set to a rotation body driving system; wherein the rotation body control parameter optimization network is a neural network for optimizing rotation body control parameters according to the rotation body deviation risk assessment matrix; the process specifically comprises: inputting the rotation body deviation risk assessment matrix into a deviation feature extraction layer of the rotation body control parameter optimization network, performing deep feature mining on the rotation body deviation risk assessment matrix through a convolutional neural network to generate a high-dimensional deviation feature vector; inputting the high-dimensional deviation feature vector into a risk grading processing layer of the rotation body control parameter optimization network, calculating the probability distribution of different deviation levels through a fully connected layer and an activation processing to determine the risk level label of the current rotation body deviation; calling a corresponding parameter adjustment rule from a preset control strategy library according to the risk level label, the parameter adjustment rule comprising an adjustment direction of the traction force adjustment coefficient and a calculation method of the rotation speed correction value; inputting the high-dimensional deviation feature vector and the parameter adjustment rule into a compensation calculation layer of the rotation body control parameter optimization network, performing nonlinear mapping processing through a nonlinear mapping layer to generate a preliminary traction force adjustment coefficient and a preliminary rotation speed correction value; performing boundary constraint processing on the preliminary traction force adjustment coefficient and the preliminary rotation speed correction value to ensure that their values fall within a preset safe value range, generating a final traction force adjustment coefficient and a rotation speed correction value; arranging the traction force adjustment coefficient and the rotation speed correction value in timestamp order to generate a rotation body positioning control instruction set aligned with the rotation body process time axis, each instruction unit of the rotation body positioning control instruction set comprising a traction force adjustment coefficient and a rotation speed correction value corresponding to a timestamp.

2. The method of claim 1, wherein, The decoupling of the rotation body movement data sequence in time and space to obtain the rotation body posture evolution features and the environmental interference coupling features in the monitoring data units comprises: performing time axis division processing on the rotation body movement data sequence, dividing the rotation body movement data sequence into multiple data segment groups with continuous time correlation according to a preset time window length, each data segment group comprising a preset number of continuous monitoring data units; The monitoring data units in the data segment group are subjected to spatial coordinate system conversion processing, and each monitoring data unit is converted from a local sensor coordinate system to a global coordinate system of a bridge turntable, thereby generating a spatial position data set under a unified coordinate reference; A rotation angle change sequence, an axis offset sequence, and a center of gravity displacement trajectory in the turntable process are extracted from the spatial position data set as basic constituent elements of a turntable posture evolution feature; Environment perception data synchronized with the turntable motion data sequence timestamp is collected, and the environment perception data includes wind speed and direction time series data, temperature field distribution data, and ground vibration conduction data; The environment perception data and the spatial position data set are subjected to timestamp alignment processing, a coupling correlation model between environmental factors and turntable posture changes is established through sliding window correlation analysis, and an environment interference coupling feature vector is generated.

3. The method of claim 2, wherein, The rotation angle change sequence, the axis offset sequence, and the center of gravity displacement trajectory in the turntable process are extracted from the spatial position data set as basic constituent elements of a turntable posture evolution feature, including: The three-dimensional coordinate data in the spatial position data set is subjected to time series difference processing, the coordinate change amount between adjacent timestamps is calculated, and an instantaneous velocity vector sequence of the turntable motion is generated; Based on the instantaneous velocity vector sequence, the rotation angle change value in the horizontal plane and the inclination angle change value in the vertical plane are obtained through angle calculation processing, and a three-dimensional angle change matrix including azimuth angle, pitch angle, and roll angle is constructed; The three-dimensional angle change matrix is subjected to filtering processing, and a smooth rotation angle change sequence is generated; The distance deviation value of each monitoring point in the spatial position data set relative to the turntable axis is calculated, and an axis offset sequence is obtained through fitting processing, and the sampling frequency of the axis offset sequence is the same as that of the rotation angle change sequence; The discrete center of gravity coordinate data is subjected to curve fitting processing, and a continuous and derivable center of gravity displacement trajectory curve is generated, and the time dimension of the center of gravity displacement trajectory curve is aligned with that of the rotation angle change sequence.

4. The method of claim 3, wherein, The environment perception data and the spatial position data set are subjected to timestamp alignment processing, a coupling correlation model between environmental factors and turntable posture changes is established through sliding window correlation analysis, and an environment interference coupling feature vector is generated, including: The environment perception data is subjected to timestamp calibration processing, and the wind speed and direction time series data, the temperature field distribution data, and the ground vibration conduction data collected by different sensors are unified to the same time reference system as the turntable motion data sequence; The calibrated environment perception data is subjected to resampling processing based on the time interval of the rotation angle change sequence, so as to ensure that the sampling frequency of the environment perception data is consistent with that of the spatial position data set; A sliding analysis window including a time window size parameter is constructed, and the resampled environment perception data and the rotation angle change sequence are input into the sliding analysis window for synchronous sliding processing; In each sliding window, a correlation index between each environmental factor in the environmental perception data and the rotation angle change value is calculated, and a multi-dimensional correlation coefficient matrix is generated; Based on the multi-dimensional correlation coefficient matrix, the main environmental interference factors affecting the rotation attitude change are extracted through principal component analysis, the time series data of the main environmental interference factors are weighted and fused with the axis offset sequence, and an environmental interference coupling feature vector is generated, the dimension of the environmental interference coupling feature vector matches the dimension of the rotation attitude evolution feature.

5. The method of claim 1, wherein, The pre-trained rotation dynamics prediction network is called to jointly model the rotation attitude evolution feature and the environmental interference coupling feature, and a state prediction sequence of the rotation process is generated, including: The rotation attitude evolution feature and the environmental interference coupling feature are input into the feature fusion layer of the rotation dynamics prediction network, the rotation attitude evolution feature and the environmental interference coupling feature are merged into a joint input feature vector through feature splicing, and the number of channels of the joint input feature vector is the sum of the number of channels of the rotation attitude evolution feature and the number of channels of the environmental interference coupling feature; The joint input feature vector is processed by the spatio-temporal correlation modeling layer of the rotation dynamics prediction network to extract dynamic characteristics, capture the time-varying topological relationship of the bridge rotation structure through a dynamic graph convolution module, and adjust the feature response strength of different stages of the rotation process through a time series attention mechanism to generate a spatio-temporal correlation feature map; The spatio-temporal correlation feature map is processed by a multi-scale feature fusion process, the feature mapping of different levels is connected across scales, the contribution of local detail features and global trend features is balanced through a feature weighting algorithm, and a fusion feature tensor is generated; The fusion feature tensor is input into the dynamic memory processing layer of the rotation dynamics prediction network, a state cache unit and a trend memory unit are constructed through a hierarchical memory mechanism to store historical state information and update current state representation during the rotation process, and a dynamic evolution feature sequence is generated; The sequence prediction layer of the rotation dynamics prediction network is called to perform future state inference processing on the dynamic evolution feature sequence, and a rotation process state prediction sequence including multiple future time points is generated through a time series prediction algorithm, each element of the state prediction sequence corresponds to a predicted attitude parameter of a time stamp in the rotation process; The state prediction sequence is processed by a confidence calibration process, each time stamp's predicted attitude parameter is given a confidence weight based on the historical prediction error distribution, and a state prediction sequence with a confidence marker is generated.

6. The method of claim 5, wherein, The joint input feature vector is reconstructed into a three-dimensional feature tensor, the dimensions of the three-dimensional feature tensor correspond to the time dimension, the space dimension, and the feature channel dimension respectively; ​ An initial topology adjacency matrix is constructed based on the design parameters of the bridge swivel structure, and the edge weights of the adjacency matrix are updated in real time according to the angle change of the swivel, to generate a time-varying topology matrix dynamically changing with the swivel process; The three-dimensional feature tensor and the time-varying topology matrix are input into a dynamic graph convolution module, the spatial dimension is modeled through graph convolution operation, and a structure correlation feature map containing structure dynamic change characteristics is generated; A time sequence attention mechanism module is constructed, the relevance weight between the current timestamp and the historical timestamp feature vector is calculated based on the time sequence constraint of the swivel process, and a time sequence attention weight matrix is generated; The time sequence attention weight matrix is applied to the structure correlation feature map, the feature contribution degree of the swivel start stage, the uniform speed stage and the deceleration stage is dynamically adjusted, and the feature response strength of the swivel start stage and the deceleration stage is enhanced; The original joint input feature vector and the attention weighted structure correlation feature map are integrated through residual connection, the channel dimension is adjusted through 1*1 convolution layer, and the space-time correlation feature map containing time-varying topology relationship and time sequence characteristics is generated.

7. The method of claim 6, wherein, The fusion feature tensor is input into the dynamic memory processing layer of the swivel dynamics prediction network, the state cache unit and the trend memory unit are constructed through hierarchical memory mechanism, the historical state information in the swivel process is stored and the current state representation is updated, and a dynamic evolution feature sequence is generated, including: The fusion feature tensor is expanded into a time sequence feature vector sequence according to the time dimension, and each time sequence feature vector corresponds to the fusion feature representation of a timestamp in the swivel process; The state cache unit and the trend memory unit of the dynamic memory processing layer are initialized, the memory window parameters and the memory decay parameters are set, the capacity of the state cache unit is consistent with the dimension of the time sequence feature vector, and the capacity of the trend memory unit is a preset multiple of the capacity of the state cache unit; The time sequence feature vectors are input into the state cache unit in sequence according to the timestamp, and the cache content is updated in real time through the sliding window mechanism, when a new feature vector is input, the earliest historical feature vector in the window is removed, and the latest feature vector in the window is retained; The output feature of the state cache unit is processed through the trend extraction module, the trend feature vector is calculated by combining the memory decay parameter, and the storage content of the trend memory unit is updated; The window feature vector output by the state cache unit and the historical trend feature vector output by the trend memory unit are weighted and fused based on the gating fusion mechanism, to generate a comprehensive state representation vector of the current timestamp, and the weighting coefficient is dynamically adjusted according to the swivel stage feature, the start stage and the deceleration stage are adjusted to be biased towards the window feature vector, and the uniform speed stage is adjusted to be biased towards the historical trend feature vector; The comprehensive state representation vectors of each timestamp are arranged in time sequence to generate a dynamic evolution feature sequence containing the window dynamic fluctuation and the historical evolution trend in the swivel process, and the length of the dynamic evolution feature sequence is consistent with the length of the input time sequence feature vector sequence.

8. The method of claim 1, wherein, The state prediction sequence is compared with the preset rotation target trajectory sequence in deviation, a rotation deviation risk evaluation matrix containing position deviation and attitude deviation is generated, and the method comprises the following steps: The state prediction sequence with confidence labels and the rotation target trajectory sequence are subjected to time axis unification processing, the time stamps of the two sequences are aligned to the same time reference system, the prediction sequence and the target sequence form a point-by-point corresponding relationship in the time dimension, and the state prediction sequence with confidence labels and the rotation target trajectory sequence are subjected to time axis unification processing; The prediction position coordinates, the prediction attitude angles and the prediction confidence vectors corresponding to each time stamp are extracted from the state prediction sequence after time axis unification, and the target position coordinates and the target attitude angles corresponding to each time stamp are extracted from the rotation target trajectory sequence after time axis unification; A three-dimensional deviation vector of the prediction position coordinates and the target position coordinates, a three-dimensional angle deviation vector of the prediction attitude angles and the target attitude angles are calculated at each time stamp, and the first-order time derivatives of the three-dimensional deviation vector are calculated as the position deviation rate, and the first-order time derivatives of the three-dimensional angle deviation vector are calculated as the attitude deviation rate; A deviation risk propagation model is constructed based on the kinetic characteristic parameters of the rotation structure, the kinetic characteristic parameters include the stiffness distribution parameters and the mass inertia parameters of the rotation structure, the prediction confidence vectors, the three-dimensional deviation vectors, the position deviation rate, the three-dimensional angle deviation vectors and the attitude deviation rate are input into the deviation risk propagation model, and a deviation risk propagation coefficient of each time stamp is output, the deviation risk propagation coefficient represents the influence degree of the current deviation on the subsequent rotation process; The deviation risk propagation coefficient, the position deviation rate and the attitude deviation rate are subjected to probability reasoning through a Bayesian network, a deviation risk transfer probability matrix is generated, and the spatial correlation features are analyzed by constructing the covariance matrix of the three-dimensional deviation vector and the three-dimensional angle deviation vector; The three-dimensional deviation vector, the three-dimensional angle deviation vector, the deviation risk propagation coefficient, the deviation risk transfer probability matrix and the spatial correlation features are integrated in the time dimension, the deviation type dimension and the risk propagation dimension, and a rotation deviation risk evaluation matrix containing dynamic risk evolution features is generated.

9. The method of claim 8, wherein, The three-dimensional deviation vector of the prediction position coordinates and the target position coordinates at each time stamp is calculated, and the method comprises the following steps: X-axis coordinate prediction values, Y-axis coordinate prediction values, Z-axis coordinate prediction values and corresponding coordinate confidence components are extracted from the prediction position coordinates of the state prediction sequence with confidence labels, and a three-dimensional prediction coordinate vector and a coordinate confidence vector are formed; X-axis coordinate target values, Y-axis coordinate target values and Z-axis coordinate target values are extracted from the target position coordinates of the rotation target trajectory sequence, and a three-dimensional target coordinate vector is formed; A position sensitivity parameter of the current stage of rotation is obtained, the position sensitivity parameter is dynamically determined according to the stiffness parameters of the rotation structure in different directions, and the position sensitivity parameter is higher in the direction with smaller stiffness parameter; The difference between the corresponding coordinate axis components of the three-dimensional prediction coordinate vector and the three-dimensional target coordinate vector is calculated, and X-axis original deviation, Y-axis original deviation and Z-axis original deviation are generated. The X-axis original deviation, Y-axis original deviation and Z-axis original deviation are respectively multiplied by the position sensitivity parameter and the coordinate confidence component of the corresponding coordinate axis to obtain X-axis weighted deviation, Y-axis weighted deviation and Z-axis weighted deviation; The X-axis weighted deviation, Y-axis weighted deviation and Z-axis weighted deviation are combined into a three-dimensional deviation vector, and each component of the three-dimensional deviation vector comprehensively reflects the size of the deviation in the direction, the structure sensitivity and the prediction reliability.

10. A computer system comprising a memory and a processor, said memory storing a computer program operable on the processor, characterised in that, The processor implements the steps in the method of any one of claims 1 to 9 when executing the program.

Images