Suspension structure construction full-period simulation optimization method based on digital twinning

Abstract

The invention relates to the technical field of digital twinning technology and suspension structure construction, and discloses a digital twinning-based suspension structure construction full-period simulation optimization method, which comprises the following steps of: constructing a multi-level nested graph to represent a suspension structure; a cross-layer attention mechanism is realized, influence transfer of structural units among different layers is calculated, and a multi-scale critical path in the structure is identified; constructing a space-time diagram neural network with double self-attention enhancement, identifying key nodes through a node self-attention layer, and capturing structural state evolution of a long time scale through a time self-attention layer; learning the dynamic characteristics of the structure by using a time sequence diagram convolutional network, and predicting the bearing capacity change and dynamic response of the structure; a structural vulnerability heat map is generated, a potential risk area is identified, and the monitoring sampling frequency of key nodes is automatically adjusted; according to the invention, high-precision simulation and optimization of the full-period construction process of the suspension structure are realized, and the construction safety and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of digital twin technology and suspension structure construction technology. More specifically, it relates to a full-cycle simulation and optimization method for suspension structure construction based on digital twin. Background Art

[0002] The construction process of large-scale suspension structures is highly complex and nonlinear. The structure consists of thousands of nodes and connections, and its construction safety and quality are crucial to the overall project.

[0003] Traditional construction monitoring methods mainly rely on manual experience judgment or simplified model calculations, and have the following technical problems: Traditional monitoring methods are difficult to accurately predict the change of structural bearing capacity and dynamic response, resulting in inaccurate construction risk assessment; Existing digital twin models lack the ability to accurately capture the nonlinear dynamic characteristics in complex construction environments, especially the transient response and structural stability changes during the multi-point synchronous lifting process; Although graph neural networks can represent the structural topological relationship, they lack the key attention to key nodes, and it is difficult to effectively model the long-term dependence in the time dimension; Traditional physical models are difficult to update in real time and cannot accurately capture the dynamic characteristic changes of suspension structures at different construction stages.

[0004] Currently, although digital twin technology has been applied in the engineering field to a certain extent, the existing technology lacks the following key capabilities: It lacks the effective expression of the multi-level relationship of suspension structures, lacks the mechanism of cross-level information fusion, cannot achieve the accurate learning and prediction of dynamic characteristics, and cannot adaptively adjust the monitoring resource allocation for key nodes and paths.

[0005] Therefore, there is an urgent need for a new method that can make full use of digital twin technology, combined with advanced artificial intelligence algorithms, to achieve the accurate simulation and optimization of the full cycle of suspension structure construction. Summary of the Invention

[0006] The present invention provides a full-cycle simulation and optimization method for suspension structure construction based on digital twin, which solves the technical problems in the related art that the existing graph neural networks are difficult to identify the key paths and multi-level coupling relationships of structures, lack of key attention to key nodes, and it is difficult to effectively model the long-term dependence in the time dimension.

[0007] The present invention provides a full-cycle simulation and optimization method for suspension structure construction based on digital twin, including the following steps:

[0008] Construct a multi-level nested graph to represent the suspension structure, including multiple levels. The high-level graph nodes represent sub-structures, and the edges represent the interaction relationships between sub-structures. The low-level graph represents the monitoring points and physical connection relationships;

[0009] Based on the constructed multi-level nested graph, implement a cross-layer attention mechanism to calculate the influence transmission of structural units between different levels and identify the multi-scale critical paths in the structure;

[0010] According to the results of the cross-layer attention mechanism, construct a spatio-temporal graph neural network with dual self-attention enhancement. Identify key nodes through the node self-attention layer and capture the evolution of the structural state at long time scales through the temporal self-attention layer;

[0011] Input the output features of the spatio-temporal graph neural network with dual self-attention enhancement into a temporal graph convolutional network and learn the dynamic characteristics of the structure to predict the change in structural bearing capacity and dynamic response;

[0012] Based on the prediction results of the temporal graph convolutional network and the multi-level attention scores, generate a structural vulnerability heat map, identify potential risk areas, and automatically adjust the monitoring sampling frequency of key nodes.

[0013] In a preferred embodiment, the steps of constructing the multi-level nested graph to represent the suspension structure include:

[0014] Represent the suspension structure as a multi-level nested graph, including a bottom layer graph, a middle layer graph, and a top layer graph;

[0015] Define the mapping relationship between levels to represent the belonging relationship of low-level nodes to high-level nodes;

[0016] Construct node feature vectors, including physical features, topological features, and temporal features.

[0017] In a preferred embodiment, the bottom layer graph represents the finest-grained structural nodes, including all monitoring points and sensor positions;

[0018] The middle layer graph is formed by aggregating bottom layer nodes and represents sub-structural units at different abstraction levels;

[0019] The top layer graph represents the structural unit at the highest abstraction level.

[0020] In a preferred embodiment, the steps of implementing the cross-layer attention mechanism include:

[0021] For nodes between adjacent layers, calculate the attention coefficient to represent the influence degree of low-level nodes on high-level nodes;

[0022] Based on the calculated attention coefficient, implement information transmission from low level to high level;

[0023] Based on the cross-layer attention coefficient, identify the multi-scale critical paths in the structure.

[0024] In a preferred embodiment, the steps of constructing the spatio-temporal graph neural network with dual self-attention enhancement include:

[0025] Construct a node self-attention layer to calculate the attention between nodes within the same layer;

[0026] Construct a time self-attention layer to capture the structural state evolution at a long time scale;

[0027] Integrate the outputs of the node self-attention layer and the time self-attention layer to obtain node representations that synthesize spatial topology and time evolution characteristics.

[0028] In a preferred embodiment, the step of inputting the output features of the spatio-temporal graph neural network with dual self-attention enhancement into a temporal graph convolutional network and learning the structural dynamic characteristics includes:

[0029] Construct a graph convolutional layer to process spatial topological relationships;

[0030] Combine graph convolution with a gated recurrent unit to construct a temporal graph convolutional network;

[0031] Based on the trained temporal graph convolutional network model, predict the dynamic response of the structure at future time steps.

[0032] In a preferred embodiment, the step of generating a structural vulnerability heat map includes:

[0033] Combine multi-level attention scores and spatio-temporal prediction results to calculate a comprehensive vulnerability score for each node;

[0034] Generate heat maps at each level of the suspended structure based on the calculated vulnerability scores;

[0035] Classify nodes based on the vulnerability scores and dynamically adjust the monitoring strategy.

[0036] In a preferred embodiment, the node classification includes high-risk nodes, medium-risk nodes, and low-risk nodes. High-frequency sampling is used for high-risk nodes, medium-frequency sampling is used for medium-risk nodes, and low-frequency sampling is used for low-risk nodes.

[0037] In a preferred embodiment, the suspended structure includes large bridges, large-span roofs, or steel structures.

[0038] In a preferred embodiment, a digital-twin-based full-cycle simulation and optimization system for the construction of suspended structures is used to execute a digital-twin-based full-cycle simulation and optimization method for the construction of suspended structures, including:

[0039] A multi-level representation module for constructing a multi-level nested graph to represent the suspended structure;

[0040] A cross-layer attention module for calculating the influence transfer of structural units between different levels and identifying multi-scale critical paths in the structure;

[0041] A dual self-attention module for identifying key nodes through a node self-attention layer and capturing the evolution of the structural state at a long time scale through a temporal self-attention layer;

[0042] A temporal graph convolutional module for learning the dynamic characteristics of the structure, predicting changes in structural bearing capacity and dynamic responses;

[0043] A risk assessment module for generating a structural vulnerability heat map, identifying potential risk areas, and automatically adjusting the monitoring sampling frequency of key nodes.

[0044] The beneficial effects of the present invention are as follows:

[0045] Significantly improved prediction accuracy: Compared with traditional finite element analysis and simplified model calculations, the present invention improves the prediction accuracy, enhances the calculation efficiency, can identify abnormal structural behaviors within a millisecond response time, and provides a real-time basis for construction decisions.

[0046] Realize multi-level risk assessment: Through a hierarchical graph attention network, it can simultaneously identify local key nodes and global key paths of the structure, provide a systematic risk assessment, and avoid the problem that traditional methods only focus on local stress concentration points and ignore system-level risk propagation.

[0047] Adaptive learning ability: It breaks through the limitation that traditional physical models are difficult to update in real time. By adaptively learning the evolution law of the structural system through a graph neural network, without presetting complex physical parameters, it can continuously optimize the prediction model according to real-time monitoring data and adapt to the characteristic changes in different construction stages.

[0048] Comprehensively capture spatio-temporal characteristics: By integrating the spatial topological structure and temporal evolution characteristics, through a dual self-attention mechanism and a temporal graph convolutional network, it can accurately simulate the dynamic change law of the suspended structure during the full-cycle construction process, especially accurately capture the transient response during the multi-point synchronous lifting process.

[0049] Provide a precise monitoring strategy: Automatically adjust the sampling frequency based on node centrality indicators and vulnerability scores, realize efficient monitoring and early warning of key nodes and key paths, optimize the allocation of monitoring resources, and reduce the data transmission and processing burden. Description of the Drawings

[0050] Figure 1 is a flowchart of a full-cycle simulation optimization method for the construction of a suspended structure based on digital twin according to the present invention;

[0051] Figure 2 is a detailed flowchart of constructing a multi-level nested graph representation of a suspended structure according to the present invention;

[0052] Figure 3It is a detailed flowchart for implementing the cross-layer attention mechanism of the present invention;

[0053] Figure 4 It is a detailed flowchart for constructing a spatio-temporal graph neural network with dual self-attention enhancement of the present invention;

[0054] Figure 5 It is a detailed flowchart for predicting the change of structural bearing capacity and dynamic response of the present invention;

[0055] Figure 6 It is a detailed flowchart for generating a structural vulnerability heat map of the present invention. Detailed implementation manners

[0056] Now, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the scope of protection of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0057] In at least one embodiment of the present invention, a full-cycle simulation and optimization method for the construction of a suspended structure based on digital twin is disclosed, as Figures 1 to 6 shown, and it includes the following steps:

[0058] Step 1: Construct a multi-level nested graph to represent the suspended structure, including multiple levels. The high-level graph nodes represent sub-structures, the edges represent the interaction relationships between sub-structures, and the low-level graph represents the monitoring points and physical connection relationships;

[0059] Specifically, it includes the following steps:

[0060] Step 1.1: Construct a multi-level nested graph network;

[0061] In this step, the suspended structure is represented as a multi-level nested graph , where , , respectively represent the , , level nested graphs, represents the number of levels;

[0062] represents the graph of the th layer, where represents the graph of the th layer, represents the node set of the th layer, represents the The edge set of a layer, representing the connection relationships between the structural units of that layer.

[0063] In this representation, the nodes in the high-level graph represent larger sub-structures (such as main girders, pylons, anchorage systems, etc.), and the edges represent the interaction relationships and force transfer paths between these sub-structures; while the low-level graph represents more fine-grained specific monitoring points and their physical connection relationships, capable of capturing the local characteristics and detailed information of the structure.

[0064] The number of levels It can be adjusted according to the specific type of suspension structure and the requirements of monitoring accuracy. For example, for a simple suspension structure, a 3-layer representation (monitoring point layer, component layer, and system layer) can be used; while for a complex large bridge or roof structure, 4 layers or more can be used for a more refined division.

[0065] In some embodiments, the construction of the multi-level nested graph can adopt a bottom-up clustering method to automatically generate a hierarchical structure based on the spatial positions and connection relationships of the nodes, without the need to pre-define the inter-layer mapping relationships. For example, spectral clustering or hierarchical clustering algorithms can be used to aggregate the underlying physical nodes into higher-level abstract nodes.

[0066] The bottom-level graph : Represents the most fine-grained structural nodes, including all monitoring points and sensor positions, where represents the set of actual physical nodes, represents the physical connection relationship;

[0067] The middle-level graph to : Formed by aggregating the bottom-level nodes, representing sub-structure units at different levels of abstraction;

[0068] The top-level graph : Represents the structural units at the highest level of abstraction, such as main functional blocks or system-level components.

[0069] Step 1.2, define the inter-layer mapping relationship;

[0070] This technology provides a mapping function between levels , used to represent the belonging relationship of low-level nodes to high-level nodes. Among them, represents the mapping function from the th layer to the th layer, represents the set of nodes in the th layer, represents the set of nodes in the th layer.

[0071] For and , if , it indicates that the lower-level node belongs to the sub-structure represented by the higher-level node . Among them, represents the th node of the th layer, represents the

[0072] Step 1.3, constructing the node feature vector;

[0073] For each layer of nodes, the feature vector constructed in this step contains the following information:

[0074] Physical features: For the bottom-layer nodes, it includes monitoring data such as coordinate position , displacement, velocity, acceleration, strain, etc.;

[0075] For the high-level nodes, it includes the geometric features and overall mechanical properties of the sub-structure;

[0076] Topological features: Graph structure features such as the connectivity and centrality index of the nodes;

[0077] Temporal features: Historical data of the change of node states over time.

[0078] Therefore, all features are combined to form the initial feature vector of the node , where represents the node index, represents the layer index.

[0079] Step 2, based on the constructed multi-level nested graph, implement the cross-layer attention mechanism, calculate the influence transfer of structural units between different layers, and identify the multi-scale key paths in the structure;

[0080] Specifically, it includes the following steps:

[0081] Step 2.1, configuring the cross-layer attention calculation unit;

[0082] In this step, first, for the nodes between two adjacent layers and , calculate the attention coefficient , indicating the influence degree of the node in the th layer on the node in the th layer:

[0083] ;

[0084] Among them, represents the hierarchical index of the figure, represents one level higher than the hierarchical level, represents the node index of the th layer, represents the node index of the th layer, represents the node of the th layer for the node of the th layer attention coefficient, the value range is , represents the node of the th layer feature vector, including the physical, topological and temporal features of the node, represents the node of the th layer feature vector, including the abstract features of the high-level node, represents the node of the th layer feature vector, including the abstract features of the high-level node, represents a trainable weight matrix for feature transformation and attention calculation, represents the weight matrix transpose, represents the vector concatenation operation, represents all high-level node sets that have an inter-layer mapping relationship with the node , represents the node index in, for normalization calculation, represents the natural exponential function, represents the leaky rectified linear unit activation function, which retains a small gradient for negative input values to avoid the vanishing gradient problem.

[0085] Step 2.2, implement cross-layer information transfer;

[0086] Based on the calculated attention coefficients, implement information transfer from the low layer to the high layer:

[0087] ;

[0088] Among them, represents the feature vector of the node of the th layer , represents all low-level node sets mapped to the high-level node , Indicates the layer node For the layer node attention coefficient, represents a trainable weight matrix for feature transformation, Indicates the layer node feature vector, represents the sigmoid function.

[0089] Reverse information transfer from high layer to low layer:

[0090] ;

[0091] Among them, Indicates the updated layer node feature vector, Indicates the original layer node feature vector, represents the control parameter of the reverse transfer intensity, Indicates all high-layer node sets that have an inter-layer mapping relationship with the node, represents the attention coefficient, measuring the influence degree of the high-layer node on the low-layer node , represents a trainable weight matrix, Indicates the layer node feature vector.

[0092] Step 2.3, identify the multi-scale critical path;

[0093] Based on the cross-layer attention coefficient, identify the multi-scale critical path in the structure:

[0094] For each layer, calculate the importance score of the node:

[0095] ;

[0096] Among them, Indicates the importance score of the layer node , Indicates all high-layer node sets that have an inter-layer mapping relationship with the node, Indicates the layer node For the layer node attention coefficient.

[0097] Select the node with the highest importance score in each layer as the key node of that layer, where represents the number of key nodes selected in the th layer, and the number of key nodes can be dynamically adjusted according to the layer characteristics and structural complexity. The key nodes represent the nodes in that layer that have the greatest impact on the structural stability and dynamic characteristics.

[0098] Connect the key nodes of different layers through the inter-layer mapping relationship to form a multi-scale critical path:

[0099] ;

[0100] where represents a critical path spanning all layers, represents the set of all identified multi-scale critical paths, , , respectively represent the , , th critical paths, represents the total number of identified critical paths, represents the th critical path, , , respectively represent the key nodes of the th path in the , , th layers, represents the number of levels.

[0101] These critical paths reflect the main channels of stress and deformation transmission in the structure and are the key areas for structural monitoring and risk assessment.

[0102] Additional constraints can be introduced for the identification of critical paths, such as path connectivity and smoothness. In one implementation, a regularization term can be introduced during the path selection process to keep adjacent nodes in the path continuous in physical space and avoid jumpy paths.

[0103] In addition, in some embodiments, the identified critical paths can be verified and optimized by combining the abnormal event records in the historical data. For example, the nodes and paths where abnormalities occurred in history can be counted and compared with the critical paths identified by the algorithm, and the accuracy of critical path identification can be improved by adjusting the weight parameters in the attention calculation.

[0104] For the number of key nodes For the selection, an adaptive method can be used instead of presetting fixed values. For example, based on the distribution characteristics of node importance scores, statistical methods (such as the quartile method or the standard deviation method) can be used to automatically determine the threshold, and nodes with scores significantly higher than the average level can be identified as key nodes.

[0105] Step 3: According to the results of the cross-layer attention mechanism, construct a dual self-attention enhanced spatio-temporal graph neural network, identify key nodes through the node self-attention layer, and the temporal self-attention layer captures the structural state evolution at long time scales;

[0106] Specifically, it includes the following steps:

[0107] Step 3.1: Configure the node self-attention layer;

[0108] To highlight the features of key nodes, in this step, a node self-attention layer is constructed to calculate the attention between nodes within the same layer:

[0109] ;

[0110] Among them, represents the query matrix, is the matrix composed of the features of all nodes in the th layer, which is used to represent the information that the current node wants to query, represents the key matrix, which is used to calculate the similarity with the query matrix to determine the relevance of the information, represents the value matrix, which contains the actual information content to be extracted, , , are trainable parameter matrices that respectively represent converting the original features to the query space, key space, and value space, is the square root of the feature dimension, which is used to scale the dot product to avoid the vanishing gradient problem, represents the dot product of the query matrix and the key matrix, which calculates the similarity between the query and the key, The function converts the similarity into a probability distribution so that the sum of all attention weights is 1,

[0111] The multi-head attention mechanism is adopted to enhance the expressive ability:

[0112] ;

[0113] Among them, represents the enhanced expressive ability of the multi-head attention mechanism, , , respectively represent the , , The output of attention heads, where represents the number of attention heads, and , , are the trainable parameter matrices from the -th attention head to the query space, key space, and value space, respectively. Each head focuses on different aspects of the features. The multi-head mechanism allows the model to simultaneously attend to information in different representation subspaces. denotes concatenating the outputs of multiple attention heads along the feature dimension, and

[0114] is the output linear transformation matrix used to map the concatenated multi-head attention output to the desired output dimension space.

[0115] Step 3.2: Construct the temporal self-attention layer; This step provides a temporal self-attention layer for capturing the structural state evolution at long time scales. Suppose there is a sequence of data with , , representing the node feature matrices at the -th, -th, and -th time steps of the -th layer, respectively, and denotes the total number of time steps.

[0116] Calculate the self-attention in the time dimension:

[0117] ;

[0118] where represents the query matrix in the time dimension, used to represent the information to be queried at the current time step, represents the key matrix in the time dimension, used to calculate the similarity with the query matrix to determine the relevance of information at different time steps, represents the value matrix in the time dimension, containing the actual time series information to be extracted, , , represent the trainable parameter matrices for transforming the original time series features into the query space, key space, and value space, respectively, is the square root of the feature dimension, used to scale the dot product to avoid the vanishing gradient problem, The function converts the similarity between different time steps into a probability distribution, making the sum of all time attention weights equal to 1. The attention calculation spans different time steps to capture long-term dependencies in the time series, rather than the spatial relationships between different nodes.

[0119] Step 3.3, integrate spatial and temporal features;

[0120] In this application, the outputs of the node self-attention layer and the temporal self-attention layer are integrated to obtain a node representation that combines the spatial topology and temporal evolution characteristics:

[0121] ;

[0122] Among them, represents the spatial graph neural network layer, which processes the spatial topology relationship between nodes and further processes the output of the multi-head attention mechanism through the graph neural network. represents the output result of applying the multi-head attention mechanism to the node features of the th layer. represents the output result of applying the temporal self-attention layer to the time series data. is the weight parameter that balances the spatial and temporal features, and its value range is , is the final node representation, which combines the spatial topology structure information and the temporal evolution characteristics.

[0123] Step 4, input the output features of the dual self-attention enhanced spatio-temporal graph neural network into the temporal graph convolutional network and learn the structural dynamic characteristics to predict the changes in structural bearing capacity and dynamic response;

[0124] Specifically, it includes the following steps:

[0125] Step 4.1, construct the graph convolutional layer;

[0126] First, construct the basic graph convolutional layer to process the spatial topology relationship:

[0127] ;

[0128] Among them, represents the graph convolutional layer, is the node feature matrix of the th layer, which contains the feature representations of all nodes. is the adjacency matrix, which represents the connection relationship between nodes. Among them, represents that there is a connection between node and node , represents that there is no connection, is the adjacency matrix with self-loops added, is the identity matrix. Adding self-loops ensures that the node's own information is also considered. is the degree matrix of which represents the node degree (number of connections). It is a diagonal matrix. is the trainable weight matrix used to transform node features. is the sigmoid function.

[0129] Step 4.2, integrating temporal features;

[0130] Combine graph convolution with gated recurrent unit (GRU) to construct a temporal graph convolutional network:

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] where is the input feature at time step representing the original feature data of all nodes at the current moment. represents the graph convolution network operation, which combines the node features at the current moment with the graph structure for processing. is the hidden state at time step representing the node representation after graph convolution processing. is the hidden state at time step containing historical information from the previous moment. is the reset gate, which controls the degree of retention of the previous moment's state. The value range is and being close to 0 means discarding most of the historical information. is the update gate, which controls the weights of the current input and the previous moment's state. The value range is and being close to 1 means retaining more of the current information. is the candidate hidden state, representing the new information at the current moment. represents concatenating the current feature and the historical feature in the feature dimension. represents the Hadamard product, that is, multiplying elements at corresponding positions to achieve selective information transfer. 、 、 respectively represent the trainable weight matrices of the reset gate, update gate, and candidate hidden state. 、 , represent the bias terms for the reset gate, update gate, and candidate hidden state respectively. is the hyperbolic tangent activation function that maps values to the interval. is the sigmoid function.

[0137] Step 4.3, predict the structural dynamic response;

[0138] Based on the trained TGCN model, predict the dynamic response of the structure at future time steps:

[0139] Input the structural state data of the current and historical time windows , where , , represent the structural state data at time steps , , respectively, is the time window size, indicating the length of historical data used for prediction, in units of time steps, represents the current time step, which is the index of the time series.

[0140] Through the forward propagation of the TGCN model, obtain the predicted future state:

[0141] ;

[0142] where represents the forward calculation process of the temporal graph convolutional network model, is the number of predicted time steps, indicating the time length that the model predicts into the future, , , represent the predicted structural states at the , , moments respectively, The symbol indicates that this is a predicted value rather than an actual observed value.

[0143] Set thresholds for key parameters (such as displacement, stress, strain, etc.). When the predicted value exceeds the threshold, trigger an early warning:

[0144] Displacement threshold : When the predicted displacement triggers a displacement early warning;

[0145] Stress threshold : When the predicted stress triggers a stress early warning;

[0146] Strain threshold : When predicting strain trigger a strain warning;

[0147] Warning time , indicating the number of time steps for early warning.

[0148] Step 5. Based on the prediction results of the temporal graph convolutional network and the multi-level attention scores, generate a structural vulnerability heat map, identify potential risk areas, and automatically adjust the monitoring sampling frequency of key nodes;

[0149] Specifically, it includes the following steps:

[0150] Step 5.1, construct a vulnerability scoring system;

[0151] Combining the multi-level attention scores and the spatio-temporal prediction results, calculate the comprehensive vulnerability score for each node:

[0152] ;

[0153] Among them, is the comprehensive vulnerability score of the th node, representing a quantitative indicator of the potential risk of this node, is the node importance score calculated based on the attention mechanism, reflecting the degree of attention of this node in the structural network, is the node anomaly probability obtained from the prediction model, indicating the probability value that this node may have an anomaly in future time steps, is the node centrality index based on the topological structure, measuring the importance and influence of the node in the overall structural network, , , respectively represent the weight coefficients of the node importance score, the node anomaly probability, and the node centrality index.

[0154] Weight coefficients , , can be obtained through training with historical data, rather than being set artificially. For example, based on the known risk event records in history, optimization methods such as gradient descent or genetic algorithms can be used to automatically adjust the weight coefficients to make the recognition accuracy of the scoring system for historical risk events the highest.

[0155] In some embodiments, in addition to the three basic factors of the node importance score, the node anomaly probability, and the node centrality index, the vulnerability score can also consider the influence of time factors. For example, a time decay function can be introduced so that the abnormal conditions in the recent time period have a greater impact on the current vulnerability score, while the abnormal conditions in the far time period have a smaller impact. This time decay function can be an exponential decay function , where represents the time interval, with the unit of time step, is the attenuation coefficient, a parameter that controls the attenuation rate, the larger the value, the faster the attenuation.

[0156] For different types of suspension structures, the vulnerability scoring system can introduce structure type-specific scoring factors. For example, for suspension bridge structures, special attention can be paid to the tension changes in the main cables and suspenders; while for reticulated shell structures, the coupling effect of joint displacements and member stresses can be monitored.

[0157] Step 5.2, generate a multi-level vulnerability heat map;

[0158] Based on the calculated vulnerability scores, generate heat maps at each level of the suspension structure:

[0159] For each level , map the vulnerability scores of the nodes to the color space to form a heat map ;

[0160] Through the interpolation algorithm, extend the heat map of discrete nodes to the entire structure area to form a continuous vulnerability heat map;

[0161] Present the heat maps at different levels, from the microscopic node level to the macroscopic system level, to provide a multi-scale risk distribution view.

[0162] Step 5.3, optimize the adaptive monitoring strategy;

[0163] According to the embodiments of the present application, based on the vulnerability scores, dynamically adjust the monitoring strategy:

[0164] Sort the vulnerability scores and classify the nodes into three categories: high risk, medium risk, and low risk;

[0165] For different risk levels, set different monitoring sampling frequencies:

[0166] High-risk nodes: high-frequency sampling (e.g., 100 Hz);

[0167] Medium-risk nodes: medium-frequency sampling (e.g., 10 Hz);

[0168] Low-risk nodes: low-frequency sampling (e.g., 1 Hz);

[0169] Dynamically adjust the classification threshold according to the real-time state to ensure the effective allocation of monitoring resources;

[0170] Generate multi-scale risk warnings:

[0171] Node-level warning: for a single high-risk node;

[0172] Path-level warning: for continuous anomalies on critical paths;

[0173] System-level warning: for changes in the overall structural stability index.

[0174] Application examples of this embodiment:

[0175] This technology has been actually applied in the construction process of a certain super-large cable-stayed bridge. The main span of this cable-stayed bridge is 1200 meters, the total length of the steel box girder is 1500 meters, the main tower is 350 meters high, and there are tens of thousands of nodes and connections in total, forming a complex suspension structure system. During the construction process, especially during the hoisting of the steel box girder and the multi-point synchronous lifting stage, the accurate prediction of the structural bearing capacity change and dynamic response is particularly crucial. Traditional monitoring methods rely on manual experience judgment and finite element simplified model calculation, and cannot meet the real-time risk assessment requirements during the construction process.

[0176] Under this background, the method of the present invention is applied to simulate and optimize the full-cycle construction process of this cable-stayed bridge. The data of each monitoring point is collected in real time, a digital twin model is constructed, and through core technologies such as multi-level nested graph networks, cross-layer attention mechanisms, spatio-temporal graph neural networks with double self-attention enhancement, and temporal graph convolutional networks, the accurate prediction and risk assessment of the structural dynamic characteristics are realized.

[0177] Implementation process examples:

[0178] Implementation of multi-level nested graph representation of the suspension structure:

[0179] At the construction site of the cable-stayed bridge, a total of 3240 sensors are arranged, including displacement sensors, strain sensors, acceleration sensors, etc., forming 3240 nodes of the bottom layer graph Each node contains 15-dimensional feature data such as position coordinates and monitoring parameters. The edges between the bottom layer nodes are established according to the actual physical connection relationships, such as steel cable connections and steel beam welds.

[0180] Through the aggregation algorithm, the bottom layer nodes are aggregated to form the middle layer graph , which contains 248 component-level nodes, such as main cable segments, stay cables, and bridge deck segments. Each component-level node contains 30-dimensional feature data such as geometric characteristics and overall stress states. Continuing to aggregate forms the middle layer graph , which contains 26 sub-structure nodes, such as main tower segments, side span segments, and middle span segments. Each node contains 50-dimensional feature data such as overall stiffness and stability index. The topmost graph contains 4 system-level nodes: the upper structure system, the lower structure system, the anchorage system, and the temporary support system.

[0181] The interlayer mapping relationship is determined through construction design drawings and the actual construction process. For example, the sensors numbered 1 - 120 belong to the first stay cable, the first to eighth stay cables belong to the north main span structure, and the north main span structure belongs to the upper structure system.

[0182] Implementation of the cross - layer attention mechanism:

[0183] In practical applications, the attention coefficients are calculated between adjacent - layer nodes. For example, the influence degree of the bottom - layer monitoring points on the middle - layer stay - cable components is calculated. During a steel box girder hoisting process, the attention calculation results show that the attention coefficient of the stay - cable monitoring points at the mid - span position to the upper - layer stay - cable components is higher than that of other positions. The average attention coefficient is 0.78, while the average of other positions is only 0.25.

[0184] Based on these attention coefficients, information transfer from the lower layer to the higher layer is realized. For example, in a situation where the wind load suddenly increases, the wind speed and displacement data at the monitoring - point level are quickly transmitted to the component level, updating the overall force - bearing state characteristics of the stay - cable components. The attention - weighted transfer enables the priority processing of key monitoring - point information (displacement changes when the wind speed reaches 12 m / s).

[0185] Through the cross - layer attention mechanism, multi - scale key paths are identified, mainly including: monitoring points at the top of the main tower → stay - cable fixing components at the top of the main tower → north main span structure → upper structure system, and mid - span monitoring points → mid - span bridge - deck components → mid - span structure → upper structure system. These key paths show obvious response characteristics in subsequent multiple wind - load and temperature - change events, verifying the accuracy of path recognition.

[0186] Implementation of the spatio - temporal graph neural network with dual self - attention enhancement:

[0187] In the implementation of the node self - attention layer, an 8 - head attention mechanism is adopted. The input feature dimension is 64, and the output feature dimension is 64. Through the node self - attention mechanism, the key nodes that have the greatest impact on the overall structural state are successfully identified, mainly concentrated in the connection area at the top of the main tower, the mid - span area, and the anchorage area.

[0188] The time self - attention layer processes a historical data sequence of 72 consecutive hours. The adopted time - window size is 24 hours, and the time step is 10 minutes. This layer particularly focuses on the recovery process of structural deformation after wind - load changes and the cumulative deformation caused by daily temperature changes, effectively capturing the evolution law of the structural state on a long - time scale.

[0189] After integrating spatial and temporal features, the dual self-attention mechanism successfully predicted the displacement change law at the mid-span of the main bridge caused by the day-night temperature difference, with an average prediction error of only 3.1 mm, while the prediction error of the traditional method was 12.5 mm. Meanwhile, the mechanism effectively identified the risk situation of abnormal increase in the tension of some stay cables on the north side when the night temperature decreased and the wind load acted together.

[0190] Implementation of the temporal graph convolutional network for learning the dynamic characteristics of the structure:

[0191] In the implementation of the temporal graph convolutional network, two layers of graph convolutional layers were constructed. The first layer contained 128 convolutional kernels, and the second layer contained 64 convolutional kernels. The hidden layer dimension of the GRU unit was set to 256, and a bidirectional structure was adopted to capture the temporal forward and backward dependencies.

[0192] Especially during the segmented hoisting process of the steel box girder, the temporal graph convolutional network collected 24-hour historical monitoring data (sampling frequency: once every 10 minutes) and predicted the structural response in the next 6 hours. During an important mid-span closure construction, the network predicted the displacement change of the main cable caused by temperature change and tension adjustment, with a maximum prediction error of 5.2 mm, while the prediction error of the traditional finite element method reached 18.7 mm.

[0193] The network also predicted the possible vibration risk of some stay cables when the wind speed increased to 15 m / s and issued a warning 3 hours in advance, enabling the construction team to have enough time to take protective measures such as adding dampers. During another main girder lifting process, the network detected the uneven stress situation that might be caused by the asynchrony of the hydraulic system and issued a warning 2 hours in advance, avoiding potential structural safety risks.

[0194] Implementation of the vulnerability heat map and risk assessment:

[0195] In practical applications, the comprehensive vulnerability score of each node was calculated by combining multi-level attention scores, spatio-temporal prediction results, and topological structure characteristics. The weight coefficients were obtained through training with historical risk events and were respectively (attention score), (prediction anomaly probability), and (topological centrality).

[0196] Based on the calculated vulnerability scores, a multi-level vulnerability heat map was generated. For example, in the heat map of the underlying monitoring points, the connection areas between the main tower and the main cable, the mid-span area, and the anchorage area showed high vulnerability (score > 0.8); in the heat map of the component layer, especially the stay cables with a length exceeding 80 meters and the main cable segments connecting multiple key nodes showed relatively high vulnerability (score > 0.75).

[0197] According to the vulnerability score, the nodes are divided into three categories: high-risk nodes (score > 0.7, accounting for 8%) adopt high-frequency sampling at 100 Hz; medium-risk nodes (score 0.4 - 0.7, accounting for 22%) adopt medium-frequency sampling at 10 Hz; low-risk nodes (score < 0.4, accounting for 70%) adopt low-frequency sampling at 1 Hz. This strategy focuses monitoring resources on key areas while reducing the data transmission and processing burden by 70%.

[0198] During a sudden strong wind event, the system first identified the group of stay cables with increased risk on the component layer heat map, then quickly located the specific anomalies of the underlying high-frequency monitoring points, and finally issued a warning 15 minutes before the risk spread.

[0199] Verification of technical effects:

[0200] Improvement in prediction accuracy:

[0201] During the 6-month construction monitoring process, this method was compared and tested with traditional finite element analysis and simplified model calculation methods. The results showed that the average error of this method in displacement prediction was 4.2 mm, while that of the traditional method was 16.8 mm, and the prediction accuracy was improved by 74.4%; in strain prediction, the average error was 32 με, while that of the traditional method was 112 με, and the accuracy was improved by 71.4%.

[0202] Especially in the complex environment under the combined action of wind load and temperature change, the prediction error of this method increased by no more than 20%, while that of the traditional method increased by more than 120%, indicating that this method has higher adaptability to complex environments than the traditional method.

[0203] In terms of computational efficiency, this method only takes 0.47 seconds for real-time processing and prediction of 3240 monitoring points, while traditional finite element analysis takes 8.2 seconds, and the computational efficiency is increased by 17.4 times. This efficiency improvement enables the system to update the state assessment and prediction of the entire bridge every 10 seconds, meeting the real-time decision-making requirements.

[0204] Systematic advantages of multi-level risk assessment:

[0205] In practical applications, the multi-level risk assessment of this method shows advantages. In 7 risk events, this method can simultaneously provide three-layer warning information at the node level, component level, and system level, while the traditional method can only provide single warning at the node level.

[0206] Multi-level risk assessment enables the construction team to have a more comprehensive understanding of the risk propagation path and system impacts. For example, in an event of abnormal cable-stayed cable tension, this method not only identified specific abnormal monitoring points (node-level early warning), but also identified the entire affected group of cable-stayed cables (component-level early warning) and possible system-level impacts (reduction in the structural stability of the north main span), enabling the construction team to formulate a more systematic response strategy.

[0207] In terms of the risk early warning time, this method issues an early warning 3.5 hours in advance on average, while the traditional method only issues an early warning 0.8 hours in advance on average. This extension of the early warning time provides sufficient time for taking preventive measures, and 3 possible construction safety accidents have been successfully avoided in practical applications, ensuring the construction progress and personnel safety.

[0208] The embodiments of the present invention have been described above, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. A full-cycle simulation optimization method for suspension structure construction based on digital twin, characterized in that: The following steps are involved: Construct a multi-level nested graph to represent the suspension structure, including multiple levels, high-level graph nodes represent substructures, edges represent the interaction between substructures, and low-level graphs represent monitoring points and physical connection relationships; Based on the constructed multi-level nested graph, the cross-layer attention mechanism is implemented to calculate the influence transfer of structural units between different levels and identify multi-scale key paths in the structure; According to the results of the cross-layer attention mechanism, a dual self-attention enhanced spatiotemporal graph neural network is constructed. The key nodes are identified through the node self-attention layer, and the temporal self-attention layer captures the evolution of the structural state over a long time scale. The output features of the dual self-attention enhanced spatiotemporal graph neural network are input into the temporal graph convolutional network to learn the structural dynamic characteristics and predict the structural bearing capacity change and dynamic response. Based on the prediction results of the temporal graph convolutional network and multi-level attention scores, a structural vulnerability heat map is generated to identify potential risk areas and automatically adjust the monitoring sampling frequency of key nodes.

2. According to claim 1, a method for full-cycle simulation and optimization of suspension structure construction based on digital twin, characterized in that: The step of constructing a multi-level nested graph to represent the suspension structure includes: The suspension structure is represented as a multi-level nested graph, including a bottom-level graph, a middle-level graph, and a top-level graph; Define the mapping relationship between levels, indicating the ownership relationship of lower-level nodes to higher-level nodes; Construct node feature vectors, including physical features, topological features, and time features.

3. According to claim 2, a method for full-cycle simulation and optimization of suspension structure construction based on digital twins is characterized in that: The bottom layer graph represents the finest-grained structural nodes, including all monitoring points and sensor locations; The intermediate layer graph is formed by aggregating the bottom layer nodes, representing sub-structure units at different abstraction levels; The top-level diagram represents the structural unit at the highest level of abstraction.

4. According to claim 1, a method for full-cycle simulation and optimization of suspension structure construction based on digital twin, characterized in that: The steps of implementing the cross-layer attention mechanism include: For nodes between two adjacent layers, the attention coefficient is calculated to indicate the influence of lower-layer nodes on higher-layer nodes. Based on the calculated attention coefficient, information transfer from low layer to high layer is realized; Based on the cross-layer attention coefficients, multi-scale critical paths in the structure are identified.

5. According to claim 1, a method for full-cycle simulation and optimization of suspension structure construction based on digital twin, characterized in that: The steps of constructing a dual self-attention enhanced spatiotemporal graph neural network include: Construct a node self-attention layer and calculate the attention between nodes in the same layer; Construct a temporal self-attention layer to capture the evolution of structural states over long time scales; The outputs of the node self-attention layer and the temporal self-attention layer are integrated to obtain a node representation that combines spatial topology and temporal evolution characteristics.

6. The method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 1 is characterized in that: The step of inputting the dual self-attention enhanced spatiotemporal graph neural network output features into the temporal graph convolutional network and learning the structural dynamic characteristics includes: Construct a graph convolution layer to process spatial topological relationships; Combine graph convolution with gated recurrent units to build a temporal graph convolutional network; Based on the trained temporal graph convolutional network model, the dynamic response of the structure in future time steps is predicted.

7. The method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 1 is characterized in that: The step of generating a structural vulnerability heat map comprises: Combining the multi-level attention scores and spatiotemporal prediction results, a comprehensive vulnerability score is calculated for each node; Based on the calculated vulnerability scores, heat maps were generated at various levels of the hanging structure; Based on the vulnerability score, nodes are classified and the monitoring strategy is adjusted dynamically.

8. The method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 7 is characterized in that: The node classification includes high-risk nodes, medium-risk nodes and low-risk nodes. High-risk nodes are sampled at a high frequency, medium-risk nodes are sampled at a medium frequency, and low-risk nodes are sampled at a low frequency.

9. The method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 1, characterized in that: The suspended structures include large bridges, long-span roofs or steel structures.

10. A full-cycle simulation optimization system for suspension structure construction based on digital twin, used to execute a full-cycle simulation optimization method for suspension structure construction based on digital twin according to any one of claims 1 to 9, characterized in that: include: Multi-level representation module, used to construct multi-level nested graph representation suspension structure; Cross-layer attention module, used to calculate the influence transfer between structural units at different levels and identify multi-scale key paths in the structure; A dual self-attention module for identifying key nodes through a node self-attention layer and a temporal self-attention layer for capturing the evolution of structural states over long time scales; The time-series graph convolution module is used to learn the dynamic characteristics of the structure and predict the changes in the structural bearing capacity and dynamic response; Risk assessment module, used to generate structural vulnerability heat maps, identify potential risk areas, and automatically adjust the monitoring sampling frequency of key nodes.

Images