A full-cycle simulation and optimization method for suspension structure construction based on digital twin

Abstract

The present invention relates to the fields of digital twin technology and suspension structure construction technology, and discloses a full-cycle simulation and optimization method for suspension structure construction based on digital twin, including: constructing a multi-level nested graph to represent the suspension structure; implementing a cross-layer attention mechanism, calculating the influence transfer of structural units between different levels, and identifying multi-scale critical paths in the structure; constructing a dual self-attention enhanced spatiotemporal graph neural network, identifying key nodes through a node self-attention layer, and capturing the long-time scale structural state evolution through a time self-attention layer; applying a temporal graph convolutional network to learn the dynamic characteristics of the structure, and predict changes in the structural bearing capacity and dynamic response; generating a structural vulnerability heat map, identifying potential risk areas, and automatically adjusting the monitoring sampling frequency of key nodes; the present invention realizes high-precision simulation and optimization of the full-cycle construction process of the suspension structure, thereby improving construction safety and efficiency.

Description

Technical Field

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

[0002] The construction process of large-scale suspended 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 changes in structural bearing capacity and dynamic responses, resulting in inaccurate construction risk assessments; existing digital twin models lack the ability to accurately capture nonlinear dynamic characteristics in complex construction environments, especially transient responses and structural stability changes during multi-point synchronous lifting; although graph neural networks can represent structural topological relationships, they lack focus on key nodes, and long-term dependencies in the time dimension are difficult to effectively model; traditional physical models are difficult to update in real time and cannot accurately capture the changes in the dynamic characteristics of suspended structures at different construction stages.

[0004] At present, although digital twin technology has been applied to a certain extent in the engineering field, the existing technology lacks the following key capabilities: lack of effective expression of the multi-level relationships of suspension structures, lack of cross-level information fusion mechanism, inability to achieve accurate learning and prediction of dynamic characteristics, and inability to adaptively adjust monitoring resource configuration to target key nodes and paths.

[0005] Therefore, there is an urgent need for a new method that can fully utilize digital twin technology and combine it with advanced artificial intelligence algorithms to achieve accurate simulation and optimization of the entire construction cycle of suspended structures. Summary of the Invention

[0006] The present invention provides a full-cycle simulation and optimization method for suspension structure construction based on digital twins, which solves the technical problems in related technologies that existing graph neural networks are difficult to identify structural critical paths and multi-level coupling relationships, lack of focus on key nodes, and difficulty in effectively modeling long-term dependencies in the time dimension.

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

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

[0009] Based on the constructed multi-level nested graph, a cross-layer attention mechanism is implemented to calculate the influence transfer between structural units at different levels and identify multi-scale key paths in the structure;

[0010] Based on the results of the cross-layer attention mechanism, a dual self-attention enhanced spatiotemporal graph neural network is constructed. The node self-attention layer is used to identify key nodes, and the temporal self-attention layer is used to capture the long-term structural state evolution.

[0011] 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;

[0012] 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.

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

[0014] The hanging structure is represented as a multi-level nested graph, including a bottom-level graph, a middle-level graph, and a top-level graph;

[0015] Define the mapping relationship between levels, indicating the ownership relationship of lower-level nodes to higher-level nodes;

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

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

[0018] The middle layer graph is formed by aggregating bottom layer nodes, representing substructure units at different abstraction levels;

[0019] The top-level diagram represents the structural unit at the highest level of abstraction.

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

[0021] For nodes between two adjacent layers, the attention coefficient is calculated to indicate the degree of influence of the lower-layer nodes on the higher-layer nodes.

[0022] Based on the calculated attention coefficient, information transfer from low layer to high layer is realized;

[0023] Identify multi-scale critical paths in the architecture based on cross-layer attention coefficients.

[0024] In a preferred embodiment, the step of constructing a dual self-attention enhanced spatiotemporal graph neural network includes:

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

[0026] Construct a temporal self-attention layer to capture the evolution of structural states over long time scales;

[0027] 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.

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

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

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

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

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

[0033] Combining the multi-level attention scores and spatiotemporal prediction results, a comprehensive vulnerability score is calculated for each node;

[0034] Based on the calculated vulnerability scores, heat maps are generated at various levels of the hanging structure;

[0035] Based on the vulnerability score, nodes are classified and the monitoring strategy is dynamically adjusted.

[0036] In a preferred embodiment, 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.

[0037] In a preferred embodiment, the suspended structure comprises a large bridge, a long-span roof or a steel structure.

[0038] In a preferred embodiment, a full-cycle simulation and optimization system for suspension structure construction based on digital twin is used to execute a full-cycle simulation and optimization method for suspension structure construction based on digital twin, including:

[0039] Multi-level representation module, used to construct multi-level nested graph representation suspension structure;

[0040] 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;

[0041] A dual self-attention module is used to identify key nodes through a node self-attention layer and a temporal self-attention layer to capture the evolution of structural states over long time scales;

[0042] A time-series graph convolution module is used to learn structural dynamic characteristics and predict structural load-bearing capacity changes and dynamic responses;

[0043] The risk assessment module is used to generate structural vulnerability heat maps, identify potential risk areas, and automatically adjust the monitoring sampling frequency of key nodes.

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

[0045] Significantly improved prediction accuracy: Compared with traditional finite element analysis and simplified model calculations, this invention improves prediction accuracy and computing efficiency, can identify abnormal structural behavior within millisecond response time, and provide real-time basis for construction decision-making.

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

[0047] Adaptive learning capability: It breaks through the limitation of traditional physical models that are difficult to update in real time. Through graph neural networks, it adaptively learns the evolution laws of structural systems without presetting complex physical parameters. It can continuously optimize the prediction model based on real-time monitoring data and adapt to changes in characteristics at different construction stages.

[0048] Comprehensive capture of spatiotemporal characteristics: By integrating 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 changes of the suspended structure during the full-cycle construction process, especially accurately capturing the transient response during multi-point synchronous lifting.

[0049] Provides precise monitoring strategies: Automatically adjusts sampling frequency based on node centrality indicators and vulnerability scores to achieve efficient monitoring and early warning of key nodes and critical paths, optimizes the allocation of monitoring resources, and reduces data transmission and processing burdens. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of a method for full-cycle simulation and optimization of suspension structure construction based on digital twins of the present invention;

[0051] Figure 2 It is a detailed flow chart of the present invention for constructing a multi-level nested graph to represent a suspension structure;

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

[0053] Figure 4 It is a detailed flow chart of the present invention for constructing a dual self-attention enhanced spatiotemporal graph neural network;

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

[0055] Figure 6 It is a detailed flow chart of generating a structural vulnerability heat map according to the present invention. DETAILED DESCRIPTION

[0056] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0057] At least one embodiment of the present invention discloses a method for simulating and optimizing the full cycle of suspension structure construction based on digital twins, such as Figures 1 to 6 As shown, the following steps are included:

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

[0059] The specific steps include:

[0060] Step 1.1, construct a multi-level nested graph network;

[0061] This step represents the suspension structure as a multi-level nested graph ,in, 、 、 Respectively represent 、 、 Hierarchical nested graph, Indicates the number of levels;

[0062] Indicates the A graph of layers, where Indicates the Layer diagram, Indicates the The node collection of the layer, Indicates the The edge set of a layer represents the connection relationship between the structural units of the layer.

[0063] In this representation, the nodes in the high-level graph represent larger substructures (such as main beams, towers, anchoring systems, etc.), and the edges represent the interactions between these substructures and the force transmission paths; while the low-level graph represents more fine-grained specific monitoring points and the physical connection relationships between them, which can capture the local characteristics and detailed information of the structure.

[0064] Number of levels The system can be adjusted according to the specific type of suspended structure and the monitoring accuracy requirements. For example, for a simple suspended structure, a three-layer representation (monitoring point layer, component layer, and system layer) can be used; while for a complex large bridge or roof structure, a four-layer or more layer representation can be used for a finer division.

[0065] In some embodiments, a multi-level nested graph can be constructed using a bottom-up clustering approach, automatically generating a hierarchical structure based on the spatial location and connectivity of nodes without pre-defining inter-level mapping relationships. For example, spectral clustering or hierarchical clustering algorithms can be used to aggregate underlying physical nodes into higher-level abstract nodes.

[0066] Bottom layer : Represents the finest-grained structural node, including all monitoring points and sensor locations, where Represents the actual physical node set, Indicates physical connection relationship;

[0067] Middle layer diagram to : formed by aggregating underlying nodes, representing substructure units at different levels of abstraction;

[0068] Top-level diagram : Represents the structural unit at the highest level of abstraction, such as a major functional block or system-level component.

[0069] Step 1.2, define the mapping relationship between layers;

[0070] This technology provides a mapping function between layers , used to indicate the ownership relationship of lower-level nodes to higher-level nodes. Indicates that from Layer to The mapping function of the layer, Indicates the The node collection of the layer, Indicates the The node collection of the layer.

[0071] for and ,if , then it represents the low-level node Belongs to high-level nodes The substructure represented by Indicates the Layer nodes, Indicates the Layer This mapping relationship clearly defines the subordinate relationship between structural units at different levels, enabling the system to track the transmission path of structural information between different abstract levels.

[0072] Step 1.3, construct node feature vector;

[0073] For each level of nodes, this step constructs a feature vector containing the following information:

[0074] Physical characteristics: For the underlying nodes, including coordinate positions , displacement, velocity, acceleration, strain and other monitoring data;

[0075] For high-level nodes, this includes the geometric characteristics and overall mechanical properties of the substructure;

[0076] Topological features: graph structural features such as node connectivity and centrality indicators;

[0077] Time characteristics: historical data of node status changes over time.

[0078] Therefore, all features are combined to form the initial feature vector of the node ,in represents the node index, Represents a hierarchical index.

[0079] Step 2: Based on the constructed multi-level nested graph, a cross-layer attention mechanism is implemented to calculate the influence transfer between structural units at different levels and identify multi-scale key paths in the structure;

[0080] The specific steps include:

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

[0082] In this step, the two adjacent layers are first and Nodes between, calculate the attention coefficient , indicating the Layer Node For the first Layer Node Degree of impact:

[0083] ;

[0084] in, Represents the level index of the graph, Indicates ratio A higher level, Indicates the The node index of the layer, Indicates the The node index of the layer, Indicates the Layer Node For the first Layer Node The attention coefficient is in the range of , Indicates the Layer Node The feature vector of contains the physical, topological and temporal characteristics of the node. Indicates the Layer Node The feature vector of contains the abstract features of the high-level node. Indicates the Layer Node The feature vector of contains the abstract features of the high-level node. Represents a trainable weight matrix for feature transformation and attention calculation, Represents the weight matrix The transpose of Represents vector concatenation operation, Representation and Node The set of all high-level nodes with inter-layer mapping relationships, express The node index in is used for normalization calculation, represents the natural exponential function, Represents a leaky linear rectifier unit activation function that retains a small gradient for negative input values ​​to avoid the gradient vanishing problem.

[0085] Step 2.2, realize cross-layer information transmission;

[0086] Based on the calculated attention coefficient, information transfer from low layer to high layer is realized:

[0087] ;

[0088] in, Indicates the Layer Node The eigenvector of Indicates mapping to high-level nodes The set of all low-level nodes, Indicates the Layer Node For the first Layer Node The attention coefficient, Represents a trainable weight matrix for feature transformation, Indicates the Layer Node The eigenvector of Represents the sigmoid function.

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

[0090] ;

[0091] in, Indicates the updated Layer Node The eigenvector of Represents the original Layer Node The eigenvector of represents the control parameter of the reverse transfer strength, Representation and Node The set of all high-level nodes with inter-layer mapping relationships, Represents the attention coefficient, which measures the high-level nodes For low-level nodes The degree of impact, represents the trainable weight matrix, Indicates the Layer Node The eigenvector of .

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

[0093] Based on the cross-layer attention coefficient, the multi-scale key paths in the identification structure are:

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

[0095] ;

[0096] in, Indicates the Layer Node The importance score of Representation and Node The set of all high-level nodes with inter-layer mapping relationships, Indicates the Layer Node For the first Layer Node The attention coefficient.

[0097] Select the one with the highest importance score in each layer Nodes are the key nodes of this layer, among which, Indicates the The number of key nodes selected by the layer can be dynamically adjusted according to the layer characteristics and structural complexity. The key nodes represent the nodes in the layer that have the greatest impact on the structural stability and dynamic characteristics.

[0098] Through the inter-layer mapping relationship, the key nodes of different layers are connected to form a multi-scale key path:

[0099] ;

[0100] in, represents a critical path across all layers, represents the set of all identified multi-scale critical paths, 、 、 Respectively represent 、 、 critical paths, represents the total number of critical paths identified, Indicates the critical paths, 、 、 Respectively represent The first path 、 、 The key nodes of the layer, Indicates the number of levels.

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

[0102] The identification of critical paths can introduce additional constraints, such as path connectivity and smoothness. In one embodiment, a regularization term can be introduced into the path selection process to ensure that adjacent nodes in the path maintain a certain degree of continuity in physical space, avoiding jumpy paths.

[0103] Furthermore, in some embodiments, the identified critical paths can be verified and optimized by combining historical data with records of abnormal events. For example, historical abnormal nodes and paths can be counted and compared with the critical paths identified by the algorithm. 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 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 value and identify nodes with scores significantly higher than the average as key nodes.

[0105] Step 3: Based on the results of the cross-layer attention mechanism, a dual self-attention enhanced spatiotemporal graph neural network is constructed. The node self-attention layer is used to identify key nodes, and the temporal self-attention layer is used to capture the long-term structural state evolution.

[0106] The specific steps include:

[0107] Step 3.1, configure the node self-attention layer;

[0108] In order to highlight the characteristics of key nodes, this step constructs a node self-attention layer and calculates the attention between nodes in the same layer:

[0109] ;

[0110] in, represents the query matrix, It is The matrix composed of all node features of the layer 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 and determine the relevance of information. Represents a value matrix, which contains the information content that needs to be extracted. 、 、 are trainable parameter matrices that transform the original features into 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, represents the dot product of the query matrix and the key matrix, and 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. Multiplied by the attention weight, it represents the information in the weighted aggregate value matrix according to the attention score.

[0111] Use multi-head attention mechanism to enhance expressive power:

[0112] ;

[0113] in, Indicates that the multi-head attention mechanism enhances the expressive power, 、 、 Respectively represent 、 、 The output of an attention head, represents the number of attention heads, Indicates the The output of the attention head, each head focuses on different aspects of the feature, and the multi-head mechanism allows the model to focus on information in different representation subspaces at the same time. 、 、 It is The trainable parameter matrix of the attention heads to the query space, key space and value space, Indicates concatenating the outputs of multiple attention heads in the feature dimension. is the output linear transformation matrix, which is used to map the concatenated multi-head attention output to the required output dimension space.

[0114] Step 3.2, construct the temporal self-attention layer;

[0115] This step provides a temporal self-attention layer to capture the long-time scale evolution of the structural state. Assume that time-step sequence data ,in, 、 、 Respectively represent 、 、 Time step The node feature matrix of the layer, Indicates the total number of time steps.

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

[0117] ;

[0118] in, The query matrix representing the time dimension is used to represent the information you want to query at the current time step. The key matrix representing the time dimension is used to calculate the similarity with the query matrix and determine the relevance of information at different time steps. The value matrix representing the time dimension contains the time series information that needs to be extracted. 、 、 Represents a trainable parameter matrix that transforms the original time series features into query space, key space, and value space. 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 so that the sum of all time attention weights is 1. The attention calculation spans different time steps and captures the long-term dependencies in the time series rather than the spatial relationships between different nodes.

[0119] Step 3.3, integrating 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 spatial topology and temporal evolution characteristics:

[0121] ;

[0122] in, Represents the spatial graph neural network layer, which processes the spatial topological relationship between nodes and further processes the output of the multi-head attention mechanism through the graph neural network. Indicates the The output result of applying the multi-head attention mechanism to the layer node features, represents the output of applying a temporal self-attention layer to time series data, It is the weight parameter that balances spatial and temporal features, and its value range is , It is the final node representation, which integrates the spatial topological structure information and time evolution characteristics.

[0123] Step 4: Input the output features of the dual self-attention enhanced spatiotemporal graph neural network into the temporal graph convolutional network to learn the structural dynamic characteristics and predict the structural bearing capacity change and dynamic response;

[0124] The specific steps include:

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

[0126] First, we build a basic graph convolution layer to process spatial topological relationships:

[0127] ;

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

[0129] Step 4.2, integrating time characteristics;

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

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] in, is the time step The input features represent the original feature data of all nodes at the current moment. Represents the graph convolutional network operation, which combines the node features at the current moment with the graph structure. is the time step The hidden state of , which represents the node representation after graph convolution processing, is the time step The hidden state contains the historical information of the previous moment. Is the reset gate, which controls the degree of retention of the previous state, and its value range is , close to 0 means discarding most of the historical information, Is the update gate, which controls the weight of the current input and the previous state, and its value range is , close to 1 means more current information is retained, is the candidate hidden state, representing the new information at the current moment, Indicates that the current features and historical features are spliced ​​in the feature dimension. Represents the Hadamard product, that is, the corresponding positions of the elements are multiplied to achieve selective information transmission, 、 、 Represent the trainable weight matrices of the reset gate, update gate, and candidate hidden state, respectively, 、 、 Represent the bias items of reset gate, update gate and candidate hidden state respectively, is the hyperbolic tangent activation function, which maps the value to 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 in the future time step:

[0139] Input structural status data for current and historical time windows ,in, 、 、 Represents the time step 、 、 Structural status data, is the time window size, which indicates the length of historical data used for prediction, in time steps. Represents the current time step, which is the index of the time series.

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

[0141] ;

[0142] in, Represents the forward calculation process of the temporal graph convolutional network model, is the number of predicted time steps, which indicates the length of time the model predicts into the future. 、 、 Respectively represent the predicted 、 、 The structural state at the moment, 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.) and trigger an alert when the predicted value exceeds the threshold:

[0144] Displacement threshold :When predicting displacement Trigger displacement warning;

[0145] Stress threshold :When predicting stress When the stress warning is triggered;

[0146] Strain threshold :When predicting strain triggering emergency warnings;

[0147] Warning time , which represents 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, a structural vulnerability heat map is generated to identify potential risk areas and automatically adjust the monitoring sampling frequency of key nodes;

[0149] The specific steps include:

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

[0151] Combining the multi-level attention scores and spatiotemporal prediction results, a comprehensive vulnerability score is calculated for each node:

[0152] ;

[0153] in, It is The comprehensive vulnerability score of each node is a quantitative indicator of the potential risk of the node. It is a node importance score calculated based on the attention mechanism, reflecting the degree of attention paid to the node in the structural network. It is the node abnormality probability obtained based on the prediction model, which indicates the probability value of abnormality that may occur in the node in the future time step. It is a node centrality index based on the topological structure, which measures the importance and influence of the node in the overall structural network. 、 、 They represent the weight coefficients of node importance score, node anomaly probability and node centrality index respectively.

[0154] Weight coefficient 、 、 This can be obtained through training with historical data rather than manually set. For example, based on known historical risk event records, optimization methods such as gradient descent or genetic algorithms can be used to automatically adjust the weight coefficients, ensuring that the scoring system has the highest accuracy in identifying historical risk events.

[0155] In some embodiments, in addition to the three basic factors of node importance score, node anomaly probability and 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 anomalies in the recent time period have a greater impact on the current vulnerability score, while the anomalies in the distant time period have a smaller impact. The time decay function can be an exponential decay function ,in, Represents the time interval, the unit is time step, is the attenuation coefficient, a parameter that controls the attenuation rate, The larger the value, the faster the decay.

[0156] For different types of suspended structures, the vulnerability scoring system can incorporate structure-specific scoring factors. For example, for a suspension bridge, special attention can be paid to the tension changes in the main cables and suspenders; while for a lattice shell, the coupling effect of node displacement and component stress can be monitored.

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

[0158] Based on the calculated vulnerability scores, heat maps are generated at various levels of the hanging structure:

[0159] For each layer , the vulnerability score of the node Map to color space to form a heat map ;

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

[0161] Heat maps are presented at different levels, from the micro node level to the macro system level, providing a multi-scale view of risk distribution.

[0162] Step 5.3, adaptive monitoring strategy optimization;

[0163] According to an embodiment of the present application, the monitoring strategy is dynamically adjusted based on the vulnerability score:

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

[0165] Set different monitoring sampling frequencies for different risk levels:

[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 classification thresholds based on real-time status to ensure 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 overall structural stability indicators.

[0174] Application examples of this embodiment:

[0175] This technology was put to practical use during the construction of a particularly large cable-stayed bridge. The bridge boasts a 1,200-meter main span, a 1,500-meter-long steel box girder, and a 350-meter-high main tower. The bridge comprises a complex suspension structure system comprised of thousands of nodes and connections. Accurately predicting changes in the structural load-bearing capacity and dynamic response is crucial during construction, particularly during the hoisting and multi-point simultaneous lifting of the steel box girder. Traditional monitoring methods, relying on manual judgment and simplified finite element model calculations, are unable to meet the real-time risk assessment requirements of the construction process.

[0176] Against this backdrop, the method presented in this paper simulated and optimized the full-cycle construction process of this cable-stayed bridge. By collecting data from each monitoring point in real time and constructing a digital twin model, the team leveraged core technologies such as multi-level nested graph networks, a cross-layer attention mechanism, a dual self-attention-enhanced spatiotemporal graph neural network, and a temporal graph convolutional network to accurately predict the structural dynamics and conduct risk assessment.

[0177] Implementation process example:

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

[0179] At the cable-stayed bridge construction site, a total of 3,240 sensors were deployed, including displacement sensors, strain sensors, acceleration sensors, etc., forming the underlying map. The system consists of 3240 nodes. Each node contains 15-dimensional feature data, including location coordinates and monitoring parameters. Edges between underlying nodes are established based on actual physical connections, such as cable connections and beam welding.

[0180] Through the aggregation algorithm, the bottom nodes are aggregated to form the middle layer graph , contains 248 component-level nodes, such as main cable segments, inclined cables, bridge deck segments, etc. Component-level nodes contain 30-dimensional feature data such as geometric characteristics and overall stress state. Continue to aggregate to form the intermediate layer graph , contains 26 substructure nodes, such as the main tower section, side span section, mid-span section, etc. Each node contains 50-dimensional feature data such as overall stiffness and stability index. It includes 4 system-level nodes: superstructure system, substructure system, anchoring system and temporary support system.

[0181] The inter-layer mapping relationship is determined through the construction design drawings and the actual construction process. For example, sensors numbered 1-120 belong to the No. 1 inclined cable, No. 1-8 inclined cables belong to the north main span structure, and the north main span structure belongs to the superstructure system.

[0182] Implementation of cross-layer attention mechanism:

[0183] In practical applications, attention coefficients are calculated between two adjacent layers of nodes. For example, the influence of a bottom-level monitoring point on a middle-level cable-stayed component can be calculated. During the installation of a steel box girder, attention calculation results showed that the cable-stayed monitoring point at midspan had a higher attention coefficient for the upper-level cable-stayed components than other locations, with an average attention coefficient of 0.78, while other locations averaged only 0.25.

[0184] Based on these attention coefficients, information transfer from lower layers to higher layers is achieved. For example, during a sudden increase in wind load, wind speed and displacement data from the monitoring point level are rapidly transmitted to the component level, updating the overall stress state characteristics of the cable-stayed component. This weighted attention transfer prioritizes key monitoring point information (displacement changes when wind speeds reach 12 m / s).

[0185] Through a cross-layer attention mechanism, multi-scale critical paths were identified, primarily including: a monitoring point at the top of the main tower → the cable-stayed structure at the top of the main tower → the north main span structure → the superstructure system; and a monitoring point at the mid-span → the mid-span deck component → the mid-span structure → the superstructure system. These critical paths demonstrated significant response characteristics in subsequent wind load and temperature change events, verifying the accuracy of path identification.

[0186] Implementation of spatiotemporal graph neural network enhanced by dual self-attention:

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

[0188] The temporal self-attention layer processes 72 consecutive hours of historical data, using a 24-hour window and a 10-minute time step. This layer specifically focuses on the recovery of structural deformations after wind load changes and the cumulative deformations caused by diurnal temperature variations, effectively capturing the long-term evolution of the structural state.

[0189] By integrating spatial and temporal features, the dual self-attention mechanism successfully predicted the displacement patterns of the middle section of the main span due to the diurnal temperature difference, with an average prediction error of only 3.1 mm, compared to 12.5 mm for traditional methods. The mechanism also effectively identified the risk of abnormal increases in tension in the northern portion of the cable stays caused by the combined effects of nighttime temperature drops and wind loads.

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

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

[0192] During the segmented hoisting of steel box girders, a time-series graph convolutional network collected 24 hours of historical monitoring data (with a sampling frequency of 10 minutes) to predict structural responses over the next six hours. During a critical mid-span closure, the network predicted main cable displacement changes caused by temperature fluctuations and tension adjustments, with a maximum prediction error of 5.2 mm, compared to 18.7 mm using traditional finite element methods.

[0193] The network also predicted the potential vibration risk of certain stay cables when wind speeds increased to 15m / s, issuing a three-hour advance warning, allowing the construction team ample time to implement protective measures such as adding dampers. During another main girder hoisting operation, the network identified the potential for uneven stress due to hydraulic system desynchronization and issued a two-hour advance warning, averting a potential structural safety risk.

[0194] Implementation of vulnerability heatmap and risk assessment:

[0195] In practical applications, the comprehensive vulnerability score of each node is calculated by combining multi-level attention scores, spatiotemporal prediction results and topological structure characteristics. The weight coefficients are obtained through historical risk event training and are (Attention score), (predicted anomaly probability) and (Topological centrality).

[0196] Based on the calculated vulnerability scores, a multi-level vulnerability heat map was generated. For example, in the bottom-level monitoring point heat map, the main tower and main cable connection area, the mid-span area, and the anchorage area showed high vulnerability (scores > 0.8). In the component-level heat map, especially the stay cables longer than 80 meters and the main cable sections connecting multiple key nodes showed high vulnerability (scores > 0.75).

[0197] Nodes are divided into three categories based on vulnerability scores: high-risk nodes (score > 0.7, accounting for 8%) use a high-frequency sampling rate of 100Hz; medium-risk nodes (score 0.4-0.7, accounting for 22%) use a medium-frequency sampling rate of 10Hz; and low-risk nodes (score < 0.4, accounting for 70%) use a low-frequency sampling rate of 1Hz. This strategy focuses monitoring resources on critical areas and reduces the data transmission and processing burden by 70%.

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

[0199] Technical effect verification:

[0200] Improved prediction accuracy:

[0201] During a six-month construction monitoring period, this method was compared with traditional finite element analysis and simplified model calculation methods. The results showed that the average error in displacement prediction for this method was 4.2 mm, compared to 16.8 mm for the traditional method, representing a 74.4% improvement in prediction accuracy. The average error in strain prediction was 32 με, compared to 112 με for the traditional method, representing a 71.4% improvement in accuracy.

[0202] Especially in the complex environment where wind load and temperature change act together, the prediction error of this method does not increase by more than 20%, while the error of the traditional method increases by more than 120%, showing that this method is more adaptable to complex environments than the traditional method.

[0203] In terms of computational efficiency, this method takes only 0.47 seconds to process and predict the condition of 3,240 monitoring points in real time, compared to 8.2 seconds for traditional finite element analysis, resulting in a 17.4-fold increase in computational efficiency. This efficiency improvement enables the system to update the entire bridge's condition assessment and prediction in a 10-second cycle, meeting real-time decision-making requirements.

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

[0205] In practical applications, this method's multi-level risk assessment demonstrates its advantages. In seven risk events, this method was able to simultaneously provide warning information at the node, component, and system levels, while traditional methods only provided a single warning at the node level.

[0206] Multi-level risk assessments enabled the construction team to gain a more comprehensive understanding of risk propagation pathways and systemic impacts. For example, in an abnormal stay-cable tension event, this method not only identified the specific abnormal monitoring point (node-level warning), but also the entire affected stay-cable group (component-level warning) and the potential system-level impact (reduced structural stability of the north main span), enabling the construction team to develop a more systematic response strategy.

[0207] In terms of risk warning time, this method issued an average of 3.5 hours in advance, while traditional methods only issued an average of 0.8 hours. This extended warning time provided ample time for preventive measures. In practice, it successfully avoided three potential construction safety accidents, ensuring construction progress and personnel safety.

[0208] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by 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. The high-level graph nodes represent substructures, the edges represent the interaction relationship between substructures, and the low-level graph represents the monitoring points and physical connection relationships; Based on the constructed multi-level nested graph, we implement the cross-layer attention mechanism, calculate the influence transfer between structural units at different levels, and identify the multi-scale key paths in the structure, including: For nodes between two adjacent layers, the attention coefficient is calculated to indicate the degree of influence of the lower-layer nodes on the higher-layer nodes. Based on the calculated attention coefficient, information transfer from low layer to high layer is realized; Identify multi-scale critical paths in the structure based on cross-layer attention coefficients; Based on the results of the cross-layer attention mechanism, a dual self-attention enhanced spatiotemporal graph neural network is constructed. The node self-attention layer is used to identify key nodes, and the temporal self-attention layer is used to capture the long-term structural state evolution. 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. A method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 1, characterized in that: The step of constructing a multi-level nested graph to represent the suspension structure includes: The hanging 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. A method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 2, characterized in that: The bottom graph represents the finest-grained structural nodes, including all monitoring points and sensor locations; The middle layer graph is formed by aggregating bottom layer nodes, representing substructure units at different abstraction levels; The top-level diagram represents the structural unit at the highest level of abstraction.

4. The method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 1, 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.

5. The method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 1, characterized in that: The step of inputting the output features of the dual self-attention enhanced spatiotemporal graph neural network 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 construct 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.

6. The method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 1, 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 are generated at various levels of the hanging structure; Based on the vulnerability score, nodes are classified and the monitoring strategy is dynamically adjusted.

7. The method for full-cycle simulation and optimization of suspension structure construction based on digital twin according to claim 6, 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.

8. 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.

9. A full-cycle simulation and optimization system for suspension structure construction based on digital twins, used to execute a full-cycle simulation and optimization method for suspension structure construction based on digital twins according to any one of claims 1 to 8, 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 is used to identify key nodes through a node self-attention layer and a temporal self-attention layer to capture the evolution of structural states over long time scales; A time-series graph convolution module is used to learn structural dynamic characteristics and predict structural load-bearing capacity changes and dynamic responses; The risk assessment module is used to generate structural vulnerability heat maps, identify potential risk areas, and automatically adjust the monitoring sampling frequency of key nodes.

Images