Deep learning method for bridge damage integrating multi-source features

Abstract

The present invention discloses a deep learning method for bridge damage integrating multi-source features, comprising the following steps: establishing a bridge damage simulation model and obtaining bridge damage simulation data; associating the bridge damage simulation data with explosion conditions and integrating them with bridge feature information to obtain a bridge damage database including a training data set and a verification data set; the bridge feature information includes bridge geometric features and bridge material features; using the training data set and the verification data set to train and optimize the constructed multi-layer composite graph neural network to obtain a bridge damage prediction model; using the bridge damage prediction model to complete the prediction of target bridge damage. The prediction method of the present invention has high accuracy and good adaptability.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a deep learning method for bridge damage integrating multi-source features. Background Art

[0002] Bridges are an important part of modern transportation systems, and their riot protection design is of great significance to improving urban resilience and ensuring public safety. Under extreme events such as explosions or impact loads, bridge structures are susceptible to serious damage, affecting their bearing capacity and service life. Therefore, bridge design must be able to withstand extreme events such as explosion shocks to ensure that structural damage is reduced when emergencies occur. In the current design process, the premise for optimizing riot protection design is to accurately simulate and analyze the damage response of bridge structures under explosion shocks, so as to formulate reinforcement measures based on the force distribution and vulnerability of the structure.

[0003] At present, the prediction of bridge structure damage is described in the Chinese patent application "A clustering-based online evaluation method for bridge bearing service performance" (application publication number: CN118051739A, publication date: 2024-05-17), which includes: analyzing the bridge temperature field based on the numerical model, selecting the optimal temperature measurement point according to the simulation results to install the temperature sensor for temperature sampling, and installing a wire displacement meter for each bearing to measure the horizontal deformation of each bearing, and performing statistical analysis on the corresponding bearing horizontal deformation data according to the bridge temperature cluster analysis results, and calculating the mean and variance of each statistical model, and evaluating the bearing service performance. This method is based on the traditional finite element method. Although it can provide relatively accurate predictions, the calculation process is cumbersome, time-consuming, and costly, and it is difficult to meet the actual needs of rapid response.

[0004] To this end, in recent years, people have tried to develop data-driven models to achieve damage prediction of bridge structures in a more efficient and convenient way. For example, in the Chinese patent application "Intelligent Monitoring and Evaluation Method of Bridge Support Structure State Based on Neural Network Model" (Application Publication Number: CN118536385A, Publication Date: 2024-08-23), the spatial relationship data of the bridge support structure network and the state parameter data transmitted by the sensor are first obtained, and after preprocessing, the data set is divided into a training set and a test set. The training set data is trained using the Transformer model self-attention mechanism, and tested by the test set to form a bridge support structure state monitoring and evaluation model trained by deep learning, and the parameter set is brought into the bridge support structure state monitoring and evaluation model for evaluation and monitoring, and the output bridge support structure state evaluation analysis results are obtained. According to the model output state evaluation analysis results, corresponding warning displays and matching preset repair plans are performed. The introduction of machine learning algorithms enables data-driven damage prediction models to output prediction results at a faster speed. However, this single-layer neural network has the problem of insufficient robustness in multi-scenario prediction, especially for the situation where the material and structural strength of different locations in the bridge (such as the bridge deck and piers) are inconsistent, it is difficult to achieve adaptive damage prediction.

[0005] In summary, the problems existing in the prior art are: low accuracy in predicting damage to bridge structures under explosion impact and limited adaptability. Summary of the invention

[0006] The purpose of the present invention is to provide a deep learning method for bridge damage that integrates multi-source features with high accuracy and good adaptability.

[0007] In order to achieve the above objectives, the technical solution of the present invention is:

[0008] In a first aspect, a deep learning method for bridge damage integrating multi-source features is provided, comprising the following steps:

[0009] Establish a bridge damage simulation model and obtain bridge damage simulation data;

[0010] The bridge damage simulation data is associated with the explosion conditions and integrated with the bridge characteristic information to obtain a bridge damage database including a training data set and a verification data set; the bridge characteristic information includes bridge geometric characteristics and bridge material characteristics;

[0011] The constructed multi-layer composite graph neural network is trained and optimized using the above training data set and validation data set to obtain a bridge damage prediction model.

[0012] The bridge damage prediction model is used to predict the damage of the target bridge.

[0013] Preferably, the step of establishing the bridge damage simulation model includes:

[0014] According to the principle of explosion mechanics, the energy release function of explosives at different equivalents and densities over time is established, and the initial conditions of the explosion center are determined in combination with the physical and chemical properties of the explosives to obtain the physical model of the explosion source; the initial conditions of the explosion center include the initial pressure, temperature, energy density and initial expansion velocity of the explosion center;

[0015] According to the type of explosion source and its surrounding environment information, a nonlinear equation based on fluid mechanics is introduced to describe the reflection, scattering, diffraction and attenuation effects of shock waves during propagation. Combined with the physical law of explosion shock wave propagation in the air, the mathematical equation of shock wave propagation of explosion shock wave in the air is obtained.

[0016] After the explosion source physical model and the shock wave propagation mathematical equation are combined, the continuous mathematical equation is converted into a finite element equation using the Galerkin method, and the field variables in the calculation domain are interpolated and calculated using a three-dimensional first-order four-node element shape function;

[0017] The finite element equation is iterated in a discrete manner to form a shock wave-bridge structure fluid-solid coupling calculation form, and the bridge stress state formed by the shock wave on the surface of each part of the bridge is obtained, including the overpressure peak value, impulse, and duration;

[0018] According to the material characteristics and mechanical properties of the bridge structure, determine the damage thresholds of different structures under different stress states;

[0019] According to the material characteristics and damage threshold of the bridge structure, combined with the stress state of the bridge, the structural response and damage of the bridge structure under various stress states are analyzed;

[0020] A bridge damage simulation model is established by combining the bridge stress state and the bridge structure damage analysis results.

[0021] Preferably, the step of establishing the bridge damage database includes:

[0022] Measuring key parameters of the bridge structure, setting explosion conditions according to the operating condition information including explosion location, explosion equivalent and explosion type, traversing sampling from the value range of the key parameters, and using the bridge damage simulation model to obtain bridge damage simulation data; the key parameters include the value range of the size and reinforcement ratio of the bridge deck and pier;

[0023] Associating the bridge damage simulation data with corresponding explosion conditions to obtain condition-associated bridge damage data, wherein the explosion condition parameters include explosion equivalent, explosion location, and explosion type;

[0024] Integrating bridge characteristic information, including bridge geometric characteristics and bridge material characteristics, with the working condition-associated bridge damage data to obtain structure-coupled bridge damage data, wherein the bridge material characteristics include elastic modulus, yield strength, and fracture toughness;

[0025] The explosion condition, bridge material characteristics, and bridge geometric characteristics in the structurally coupled bridge damage data are subjected to data pruning, normalization, and feature extraction to obtain an explosion feature vector that integrates multi-source information;

[0026] For the explosion conditions and bridge objects involved in different scenarios, the explosion feature vector of comprehensive multi-source information in each scenario is obtained, and all scenarios and corresponding bridge damage simulation data are stacked to form a set with the explosion feature vector of comprehensive multi-source information as input and the bridge damage simulation data as output, which is entered into the library system;

[0027] The set is divided to obtain a bridge damage database including a training data set and a validation data set.

[0028] Preferably, the multi-layer composite graph neural network includes a hierarchical attention module, a hierarchical feature interaction module and an upsampling prediction module:

[0029] The hierarchical attention module includes a geometric feature attention layer, a load feature attention layer and a material feature attention layer, each of which is composed of a graph attention network and a context information extractor; the graph nodes of the geometric feature attention layer are embedded with bridge geometric features; the graph nodes of the load feature attention layer are embedded with explosion condition features; the graph nodes of the material feature attention layer are embedded with bridge material features;

[0030] The hierarchical feature interaction module uses the FM model to perform feature interaction on the outputs of the geometric feature attention layer, the load feature attention layer, and the material feature attention layer to obtain graph data;

[0031] The upsampling prediction module includes m GN layers and a graph decoding layer, wherein, except for the first GN layer input, the inputs of the remaining GN layers are the superposition of the input and input of the previous GN layer, and the input and output of the last GN layer are superimposed as the input of the graph decoding layer, and the output of the graph decoding layer is the bridge damage data.

[0032] Preferably, the multi-layer composite graph neural network is trained using a distributed subgraph training method, and the specific steps include:

[0033] Step 1: Adaptively divide the subgraph according to the density of nodes and edges in the input data: when the degree of a node is less than the set degree threshold, the node is considered to be the segmentation edge of the subgraph, otherwise the node is considered not to be the segmentation edge of the subgraph; complete the subgraph division according to the determined segmentation edge;

[0034] Step 2: During the training process, when the update frequency of node features in a subgraph in a training epoch exceeds the set frequency threshold, k-mean clustering is performed on the subgraph node information to achieve gradual refinement of the subgraph structure and node features;

[0035] Step 3: Monitor the calculated gradient changes of the subgraph during the training process, and compensate and adjust the gradient according to the update speed of the subgraph nodes; the gradient compensation formula is expressed as:

[0036] ,

[0037] Where, is the gradient after compensation, represents the gradient size of subgraph node i at time t, is the gradient size of subgraph node i at time t-1, represents the time difference, is the constant that controls the influence of the compensation gradient, and T is the global time constant;

[0038] Step 4: Through repeated cycles of step 1, step 2, and step 3, the training of the multi-layer composite graph neural network is realized.

[0039] Preferably, the target bridge damage prediction step includes:

[0040] Obtain the bridge characteristic information of the target bridge and the on-site explosion conditions, and construct an on-site explosion input data set;

[0041] The on-site explosion input data set is input into the bridge damage prediction model, and the prediction result of the target bridge damage is output.

[0042] In the second aspect, a bridge damage deep learning system integrating multi-source features is also provided, including:

[0043] A simulation unit, used to establish a bridge damage simulation model and obtain bridge damage simulation data;

[0044] A database unit is used to associate the bridge damage simulation data with the explosion condition and integrate it with the bridge characteristic information to obtain a bridge damage database including a training data set and a verification data set;

[0045] A model unit, used to train and optimize the constructed multi-layer composite graph neural network using the above training data set and verification data set to obtain a bridge damage prediction model;

[0046] The prediction unit is used to use the bridge damage prediction model to complete the prediction of target bridge damage.

[0047] According to a third aspect, a computer-readable storage medium storing one or more programs is provided. The one or more programs include instructions. When the instructions are executed by a computing device, the computing device executes the method as described above.

[0048] In a fourth aspect, an electronic device is also provided, comprising one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the method as described above.

[0049] Compared with the prior art, the present invention has the following significant advantages:

[0050] 1. High prediction accuracy: Based on the information of geometric characteristics, material characteristics, explosion load characteristics and other aspects of the bridge structure, the present invention constructs a multi-level graph neural network model and integrates the coupling influence between complex multi-source data, thereby greatly improving the accuracy of the model in predicting the damage of the bridge structure under the impact of explosion.

[0051] 2. Good adaptability: The present invention aims at different locations of bridges with inconsistent bridge materials and structural strength (such as bridge decks and piers), and introduces a multi-scenario and multi-source information damage database to improve the robustness of model training and achieve adaptive and accurate prediction of damage at different bridge locations.

[0052] 3. High model training efficiency: To address the problem of model training efficiency of large-scale data sets, this paper proposes a distributed subgraph training strategy. By scientifically dividing subgraphs, it breaks through the training upper limit and efficiency of graph neural networks on large-scale data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is the main flow chart of the present invention.

[0054] Figure 2 It is a flow chart of the steps for establishing a bridge damage simulation model.

[0055] Figure 3 It is a flow chart of the steps for establishing a bridge damage database.

[0056] Figure 4 It is a flow chart of the steps for establishing a bridge damage prediction model.

[0057] Figure 5A It is the hierarchical attention module and hierarchical feature interaction module of the bridge damage prediction model.

[0058] Figure 5B It is the upsampling prediction module.

[0059] Figure 6 It is a flow chart of the target bridge damage prediction steps. DETAILED DESCRIPTION

[0060] like Figure 1 As shown, the bridge structure damage prediction method based on the multi-layer composite graph neural network of the present invention comprises the following steps:

[0061] S10. Establishment of bridge damage simulation model: Establish a physical model of the explosion source, analyze the explosion damage of the bridge structure by solving the shock wave propagation mathematical equation expressed by the finite element equation and the mathematical description of the material response, and obtain a bridge damage simulation model;

[0062] The present invention uses the finite element analysis method to establish a numerical simulation model of the damage of the bridge structure under the impact of the explosion, providing accurate basic data for the training of the graph neural network prediction method of the adaptive explosion equivalent and position. Specifically, it includes: simulating and calculating to determine the stress state formed on the bridge surface by the explosion impact, determining the component damage threshold according to the material characteristics and mechanical properties of the bridge structure, analyzing the mechanical response of the bridge structure under stress, and obtaining the damage data of the bridge under different explosion conditions.

[0063] like Figure 2 As shown, the steps of establishing the bridge damage simulation model include:

[0064] S11. Establishing a physical model of the explosion source: The present invention conducts research on chemical explosion sources, establishes the energy release function of explosives at different equivalents and densities over time based on the principles of explosion mechanics, and determines the initial conditions of the explosion center point in combination with the physical and chemical properties of the explosives, including the initial pressure, temperature, energy density and initial expansion velocity of the explosion center;

[0065] S12. Mathematical description of shock wave propagation: According to the type of explosion source and its surrounding environment information, a nonlinear equation based on fluid mechanics (Euler equation) is introduced to describe the reflection, scattering, diffraction and attenuation effects of shock waves during propagation. Combined with the physical law of explosion shock wave propagation in the air, the mathematical equation of shock wave propagation of explosion shock wave in the air is obtained;

[0066] S13, mathematical equation finite element calculation: after the explosion source physical model and the shock wave propagation mathematical equation are combined, the continuous mathematical equation is converted into a finite element equation using the Galerkin method, and the field variables in the calculation domain are interpolated and calculated using a three-dimensional first-order four-node element shape function;

[0067] The simultaneous equations are as follows:

[0068] (1) Mathematical equations of the explosion source physical model:

[0069] ,

[0070] Where Δρ(r) is the local change determined by the mass distribution of the material released by the explosion, and r is the distance from the center of the explosion (source) to any point in space.

[0071] ,

[0072] In the formula, R 0 is the characteristic radius of the explosion source.

[0073] ,

[0074] In the formula, is the normalized radial distance.

[0075] (2) Mathematical equation of shock wave propagation:

[0076] ,

[0077] Where ρ is the air density, is the velocity vector and t is the time.

[0078] ,

[0079] Where p is the pressure.

[0080] ,

[0081] In the formula, , total energy density, e is the internal energy per unit mass.

[0082] ,

[0083] In the formula, is the specific heat ratio (γ=1.4 is usually taken in air).

[0084] S14. Determine the stress state of the bridge: iterate the finite element equation to form a shock wave-bridge structure fluid-solid coupling calculation form, and obtain the stress state of the bridge formed by the shock wave on the surface of each part of the bridge, including the overpressure peak, impulse, and duration.

[0085] S15. Determination of bridge damage threshold: Determine the damage threshold of different structures under different stress states based on the material characteristics and mechanical properties of the bridge structure.

[0086] Study the material characteristics of the main components of the facility (such as bridge decks and bridge piers), including their basic mechanical properties such as compression, tension, and shear. Based on the material mechanical parameters, determine the damage threshold of the bridge structure under different stress states to provide a physical basis for subsequent damage analysis.

[0087] S16. Bridge damage analysis: Based on the material properties and damage threshold of the bridge structure and combined with the stress state of the bridge, analyze the structural response of the bridge under various stress states and the damage of the bridge structure.

[0088] S17. Simulation model establishment: Based on the bridge stress state and the bridge structure damage analysis results, a bridge damage simulation model is established.

[0089] S20. Establishing a bridge damage database: associating the bridge damage data calculated by the bridge damage simulation model with the explosion condition data, and integrating it with the bridge characteristic information, including bridge geometric characteristics and bridge material characteristics, extracting multi-source data features, storing the data, and obtaining a bridge structure explosion damage database including a training data set and a verification data set.

[0090] Based on the damage data calculated by the bridge damage simulation model, key damage data of the bridge structure are obtained, including the stress distribution, failure mode, damage degree of components, and residual strength of the bridge structure. These data are structured and preprocessed, and coupled with the corresponding explosion conditions, bridge structure geometry, and material characteristics to form a comprehensive bridge damage database, providing high-quality basic data for subsequent model training.

[0091] like Figure 3 As shown, the steps of establishing the bridge damage database include:

[0092] S21. Damage data collection: According to the actual situation of the bridge structure, different value ranges of key parameters are determined, including the value ranges of the bridge deck, bridge pier size and reinforcement ratio; explosion conditions are set according to the explosion location, explosion equivalent and explosion type; samples are traversed from the value ranges of the key parameters; and the bridge damage data is calculated using the bridge damage simulation model.

[0093] S22. Explosion condition data association: Associating the bridge damage data with corresponding explosion condition parameters, wherein the key explosion condition parameters include explosion equivalent, explosion location and explosion type.

[0094] S23. Integration of geometry and material information: Integrate the bridge characteristic information, including bridge geometry characteristics and bridge material characteristics, with the bridge damage data after associating them with the explosion condition data to obtain structure-coupled bridge damage data. The bridge material characteristics include elastic modulus, yield strength, and fracture toughness.

[0095] Steps S22 and S23 complete the preliminary association of data features, forming a corresponding relationship between input data such as bridge geometry and explosion conditions and bridge damage data, ensuring that the subsequent damage prediction model can perform damage prediction based on bridge structural characteristics and explosion conditions.

[0096] S24. Multi-source data preprocessing: The explosion key operating parameters after the structural coupling bridge damage data is associated with the bridge damage data are coupled with the bridge material characteristics and bridge geometric characteristics. Through data pruning, normalization, and feature extraction, the explosion feature vector of comprehensive multi-source information is obtained.

[0097] The multi-source data preprocessing step includes:

[0098] S241, data pruning: pruning the data of each vector of material characteristics, explosion equivalent, explosion location and damage degree, and integrating the information of different dimensions in the explosion feature vector into the same dimension;

[0099] S242, data normalization: After data pruning is completed, the data needs to be normalized to the minimum and maximum values. This method scales all features to the range of 0, 1. The calculation formula can be expressed as:

[0100] ,

[0101] In the formula Represents the original data, represents the normalized data;

[0102] S243, feature extraction: Use a multi-layer convolutional neural network to downsample and encode the normalized data, remove redundant information, and extract the potential invariant features of the data.

[0103] S25. Data entry: For the explosion conditions and bridge structures involved in different scenarios, each scenario is preprocessed in turn through a multi-source data preprocessing method to obtain the explosion feature vector of the corresponding scenario with comprehensive multi-source information, and all scenarios and corresponding bridge structure response results are stacked to form a set with multi-scenario multi-source information explosion feature vectors as input features and a bridge structure response set as an output result set. The feature vector set is divided into a training data set and a verification data set, which are entered into the library system to obtain a bridge structure explosion damage database including the training data set and the verification data set.

[0104] S30. Establishing a bridge damage prediction model: cascading a hierarchical attention module, a hierarchical feature interaction module, and an upsampling prediction module to obtain a multi-layer composite graph neural network, and using the bridge damage database to train and optimize the multi-layer composite graph neural network to obtain a bridge damage prediction model based on the multi-layer composite graph neural network.

[0105] like Figure 4 As shown, the steps of establishing the bridge damage prediction model include:

[0106] S31, hierarchical attention module establishment: Figure 5AAs shown in the figure, based on the graph attention neural network and the context information extractor, a separate feature attention layer is formed, and a hierarchical attention module is formed by constructing multiple separate feature attention layers. The hierarchical attention module is designed to better learn the geometric features, load features, and material features of the bridge, and to compound the geometric features, load features, and material features of the bridge through an efficient attention aggregation strategy with low computational complexity.

[0107] The step of establishing the hierarchical attention module in S31 includes:

[0108] S311. Establishment of geometric feature attention layer: In the input data of a separate feature attention layer, the graph nodes are embedded with the length, width, and thickness information of the bridge to form a geometric feature attention layer.

[0109] S312, establishing a load feature attention layer: In the input data of a separate feature attention layer, graph nodes are embedded with explosion equivalent, explosion location, and explosion type to form a load feature attention layer;

[0110] S313. Establishment of material feature attention layer: In the input data of a separate feature attention layer, graph nodes are embedded with concrete properties, reinforcement ratio, and elastic modulus to form a material feature attention layer.

[0111] S32, hierarchical feature interaction: by introducing two random initialization vectors and multiplying them with the output of the hierarchical attention module, the second-order cross features of the modeling are explicitly modeled, and the geometric features, load features and material features of the hierarchical feature interaction are obtained as the input data of the graph neural network module;

[0112] The feature single interaction calculation process is:

[0113] ,

[0114] In the formula, y is the result after feature interaction, x i is the value of the i-th feature, n represents the number of features in vector x, and w 0 ,w i ,w ij These are all model parameters of the feature attention layer.

[0115] S33, upsampling prediction module establishment: Figure 5B As shown in Figure 1, an upsampling prediction module is established by stacking m graph network (GN) layers and then connecting the graph decoding layer.

[0116] Graph data obtained by layered feature interaction Iterative processing is performed through m GN layers of the same structure to achieve in-depth analysis and feature extraction of data; each GN block iteration includes the update of two key information, namely edge information and point information, to obtain abstract features , and finally through the graph decoding layer analysis Obtain bridge damage data and damage effect data of bridge structure.

[0117] S34. Model training: Use the distributed subgraph training method to train the multi-layer composite graph neural network. The specific steps include:

[0118] S341, adaptive subgraph partitioning: adaptively partition the subgraph size according to the density of nodes and edges in the graph data; determine whether the node has more connections by calculating the degree of the node. The judgment threshold set by the present invention is 7. When the node degree is less than the threshold, it is considered to be a subgraph segmentation edge.

[0119] S342, progressive subgraph refinement: During the training process, the information update frequency of the nodes in each subgraph is monitored, and the subgraphs with high activity are refined; when the information update frequency of the subgraph nodes in a training epoch exceeds the set frequency threshold (0.6 in this embodiment), k-mean clustering is performed on the nodes in the subgraph, and the subgraph is subdivided again according to the clustering results, so as to achieve progressive refinement of the subgraph structure and node information.

[0120] Specifically, the calculation formula for the update frequency is:

[0121] Update frequency = number of nodes with updated information in the subgraph / total number of nodes in the subgraph.

[0122] Specifically, in the process of k-mean clustering of nodes in the subgraph, the k value is optimized according to the elbow rule to divide the nodes in the subgraph into k clusters. After each clustering, the subgraph nodes are assigned to different clusters, and the nodes of each cluster will share a new cluster center, which is the mean of the position coordinates of all nodes in the cluster. Finally, the subgraph is divided again according to the clustering results, so that the nodes of the same cluster are in the same subgraph to enhance the efficiency of information transmission between nodes.

[0123] Specifically, the node allocation criterion uses minimizing the sum of squared errors (SSE) as the objective function, and the calculation formula is:

[0124] ,

[0125] In the formula, is the number of clusters, is the jth cluster, Indicates that it belongs to a cluster Nodes, Representation Cluster The cluster centers are used to assign nodes to different clusters.

[0126] Furthermore, after the subgraph is divided again, the node information needs to be updated by weighting the original information of the clustered nodes to update the existing node information instead of directly replacing it, so as to ensure the gradualness of the information update. At the same time, the clustered node information here refers to the parameters carried by the graph nodes during the training process.

[0127] S343, gradient compensation asynchronous update mechanism: by monitoring the calculated gradient changes of the distributed subgraph during the training process, the gradient of the asynchronous update is compensated and adjusted according to the update speed of the subgraph node; the gradient compensation formula can be expressed as:

[0128] ,

[0129] In the formula, Indicates the current gradient size of subgraph node i, is the corresponding historical cache gradient, represents the time difference between the two. is a constant that controls the degree of influence of the compensation gradient, and T is a global time constant.

[0130] S344. Through repeated cycles of S341, S342, and S343, the training of the multi-layer composite graph neural network is realized.

[0131] S35. Obtain a fully trained target model: Use the training data set and the above-mentioned training method to train the multi-layer composite graph neural network model and update the parameters. Use the verification data set to evaluate and optimize the trained multi-layer composite graph neural network to obtain a multi-layer composite bridge damage prediction model that can comprehensively analyze multi-source characteristics.

[0132] S40, bridge damage prediction: construct an on-site explosion input data set based on target bridge feature information, including bridge geometry features and bridge material features, as well as on-site explosion data, input the on-site explosion input data set data into the multi-layer composite bridge damage prediction model, and the output is the target bridge damage prediction result.

[0133] like Figure 6 As shown, the bridge damage prediction step includes:

[0134] S41, constructing a prediction input data set: obtaining geometric features and material features of a target bridge, as well as on-site explosion data, and constructing an on-site explosion input data set;

[0135] The step of constructing the prediction input data set in S41 is specifically as follows:

[0136] According to the parameters of the target bridge structure, material type, explosion equivalent, and explosion distance, a prediction input data set is constructed through data preprocessing, including data pruning, normalization, and feature extraction.

[0137] S42, bridge structure damage prediction: using the on-site explosion prediction input data set data as the input of the optimized bridge damage prediction model, and the output thereof is the target bridge damage prediction result.

[0138] Through the above steps, a rapid prediction method for explosion damage to bridge structures is established. Through numerical simulation, the overpressure distribution and damage data set of bridge facilities under multi-scenario explosions are obtained. Through data processing and conversion, a training database is constructed, and a multi-layer composite graph neural network digital intelligence prediction model is established. The distributed subgraph method is used for training, thereby obtaining a rapid prediction model for explosion damage to bridge structures with fast learning ability and strong generalization ability. Error evaluation and parameter tuning of the model are performed to improve the prediction ability of the digital intelligence prediction model, and to achieve real-time and accurate prediction of the damage of bridge facilities under various types of explosion impacts.

[0139] The present invention also designs a bridge damage deep learning system integrating multi-source features, including:

[0140] A simulation unit, used to establish a bridge damage simulation model and obtain bridge damage simulation data;

[0141] A database unit is used to associate the bridge damage simulation data with the explosion condition and integrate it with the bridge characteristic information to obtain a bridge damage database including a training data set and a verification data set;

[0142] A model unit, used to train and optimize the constructed multi-layer composite graph neural network using the above training data set and verification data set to obtain a bridge damage prediction model;

[0143] The prediction unit is used to use the bridge damage prediction model to complete the prediction of target bridge damage.

[0144] The technical solution of the above-mentioned bridge damage deep learning system is similar to the aforementioned bridge damage deep learning method, which will not be repeated here.

[0145] Based on the same technical solution, the present invention also discloses a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, enable the computing device to execute the above-mentioned deep learning method for bridge damage that integrates multi-source features.

[0146] Based on the same technical solution, the present invention also discloses a computing device, including one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the above-mentioned deep learning method for bridge damage that integrates multi-source features.

[0147] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0149] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0150] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0151] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A deep learning method for bridge damage integrating multi-source features, characterized in that: The steps include: Establish a bridge damage simulation model and obtain bridge damage simulation data; The bridge damage simulation data is associated with the explosion conditions and integrated with the bridge characteristic information to obtain a bridge damage database including a training data set and a validation data set; The constructed multi-layer composite graph neural network is trained and optimized using the training data set and the validation data set to obtain a bridge damage prediction model. Use the bridge damage prediction model to predict the damage of the target bridge; The steps to establish a bridge damage database include: Measure the key parameters of the bridge structure, set the explosion conditions according to the working condition information including the explosion location, explosion equivalent and explosion type, traverse the sampling from the key parameter value range, and use the bridge damage simulation model to obtain the bridge damage simulation data; the key parameters include the size of the bridge deck and piers and the value range of the reinforcement ratio; Associating the bridge damage simulation data with corresponding explosion conditions to obtain condition-associated bridge damage data, wherein the explosion conditions include explosion equivalent, explosion location, and explosion type; The bridge characteristic information is integrated with the working condition-related bridge damage data to obtain the structural coupled bridge damage data; the bridge characteristic information includes bridge geometric characteristics and bridge material characteristics; the bridge material characteristics include elastic modulus, yield strength, and fracture toughness; The explosion conditions, bridge material characteristics, and bridge geometric characteristics in the structural coupled bridge damage data are trimmed, normalized, and feature extracted to obtain the explosion feature vector that integrates multi-source information. For the explosion conditions and bridge objects involved in different scenarios, the explosion feature vector of comprehensive multi-source information in each scenario is obtained, and all scenarios and corresponding bridge damage simulation data are stacked to form a set with the explosion feature vector of comprehensive multi-source information as input and the bridge damage simulation data as output, which is entered into the library system; Dividing the set to obtain a bridge damage database including a training data set and a validation data set; The distributed subgraph training method is used to train the multi-layer composite graph neural network. The specific steps include: Step 1: Adaptively divide the subgraph according to the density of nodes and edges in the input data: when the degree of a node is less than the set degree threshold, the node is considered to be the segmentation edge of the subgraph, otherwise the node is considered not to be the segmentation edge of the subgraph; complete the subgraph division according to the determined segmentation edge; Step 2: During the training process, when the update frequency of node information in a subgraph in a training epoch exceeds the set frequency threshold, k-mean clustering is performed on the nodes in the subgraph, and the subgraph is divided again according to the clustering result. The principle of the subgraph division is that the nodes in the same cluster are in the same subgraph; the node information refers to the parameters carried by the node during the training process; Step 3: Monitor the calculated gradient changes of the subgraph during the training process, and compensate and adjust the gradient according to the update speed of the subgraph nodes; the gradient compensation formula is expressed as: ， In the formula, is the gradient after compensation, Indicates the gradient size of subgraph node i at time t, is the gradient size of subgraph node i at time t-1, represents the time difference, which is a constant for controlling the influence of the compensation gradient, and T is the global time constant; Step 4: Through repeated cycles of step 1, step 2, and step 3, the training of the multi-layer composite graph neural network is realized.

2. The method according to claim 1, characterized in that The steps to establish the bridge damage simulation model include: According to the principle of explosion mechanics, the energy release function of explosives with different equivalents and densities over time is established. Combined with the physical and chemical properties of explosives, the initial conditions of the explosion center are determined to obtain the physical model of the explosion source; the initial conditions of the explosion center include the initial pressure, temperature, energy density and initial expansion velocity of the explosion center; According to the type of explosion source and its surrounding environment information, a nonlinear equation based on fluid mechanics is introduced to describe the reflection, scattering, diffraction and attenuation effects of shock waves during propagation. Combined with the physical law of explosion shock wave propagation in the air, the mathematical equation of shock wave propagation of explosion shock wave in the air is obtained. After combining the explosion source physical model and the shock wave propagation mathematical equation, the continuous mathematical equation is transformed into a finite element equation using the Galerkin method, and the field variables in the calculation domain are interpolated through the three-dimensional first-order four-node element shape function. The finite element equations are iterated in a discrete manner to form a shock wave-bridge structure fluid-solid coupling calculation form, and the stress state of the bridge formed by the shock wave on the surface of each part of the bridge is obtained, including the overpressure peak, impulse, and duration; According to the material characteristics and mechanical properties of the bridge structure, determine the damage thresholds of different structures under different stress states; According to the material characteristics and damage threshold of the bridge structure, combined with the stress state of the bridge, the structural response and damage of the bridge structure under various stress states are analyzed; A bridge damage simulation model is established by combining the bridge stress state and the bridge structure damage analysis results.

3. The method according to claim 1, characterized in that The multi-layer composite graph neural network includes a hierarchical attention module, a hierarchical feature interaction module, and an upsampling prediction module: The hierarchical attention module includes a geometric feature attention layer, a load feature attention layer, and a material feature attention layer. Each of the geometric feature attention layer, the load feature attention layer, and the material feature attention layer is composed of a graph attention network and a context information extractor. The graph nodes of the geometric feature attention layer embed the geometric features of bridges; The graph nodes of the load feature attention layer are embedded with explosion condition features; the graph nodes of the material feature attention layer are embedded with bridge material features; The hierarchical feature interaction module uses the FM model to perform feature interaction on the outputs of the geometric feature attention layer, the load feature attention layer, and the material feature attention layer to obtain graph data ; The upsampling prediction module includes m GN layers and image decoding layers, where the first GN layer input is The inputs of the remaining GN layers are the superposition of the input and output of the previous GN layer. The input and output of the last GN layer are superimposed as the input of the image decoding layer. The output of the image decoding layer is the bridge damage data.

4. The method according to claim 1, characterized in that The expression of update frequency is as follows: Update frequency = number of nodes with updated information in the subgraph / total number of nodes in the subgraph.

5. The method according to claim 1, characterized in that When sub-dividing again, it also includes: The weight is set according to the distance between each node in each cluster and the cluster center, and the information of each node is weighted updated; the smaller the distance to the cluster center, the larger the weight is set.

6. A bridge damage deep learning system based on the method as claimed in any one of claims 1 to 5, characterized in that: include: A simulation unit, used to establish a bridge damage simulation model and obtain bridge damage simulation data; A database unit is used to associate the bridge damage simulation data with the explosion condition and integrate it with the bridge characteristic information to obtain a bridge damage database including a training data set and a verification data set; A model unit is used to train and optimize the constructed multi-layer composite graph neural network using a training data set and a validation data set to obtain a bridge damage prediction model; The prediction unit is used to complete the prediction of target bridge damage by using the bridge damage prediction model.

7. A computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, characterized in that When the instructions are executed by a computing device, the computing device is caused to perform the method as claimed in any one of claims 1 to 5.

8. An electronic device, characterized in that: The method comprises one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the method as claimed in any one of claims 1 to 5.

Images