Concrete penetration response prediction method
By combining graph neural networks with Bayesian uncertainty assessment, the problem of low computational efficiency in predicting the penetration response of concrete structures is solved, achieving efficient and accurate prediction results and reliable support, which is applicable to damage assessment of concrete structures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies have low computational efficiency in predicting the penetration response of concrete structures, making it difficult to meet the needs of rapid engineering assessment. Furthermore, traditional deep learning models do not fully utilize finite element mesh information and lack descriptions of result reliability.
A response prediction model based on a graph attention network is constructed using graph neural networks and Bayesian uncertainty assessment. This model is then combined with a finite element simulation model to predict the concrete penetration response. The reliability of the prediction results is provided through Dropout operation and Bayesian uncertainty assessment.
It improves computational efficiency, can more accurately describe the concrete penetration process, provides high-precision prediction results and reliability support, and saves computational resources.
Smart Images
Figure CN122065591A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of concrete structure damage prediction technology, and in particular relates to a method for predicting concrete penetration response. Background Technology
[0002] Concrete structures are widely used in protective engineering, such as underground bunkers, airport runways, and bridges. Under extreme loads such as projectile penetration and explosive impact, the dynamic response and damage evolution of concrete structures are important foundations for damage assessment.
[0003] Currently, the response prediction of concrete structures under penetration loading typically employs the finite element method (FEM), which can consider complex geometries, reinforcement configurations, and boundary conditions, and is capable of handling instantaneous dynamic responses, large deformations, material nonlinearities, and complex inter-structural interactions. However, these high-precision FEM simulations usually require fine meshes and extremely small time steps, resulting in large computational scales, long processing times, and high computational resource requirements. When multi-condition parameter analysis is needed, the overall computation cycle often fails to meet the demands of rapid evaluation in engineering projects.
[0004] In recent years, to overcome the limitations of traditional numerical simulation in terms of computational efficiency, researchers have begun to introduce deep learning techniques for approximating complex physical processes. However, traditional black-box regression models often simplify the input to vectors or image features, rarely making full use of the structural information contained in the finite element mesh, and generally lack a quantitative description of the reliability of the prediction results.
[0005] To address the aforementioned issues, this invention proposes an accelerated prediction method for concrete penetration response based on graph neural networks and Bayesian uncertainty assessment. Summary of the Invention
[0006] The purpose of this invention is to provide a method for predicting concrete penetration response, so as to solve the problems existing in the prior art.
[0007] To achieve the above objectives, the present invention provides a method for predicting concrete penetration response, comprising:
[0008] S1: Construct a finite element simulation model of the concrete structure based on the working condition data in the penetration test;
[0009] S2: Input the simulation data of the finite element simulation model into the response prediction model for classification and prediction, and output the node response data of each node; wherein, the response prediction model is based on graph neural network and includes a multi-layer graph attention network and a linear output layer connected in sequence, and a Dropout operation is applied between adjacent graph attention networks;
[0010] S3: Perform Bayesian uncertainty assessment on the response data of each node and output the credibility of the response data of each node.
[0011] Optionally, the process of constructing the finite element simulation model in step S1 specifically includes:
[0012] S11: Based on the working condition data in the penetration test, establish a three-dimensional geometric model corresponding to the concrete material, projectile and steel reinforcement, and divide it into finite element meshes;
[0013] S12: Define the constitutive model and parameters of the concrete material, set the contact boundary conditions and analysis step size;
[0014] S13: Conduct numerical simulation analysis to obtain the finite element simulation model of the concrete structure.
[0015] Optionally, the training process of the response prediction model in step S2 specifically includes:
[0016] S21: Obtain training data: Construct a node feature dataset based on historical time series using simulation data from the finite element simulation model, and generate a node adjacency graph structure based on the node feature dataset and K nearest neighbors. The node adjacency graph structure includes node coordinates, node features, and edge index matrices corresponding to each time frame.
[0017] S22: Model Training: Construct an initial graph neural network model based on the graph attention mechanism, input the training data into the initial graph neural network model for classification prediction, and train it according to the target loss function to obtain the trained response prediction model.
[0018] Optionally, the process of constructing the node feature dataset in step S21 specifically includes:
[0019] Extract the position information and node response of each node at each time step from the simulation data of the finite element simulation model;
[0020] Standardize and normalize the node responses of all nodes across the entire time sequence.
[0021] Select the number of historical frames num_history. For the time frame t to be predicted, select the data of num_history time steps before time frame t to form the time series history; where t≥num_history.
[0022] For each node, extract its 3D coordinates and normalized response data at each time frame of the time-series history. Concatenate the extracted data in chronological order to obtain the node's temporal feature vector.
[0023] Stack the node time sequence feature vectors of each node in the current time frame row by row to obtain the node feature matrix of the current time frame.
[0024] Optionally, the process of constructing the node adjacency graph structure in step S21 specifically includes:
[0025] The node coordinates of the current time frame are selected as the spatial features of the nearest neighbor search. The k-nearest neighbor algorithm is used to find the nearest neighbor node of each node at a preset Euclidean distance. An edge relationship is established between each node and its neighboring nodes, and the edge index matrix is obtained by summarizing the results.
[0026] The node feature matrix, edge index matrix, and node coordinates corresponding to the current time frame are encapsulated into a graph data object for use in the training and application of graph neural networks.
[0027] Optionally, the processing procedure of the response prediction model in step S2 specifically includes:
[0028] The response prediction model was set to Dropout enabled mode, and forward propagation was repeated 30 times for each test time frame and the corresponding graph data to obtain 30 prediction results for each node.
[0029] Optionally, step S3 specifically includes: calculating the mean and standard deviation of each prediction result for each node, using the mean as the predicted value of the node response, and using the standard deviation as the uncertainty of the prediction result.
[0030] The technical effects of this invention are as follows:
[0031] This invention employs a graph neural network based on a finite element mesh, which can more accurately describe the physical mechanism of concrete penetration compared to a regression model that only uses simplified features.
[0032] The graph neural network model of this invention can be used as an acceleration module for traditional finite element simulation analysis, which can save computation time and resources to the greatest extent while maintaining high computational accuracy.
[0033] This invention introduces Bayesian uncertainty assessment to provide the predicted mean and standard deviation of the node response, which can support the credibility of the prediction results. Attached Figure Description
[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0035] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0036] Figure 1 This is a flowchart illustrating the implementation of this invention.
[0037] Figure 2 This is a schematic diagram of the k-nearest neighbor algorithm in an embodiment of the present invention;
[0038] Figure 3 This is a diagram of the graph neural network framework in an embodiment of the present invention;
[0039] Figure 4 This is a Bayesian uncertainty histogram in an embodiment of the present invention;
[0040] Figure 5 This is an example diagram of the prediction results in an embodiment of the present invention. Detailed Implementation
[0041] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as a limitation of the present invention, but rather as a more detailed description of certain aspects, features, and embodiments of the present invention.
[0042] It should be understood that the terminology used in this invention is merely for describing particular embodiments and is not intended to limit the invention. Furthermore, with respect to numerical ranges in this invention, it should be understood that each intermediate value between the upper and lower limits of the range is also specifically disclosed. Every smaller range between any stated value or intermediate value within a stated range, and any other stated value or intermediate value within said range, is also included in this invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.
[0043] Various modifications and variations can be made to the specific embodiments described in this specification without departing from the scope or spirit of the invention, as will be apparent to those skilled in the art. Other embodiments derived from this specification will also be obvious to those skilled in the art. This application specification and embodiments are merely exemplary.
[0044] The terms “include,” “including,” “have,” “contain,” etc., used in this article are all open-ended terms, meaning that they include but are not limited to.
[0045] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0046] Example 1
[0047] like Figure 1 - Figure 5 As shown, this embodiment provides a method for predicting concrete penetration response, including:
[0048] Step 1: Finite element simulation and data extraction;
[0049] Step 2: Construct a node feature dataset based on historical time series;
[0050] Step 3: Generate a node adjacency graph structure based on K-nearest neighbors;
[0051] Step 4: Construct and train the graph neural network model;
[0052] Step 5: Bayesian uncertainty estimation and prediction.
[0053] The feasible step 1 is as follows:
[0054] Step 1.1: Based on the actual working conditions in the penetration test, establish three-dimensional geometric models of the target, projectile, and reinforcing steel, divide them into finite element meshes, define the constitutive models and parameters of each material, set the contact boundary conditions and analysis step size, and carry out numerical simulation analysis.
[0055] Step 1.2: After obtaining the numerical calculation results, compare the simulation data with the penetration test data to verify the correctness of the simulation results.
[0056] Step 1.3: Based on the numerical simulation results, extract the position information and response of each node at multiple time steps, such as equivalent plastic strain. Extract the node coordinates at each time step t. Store as The array will correspond to the strain Store as The array, obtained by stacking data from multiple time steps, is as follows:
[0057]
[0058] Where N is the total number of nodes and T is the number of time frames.
[0059] The above data is stored in an array file (.npz) for easy retrieval later.
[0060] The feasible step 2 is as follows:
[0061] Step 2.1: Strain analysis of the entire time series and all nodes. Standardization is performed. The strain across all time frames is unfolded into a one-dimensional array, and the global mean is calculated. and standard deviation And normalize according to the following formula:
[0062]
[0063] Step 2.2: Select the number of historical frames num_history. In this embodiment, it is set to 4. For the t-th frame to be predicted (t≥num_history), take the data of the first 4 time steps to form the time series history.
[0064] Step 2.3: For each node i, extract the 3D coordinates and normalized strain of that node at each time frame in the history. The node temporal feature vector is obtained by concatenating them in chronological order. :
[0065]
[0066] when At that time, the feature dimension of a single node is All The features of each node are stacked row by row to obtain the node feature matrix of the current time frame:
[0067]
[0068] Step 2.4: Randomly shuffle the time series data and divide it into training, validation, and test sets according to a preset ratio of 70%:15%:15%. This ensures that the sampling along the time dimension is representative.
[0069] The feasible step 3 is as follows:
[0070] Step 3.1: Select the node coordinates of the current time frame , as a spatial feature of nearest neighbor search.
[0071] Step 3.2: Use the k-nearest neighbor algorithm with k=10 to find the 10 nearest Euclidean neighbors for each node i.
[0072] Step 3.3: Establish edge relationships between each node i and its neighboring nodes, and summarize them to obtain the edge index matrix edge_index.
[0073] Step 3.4: Convert the node feature matrix The edge index matrix edge_index and the node coordinates corresponding to the current frame are encapsulated into a graph data object for use in the training and use of graph neural networks.
[0074] The feasible step 4 is as follows:
[0075] Step 4.1: Construct a graph neural network model based on graph attention mechanism. In this embodiment, a 4-layer graph attention network is used. Each layer includes graph attention convolution, normalization, and ELU activation, and a Dropout operation is applied between layers with a preset probability of 0.15. The network has a hidden layer dimension of 128 and 6 attention heads. After the output of the last graph attention layer, a linear output layer is connected to obtain the predicted strain corresponding to each node.
[0076] Step 4.2: When training the network model, randomly shuffle N nodes in each time frame and divide them into multiple batches of 6000 nodes each.
[0077] Step 4.3: For each batch, generate the corresponding graph data object, input it into the graph neural network, and obtain the prediction results. Calculate the smoothed L1 loss based on the actual strain values, and use the AdamW optimizer for parameter backpropagation and update.
[0078] Step 4.4: Traverse all time frames and node samples in the training set, record the changes in loss and error between the training set and the validation set, adjust the learning rate based on the performance of the validation set, and determine the number of training rounds.
[0079] The feasible step 5 is as follows:
[0080] Step 5.1: Put the trained graph neural network into Dropout enabled mode, and repeat the forward propagation 30 times for each test time frame and its corresponding graph data to obtain 30 prediction samples for each node.
[0081] Step 5.2: Calculate the mean and standard deviation of the predicted samples for each node. Use the mean as the predicted value of the node strain and the standard deviation as the uncertainty of the prediction result.
[0082] In summary, this embodiment constructs a spatiotemporal graph dataset based on finite element numerical simulation and experimental data. It further trains a graph neural network to quickly predict the nodal response of the structure and uses Monte Carlo Dropout to perform Bayesian uncertainty assessment. It also outputs various statistical and visualization results. This embodiment can provide a technical solution that balances efficiency and reliability for damage assessment of concrete structures.
[0083] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for predicting the penetration response of concrete, characterized in that, include: S1: Construct a finite element simulation model of the concrete structure based on the working condition data in the penetration test; S2: Input the simulation data of the finite element simulation model into the response prediction model for classification and prediction, and output the node response data of each node; wherein, the response prediction model is based on graph neural network and includes a multi-layer graph attention network and a linear output layer connected in sequence, and a Dropout operation is applied between adjacent graph attention networks; S3: Perform Bayesian uncertainty assessment on the response data of each node and output the credibility of the response data of each node.
2. The method for predicting concrete penetration response according to claim 1, characterized in that, The process of constructing the finite element simulation model in step S1 specifically includes: S11: Based on the working condition data in the penetration test, establish a three-dimensional geometric model corresponding to the concrete material, projectile and steel reinforcement, and divide it into finite element meshes; S12: Define the constitutive model and parameters of the concrete material, set the contact boundary conditions and analysis step size; S13: Conduct numerical simulation analysis to obtain the finite element simulation model of the concrete structure.
3. The method for predicting concrete penetration response according to claim 1, characterized in that, The training process of the response prediction model in step S2 specifically includes: S21: Obtain training data: Construct a node feature dataset based on historical time series using simulation data from the finite element simulation model, and generate a node adjacency graph structure based on the node feature dataset and K nearest neighbors. The node adjacency graph structure includes node coordinates, node features, and edge index matrices corresponding to each time frame. S22: Model Training: Construct an initial graph neural network model based on the graph attention mechanism, input the training data into the initial graph neural network model for classification prediction, and train it according to the target loss function to obtain the trained response prediction model.
4. The method for predicting concrete penetration response according to claim 3, characterized in that, The process of constructing the node feature dataset in step S21 specifically includes: Extract the position information and node response of each node at each time step from the simulation data of the finite element simulation model; Standardize and normalize the node responses of all nodes across the entire time sequence. Select the number of historical frames num_history. For the time frame t to be predicted, select the data of num_history time steps before time frame t to form the time series history; where t≥num_history. For each node, extract its 3D coordinates and normalized response data at each time frame of the time-series history. Concatenate the extracted data in chronological order to obtain the node's temporal feature vector. Stack the node time sequence feature vectors of each node in the current time frame row by row to obtain the node feature matrix of the current time frame.
5. The method for predicting concrete penetration response according to claim 4, characterized in that, The process of constructing the node adjacency graph structure in step S21 specifically includes: The node coordinates of the current time frame are selected as the spatial features of the nearest neighbor search. The k-nearest neighbor algorithm is used to find the nearest neighbor node of each node at a preset Euclidean distance. An edge relationship is established between each node and its neighboring nodes, and the edge index matrix is obtained by summarizing the results. The node feature matrix, edge index matrix, and node coordinates corresponding to the current time frame are encapsulated into a graph data object for use in the training and application of graph neural networks.
6. The method for predicting concrete penetration response according to claim 5, characterized in that, The processing procedure of the response prediction model in step S2 specifically includes: The response prediction model was set to Dropout enabled mode, and forward propagation was repeated 30 times for each test time frame and the corresponding graph data to obtain 30 prediction results for each node.
7. The method for predicting concrete penetration response according to claim 6, characterized in that, Step S3 specifically includes: calculating the mean and standard deviation of each prediction result for each node, using the mean as the predicted value of the node response, and using the standard deviation as the uncertainty of the prediction result.