A 3D Prediction Method and System for Explosion Damage to RC Structures Based on Voxelization Modeling

By using voxel modeling and local block segmentation, the problem of preserving global and local information in the three-dimensional prediction of RC structure explosion damage was solved, achieving high-precision three-dimensional damage morphology restoration and evaluation.

CN123069802APending Publication Date: 2026-11-10BEIJING INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611571348.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-12
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies struggle to simultaneously preserve global geometric information and local damage details of RC structure explosion damage under conditions of scarce samples. Direct whole-structure 3D modeling can easily lead to over-smoothing of high-strain zones and local damage zones, and lacks a reconstruction mechanism to bring local prediction results back to the spatial morphology of the whole structure.

Method used

By mapping the finite element level results to a unified three-dimensional voxel field and dividing them into local block samples, a prediction network for the working conditions and the local block strain voxel field is constructed. By utilizing the local block position encoding and reconstruction mechanism, high-precision prediction of the three-dimensional strain field of RC structures is achieved.

Benefits of technology

It outputs a complete three-dimensional strain field of the structure, which is rich in information and suitable for explosion damage assessment. It significantly expands the training sample size, preserves the details of local high strain areas, is suitable for three-dimensional convolutional network processing, and restores the true damage morphology.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069802A_ABST
    Figure CN123069802A_ABST
Patent Text Reader

Abstract

This invention belongs to the fields of explosion impact engineering, reinforced concrete structure damage assessment, structural intelligent computing, and artificial intelligence-assisted simulation technology. Specifically, it relates to a three-dimensional prediction method and system for explosion damage of RC structures based on voxel modeling. The specific process of this method is as follows: voxel index mapping is performed on the finite element elements of the RC structure. According to the voxel index mapping, all element-level strain fields are mapped to the strain element field of the whole structure. The strain element field of the whole structure is slidably cut into blocks according to the preset local block size and inter-block step size. A local block strain prediction network is constructed. The network input is the local block input feature composed of the spatial position code of the local block and the working condition. The network output is the predicted local block strain element field. The trained network is used to predict and map the local block strain element field to realize the prediction of the three-dimensional strain field.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the fields of explosion impact engineering, damage assessment of reinforced concrete structures, intelligent structural computing and artificial intelligence-assisted simulation technology, specifically involving a three-dimensional prediction method and system for explosion damage of RC structures based on voxel modeling. Background Technology

[0002] Current damage analysis of reinforced concrete (RC) structures under explosive loading mainly relies on live-fire tests, approximate empirical formulas, and high-fidelity finite element numerical simulations. Live-fire tests can directly obtain displacement response, failure area, and damage morphology, but they are costly, time-consuming, have limited coverage of operating conditions, and it is difficult to generate large-scale data samples. Finite element simulation can obtain fine-grained information such as element strain, displacement, and failure state under a large number of operating conditions at a relatively low cost, but directly using irregular finite element mesh results for intelligent algorithm modeling will face problems such as extremely high data dimensionality, inconsistent mesh topology, and difficulty in aligning samples for different operating conditions.

[0003] In existing technologies, one approach simplifies the RC structure into several global response indices, such as peak mid-span displacement, residual displacement, or total damage volume fraction, and then uses machine learning regression models to predict one or a few indices. While this approach is computationally simple, it cannot output three-dimensional damage morphology or characterize spatial damage distribution. Another approach attempts to learn directly from finite element-level results, but due to issues such as irregular mesh representations, large number of elements, and scarce samples in different working conditions, the models are prone to training instability, overly smooth prediction results, and difficulty in generalization.

[0004] For 3D image data, a common approach is to represent the structure as a regular voxel mesh. However, most existing publicly available solutions only treat voxelization as a simple preprocessing step, without addressing the following key challenges: First, the dimensionality of the whole-structure 3D voxel field is extremely high, and direct whole-field regression under limited sample size easily leads to meanification and smoothing phenomena; second, damage under explosion conditions is often highly localized and discontinuous, making it difficult for the whole-field model to effectively learn local high-gradient regions; third, there is a lack of a unified coordinate representation and reconstruction mechanism between local modeling and whole-structure morphology restoration.

[0005] In addition, although the idea of ​​image blocks exists in existing research, most of them are used for natural image classification or segmentation tasks. A complete technical link suitable for predicting explosion damage to RC structures has not yet been formed, including: extracting effective structural information from unit-level simulation results, constructing a unified voxel representation, dividing local blocks according to spatial location, using working conditions and local location to jointly predict the strain field, and then reconstructing the three-dimensional damage morphology of the whole structure from the local blocks.

[0006] In summary, existing technologies have at least the following drawbacks: First, it is difficult to simultaneously preserve the global geometric information and local damage details of RC structure explosion damage under conditions of scarce samples; second, directly modeling on the entire structure in a three-dimensional field can easily lead to over-smoothing of high-strain zones and local damage zones, and the output results can only reflect the average trend, making it difficult to recover the true damage morphology; third, finite element results are usually expressed as irregular elements, which cannot be directly and efficiently utilized by three-dimensional convolutional neural networks or other three-dimensional deep networks; fourth, there is a lack of feasible reconstruction mechanisms to return from local prediction results to the spatial morphology of the entire structure. Summary of the Invention

[0007] In view of this, the present invention proposes a three-dimensional prediction method and system for explosive damage of RC structures based on voxel modeling. By mapping the finite element level results to a unified three-dimensional voxel field and further dividing them into local block samples with spatial location codes, a technical route of "working conditions - prediction of local block strain voxel field - reconstruction of the three-dimensional strain field of the whole structure" is constructed, which realizes high-precision prediction of the three-dimensional strain field of RC structures under the condition of scarce samples.

[0008] The technical solution for implementing the present invention is as follows: In a first aspect, the present invention provides a three-dimensional prediction method for explosive damage to RC structures based on voxel modeling, the specific process of which is as follows: Step 1: Obtain the finite element simulation results of the reinforced concrete (RC) structure under different damage conditions, and extract the element-level strain field of each concrete element from the simulation results. Step 2: Set the number of regular voxel meshes for the whole structure, and uniformly divide the envelope of the RC structure in three-dimensional space into small cubic elements with the corresponding number of voxel meshes; perform voxel index mapping on the finite element elements of the RC structure, and map all element-level strain fields to whole structure strain voxel fields according to the voxel index mapping, and determine the mask value occupied by each voxel. Step 3: The entire structure strain element field is slidably cut into multiple local blocks according to the preset local block size and inter-block step size, and the local blocks are filtered according to the occupancy mask value. Step 4: Construct a local block strain prediction network. The network input is the local block input feature composed of the spatial location code of the selected local block and the working condition. The network output is the predicted local block strain element field. The network is trained using the real local block strain element field as the supervision signal and a loss function with occupancy mask value. Step 5: The working conditions of the damage condition to be predicted and the spatial location of the local block are encoded and spliced ​​together to form the local block input features. The trained network is used to predict the strain element field of the local block. According to the spatial index of the local block in the whole structure, the predicted strain element field of the local block is fused to obtain the strain element field of the whole structure. The strain element field of the whole structure is further mapped back to the finite element space to realize the prediction of the three-dimensional strain field.

[0009] Optionally, the present invention performs voxel index mapping on the RC structure finite element elements, the specific process of which is as follows: Calculate the envelope box of the RC structure in three-dimensional space, and add a boundary blank space of a preset proportion outside the envelope box; Given a set number of regular voxel grids for the whole structure, the envelope box is uniformly divided into small cubic units along three directions, corresponding to the number of regular voxel grids for the whole structure. The physical dimensions of each voxel are as follows: ; Let the center coordinates of each finite element of the RC structure be... Calculate the voxel subscript where it falls. Complete the voxel index mapping of the RC structure finite element elements;

[0010] in, This indicates a round-down operation. These are the minimum coordinates of the envelope box.

[0011] Optionally, in step two of the present invention, when multiple finite element elements fall into the same voxel, the strain voxel field of the corresponding voxel is taken as the average value of the element-level strain field of the contained elements, and its occupancy mask value is determined according to whether the voxel contains real structural material.

[0012] Optionally, in this invention, if the voxel contains real structural material, the occupancy mask value is 1; otherwise, the occupancy mask value is 0. In step three, the local block filtering based on the occupancy mask value is performed as follows: filtering out local blocks containing voxel occupancy mask values ​​whose sum is not 0.

[0013] Optionally, the present invention provides that the first in the local block The occupancy mask value of the individual element is The strain field predicted by the network is The real strain field is The training loss function is then defined as: , Among them, the weight term is , , Indicates the first Individual elements belong to the RC structural entity region. Indicates the first Individual elements belong to the spatial region of the RC structure. This is the weighting adjustment coefficient for the high-strain region. To prevent extremely small constants with a denominator of zero, This represents the L2 norm.

[0014] Optionally, the operating conditions of the present invention include operating parameters and derived parameters, wherein the derived parameters are at least one dimensionless combination of operating parameters, and the spatial position encoding includes at least the normalized coordinates of the start and end points of the local block on the whole structural dimensions.

[0015] Optionally, the operating parameters of the present invention include: explosive equivalent. Explosion distance And the offset of the explosion point relative to the centerline of the structure Derived parameters include: , as well as .

[0016] Optionally, the present invention sets the local block size and the inter-block step size so that there is overlap between local blocks.

[0017] Optionally, the present invention specifies the center coordinates of each finite element of the RC structure. The calculation process is as follows: Let the first The coordinates of the 8 nodes associated with each finite element are denoted as follows: Then the center coordinates of the finite element are: .

[0018] Secondly, the present invention provides a three-dimensional prediction system for explosive damage to RC structures based on voxel modeling, comprising: The data extraction module is used to obtain the finite element simulation results of reinforced concrete (RC) structures under different damage conditions, and extract the element-level strain field of each concrete element from them. The voxelization module is used to set the number of regular voxel meshes for the whole structure, uniformly divide the envelope of the RC structure in three-dimensional space into small cubic elements with the corresponding number of voxel meshes; perform voxel index mapping on the finite element elements of the RC structure, map all element-level strain fields to the whole structure strain voxel field according to the voxel index mapping, and determine the occupancy mask value of each voxel. The local block generation module is used to slide and cut the entire structure strain element field into multiple local blocks according to the preset local block size and inter-block step size, and to filter the local blocks according to the occupancy mask value. The local block strain prediction module is a local block strain prediction network, which is used to output the predicted local block strain element field based on the local block input features formed by splicing the spatial location codes of the selected and retained local blocks with the working conditions. The whole structure reconstruction and topography display module is used to fuse the predicted local block strain element fields to obtain the whole structure strain element field based on the spatial index of the local blocks in the whole structure, and further map the whole structure strain element field back to the finite element space to realize the prediction of the three-dimensional strain field.

[0019] Beneficial effects: First, compared with existing methods that only predict peak displacement or overall damage indicators, the present invention can output the three-dimensional strain field of the entire structure, which is richer in information and more suitable for explosion damage assessment and scheme comparison.

[0020] Second, compared with the direct whole-field modeling scheme, the present invention significantly expands the training sample size by local block segmentation, reduces the learning difficulty of high-dimensional field prediction under small sample conditions, and can more effectively preserve the details of local high strain zones and damage boundaries.

[0021] Third, compared with the method of directly manipulating irregular finite element meshes, the present invention adopts a unified voxelization representation, which is convenient for three-dimensional convolutional network processing, while retaining the ability to backmap voxel results to cell-level display.

[0022] Fourth, through local block position encoding and reconstruction mechanisms, this invention takes into account both global geometric positional relationships and local damage detail representation, avoiding the problem of traditional local models only seeing the part and not the whole. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, 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.

[0024] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0027] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0028] This invention comprises two parts: a method and a system. The method is used to implement data extraction, voxelization, local block modeling, and overall structure reconstruction; the system supports the modular implementation of the method. For ease of understanding, the invention will be described in detail below according to the implementation process.

[0029] The technical approach of this invention is as follows: First, information such as the coordinates of eight nodes associated with each concrete element, the maximum equivalent plastic strain of the element, and the element failure state are extracted from the finite element simulation results of the RC structure under different damage conditions. Second, irregular elements are mapped into regular three-dimensional voxels based on a unified spatial envelope box to obtain the strain voxel field and occupancy mask value of the entire structure. Third, the voxel field of the entire structure is slidably cut into blocks according to a preset local block size and inter-block step size, and the working conditions and the spatial position encoding of the local blocks are combined to form the local block input features. Then, the strain voxel field of each local block is predicted and output using a local block strain prediction network. Finally, the predicted local block strain voxel field is fused back into the three-dimensional voxel field of the entire structure based on the spatial index of the local block in the entire structure, thereby restoring the damage morphology of the entire structure.

[0030] Input data definition: For any RC structure explosion condition sample, the input parameters must include at least the explosion equivalent. Explosion distance Offset of the explosion point relative to the centerline of the structure The operating condition parameters and at least one dimensionless combination of derived parameters, such as The output labels should include at least the element-level strain field of each concrete unit. The RC structure can correspond to beams, slabs, columns, joints, frame components, or an integral structure composed of multiple components; in this invention, all are uniformly abstracted as RC structures.

[0031]

[0032] like Figure 1 As shown in the figure, this embodiment presents a three-dimensional prediction method for RC structure explosion damage based on voxelization modeling. The specific process is as follows: Step 1: Element-level data extraction: Obtain the finite element simulation results of the RC structure under different damage conditions, and extract the element-level strain field of each concrete element. The specific process is as follows: Step S1: Perform finite element simulations of the RC structure under different damage conditions, and read the finite element simulation result file of the RC structure corresponding to each damage condition.

[0033] Step S2: Extract the identifier, element center coordinates, time history strain data, and element survival status data of each concrete element from the finite element simulation result file. The element center coordinates are obtained by averaging the coordinates of its eight associated nodes. Let the first node be... The coordinates of the 8 nodes associated with each unit are denoted as follows: Then the center coordinates of the finite element are:

[0034] Step S3: Extract the maximum equivalent plastic strain value along the time dimension from the time history strain data of each concrete unit to form a unit-level strain field as a unit-level label.

[0035] Step S4: Store the working conditions of the damage condition and the unit-level tags in a unified manner to provide a basis for subsequent spatial mapping.

[0036] Because a unified finite element mesh or a unified geometric representation is used under different damage conditions, a stable element sequence index can be established, making the same spatial location under different conditions comparable in the subsequent voxelization process.

[0037] Step 2: Voxelization Representation of the Whole Structure: Set the number of regular voxel meshes for the whole structure, and uniformly divide the envelope of the RC structure in 3D space into small cubic elements corresponding to the number of voxel meshes. Perform voxel index mapping on the finite element elements of the RC structure, mapping all element-level strain fields to the whole structure strain voxel field according to the voxel index mapping, and determine the mask value for each voxel. Specifically: Step S5: Calculate the envelope box of the RC structure in three-dimensional space, and denote its minimum coordinates as... The maximum coordinate is Furthermore, a pre-defined margin is added outside the envelope box to prevent boundary units from being too close to voxel boundaries, which could lead to mapping distortion.

[0038] Step S6: Set the number of voxel meshes for the whole structure to [value]. For example, 96×48×48, the envelope box is evenly divided into corresponding numbers along three directions. The small cube unit is used to obtain the physical size of each voxel, voxel_size. The size of the small cube unit here is the physical size of each voxel. The specific calculation is as follows: , , Then the physical size of each voxel is: .

[0039] Step S7: Let the center coordinates of each finite element of the RC structure be... Calculate the voxel subscript where it falls. Complete the voxel index mapping of the RC structure finite element elements;

[0040] in, This indicates the floor function.

[0041] Step S8: Since the size of a finite element element is related to the finite element meshing strategy, and the voxel mesh is a regular mesh established to construct the input of a deep learning model, its resolution is usually preset. Therefore, the physical size of a voxel may be larger than the size of a finite element element. When the finite element mesh is finer and the voxel mesh is relatively coarser, the center points of multiple finite element elements may fall into the same voxel simultaneously.

[0042] If multiple finite element elements fall into the same voxel, the strain modulus field of the corresponding voxel is the average of the element-level strain fields of the contained elements; finally, a whole-structure strain modulus field and an occupancy value are generated; the occupancy value is used to indicate whether the voxel contains real structural material, and is used to mask the spatial domain during training. If a finite element element is mapped into the voxel, the voxel is considered to contain real structural material, and occupancy=1; otherwise, occupancy=0.

[0043] The essence of this step is to uniformly resample the irregular finite element mesh into a regular three-dimensional voxel space, so that the cell-level data that is originally difficult to use directly for deep learning is transformed into a regular tensor that can be processed by a three-dimensional convolutional network.

[0044] Step 3: Local Block Segmentation and Filtering: The entire structure strain element field is segmented into multiple local blocks according to a preset local block size and inter-block step size. These local blocks are then filtered based on the occupancy mask value. The specific process is as follows: Step S9: Set the local block size to Px×Py×Pz, for example, 32×32×32, and set the inter-block step size to Sx×Sy×Sz, for example, 16×16×16, with overlap between local blocks.

[0045] Step S10: Traverse the entire structure strain element field using a sliding window method to generate multiple local blocks. Each local block is uniquely determined by its start and end coordinates (x0, y0, z0, x1, y1, z1).

[0046] Step S11: If the sum of the occupancy mask values ​​of the voxels contained in a local block is 0, it means that the local block is completely located in the spatial domain and is directly discarded; otherwise, the local block is retained, and the strain voxel field, occupancy mask value and local block spatial location code of the local block are extracted.

[0047] Through this design, the present invention expands a single whole structure sample into a large number of local block samples, which significantly increases the number of training samples in the case of scarce samples and enhances the model's ability to learn local high gradient strain regions.

[0048] Step 4: Local Block Strain Prediction Network: Construct a local block strain prediction network. The network input is the local block input feature, which is formed by concatenating the location codes of the selected local blocks with the working conditions. The network output is the predicted local block strain element field. The network is trained using a loss function with occupancy mask, with the actual local block strain element field as the supervision signal. The specific process is as follows: Step S12: Concatenate the working conditions (including the working parameters and derived parameters in the table) with the normalized local block spatial location codes that have been filtered and retained to form local block input features; wherein, the spatial location codes include at least the normalized coordinates of the start and end points of the local block on the whole structure size, so that the model can distinguish the response patterns of different spatial locations under the same working condition.

[0049] Step S13: Construct a local block strain prediction network. The network input consists of local block input features composed of working conditions and local block location encoding, and the output is the predicted local block strain element field. The network can adopt a structure combining a multilayer perceptron and a 3D decoder, or other neural network structures that can achieve "low-dimensional condition input - 3D field output".

[0050] Step S14: Using the real local block strain element field as the supervision signal, train the network using a loss function with occupancy mask value.

[0051] Preferably, the loss is calculated only on voxels with an occupancy of 1 to avoid the model being dominated by a large number of spatial voxels; more preferably, higher loss weights are assigned to high strain regions to mitigate the problem of over-smoothing of the prediction results.

[0052] Let the first local block be... The occupancy mask value of an individual element is The predicted strain field is The real strain field is The training loss function is then defined as: , Among them, the weight term is , , Indicates the first Individual elements belong to the RC structural entity region. Indicates the first Individual elements belong to the spatial region of the RC structure. This is the weighting adjustment coefficient for the high-strain region. To prevent extremely small constants with a denominator of zero, the loss function is trained to calculate prediction errors only in structural entities, and higher penalty weights are assigned to high-strain regions. This avoids a large number of spatial voxels dominating the training process, while improving the network's ability to predict regions with concentrated local strain.

[0053] Step 5: Overall Structure Reconstruction and Damage Morphology Restoration: The working conditions of the damage condition to be predicted and the spatial location codes of local blocks are concatenated to form the local block input features. The trained network is used to predict the strain element field of the local blocks. Based on the spatial index of the local blocks in the overall structure, the predicted local block strain element fields are fused to obtain the overall structure strain element field. The overall structure strain element field is further mapped back to the finite element space, and the damage morphology of the overall structure is restored. The specific process is as follows: Step S15: After training, the working conditions of the damage condition to be predicted and the spatial location codes of the local block are concatenated to form the local block input features, and the trained network is used to predict the strain element field of the local block.

[0054] Step S16: Based on the spatial index of each local block in the whole structure, write the predicted local block strain element field back to the corresponding position in the whole structure; for overlapping regions, use weighted averaging or simple averaging to fuse them to generate the strain element field of the whole structure.

[0055] Step S17: Further map the entire structure strain voxel field back to the original finite element space. Specifically, according to the pre-established voxel index mapping, assign the corresponding voxel value to the finite element falling into that voxel, thereby recovering the element-level predicted strain value.

[0056] Step S18: Construct the three-dimensional damage morphology display result of the whole structure based on the element-level predicted strain value. It can be displayed in regular voxel space, or in the form of an element cube or equivalent geometry under the original element center coordinates.

[0057] This application provides a three-dimensional prediction system for RC structure explosion damage based on voxel modeling, comprising: The data extraction module is used to obtain the finite element simulation results of reinforced concrete (RC) structures under different damage conditions, and extract the element-level strain field of each concrete element from them. The voxelization module is used to set the number of regular voxel meshes for the whole structure, uniformly divide the envelope of the RC structure in three-dimensional space into small cubic elements with the corresponding number of voxel meshes; perform voxel index mapping on the finite element elements of the RC structure, map all element-level strain fields to the whole structure strain voxel field according to the voxel index mapping, and determine the occupancy mask value of each voxel. The local block generation module is used to slide and cut the entire structure strain element field into multiple local blocks according to the preset local block size and inter-block step size, and to filter the local blocks according to the occupancy mask value. The local block strain prediction module is a local block strain prediction network, which is used to output the predicted local block strain element field based on the local block input features formed by splicing the spatial location codes of the selected and retained local blocks with the working conditions. The whole structure reconstruction and topography display module is used to fuse the predicted local block strain element fields to obtain the whole structure strain element field based on the spatial index of the local blocks in the whole structure, and further map the whole structure strain element field back to the finite element space to realize the prediction of the three-dimensional strain field.

[0058] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-dimensional prediction method for explosive damage to RC structures based on voxel modeling, characterized in that, The specific process is as follows: Step 1: Obtain the finite element simulation results of the reinforced concrete (RC) structure under different damage conditions, and extract the element-level strain field of each concrete element from the simulation results. Step 2: Set the number of regular voxel meshes for the whole structure, and uniformly divide the envelope box of the RC structure in three-dimensional space into small cubic units with the corresponding number of voxel meshes. Voxel index mapping is performed on the finite element elements of the RC structure. Based on the voxel index mapping, all element-level strain fields are mapped to the whole structure strain voxel fields, and the occupancy mask value of each voxel is determined. Step 3: Perform sliding cutting on the strain element field of the whole structure according to the preset local block size and inter-block step size to obtain multiple local blocks; filter the local blocks according to the occupancy mask value; Step 4: Construct a local block strain prediction network. The network input is the local block input feature composed of the spatial location code of the selected local block and the working condition. The network output is the predicted local block strain element field. The network is trained using the real local block strain element field as the supervision signal and a loss function with occupancy mask value. Step 5: Use the trained network to predict the local block strain element field. Based on the spatial index of the local block in the whole structure, fuse the local block strain element fields to obtain the whole structure strain element field and further map it back to the finite element space to realize the prediction of the three-dimensional strain field.

2. The method for three-dimensional prediction of RC structure explosion damage based on voxelization modeling according to claim 1, characterized in that, The specific process of performing voxel index mapping on the finite element elements of the RC structure is as follows: Calculate the envelope box of the RC structure in three-dimensional space, and add a boundary blank space of a preset proportion outside the envelope box; Given a set number of regular voxel grids for the whole structure, the envelope box is uniformly divided into small cubic units along three directions, corresponding to the number of regular voxel grids for the whole structure. The physical dimensions of each voxel are as follows: ; Let the center coordinates of each finite element of the RC structure be... Calculate the voxel subscript where it falls. Complete the voxel index mapping of the RC structure finite element elements; in, This indicates a round-down operation. These are the minimum coordinates of the envelope box.

3. The method for three-dimensional prediction of RC structure explosion damage based on voxelization modeling according to claim 2, characterized in that, In step two, during the voxel indexing process of the RC structure finite element elements, when multiple finite element elements fall into the same voxel, the strain voxel field of the corresponding voxel is taken as the average value of the element-level strain field of the contained elements, and its occupancy mask value is determined according to whether the voxel contains real structural material.

4. The three-dimensional prediction method for RC structure explosion damage based on voxelization modeling according to claim 3, characterized in that, If a voxel contains real structural material, the occupancy mask value is 1; otherwise, the occupancy mask value is 0. In step three, the local block filtering is performed based on the occupancy mask value: local blocks containing voxel occupancy mask values ​​whose sum is not 0 are filtered out.

5. The three-dimensional prediction method for RC structure explosion damage based on voxelization modeling according to claim 4, characterized in that, let... The first in the local block The occupancy mask value of the individual element is The strain field predicted by the network is The real strain field is The training loss function is then defined as: , Among them, the weight term is , , Indicates the first Individual elements belong to the RC structural entity region. Indicates the first Individual elements belong to the spatial region of the RC structure. This is the weighting adjustment coefficient for the high-strain region. To prevent extremely small constants with a denominator of zero, This represents the L2 norm.

6. The three-dimensional prediction method for RC structure explosion damage based on voxelization modeling according to claim 5, characterized in that, The operating conditions include operating parameters and derived parameters. The derived parameters are at least one dimensionless combination of the operating parameters. The spatial location coding includes at least the normalized coordinates of the start and end points of the local block on the overall structural dimensions.

7. The method for three-dimensional prediction of RC structure explosion damage based on voxelization modeling according to claim 6, characterized in that, Operating parameters include: explosive yield Explosion distance And the offset of the explosion point relative to the centerline of the structure Derived parameters include: , as well as .

8. The three-dimensional prediction method for RC structure explosion damage based on voxelization modeling according to claim 5, characterized in that, Set the local block size and the inter-block step size so that there is overlap between local blocks.

9. The method for three-dimensional prediction of RC structure explosion damage based on voxelization modeling according to claim 5, characterized in that, For the center coordinates of each finite element of the RC structure The calculation process is as follows: Let the first The coordinates of the 8 nodes associated with each finite element are denoted as follows: Then the center coordinates of the finite element are: 。 10. A three-dimensional prediction system for explosive damage to RC structures based on voxel modeling, characterized in that, include: The data extraction module is used to obtain the finite element simulation results of reinforced concrete (RC) structures under different damage conditions, and extract the element-level strain field of each concrete element from them. The voxelization module is used to set the number of regular voxel meshes for the whole structure, and to uniformly divide the envelope of the RC structure in three-dimensional space into small cubic units with the corresponding number of voxel meshes. Voxel index mapping is performed on the finite element elements of the RC structure. Based on the voxel index mapping, all element-level strain fields are mapped to the whole structure strain voxel fields, and the occupancy mask value of each voxel is determined. The local block generation module is used to slide and cut the entire structure strain element field into multiple layout blocks according to the preset local block size and inter-block step size, and to filter the local blocks according to the occupancy mask value. The local block strain prediction module is a local block strain prediction network, which is used to output the predicted local block strain element field based on the local block input features formed by splicing the spatial location codes of the selected and retained local blocks with the working conditions. The whole structure reconstruction and topography display module is used to fuse the predicted local block strain element fields to obtain the whole structure strain element field based on the spatial index of the local blocks in the whole structure, and further map the whole structure strain element field back to the finite element space to realize the prediction of the three-dimensional strain field.