A condition-driven RC structure explosion damage intelligent evaluation method and device

Abstract

The present application belongs to the technical field of explosion damage engineering evaluation, and particularly relates to a condition-driven RC structure explosion damage intelligent evaluation method and device. The method expands samples through data enhancement, establishes a phased evaluation framework from working conditions to a three-dimensional voxel strain field, from the three-dimensional voxel strain field to a damage probability map, and from the damage probability map to adaptive binary determination, so as to realize more robust and more interpretable RC structure damage evaluation results.

Description

Technical Field

[0001] This invention belongs to the field of explosion damage engineering assessment technology, specifically relating to a condition-driven intelligent assessment method and device for explosion damage of RC structures. Background Technology

[0002] In existing studies on the damage caused by explosions to reinforced concrete (RC) structures, experimental samples are typically very limited. Live-fire tests are not only costly, but also difficult to obtain large-sample data covering different yields, blast distances, and eccentricities due to limitations in safety conditions, site conditions, and measurement methods. In contrast, finite element simulation can generate information such as displacement, strain, and element failure states in batches within a larger parameter space, thus serving as an important data source for constructing intelligent evaluation models.

[0003] However, existing technologies have many technical shortcomings when using simulation data for intelligent damage assessment: First, although the simulation covers a wide range of working conditions, the damage modes vary greatly under different working conditions, and the sample size is still insufficient compared to deep learning methods. Secondly, the geometric and physical symmetry of real structures are often not fully utilized, leading to an underestimation of the available sample size. Third, existing methods often directly use a fixed threshold to binarize the damage probability map. The threshold selection depends on human experience, which is sensitive and unstable. Fourth, many methods confuse strain prediction with damage assessment, failing to fully utilize the physical fact that strain is a continuous physical quantity while damage is a discrete result. Here, damage refers to whether the RC element has failed.

[0004] Furthermore, in existing explosion damage prediction methods, a common approach is to directly train an end-to-end model from the working condition to the damage label. While this method is simple, the model training is often unstable due to the highly sparse and discontinuous nature of the damage labels. The output probability map is prone to problems such as blurred classification boundaries, extreme sensitivity to thresholds, and distortion of local morphology. Especially under small sample conditions, the failure area may change drastically under different thresholds, seriously affecting the interpretability of the results.

[0005] In summary, the main drawbacks of existing technologies include: First, they do not fully utilize the geometric symmetry of RC structures for sample expansion, resulting in high simulation costs and prominent small sample size issues; Second, they directly predict damage labels from working conditions, ignoring the leading role of the strain field in damage evolution, leading to poor model stability and interpretability; Third, the binary determination of the damage probability map relies on a fixed threshold, and even a small change in the threshold can cause a significant change in the failure area, making it difficult to apply stably in engineering; Fourth, there is a lack of a complete technical solution that takes into account the utilization of simulation data, probability output calibration, adaptive threshold selection, and three-dimensional damage morphology restoration. Summary of the Invention

[0006] In view of this, the present invention provides a condition-driven intelligent assessment method and device for explosive damage of RC structures. By expanding the sample through data augmentation, a phased assessment framework is established, which ranges from working conditions to a three-dimensional voxel strain field, from the three-dimensional voxel strain field to a damage probability map, and from the damage probability map to an adaptive binary determination, thereby achieving more robust and interpretable damage assessment results for RC structures.

[0007] The technical solution for implementing the present invention is as follows: Firstly, the present invention provides a condition-driven intelligent assessment method for explosion damage to RC structures, the specific process of which is as follows: Step 1: Obtain the finite element simulation results of the reinforced concrete (RC) structure under different original damage conditions, and extract the three-dimensional strain field and three-dimensional failure field of each concrete element as the original sample; construct the mirror damage condition of the original damage condition, and rearrange the three-dimensional strain field and three-dimensional failure field of each concrete element according to the mirror index to obtain the mirror enhanced sample. Step 2: Construct a strain field prediction model based on a conditional generative neural network. The strain field prediction model is used to nonlinearly map the input damage condition parameters to extract high-dimensional feature representations, and then decode the high-dimensional feature representations into a three-dimensional strain element field output. Step 3: Construct a damage prediction model by encoding the three-dimensional strain element field, occupancy mask value, and damage condition parameters, and then expanding the encoding results into a condition feature body consistent with the three-dimensional space and fusing them as the input of the damage prediction model. The damage prediction model output is a damage probability map. Step 4: Set the adaptive damage threshold A three-dimensional voxel failure field is obtained by performing a binary transformation on the damage probability map. Step 5: Voxel mapping is performed on the three-dimensional strain field and the three-dimensional failure field of the original sample and the enhanced sample. The mapping results are used as training labels for the strain field prediction model and the damage prediction model, respectively. The trained strain field prediction model and the damage prediction model are then used to realize intelligent assessment of explosion damage.

[0008] Optionally, the present invention constructs a mirror image of the original damage condition, and rearranges the three-dimensional strain field and three-dimensional failure field of each concrete unit according to the mirror index to obtain a mirror-enhanced sample; the specific process is as follows: First, regarding the offset of the explosion point relative to the structural centerline: The original damage condition, constructing the mirror offset All other operating parameters remain unchanged; Secondly, establish a unit-level mirror mapping relationship: for the center coordinates of each finite element element, construct the mirror coordinates of the center coordinates about the plane of symmetry, and find the corresponding element in the reference mesh that is closest to the mirror coordinates and whose distance does not exceed a preset threshold to form a mirror index mapping; Finally, based on the mirror index mapping, the unit-level three-dimensional strain field and three-dimensional failure field in the original sample are rearranged to obtain the mirror-enhanced sample.

[0009] Optionally, the specific process of step two of the present invention is as follows: First, the number of regular voxel meshes for the whole structure is set, and the envelope of the RC structure in three-dimensional space is uniformly divided into small cubic elements with the corresponding number of voxel meshes. Voxel index mapping is performed on the finite element elements of the RC structure. According to the voxel index mapping, the three-dimensional strain field of each concrete element is mapped to the strain voxel field of the whole structure, and the mask value occupied by each voxel is determined. Secondly, the entire structure strain element field is slid-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. Finally, a strain field prediction model is constructed based on a conditional generative neural network. The input of the strain field prediction model is the local block features formed by splicing the spatial location codes of the selected local blocks with the damage condition parameters. The strain field prediction model is used to perform nonlinear mapping on the input local block features to obtain the latent condition vector, and then decode the latent condition vector to output a three-dimensional strain element field.

[0010] Optionally, the present invention provides that the first in the local block The structure occupancy mask value of individual elements is The strain field prediction model predicts the strain element field as follows: 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. To prevent extremely small constants with a denominator of zero, This represents the L2 norm.

[0011] Optionally, the damage prediction model of the present invention employs a combined loss function during training. ;

[0012] in, For the focus of loss, For Dice's loss, Weights to balance the losses;

[0013]

[0014] in, 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. Indicates the first The probability that an individual element is predicted as damage. Indicates the first The true damage label of individual elements, Indicates the first Individuals are damaged individuals. Indicates the first The voxels are undamaged voxels; Indicates the first The probability that an individual element is predicted as the true class, when hour, ,when hour, ; Represents the category balance factor. Indicates the focus parameter, This is a smoothing term.

[0015] Optionally, the damage parameters of the present invention 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 .

[0016] Optionally, the present invention provides an adaptive damage threshold. Set it up through the following process: Let the damage prediction model output validation set be the first... The damage probability map for each sample is as follows ,in Indicates the first In the nth sample The probability that an individual element is predicted as damage; a multi-candidate threshold scan is performed on the damage probability map, and the candidate threshold set is... ; For each candidate threshold , Binarize the damage probability map into damage labels:

[0017] Damage labels obtained from binarization and true damage label Calculate the Dice coefficient or Intersection over Union (IoU), and select the candidate threshold that maximizes the Dice coefficient or IoU as the adaptive damage threshold. .

[0018] Optionally, the present invention uses a candidate threshold. The Dice coefficient is calculated as follows:

[0019] in, To determine the number of samples in the validation set, To prevent extremely small constants with a denominator of zero.

[0020] Optionally, the present invention uses a candidate threshold. The intersection-union ratio (IoU) is calculated as follows:

[0021] in, To determine the number of samples in the validation set, To prevent extremely small constants with a denominator of zero.

[0022] Secondly, the present invention provides a condition-driven intelligent assessment device for RC structure explosion damage, comprising: The strain field prediction model is used to perform nonlinear mapping on the input damage condition parameters to obtain the potential condition vector, and then decode the potential condition vector to output a three-dimensional strain element field. The damage prediction model is used to predict the damage probability map based on the three-dimensional strain element field, occupancy mask value, and the encoded, expanded, and fused features of the damage condition parameters. The binary conversion module is used to convert values ​​based on an adaptive damage threshold. A three-dimensional voxel failure field is obtained by performing a binary transformation on the damage probability map. The damage intelligent assessment module is used to obtain the three-dimensional strain field and the three-dimensional failure field based on the three-dimensional strain voxel field and the three-dimensional voxel failure field.

[0023] Beneficial effects: First, this invention integrates a phased damage assessment architecture that combines "conditional generation network" and "strain-driven" approaches. It adopts a phased assessment framework that includes three-dimensional voxel strain field prediction, damage probability map inference, and threshold adaptive determination, rather than a direct end-to-end determination from working condition to damage label.

[0024] Second, adaptive threshold determination: The optimal damage determination threshold is automatically searched on the validation set, and the threshold is determined based on the maximum Dice coefficient or IoU, thereby improving the stability of damage assessment and engineering interpretability.

[0025] Third, intelligent damage assessment: It outputs the continuous damage probability map and the binary damage morphology simultaneously, and can restore it to the original finite element space, thereby realizing rapid damage assessment. Attached Figure Description

[0026] 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.

[0027] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0028] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] 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.

[0030] 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.

[0031] The overall approach of this invention is as follows: Instead of directly mapping explosion conditions to damage labels, this invention first predicts the three-dimensional strain coefficient field of the RC structure based on the damage conditions. Then, it uses the three-dimensional strain coefficient field as the core input for damage assessment, outputting a continuous damage probability map. The optimal binary damage distribution is obtained by setting an adaptive damage threshold and performing a binary transformation. Simultaneously, the geometric features of the RC structure being symmetrical about the central plane are utilized to mirror and enhance existing damage conditions, thereby expanding the sample size and improving the model's generalization ability under eccentric explosion conditions.

[0032] For any RC structure damage case sample, the input parameters should include at least the explosive equivalent. Explosion distance And the offset of the explosion point relative to the centerline of the structure And at least one dimensionless combination parameter, for example: The output labels should include at least: the maximum equivalent plastic strain of each concrete element and the failure state of each element. "RC structure" can refer to beams, slabs, columns, joints, frame members, or an integral structure composed of multiple components; in this invention, all are uniformly abstracted as RC structures.

[0033]

[0034] like Figure 1 As shown in the figure, an embodiment of this application provides a condition-driven intelligent assessment method for RC structure explosion damage, which specifically includes the following steps: Step 1: Construction of symmetrical reinforcement data: Obtain finite element simulation results of the RC structure under different original damage conditions, and extract the three-dimensional strain field (i.e., maximum equivalent plastic strain) and three-dimensional failure field of each concrete element as original samples; construct a mirror damage condition of the original damage condition, and rearrange the three-dimensional strain field and three-dimensional failure field of each concrete element according to the mirror index to obtain mirror reinforcement samples. The specific process is as follows: Step S0: Obtain the finite element simulation results of the RC structure under different original damage conditions, wherein the original damage conditions include: explosive equivalent. Explosion distance And the offset of the explosion point relative to the centerline of the structure The maximum equivalent plastic strain field (i.e., three-dimensional strain field) and the failure state of each element are extracted from the simulation results. The element failure states with non-zero values ​​are assigned a value of 0 to indicate that the element is not damaged, and the element failure states with zero values ​​are assigned a value of 1 to indicate that the element is damaged, thereby constructing a three-dimensional failure field.

[0035] Step S1: Determine the plane of symmetry of the RC structure, for example, take the center plane of the structure x=0 as the plane of symmetry.

[0036] Step S2: For the offset in the original damage condition For the sample, construct the mirror offset Explosion equivalent in operating parameters Explosion distance The values ​​remain unchanged, only the offset sign is reversed.

[0037] Step S3: Establish element-level mirror mapping relationship: For the center coordinates (x,y,z) of each finite element element, construct the mirror coordinates (-x,y,z) of the center coordinates (x,y,z) about the plane of symmetry. Then, find the corresponding element in the reference mesh that is closest to the mirror coordinates (-x,y,z) and the distance does not exceed the preset threshold to form a mirror index mapping.

[0038] Step S4: Based on the mirror index mapping, rearrange the three-dimensional strain field and three-dimensional failure field at the unit level in the original sample to obtain the mirror-enhanced sample.

[0039] Step S5: Construct a training dataset: Use the damage condition parameters of the original sample and the enhanced sample as input data, the three-dimensional strain field as the label data of the strain prediction model, and the three-dimensional failure field as the label data of the damage prediction model.

[0040] Step 2: Constructing a strain field prediction model to achieve three-dimensional voxel strain field prediction: A strain field prediction model is constructed based on a conditional generative neural network. This model is used to perform nonlinear mapping on the input damage condition parameters to obtain a latent condition vector, which is then decoded and output as a three-dimensional strain voxel field. Specific process: Step S6: Construct a strain field prediction model. This model employs a conditional generative neural network structure to realize the mapping relationship between damage condition parameters and the three-dimensional strain element field of the RC structure. This conditional generative neural network includes: The conditional input module is used to input the explosive equivalent. Explosion distance And the offset of the explosion point relative to the centerline of the structure Operating condition parameters and derived parameters, including: , as well as .

[0041] The feature mapping module uses a fully connected neural network to perform nonlinear mapping on low-dimensional damage condition parameters and extract high-dimensional feature representations. The spatial decoding module uses a three-dimensional convolutional neural network or a three-dimensional deconvolutional network to decode high-dimensional features into a three-dimensional voxel strain field. The output module outputs the three-dimensional strain element field of the RC structure under the corresponding damage condition.

[0042] This model essentially constructs a generative mapping relationship of "condition variable → high-dimensional physical field".

[0043] Step S7: When the number of samples is insufficient, a local block modeling approach is also adopted. The overall structure is divided into multiple three-dimensional local blocks, and then the three-dimensional strain element field of the overall structure is reconstructed by overlapping and stitching. Specifically: First, the number of regular voxel meshes for the whole structure is set, and the envelope of the RC structure in three-dimensional space is uniformly divided into small cubic elements with the corresponding number of voxel meshes. Voxel index mapping is performed on the finite element elements of the RC structure. According to the voxel index mapping, the three-dimensional strain field of each concrete element is mapped to the strain voxel field of the whole structure, and the mask value occupied by each voxel is determined. Secondly, the entire structure strain element field is slid-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. Finally, a strain field prediction model is constructed based on a conditional generative neural network. The input of the strain field prediction model is the local block features formed by splicing the spatial location codes of the selected local blocks with the damage condition parameters. The strain field prediction model is used to perform nonlinear mapping on the input local block features to obtain the latent condition vector, and then decode the latent condition vector to output a three-dimensional strain element field.

[0044] In this embodiment, the training loss function for the strain field prediction model is designed as follows: Let the first local block be... The structure occupancy mask value of individual elements is The strain field prediction model predicts the strain element field as follows: 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. To prevent extremely small constants with a denominator of zero, This represents the L2 norm.

[0045] Step 3: Construct a damage prediction model to achieve damage inference: Construct a damage prediction model by encoding the three-dimensional strain element field, occupancy mask value, and damage condition parameters. Expand the encoded results into a conditional feature volume consistent with the three-dimensional space and fuse them as input to the damage prediction model. The damage prediction model outputs a damage probability map. The specific process of this step is as follows: Step S8: Construct a damage prediction model to predict the three-dimensional damage distribution based on the input three-dimensional strain element field, occupancy mask value, and damage condition parameters.

[0046] The model first encodes the input features, expanding the three-dimensional strain element field, occupancy mask value, and damage condition parameters into a condition feature body consistent with the three-dimensional space, and then fuses them as input to the damage prediction model. The damage prediction model predicts and outputs a damage probability map.

[0047] Damage prediction models, by integrating global operating condition information and local strain response information, can effectively characterize the evolution of damage in three-dimensional space, thereby improving the accuracy and stability of damage prediction.

[0048] Step S9: When training the damage prediction model, a combined loss function is used. Including focal loss (FocalLoss) and Dice loss function (DiceLoss).

[0049] FocalLoss reduces the loss weight of easily classified samples through a balancing factor, making the model training focus more on difficult-to-classify or minority class samples. Here, easily classified samples refer to voxels in the non-failed region, and difficult-to-classify samples refer to voxels in the failed region. DiceLoss is used to optimize the overall overlap of damaged regions, thereby solving the problems of sparse damaged regions and class imbalance.

[0050]

[0051]

[0052]

[0053] in, 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. Indicates the first The probability that an individual element is predicted as damage. Indicates the first The true damage label of individual elements, Indicates the first Individuals are damaged individuals. Indicates the first The voxels are undamaged voxels; Indicates the first The probability that an individual element is predicted as the true class, when hour, ,when hour, In other words, when The larger the value, the more accurate the model prediction. The smaller the value, the more difficult it is to classify the sample; Represents the category balance factor. Indicates the focus parameter, It is a very small positive number and is a smooth term.

[0054] Step 4: Adaptive Damage Threshold Setting and Binary Judgment: Set the adaptive damage threshold. Binarize the damage probability map to obtain a three-dimensional voxel failure field; adaptive damage threshold. The acquisition process is as follows: Step S10: On the validation set, perform a multi-candidate threshold scan on the damage probability map. Let the candidate threshold set be... The threshold range is preferably set to cover the minimum to maximum value of the damage probability map, or it can be set to uniform sampling between 0 and 1 based on experience.

[0055] Let the damage prediction model output validation set be the first... The damage probability map of each validation set sample is as follows ,in Indicates the first The first sample The probability that an individual element is predicted as damage. Let the candidate threshold set be... Candidate thresholds can be uniformly sampled between the minimum and maximum values ​​of the damage probability map.

[0056] ,

[0057] in, The total number of candidate thresholds. This is the index of the current candidate threshold.

[0058] Alternatively, the threshold set can be set based on experience as follows:

[0059] That is, uniform sampling within the interval [0,1].

[0060] Step S11: For each candidate threshold The damage probability map is binarized into damage labels, and at least one evaluation index is calculated with the true labels, preferably the Dice coefficient and the cross-union ratio (IoU).

[0061]

[0062] in, Indicates the first In the nth sample The individual was identified as a damaged individual. In the m-th sample, the first... The voxels were determined to be undamaged.

[0063] Let the true damage label be Then at the candidate threshold The Dice coefficient can be calculated below:

[0064] in, To determine the number of samples in the validation set, To prevent extremely small constants with a denominator of zero.

[0065] The intersection-union ratio (IoU) is calculated as follows:

[0066] Step S12: Select the threshold that maximizes either the Dice coefficient or the IoU (these two values ​​can be converted to each other; the larger the Dice coefficient, the larger the IoU) as the adaptive damage threshold for the current model. .

[0067]

[0068] This step overcomes the problem of subjectively setting fixed thresholds and can automatically select a more reasonable engineering judgment threshold based on the actual output distribution of the model.

[0069] The above system realizes the physical modeling chain: operating conditions → three-dimensional voxel strain field (continuous physical quantity) → damage probability map (continuous field) → damage morphology (binary field). This avoids the instability problem caused by directly predicting discrete damage labels from operating conditions, and improves the physical consistency and interpretability of the model.

[0070] Step 5: Model Training and Damage Assessment: Voxel mapping is performed on the three-dimensional strain field and three-dimensional failure field of the original and enhanced samples. The mapping results are used as training labels for the strain field prediction model and the damage prediction model, respectively. The trained models are then used to achieve intelligent damage assessment.

[0071] Step S13: During the test set or engineering inference phase, an adaptive damage threshold is adopted. The damage probability map output by the damage prediction model is binarized to output the final damaged area. For the voxel-level damage probability map and the binary damage field, the results are restored to the original finite element space based on the voxel-to-finite element index mapping.

[0072] Step S14: Output at least one of the following engineering results: three-dimensional binary damage morphology, three-dimensional continuous damage probability map, damage volume fraction, number of failure elements, center location of the failure region, and peak strain distribution. Predicting damage morphology can be used to assist in model evaluation and engineering decision-making.

[0073] Compared with the prior art, the present invention has the following technical advantages: First, this invention makes full use of the symmetry of the RC structure for mirror enhancement, which can expand the effective sample size without increasing simulation costs and alleviate the small sample size problem.

[0074] Secondly, compared with the method of directly predicting damage labels from working conditions, the present invention first predicts the strain field and then infers the damage probability, which is more in line with the law of material damage evolution and the model has stronger interpretability.

[0075] Third, by introducing the BCE and Dice combined loss into the damage training, this invention can simultaneously take into account point-by-point classification and overall region overlap, thereby improving the prediction bias of the damage region.

[0076] Fourth, by using adaptive search of loss threshold, this invention avoids the extreme sensitivity problem caused by manually fixing the threshold, and can automatically determine a more reasonable engineering judgment threshold based on the model output distribution.

[0077] 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 condition-driven intelligent assessment method for explosion damage to RC structures, 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 original damage conditions, and extract the three-dimensional strain field and three-dimensional failure field of each concrete element as the original sample. A mirrored damage condition is constructed from the original damage condition. The three-dimensional strain field and three-dimensional failure field of each concrete unit are rearranged according to the mirror index to obtain a mirrored enhanced sample. Step 2: Construct a strain field prediction model based on a conditional generative neural network. The strain field prediction model is used to nonlinearly map the input damage condition parameters to extract high-dimensional feature representations, and then decode the high-dimensional feature representations into a three-dimensional strain element field output. Step 3: Construct a damage prediction model by encoding the three-dimensional strain element field, occupancy mask value, and damage condition parameters, and then expanding the encoding results into a condition feature body consistent with the three-dimensional space and fusing them as the input of the damage prediction model. The damage prediction model output is a damage probability map. Step 4: Set the adaptive damage threshold A three-dimensional voxel failure field is obtained by performing a binary transformation on the damage probability map. Step 5: Voxel mapping is performed on the three-dimensional strain field and the three-dimensional failure field of the original sample and the enhanced sample. The mapping results are used as training labels for the strain field prediction model and the damage prediction model, respectively. The trained models are then used to achieve intelligent assessment of explosion damage.

2. The condition-driven intelligent assessment method for RC structure explosion damage according to claim 1, characterized in that, A mirror-image damage condition is constructed based on the original damage condition. The three-dimensional strain field and three-dimensional failure field of each concrete element are rearranged according to the mirror index to obtain a mirror-enhanced sample. The specific process is as follows: First, regarding the offset of the explosion point relative to the structural centerline: The original damage condition, constructing the mirror offset All other operating parameters remain unchanged; Secondly, establish a unit-level mirror mapping relationship: for the center coordinates of each finite element element, construct the mirror coordinates of the center coordinates about the plane of symmetry, and find the corresponding element in the reference mesh that is closest to the mirror coordinates and whose distance does not exceed a preset threshold to form a mirror index mapping; Finally, based on the mirror index mapping, the unit-level three-dimensional strain field and three-dimensional failure field in the original sample are rearranged to obtain the mirror-enhanced sample.

3. The condition-driven intelligent assessment method for RC structure explosion damage according to claim 2, characterized in that, The specific process of step two is as follows: First, the number of regular voxel meshes for the whole structure is set, and the envelope of the RC structure in three-dimensional space is uniformly divided 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, the three-dimensional strain field of each concrete element is mapped to the strain voxel field of the whole structure, and the occupancy mask value of each voxel is determined. Secondly, the entire structure strain element field is slid-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. Finally, a strain field prediction model is constructed based on a conditional generative neural network. The input of the strain field prediction model is the local block features formed by splicing the spatial location codes of the selected local blocks with the damage condition parameters. The strain field prediction model is used to perform nonlinear mapping on the input local block features to obtain the latent condition vector, and then decode the latent condition vector to output a three-dimensional strain element field.

4. The condition-driven intelligent assessment method for RC structure explosion damage according to claim 3, characterized in that, ... The first in the local block The structure occupancy mask value of individual elements is The strain field prediction model predicts the strain element field as follows: 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. To prevent extremely small constants with a denominator of zero, This represents the L2 norm.

5. The condition-driven intelligent assessment method for RC structure explosion damage according to claim 4, characterized in that, During the training of the damage prediction model, a combined loss function is used. ; in, For the focus of loss, For Dice's loss, Weights to balance the losses; in, 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. Indicates the first The probability that an individual element is predicted as damage. Indicates the first The true damage label of individual elements, Indicates the first Individuals are damaged individuals. Indicates the first The voxels are undamaged voxels; Indicates the first The probability that an individual element is predicted as the true class, when hour, ,when hour, ; Represents the category balance factor. Indicates the focus parameter, This is a smoothing term.

6. The condition-driven intelligent assessment method for RC structure explosion damage according to claim 5, characterized in that, Damage 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 .

7. The condition-driven intelligent assessment method for RC structure explosion damage according to claim 5, characterized in that, Adaptive damage threshold Set it up through the following process: Let the damage probability map of the m-th sample in the output validation set of the damage prediction model be... ,in In the m-th sample, the first... The probability that an individual element is predicted as damage; a multi-candidate threshold scan is performed on the damage probability map, and the candidate threshold set is... ; For each candidate threshold , Binarize the damage probability map into damage labels: Damage labels obtained from binarization and true damage label Calculate the Dice coefficient or Intersection over Union (IoU), and select the candidate threshold that maximizes the Dice coefficient or IoU as the adaptive damage threshold. .

8. The condition-driven intelligent assessment method for RC structure explosion damage according to claim 7, characterized in that, In candidate threshold The Dice coefficient is calculated as follows: in, To determine the number of samples in the validation set, To prevent extremely small constants with a denominator of zero.

9. The condition-driven intelligent assessment method for RC structure explosion damage according to claim 7, characterized in that, In candidate threshold The intersection-union ratio (IoU) is calculated as follows: in, To determine the number of samples in the validation set, To prevent extremely small constants with a denominator of zero.

10. A condition-driven intelligent assessment device for RC structure explosion damage, comprising: The strain field prediction model is used to perform nonlinear mapping on the input damage condition parameters to obtain the potential condition vector, and then decode the potential condition vector to output a three-dimensional strain element field. The damage prediction model is used to predict the damage probability map based on the three-dimensional strain element field, occupancy mask value, and the encoded, expanded, and fused features of the damage condition parameters. The binary conversion module is used to convert values ​​based on an adaptive damage threshold. A three-dimensional voxel failure field is obtained by performing a binary transformation on the damage probability map. The damage intelligent assessment module is used to obtain the three-dimensional strain field and the three-dimensional failure field based on the three-dimensional strain voxel field and the three-dimensional voxel failure field.

Images