A method and system for damage prediction of a concrete structure subjected to blast

Abstract

The present application relates to a kind of concrete structure damage prediction method and system of explosion, belong to concrete explosion technology field, solve the problem that existing concrete damage prediction is limited by insufficient data, high numerical simulation calculation cost.S1, initial first graph construction data is obtained;S2, current first adjacent graph is constructed;S3, if not meet the initiation condition, then based on current first adjacent graph, first prediction model trained, update next time window and its first graph construction data and return S2;If meet the initiation condition, then obtain explosion center position and execute S4;S4, current second adjacent graph is constructed;S5, if not meet the termination condition, then based on current second adjacent graph, second prediction model trained, update next time window and second graph construction data and return S4;If meet the termination condition, then based on current concrete deformation data obtains damage prediction result. Efficient concrete structure damage prediction of explosion is realized.

Description

Technical Field

[0001] This invention relates to the field of concrete structure blasting technology, and in particular to a method and system for predicting damage caused by concrete structure blasting. Background Technology

[0002] Concrete structures are a critical component of infrastructure systems and are frequently subjected to complex external loads, including high-speed penetration loads from warheads and explosive loads after penetrating the target. Under extreme load conditions, accurately predicting the dynamic response of concrete structures is not only crucial for ensuring structural safety and integrity but also provides important information for developing reliable protective designs, guiding emergency response strategies, and conducting damage assessments.

[0003] Traditional methods for predicting the response of concrete structures under penetration and blast loads mainly rely on high-precision numerical simulation methods, primarily using commercial software (such as LS-DYNA) to perform finite element analysis calculations. These numerical simulation tools are capable of handling highly nonlinear, transient dynamic responses, large deformations, material nonlinearities, and complex interactions between structures, and can obtain extremely accurate prediction results. However, their high computational cost, high resource consumption, and long processing time limit their application in practical scenarios.

[0004] In recent years, the rapid development of deep learning technology has provided new ideas for predicting concrete structure damage under blasting. However, existing methods for predicting concrete damage using neural networks are mostly limited by the small amount of available data. When using numerical simulations to expand the dataset, especially for large models, the computation time increases exponentially. Furthermore, the loading conditions are limited to a single penetration or blast, which restricts the feasibility of applying deep learning models in practical engineering scenarios. Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide a method and system for predicting the damage of concrete structures by blasting, in order to solve the problems that existing concrete damage prediction is limited by insufficient data, high cost of numerical simulation calculation, and only applicable to a single loading condition.

[0006] On one hand, embodiments of the present invention provide a method for predicting damage from concrete blasting, comprising the following steps:

[0007] S1. Obtain the first map construction data of the initial time window under the intrusion parameters to be predicted; the first map construction data includes concrete deformation data, intrusion parameters and projectile pose data. S2. Construct the current first adjacency graph based on the first graph in the current time window; S3. Determine whether the detonation conditions are met. If not, update the next time window and its first graph construction data based on the current first adjacency graph and the trained first prediction model, and return to step S2. If the conditions are met, obtain the explosion center location and execute step S4. S4. Based on the current time window, construct the second graph data to build the current second adjacency graph; the second graph construction data includes concrete deformation data and the location of the explosion center; S5. Determine whether the termination condition is met. If not, update the next time window and its second graph construction data based on the current second adjacency graph and the trained second prediction model, and return to step S4. If the condition is met, obtain the concrete damage prediction result based on the concrete deformation data of the current time window.

[0008] Furthermore, the time window is a continuously preset number of time steps; the first graph construction data includes concrete deformation data for each time step of the time window, projectile pose data for the last time step, and penetration parameters; the second graph construction data includes concrete deformation data for each time step of the time window, and the location of the explosion center; the concrete deformation data includes concrete coordinate data and strain data; the penetration parameters include penetration condition parameters and projectile size parameters, the penetration condition parameters including initial penetration velocity, initial incident angle, and concrete strength; the projectile pose data includes the projectile's center of mass position and the unit vector of the axial direction.

[0009] Furthermore, the first adjacency graph includes nodes and their adjacency relationships, wherein each node corresponds to a concrete element, and the node features include the coordinate data, strain data, and erosion parameters of the corresponding concrete element at each time step within the time window; the first adjacency graph is constructed in the following manner: Obtain concrete deformation data, projectile pose data, and projectile size parameters at the last time step of the current time window; The concrete coordinate data is transformed from the global coordinate system to the projectile's local coordinate system, and the projectile shape function is used to obtain the candidate neighbor node set; Based on the candidate neighbor node set, a distance-constrained local neighborhood search is used to determine the neighbor node set of each node and obtain the adjacency relationship of each node; then, based on the first graph, data is constructed to obtain the node features of each node, and the construction of the first adjacency graph is completed.

[0010] Furthermore, in step S3, if the detonation conditions are not met, then Input the current first adjacency graph into the trained first prediction model to obtain the predicted concrete strain data and projectile pose data for the next time step; Based on the predicted concrete strain data for the next time step, update the concrete element coordinate data for the next time step to obtain the concrete deformation data for the next time step. Update the time window by setting the next time step as the last time step, and obtain the next time window and its first graph construction data.

[0011] Furthermore, the first prediction model includes a feature projection layer, a graph encoder, a strain decoder, and a pose decoder; The feature projection layer is used to map the node features of the received first adjacency graph into an initial node feature matrix of high-dimensional features. The graph encoder includes a multi-layered stacked graph attention layer, which is used to exchange and aggregate information between each node in the received initial node feature matrix and its neighboring nodes based on the adjacency relationship of the received first adjacency graph, to obtain a high-level node feature matrix. The strain decoder is used to map the high-dimensional features of each node in the high-level node feature matrix to a scalar value, thereby obtaining the predicted concrete strain data for the next time step. The pose decoder is used to aggregate the high-level node feature matrix into a global feature vector, and then decode it to obtain the predicted projectile pose data for the next time step.

[0012] Furthermore, the second adjacency graph includes nodes and their adjacency relationships, wherein each node corresponds to a concrete element, and the node features include the coordinate data and strain data of the corresponding concrete element at each time step within the time window; the second adjacency graph is constructed in the following manner: Based on the coordinates of the explosion center, the radial distance and radial unit vector of each concrete element from the explosion center are obtained; For each concrete element, other concrete elements that simultaneously satisfy the perimeter condition, orientation alignment condition, and state continuity condition are taken as its candidate neighbor node set, and the nearest nodes are selected from the candidate neighbor set as its final neighbor node set to obtain the adjacency relationship of each node. Then, based on the second graph, data is constructed to obtain the node characteristics of each node, thus completing the construction of the second adjacency graph.

[0013] Furthermore, the second prediction model includes a feature projection layer, a graph encoder, and a strain decoder; The feature projection layer is used to map the node features of the received second adjacency graph into an initial node feature matrix of high-dimensional features. The graph encoder includes a multi-layered stacked graph attention layer, which is used to exchange and aggregate information between each node in the received initial node feature matrix and its neighboring nodes based on the adjacency relationship of the received second adjacency graph to obtain a high-level node feature matrix. The strain decoder is used to map the high-dimensional features of each node in the high-level node feature matrix to a scalar value, thereby obtaining the predicted concrete strain data for the next time step.

[0014] Furthermore, the loss function used during the training of the first prediction model... Represented as: ; in, ; ; ; In the formula, , , These represent the strain prediction loss, pose prediction loss, and physical constraint loss before detonation, respectively. , , These represent the weighting coefficients for strain prediction loss, pose prediction loss, and physical constraint loss before detonation, respectively. , These represent the weighted mean square error and Smooth L1 loss before detonation, respectively. This indicates the initial balancing weights before detonation. , These represent the loss of the center of mass position and the loss in the axial direction before detonation, respectively. This represents the second balancing weight before detonation. , These represent the strain monotonicity constraint loss and spatial continuity constraint loss before detonation, respectively. This represents the third balancing weight before detonation.

[0015] Furthermore, the loss function used during the training of the second prediction model is... Represented as: ; in, ; ; In the formula, , These represent the predicted strain loss and physical constraint loss after detonation, respectively. , These represent the weighting coefficients for the strain prediction loss and physical constraint loss after detonation, respectively. , These represent the weighted mean square error and the Smooth L1 loss after detonation, respectively. This indicates the initial balancing weight after detonation. , These represent the wave propagation direction smoothness constraint loss and wavefront characteristic constraint loss after detonation, respectively. This represents the third balancing weight after detonation.

[0016] On the other hand, embodiments of the present invention provide a damage prediction system for concrete structure blasting, comprising: The initial data acquisition module is used to acquire the first map construction data of the initial time window under the predicted blasting parameters; the first map construction data includes concrete deformation data, blasting parameters and projectile pose data. The penetration phase prediction module is used to iterate the following process: construct the current first adjacency graph based on the first graph construction data of the current time window; determine whether the detonation conditions are met; if not, update the next time window and its first graph construction data based on the current first adjacency graph and the trained first prediction model, and continue iterating; if met, obtain the explosion center location and the second graph construction data, and transmit them to the explosion phase prediction module. The explosion stage prediction module iterates through the following process: constructing the current second adjacency graph based on the second graph construction data of the current time window; the second graph construction data includes concrete deformation data and the explosion center location; determining whether the termination condition is met; if not, updating the next time window and its second graph construction data based on the current second adjacency graph and the trained second prediction model, and continuing the iteration; if the condition is met, obtaining the concrete damage prediction result based on the concrete deformation data of the current time window.

[0017] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: This invention provides a method and system for predicting damage to concrete structures under blast loading. By acquiring data from an initial time window, an initial first adjacency graph is obtained. Based on the construction of the first adjacency graph and a trained first prediction model, concrete deformation data during the blast process is predicted. Then, based on the construction of a second adjacency graph and a trained second prediction model, concrete deformation data during the blast process is predicted, resulting in final concrete deformation data and a predicted damage result. This achieves rapid and high-precision prediction of the damage and response of concrete structures under blast loading. Based on the initial simulation results, the damage of concrete structures under blast loading is predicted. The prediction framework, combining numerical simulation and data-driven modeling, can efficiently assess the damage modes of concrete structures under penetration and explosion, achieving efficient and high-precision prediction of concrete penetration and explosion damage processes. While fully leveraging the advantages of efficient modeling in deep learning, it introduces physical constraints to ensure the physical rationality and engineering interpretability of the prediction results, providing a new technical path for structural protection design, emergency response, and rapid damage assessment. It can successfully capture key damage features such as crater size and penetration depth, providing a feasible approach for rapid damage assessment of complex structures under extreme loads, and has significant application value for protection design and operational planning.

[0018] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0020] Figure 1 This is a flowchart illustrating the method for predicting damage from concrete structure blasting provided in Embodiment 1 of the present invention. Figure 2 This is a finite element diagram of the concrete structure blasting process provided in Embodiment 1 of the present invention. Detailed Implementation

[0021] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0022] Example 1 A specific embodiment of the present invention discloses a method for predicting damage from concrete blasting, such as... Figure 1 As shown, it includes the following steps S1 to S5: S1. Obtain the first map construction data for the initial time window under the predicted blast parameters; the first map construction data includes concrete deformation data, blast parameters, and projectile pose data.

[0023] Specifically, the time window is a continuously preset number of time steps; the penetration parameters include penetration condition parameters and projectile size parameters, the penetration condition parameters include initial penetration velocity, initial incident angle, and concrete strength; the first graph construction data includes concrete deformation data for each time step of the time window, projectile pose data for the last time step, and penetration parameters; the concrete deformation data includes concrete coordinate data and strain data; the projectile pose data includes the projectile's center of mass position and axial direction unit vector.

[0024] Specifically, based on the predicted penetration parameters, numerical simulation data from multiple consecutive time steps within the initial time window (i.e., the initial penetration and explosion) of the projectile were obtained using LS-DYNA software. In detail, finite element models of the projectile and concrete were established, material parameters were input, contact, boundary conditions, and initial conditions were set, and the projectile penetration and explosion process was realized through a complete restart method. The established model is as follows: Figure 2 As shown; for finite element calculation, the concrete and the projectile need to be discretized and represented by individual elements, which contain coordinate and strain data, thereby obtaining the projectile pose data; during the data acquisition process, a method of constructing data in batches was adopted to avoid the occurrence of insufficient hardware memory.

[0025] It should be noted that the concrete strain data in the concrete deformation data needs to be normalized according to the mean and standard deviation of the concrete elements saved in the training phase, so that the models used later can be trained stably and efficiently and achieve good generalization performance.

[0026] S2. Construct the data based on the first graph in the current time window, and construct the current first adjacency graph.

[0027] In implementation, the first adjacency graph includes nodes and their adjacency relationships, wherein one node corresponds to one concrete element, and the node features include the coordinate data and strain data of the corresponding concrete element at each time step of the time window, as well as the erosion parameters. During implementation, the first adjacency graph is constructed in the following manner: S21. Obtain the concrete deformation data, projectile pose data, and projectile size parameters at the last time step of the current time window.

[0028] S22. Transform the concrete coordinate data from the global coordinate system to the projectile's local coordinate system, and use the projectile shape function to obtain the candidate neighbor node set.

[0029] Specifically, the projectile's local coordinate system is a coordinate system with the projectile's center of mass as the origin and its axis as the reference direction. In other words, it is a local coordinate system with the projectile's center of mass as the origin and the projectile's axis as the Z-axis.

[0030] In practice, the candidate neighbor node set is obtained in the following way: S221. Based on the axial coordinates of each concrete element in the local coordinate system of the projectile and the shape function of the projectile, the radial distance threshold of each concrete element in the local coordinate system of the projectile is obtained.

[0031] Specifically, the projectile shape function is constructed based on the projectile shape type, and the axial coordinates of the concrete element in the projectile local coordinate system are substituted to output the radial distance threshold of the corresponding concrete element in the projectile local coordinate system.

[0032] Specifically, if the projectile shape is oval, then the projectile dimensional parameters are the projectile radius of curvature, the projectile radius, the projectile length, and the total projectile length. The projectile shape function is expressed as: ; In the formula, The axial coordinates of the concrete element in the projectile's local coordinate system are: Radial distance threshold at time, , , , These represent the warhead radius of curvature, the projectile radius, the warhead length, and the total projectile length, respectively.

[0033] Specifically, if the projectile shape is a pointed cone, then the projectile's dimensional parameters are the cone half-angle, projectile radius, projectile length, and projectile total length, and the projectile shape function is expressed as: ; In the formula, The axial coordinates of the concrete element in the projectile's local coordinate system are: Radial distance threshold at time, , , , These represent the warhead radius of curvature, the projectile radius, the warhead length, and the total projectile length, respectively.

[0034] Specifically, if the projectile shape is flat-topped, then the projectile size parameters are the projectile radius and the total projectile length, and the projectile shape function is expressed as: ; In the formula, The axial coordinates of the concrete element in the projectile's local coordinate system are: Radial distance threshold at time, , These represent the projectile radius and the total length of the projectile, respectively.

[0035] Specifically, if the projectile shape is biconical, then the projectile dimensional parameters are the projectile radius, the front cone half-angle, the rear cone half-angle, the front cone length, the suffix total length, and the total projectile length. The projectile shape function is expressed as: ; In the formula, The axial coordinates of the concrete element in the projectile's local coordinate system are: Radial distance threshold at time, , , , , , These represent the projectile radius, the front cone half-angle, the rear cone half-angle, the front cone length, the suffix total length, and the projectile total length, respectively.

[0036] S222. Based on the abscissa and ordinate of each concrete element in the local coordinate system of the projectile, the radial distance of each concrete element is obtained; if the radial distance of a concrete element is less than or equal to the radial distance threshold of that concrete element, then the concrete element is regarded as a neighbor node; all neighbor nodes constitute a candidate neighbor node set.

[0037] In practical implementation, the radial distance of the concrete element is the radial distance from the concrete element to the axis of the projectile, expressed as: ; In the formula, Indicates the first The radial distance of each concrete element in the projectile's local coordinate system. , They represent the first The horizontal and vertical coordinates of a concrete element in the local coordinate system of the projectile.

[0038] S23. Based on the candidate neighbor node set, use distance-constrained local neighborhood search to determine the neighbor node set of each node and obtain the adjacency relationship of each node; then construct data based on the first graph to obtain the node features of each node and complete the construction of the first adjacency graph.

[0039] In practice, based on the candidate neighbor node set, a distance-constrained local neighborhood search is used to determine the neighbor node set of each node, including: S231. Determine the maximum connection distance based on the projectile size parameters. , represented as: ; In the formula, Indicates the attenuation coefficient. This represents the projectile radius. For example, based on the stress attenuation law, the attenuation coefficient is set to 2 or 3.

[0040] S232. Using the coordinates of the current node as the center, find all nodes in the candidate neighbor node set that satisfy the condition that the Euclidean distance is less than the maximum connection distance, and select the closest nodes as the neighbor node set of the current node.

[0041] Specifically, the 8 or 9 nodes with the largest distance are selected as the set of neighbor nodes of the current node. If the number is insufficient, all nodes are selected. This setting ensures the sparsity of the graph and computational efficiency.

[0042] Understandably, the connectivity in the graph construction is directly defined by the physical interaction between the projectile and the concrete, rather than purely geometric relationships, thus imbuing the graph structure with prior physical knowledge. The search range is adaptively adjusted through shape functions and axial positions, making it more realistic. Rapid initial screening using shape functions greatly reduces the set of candidate elements requiring precise distance calculations, improving the speed of graph construction. The search range for neighboring nodes is a dynamic range defined based on the projectile's size parameters, ensuring graph sparsity and computational efficiency while avoiding isolated nodes or extreme degree distribution problems that may result from uneven element distribution.

[0043] S3. Determine whether the detonation conditions are met. If not, update the next time window and its first graph construction data based on the current first adjacency graph and the trained first prediction model, and return to step S2. If the conditions are met, obtain the explosion center location and execute step S4.

[0044] Optionally, the detonation condition is that the number of initial time steps of the final time step of the current window is greater than a set detonation number threshold; wherein, the detonation number threshold is set according to the specific concrete parameters and projectile parameters.

[0045] Optionally, the detonation condition is that the predicted value of the current first adjacency graph input to the detonation prediction model is greater than or equal to a set detonation threshold.

[0046] Specifically, the detonation prediction model is a binary classification neural network, for example, a small GNN or an MLP with global pooling features, which takes a first adjacency graph as input and outputs a predicted probability value of the explosion.

[0047] Specifically, the detonation threshold is set according to the detonation law of the projectile; for example, it is set to 0.7.

[0048] During implementation, if the detonation conditions are not met in step S3, then S31. Input the current first adjacency graph into the trained first prediction model to obtain the predicted concrete strain data and projectile pose data for the next time step.

[0049] Specifically, the first adjacency graph is represented using a standard graph data structure, such as an adjacency matrix or an edge index list.

[0050] In specific implementation, the first prediction model includes a feature projection layer, a graph encoder, a strain decoder, and a pose decoder. The feature projection layer is used to map the node features of the received first adjacency graph into an initial node feature matrix of high-dimensional features. The graph encoder includes a multi-layered stacked graph attention layer, which is used to exchange and aggregate information between each node in the received initial node feature matrix and its neighboring nodes based on the adjacency relationship of the received first adjacency graph, to obtain a high-level node feature matrix. The strain decoder is a node-level regressor used to map the high-dimensional features of each node in the high-level node feature matrix to a scalar value, thereby obtaining the predicted concrete strain data for the next time step; it includes applying a fully connected layer to the features of each node. The pose decoder is a graph-level regressor used to aggregate high-level node feature matrices into a global feature vector, and then decode it to obtain the predicted projectile pose data for the next time step; it includes global pooling and fully connected decoding.

[0051] Specifically, the graph encoder includes four stacked graph attention layers. Each graph attention layer includes attention coefficient calculation, normalized attention weights, weighted feature aggregation, and multi-head attention output.

[0052] S32. Based on the predicted concrete strain data for the next time step, update the concrete coordinate data for the next time step to obtain the concrete deformation data for the next time step.

[0053] Specifically, the strain mapping method or the coordinate invariance assumption is used to obtain concrete coordinate data based on concrete strain data. The strain mapping method estimates the displacement increment of the element based on the predicted strain increment using simplified constitutive or geometric relations, thereby updating the element coordinates. The coordinate invariance assumption assumes that the element coordinate changes are negligible within a short prediction step; both are existing technologies and will not be elaborated upon here.

[0054] S33. Update the time window by setting the next time step as the last time step, and obtain the next time window and its first graph construction data. Specifically, this involves moving the time window by one time step.

[0055] During implementation, in step S3, if the detonation conditions are met, the projectile pose data of the current time window is used as the explosion center position.

[0056] S4. Based on the data of the second graph constructed in the current time window, construct the current second adjacency graph; the data of the second graph construction includes concrete deformation data and the location of the explosion center.

[0057] Specifically, the data used to construct the second graph includes concrete deformation data at each time step of the time window, as well as the location of the explosion center.

[0058] In implementation, the second adjacency graph includes nodes and their adjacency relationships, wherein each node corresponds to a concrete element, and the node features include the coordinate data and strain data of the corresponding concrete element at each time step of the time window.

[0059] During implementation, the second adjacency graph is constructed in the following manner: S41. Based on the coordinates of the explosion center, obtain the radial distance and radial unit vector of each concrete element from the explosion center.

[0060] Specifically, the radial distance and radial unit vector from each concrete element to the explosion center are expressed as follows: ; ; In the formula, Indicates the first The radial distance from each concrete unit to the explosion center Indicates the first The coordinates of each concrete element The coordinates representing the location of the explosion's center. Indicates the first The radial unit vector from a concrete element to the explosion center. This indicates taking the modulus.

[0061] S42. For each concrete element, other concrete elements that simultaneously satisfy the external conditions, orientation alignment conditions, and state continuity conditions are taken as its candidate neighbor node set, and the nearest nodes are selected from the candidate neighbor set as its final neighbor node set, thus obtaining the adjacency relationship of each node.

[0062] Specifically, for the first One concrete unit, The external conditions are expressed as follows: ; In the formula, Indicates the first The radial distance from each concrete unit to the explosion center.

[0063] The orientation alignment condition is from the first The concrete unit points to the first The unit vector of the first concrete element and the first... The angle between the radial unit vector of each concrete element and the explosion center is less than a preset angle threshold. The condition for state continuity is the first The concrete unit and the first The absolute value of the difference in strain data among individual concrete units is less than the preset strain threshold.

[0064] Specifically, the included angle threshold is set according to the acceptable range of propagation directions, for example, it is set to 60 degrees.

[0065] Specifically, based on the fundamental principle of continuous deformation of adjacent material points in continuum mechanics, the maximum strain data of the current time step is set to 5% to 10% to filter out non-physical abrupt changes caused by numerical noise or isolated failure points, ensuring that the connection mainly occurs in damage areas with similar mechanical properties.

[0066] Specifically, the nearest nodes are selected from the candidate neighbor set as its final neighbor set, including: S421. Determine the search radius based on the radial distance from the current node to the explosion center; Specifically, the search radius is expressed as: ; in, ; In the formula, Indicates the first The search radius of each node. This represents the search range coefficient. , Let represent the first search coefficient and the second search coefficient, respectively. , ;in, The simplified function indicates that the search radius increases with distance, but the rate of increase decreases, simulating wavefront expansion.

[0067] S422. Using the coordinates of the current node as the center, find all nodes in the candidate neighbor set of the current node whose Euclidean distance is less than its search radius, and select several nodes with the smallest radial distance difference as the final neighbor set of the current node.

[0068] Specifically, the set of 8 or 9 nodes with the smallest radial distance difference is selected as the final neighbor set of the current node; if the number is insufficient, all nodes are selected. This setting ensures both graph sparsity and computational efficiency. S43. Then, based on the second graph, construct the data to obtain the node characteristics of each node, and complete the construction of the second adjacency graph.

[0069] S5. Determine whether the termination condition is met. If not, update the next time window and its second graph construction data based on the current second adjacency graph and the trained second prediction model, and return to step S4. If the condition is met, obtain the concrete damage prediction result based on the concrete deformation data of the current time window.

[0070] Specifically, the termination condition is that the number of initial time steps of the current window's final time step intrusion is greater than a set termination number threshold; whereby the termination number threshold is set according to specific concrete parameters and projectile parameters.

[0071] During implementation, if the termination condition is not met in step S5, then S51. Input the current second adjacency graph into the trained second prediction model to obtain the predicted concrete strain data for the next time step.

[0072] Specifically, the second adjacency graph is represented using standard graph data structures, such as an adjacency matrix or an edge index list.

[0073] In specific implementation, the second prediction model includes a feature projection layer, a graph encoder, a strain decoder, and a pose decoder; The feature projection layer is used to map the node features of the received second adjacency graph into an initial node feature matrix of high-dimensional features. The graph encoder includes a multi-layered stacked graph attention layer, which is used to exchange and aggregate information between each node in the received initial node feature matrix and its neighboring nodes based on the adjacency relationship of the received second adjacency graph to obtain a high-level node feature matrix. The strain decoder is a node-level regressor used to map the high-dimensional features of each node in the high-level node feature matrix to a scalar value, thereby obtaining the predicted concrete strain data for the next time step; it includes applying a fully connected layer to the features of each node. Specifically, the graph encoder includes four stacked graph attention layers. Each graph attention layer includes attention coefficient calculation, normalized attention weights, weighted feature aggregation, and multi-head attention output.

[0074] S52. Based on the predicted concrete strain data for the next time step, update the concrete coordinate data for the next time step to obtain the concrete deformation data for the next time step.

[0075] Specifically, the strain mapping method or the coordinate invariance assumption is used to obtain concrete coordinate data based on concrete strain data. The strain mapping method estimates the displacement increment of the element based on the predicted strain increment using simplified constitutive or geometric relations, thereby updating the element coordinates. The coordinate invariance assumption assumes that the element coordinate changes are negligible within a short prediction step; both are existing technologies and will not be elaborated upon here.

[0076] S53. Update the time window by setting the next time step as the last time step, and obtain the next time window and its first graph construction data. Specifically, this involves moving the time window by one time step.

[0077] During implementation, the concrete damage prediction results include penetration depth and crater diameter; based on the concrete deformation data of the current time window, the concrete damage prediction results include: Based on concrete deformation data and a set critical damage strain threshold, the set of concrete elements that have failed is determined; the penetration depth is determined by calculating the maximum axial displacement difference between the original and final positions of all failed elements in the direction of projectile penetration; the crater diameter is determined by extracting the final coordinates of all failed elements on the target surface and calculating twice the maximum lateral distance of their projection points on a plane perpendicular to the penetration direction; among these, the concrete strain data is restored to the original scale through an inverse normalization operation.

[0078] Specifically, the critical damage strain threshold is usually a commonly used empirical value of 0.2, which is set based on dynamic mechanical property experiments of concrete materials, experience in related fields, or failure criteria of specific constitutive models. It is used to reasonably distinguish whether the material has undergone severe plastic deformation or failure in numerical analysis.

[0079] It should be noted that the method in this embodiment predicts the damage to concrete structures under different blasting parameters in the same blasting scenario. The blasting scenario includes the shape and size of the concrete, as well as other parameter settings of the projectile besides the size parameter, such as the projectile shape type. Different blasting scenarios require adaptive adjustments to the steps, such as the projectile shape function in the first adjacency graph construction, and the retraining of the first and second prediction models.

[0080] Specifically, in this embodiment, the time window is set to 3 or 4 consecutive time steps, which achieves the best balance between capturing sufficient time dependence and controlling model complexity and training cost. This not only reflects the trend of damage evolution but also avoids redundant information and computational burden caused by excessively long sequences.

[0081] During implementation, the datasets for training the first and second prediction models are constructed as follows: Based on the set range of each parameter in the blasting parameters, a parametric design space is defined; wherein, the range of each parameter in the blasting parameters is set according to the actual situation, such as velocity 200~400m / s, incident angle 0~20°, projectile radius 20~40mm, and concrete strength 30~50MPa. The Latin hypercube sampling method is used to sample in the parametric design space composed of the intrusion and explosion parameters to obtain the combination of parameters. For each set of parameters, parametric finite element simulation is performed to obtain numerical simulation data of the penetration and explosion process of concrete structure; among them, LS-DYNA is run to complete a complete penetration-explosion process simulation to obtain the time series d3plot file, which is the numerical simulation data. The first and second data sample sets are constructed based on numerical simulation data of the penetration and explosion processes, respectively. The concrete strain data is normalized according to the mean and variance of each global element.

[0082] Specifically, a first data sample set and a second data sample set are constructed based on numerical simulation data of the penetration process and the explosion process, respectively, including: For the numerical simulation data of the penetration process of each concrete structure, samples are extracted in the form of a sliding time window. The samples include concrete deformation data within the time window, and the sample labels are concrete strain data and projectile pose data of the next time step. The order of each sample is shuffled and a graph is constructed for each sample to obtain the corresponding first adjacency graph, and then the first data sample set is obtained. For the numerical simulation data of the explosion process of each concrete structure, samples are extracted in the form of a sliding time window. The samples include concrete deformation data within the time window and the sample label is the concrete strain data of the next time step. The order of each sample is shuffled and a graph is constructed for each sample to obtain the corresponding second adjacency graph, and then the second data sample set is obtained.

[0083] Specifically, the first and second data sample sets are divided into training, validation, and test sets proportionally, such as 70%, 15%, and 15%.

[0084] In practice, the loss function used during the training of the first prediction model is... Represented as: ; in, ; ; ; In the formula, , , These represent the strain prediction loss, pose prediction loss, and physical constraint loss before detonation, respectively. , , The weight coefficients of strain prediction loss, pose prediction loss and physical constraint loss before detonation are respectively represented. The combination that optimizes the overall performance is selected by mesh search or Bayesian optimization through the validation set. , These represent the weighted mean square error and Smooth L1 loss before detonation, respectively. This represents the first balancing weight before detonation, a hyperparameter between 0 and 1, used to balance the two error metrics, and is determined by adjustment through the validation set; , These represent the loss of the center of mass position and the loss in the axial direction before detonation, respectively. This indicates the second balancing weight before detonation; , These represent the strain monotonicity constraint loss and spatial continuity constraint loss before detonation, respectively. This represents the third balancing weight before detonation.

[0085] Specifically, the weighted mean square error before detonation Represented as: ; in, ; In the formula, N Indicates the total number of concrete structural units; The model predicts the first Strain data for each concrete unit; Indicates the first Real strain data for each concrete unit; Indicates the first The weight of each concrete unit; , These represent the weighting coefficients for the high and low strain regions, respectively. They are hyperparameters determined through adjustment using the validation set. For example, , They are 2 and 0.5 respectively; The threshold value that distinguishes between high and low strain regions is a hyperparameter, and for example, it is set to 0.1.

[0086] Specifically, the Smooth L1 loss before detonation Represented as: ; in, ; In the formula, The threshold for the loss function to shift from the quadratic region to the linear region is a hyperparameter, usually set to 1.

[0087] Specifically, the loss of the center of mass position before detonation Represented as: ; In the formula, This represents the coordinates of the projectile's center of mass predicted by the model. This represents the actual coordinates of the projectile's center of mass. This represents the L2 norm.

[0088] Specifically, axial loss before detonation They are represented as follows: ; In the formula, This represents the projectile axial vector predicted by the model. This represents the actual axial vector of the projectile.

[0089] Specifically, the strain monotonicity constraint loss before detonation Represented as: ; In the formula, Indicates the first The time steps predicted by the concrete unit model t Strain data, Indicates the first The concrete unit's actual previous time step t The strain data is -1. Understandably, this loss penalizes the reduction in the strain prediction relative to the actual value at the previous moment, since plastic strain is typically irreversible during penetration, i.e., monotonically unreduced.

[0090] Specifically, the loss of spatial continuity constraint before detonation Represented as: ; In the formula, , They represent the first The concrete unit and its neighbors Strain data predicted by a concrete element model; Indicates the first The set of neighbor nodes of a concrete element. The loss represents modulo; understandably, this loss penalizes the sharp difference between the node's predicted strain and the average strain of its neighbors, encouraging the damage field to propagate smoothly in space, which is in line with physical intuition.

[0091] It should be noted that in each training iteration, forward propagation obtains the predicted strain data and the predicted projectile pose data. Then, using the real labels obtained from the training samples and the state of the previous time step, combined with the neighbor relationships obtained from the current adjacency graph structure, the above total loss is calculated. Finally, the model parameters are updated through backpropagation.

[0092] In practice, the loss function used during the training of the second prediction model is... Represented as: ; in, ; ; In the formula, , These represent the predicted strain loss and physical constraint loss after detonation, respectively. , These represent the weighting coefficients for the strain prediction loss and physical constraint loss after detonation, respectively. , These represent the weighted mean square error and the Smooth L1 loss after detonation, respectively. This indicates the initial balancing weight after detonation. , These represent the wave propagation direction smoothness constraint loss and wavefront characteristic constraint loss after detonation, respectively. This represents the third balancing weight after detonation.

[0093] Specifically, the weighted mean square error after detonation Represented as: ; in, ; In the formula, Indicates the first day after detonation The weight of each concrete unit; Indicates the first The distance from each concrete unit to the explosion center. This indicates the maximum distance from each concrete unit to the explosion center. This indicates an indicator function that takes the value 1 if the condition is true and 0 otherwise. , , , All of these are hyperparameters. Used to enhance accuracy in high-strain regions (typically areas of severe damage near the explosion core). This is used to enhance accuracy in the near-field region, because the gradient is large and the error has a significant impact in this region.

[0094] Specifically, the loss of Smooth L1 after detonation Losses of Smooth L1 before detonation Indicates the same.

[0095] Specifically, the smoothness constraint loss of wave propagation direction after detonation. Loss of spatial continuity constraints before detonation Indicates the same.

[0096] Specifically, wavefront characteristic constraint loss after detonation Represented as: ; in, and .

[0097] In the formula, Indicates the first The distance from each concrete unit to the explosion center.

[0098] It is understandable that the above loss function is for a single sample. The loss function combines two types of components: data-driven and physical constraints, in order to simultaneously ensure prediction accuracy and physical consistency. Among the physical constraint losses, the strain monotonicity constraint loss is used to penalize the non-monotonic decrease of predicted strain that does not conform to physical laws; the spatial continuity constraint loss is used to penalize the drastic changes in predicted strain that do not conform to physical laws between adjacent units.

[0099] Specifically, each model was trained using the AdamW optimizer, which suppresses overfitting through weight decay. The ReduceLROnPlateau learning rate scheduler was used to dynamically adjust the learning rate based on changes in the validation set loss. The training loop iteratively processes historical strain and location information in mini-batch format to fully capture the spatiotemporal dynamics of the structural response. Simultaneously, an early stopping strategy was introduced to control the training process based on the validation set loss, achieving optimal convergence and avoiding overfitting. Model evaluation used standard metrics such as mean absolute error and coefficient of determination to quantify prediction accuracy. To facilitate physical comparison and verification with actual experimental data, the model prediction results were restored to their original scale through inverse normalization. After training, the prediction results were saved in structured .npz file format for subsequent analysis.

[0100] Compared with existing technologies, this embodiment provides a method for predicting the damage of concrete structures under blasting. It obtains an initial first adjacency graph based on the data acquired in the initial time window. Then, based on the construction of the first adjacency graph and a trained first prediction model, it predicts the concrete deformation data during the blasting process. Finally, based on the construction of a second adjacency graph and a trained second prediction model, it predicts the concrete deformation data during the blasting process, obtaining the final concrete deformation data and thus the concrete damage prediction result. This achieves rapid and high-precision prediction of the damage and response of concrete structures under blasting-blasting coupling. Based on the initial simulation results of several iterations, it predicts the concrete structure damage under blasting loading. The prediction framework, combining numerical simulation and data-driven modeling, can efficiently assess the damage modes of concrete structures under penetration and explosion, achieving efficient and high-precision prediction of concrete penetration and explosion damage processes. While fully leveraging the advantages of efficient modeling in deep learning, it introduces physical constraints to ensure the physical rationality and engineering interpretability of the prediction results, providing a new technical path for structural protection design, emergency response, and rapid damage assessment. It can successfully capture key damage features such as crater size and penetration depth, providing a feasible approach for rapid damage assessment of complex structures under extreme loads, and has significant application value for protection design and operational planning.

[0101] Example 2 A specific embodiment of the present invention discloses a damage prediction system for concrete structure blasting, comprising: The initial data acquisition module is used to acquire the first map construction data of the initial time window under the predicted blasting parameters; the first map construction data includes concrete deformation data, blasting parameters and projectile pose data. The penetration phase prediction module is used to iterate the following process: construct the current first adjacency graph based on the first graph construction data of the current time window; determine whether the detonation conditions are met; if not, update the next time window and its first graph construction data based on the current first adjacency graph and the trained first prediction model, and continue iterating; if met, obtain the explosion center location and the second graph construction data, and transmit them to the explosion phase prediction module. The explosion stage prediction module iterates through the following process: constructing the current second adjacency graph based on the second graph construction data of the current time window; the second graph construction data includes concrete deformation data and the explosion center location; determining whether the termination condition is met; if not, updating the next time window and its second graph construction data based on the current second adjacency graph and the trained second prediction model, and continuing the iteration; if the condition is met, obtaining the concrete damage prediction result based on the concrete deformation data of the current time window.

[0102] The specific implementation process of this invention can be found in the above method embodiments, and will not be repeated here.

[0103] Since this embodiment is based on the same principle as the above-described method embodiments, this system also has the corresponding technical effects of the above-described method embodiments.

[0104] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0105] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting damage from concrete blasting, characterized in that, Includes the following steps: S1. Obtain the first map construction data of the initial time window under the intrusion parameters to be predicted; the first map construction data includes concrete deformation data, intrusion parameters and projectile pose data. S2. Construct the current first adjacency graph based on the first graph in the current time window; S3. Determine whether the detonation conditions are met. If not, update the next time window and its first graph construction data based on the current first adjacency graph and the trained first prediction model, and return to step S2. If the conditions are met, obtain the explosion center location and execute step S4. S4. Based on the current time window, construct the second graph data to build the current second adjacency graph; the second graph construction data includes concrete deformation data and the location of the explosion center; S5. Determine whether the termination condition is met. If not, update the next time window and its second graph construction data based on the current second adjacency graph and the trained second prediction model, and return to step S4. If the condition is met, obtain the concrete damage prediction result based on the concrete deformation data of the current time window.

2. The method for predicting damage from concrete structure blasting according to claim 1, characterized in that, The time window is a continuously preset number of time steps; the first graph construction data includes concrete deformation data for each time step of the time window, projectile pose data for the last time step, and penetration parameters; the second graph construction data includes concrete deformation data for each time step of the time window, and the location of the explosion center; the concrete deformation data includes concrete coordinate data and strain data; the penetration parameters include penetration condition parameters and projectile size parameters, the penetration condition parameters including initial penetration velocity, initial incident angle, and concrete strength; the projectile pose data includes the projectile's center of mass position and the unit vector of the axial direction.

3. The method for predicting damage from concrete structure blasting according to claim 2, characterized in that, The first adjacency graph includes nodes and their adjacency relationships, wherein each node corresponds to a concrete element, and the node features include the coordinate data, strain data, and blasting parameters of the corresponding concrete element at each time step of the time window; the first adjacency graph is constructed in the following manner: Obtain concrete deformation data, projectile pose data, and projectile size parameters at the last time step of the current time window; The concrete coordinate data is transformed from the global coordinate system to the projectile's local coordinate system, and the projectile shape function is used to obtain the candidate neighbor node set; Based on the candidate neighbor node set, a distance-constrained local neighborhood search is used to determine the neighbor node set of each node and obtain the adjacency relationship of each node; then, based on the first graph, data is constructed to obtain the node features of each node, and the construction of the first adjacency graph is completed.

4. The method for predicting damage from concrete blasting according to claim 3, characterized in that, In step S3, if the detonation conditions are not met, then Input the current first adjacency graph into the trained first prediction model to obtain the predicted concrete strain data and projectile pose data for the next time step; Based on the predicted concrete strain data for the next time step, update the concrete element coordinate data for the next time step to obtain the concrete deformation data for the next time step. Update the time window by setting the next time step as the last time step, and obtain the next time window and its first graph construction data.

5. The method for predicting damage from concrete structure blasting according to claim 4, characterized in that, The first prediction model includes a feature projection layer, a graph encoder, a strain decoder, and a pose decoder; The feature projection layer is used to map the node features of the received first adjacency graph into an initial node feature matrix of high-dimensional features. The graph encoder includes a multi-layered stacked graph attention layer, which is used to exchange and aggregate information between each node in the received initial node feature matrix and its neighboring nodes based on the adjacency relationship of the received first adjacency graph, to obtain a high-level node feature matrix. The strain decoder is used to map the high-dimensional features of each node in the high-level node feature matrix to a scalar value, thereby obtaining the predicted concrete strain data for the next time step. The pose decoder is used to aggregate the high-level node feature matrix into a global feature vector, and then decode it to obtain the predicted projectile pose data for the next time step.

6. The method for predicting damage from concrete structure blasting according to claim 4, characterized in that, The second adjacency graph includes nodes and their adjacency relationships, wherein each node corresponds to a concrete element, and the node features include the coordinate data and strain data of the corresponding concrete element at each time step within the time window; the second adjacency graph is constructed in the following manner: Based on the coordinates of the explosion center, the radial distance and radial unit vector of each concrete element from the explosion center are obtained; For each concrete element, other concrete elements that simultaneously satisfy the perimeter condition, orientation alignment condition, and state continuity condition are taken as its candidate neighbor node set, and the nearest nodes are selected from the candidate neighbor set as its final neighbor node set to obtain the adjacency relationship of each node. Then, based on the second graph, data is constructed to obtain the node characteristics of each node, thus completing the construction of the second adjacency graph.

7. The method for predicting damage from concrete blasting according to claim 6, characterized in that, The second prediction model includes a feature projection layer, a graph encoder, and a strain decoder; The feature projection layer is used to map the node features of the received second adjacency graph into an initial node feature matrix of high-dimensional features. The graph encoder includes a multi-layered stacked graph attention layer, which is used to exchange and aggregate information between each node in the received initial node feature matrix and its neighboring nodes based on the adjacency relationship of the received second adjacency graph to obtain a high-level node feature matrix. The strain decoder is used to map the high-dimensional features of each node in the high-level node feature matrix to a scalar value, thereby obtaining the predicted concrete strain data for the next time step.

8. The method for predicting damage from concrete structure blasting according to claim 5, characterized in that, The loss function used during the training of the first prediction model Represented as: ； in, ； ； ； In the formula, , , These represent the strain prediction loss, pose prediction loss, and physical constraint loss before detonation, respectively. , , These represent the weighting coefficients for strain prediction loss, pose prediction loss, and physical constraint loss before detonation, respectively. , These represent the weighted mean square error and Smooth L1 loss before detonation, respectively. This indicates the initial balancing weights before detonation. , These represent the loss of the center of mass position and the loss in the axial direction before detonation, respectively. This indicates the second balancing weight before detonation. , These represent the strain monotonicity constraint loss and spatial continuity constraint loss before detonation, respectively. This represents the third balancing weight before detonation.

9. The method for predicting damage from concrete blasting according to claim 7, characterized in that, The loss function used during the training of the second prediction model Represented as: ； in, ； ； In the formula, , These represent the predicted strain loss and physical constraint loss after detonation, respectively. , These represent the weighting coefficients for the strain prediction loss and physical constraint loss after detonation, respectively. , These represent the weighted mean square error and the Smooth L1 loss after detonation, respectively. This indicates the initial balancing weight after detonation. , These represent the wave propagation direction smoothness constraint loss and wavefront characteristic constraint loss after detonation, respectively. This represents the third balancing weight after detonation.

10. A damage prediction system for blasting of concrete structures, characterized in that, include The initial data acquisition module is used to acquire the first map construction data of the initial time window under the predicted blasting parameters; the first map construction data includes concrete deformation data, blasting parameters and projectile pose data. The penetration phase prediction module is used to iterate the following process: construct the current first adjacency graph based on the first graph construction data of the current time window; determine whether the detonation conditions are met; if not, update the next time window and its first graph construction data based on the current first adjacency graph and the trained first prediction model, and continue iterating; if met, obtain the explosion center location and the second graph construction data, and transmit them to the explosion phase prediction module. The explosion stage prediction module iterates through the following process: constructing the current second adjacency graph based on the second graph construction data of the current time window; the second graph construction data includes concrete deformation data and the explosion center location; determining whether the termination condition is met; if not, updating the next time window and its second graph construction data based on the current second adjacency graph and the trained second prediction model, and continuing the iteration; if the condition is met, obtaining the concrete damage prediction result based on the concrete deformation data of the current time window.

Images