Multi-step continuous prediction method based on cross-step state transition and double field coupling

Abstract

The application provides a multi-step continuous prediction method based on cross-step state migration and double field coupling, and belongs to the technical field of structural dynamics numerical simulation. In the method, the explicit dynamics simulation only needs to simulate and record the field variable simulation data of the initial stage of the first step or the initial stage of each step, and the evolution data after the initial stage of each step is obtained through prediction. The prediction process adopts geometry-strain double field coupling prediction, wherein the strain field prediction adopts a graph neural network; the geometry field prediction is based on the position field data and the strain field prediction result in the historical time window to predict the position field at the next time; when the step is switched, if the mesh division of the entity unit before and after the switching is consistent, the next step inherits the data of the previous step, otherwise the data of the previous step is mapped to the next step to complete the historical time window. Using the application can realize continuous time sequence propagation prediction of damage state and improve the prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of structural dynamics numerical simulation technology, specifically involving a multi-step continuous prediction method based on cross-step state transition and dual-field coupling. Background Technology

[0002] In key engineering fields such as aerospace and rail transportation, structures will exhibit progressive damage evolution behaviors such as large deformation, strain accumulation, crack propagation, and local failure under extreme dynamic conditions such as explosive impact, collision penetration, and complex variable loads.

[0003] Structural damage evolution analysis is a key method for evaluating the load-bearing capacity and service safety of structures, aiming to reveal the entire process of structural damage evolution under complex loading. Structural responses under extreme conditions exhibit strong nonlinearity and transient abrupt changes, making it difficult for traditional theoretical analytical methods to accurately characterize their evolution.

[0004] Explicit dynamic simulation, with its high-precision solution capability for strongly nonlinear, instantaneous, and dynamically large deformation problems, has become the mainstream numerical analysis method in this field. It iteratively solves the dynamic equilibrium equations time-by-time with extremely small time steps, combining a refined mesh discretization model and material damage constitutive criteria to solve the structural deformation, strain, and damage state element by element and time-by-time, fully reproducing the temporal evolution, local concentration, and progressive failure process of structural damage under complex loading. This simulation method requires the continuous output of massive amounts of time-series frame data, and each frame contains the spatial geometric position, deformation state, and various material state variables of all solid elements, enabling accurate reproduction of the entire damage evolution process of the structure.

[0005] While explicit dynamic simulations offer high accuracy, they are computationally expensive and time-consuming, which hinders efficient R&D needs such as rapid iterative analysis, parameter scanning, and real-time approximate evaluation. This severely restricts the R&D efficiency and application flexibility of dynamic damage analysis for complex structures.

[0006] To address the aforementioned issues, the industry commonly employs data-driven surrogate models to replace explicit dynamic simulations. These models utilize intelligent algorithms such as convolutional networks, recurrent networks, and ordinary graph neural networks to achieve rapid prediction of field variables, thereby reducing computational costs and improving analytical efficiency. However, existing mainstream surrogate model technologies still suffer from several shortcomings, failing to meet the demands for high-precision and robust prediction of complex nonlinear dynamic damage evolution. These specific shortcomings are as follows: First, the model suffers from poor adaptability to temporal evolution, failing to achieve the transfer of historical states across multiple work steps. Explicit dynamic simulations such as LS-DYNA calculate the dynamic processes of different loading types, load magnitudes, and physical mechanisms in stages, i.e., multi-step simulations. Most existing models only predict field variables for a single work step and a single fixed mesh structure, limiting modeling and prediction dimensions to a static, single scenario. Faced with the complexities of continuous loading across multiple work steps and dynamic mesh updates, these models cannot effectively handle historical damage data before and after work step transitions, making it difficult to achieve continuous temporal propagation prediction of damage states and resulting in a disconnect from the actual dynamic loading process in engineering.

[0007] Second, the lack of multi-field coupling modeling limits the predictive dimensions. Current models mostly focus on the independent prediction of single material state quantities such as stress, strain, and damage, ignoring the strong coupling relationship between geometric evolution and material state response during structural dynamic deformation, and thus failing to accurately reproduce the physical mechanism of structural damage evolution.

[0008] Third, the response accuracy in key areas is insufficient, and the targeted prediction capability is weak. Under explosive impact and collision penetration conditions, structural damage and strain gradients exhibit significant localized concentration characteristics, and the core area of ​​high-strain local damage is crucial for condition analysis and performance evaluation. Existing models lack weight enhancement and feature focusing mechanisms for core key areas, resulting in low prediction accuracy for key areas and failing to meet the needs of refined analysis for key engineering areas.

[0009] Fourth, the lack of physical constraints leads to insufficient reasonableness in prediction results. However, most existing data-driven models are pure data fitting paradigms that do not embed basic physical constraint rules. Without physical constraints in the prediction process, they are prone to producing results that violate objective physical laws, significantly reducing the reliability and engineering applicability of the model predictions. Summary of the Invention

[0010] In view of this, the present invention provides a multi-step continuous prediction method based on cross-step state transition and dual-field coupling, which is applicable to complex situations of continuous loading in multiple steps and heterogeneous meshes between steps. It can effectively inherit the damage history data before and after step switching and realize the continuous temporal propagation prediction of damage state. At the same time, it adopts geometry-strain dual-field coupling prediction, which incorporates the coupling correlation characteristics between geometric morphology evolution and material state response during the dynamic deformation of the structure, so as to truly restore the physical mechanism of structural damage evolution and improve the prediction accuracy.

[0011] To solve the above-mentioned technical problems, the present invention is implemented as follows.

[0012] A multi-step continuous prediction method based on cross-step state transition and dual-field coupling is proposed for multi-step structural damage evolution analysis; the method includes: Step 1: Collect data on the structural damage evolution process through dynamic simulation. The collected data should at least cover the field variable simulation data of each entity unit in the initial stage of the first step and be stored in the evolution dataset. Step 2: At time Constructing unit features for each entity based on the current evolutionary dataset. ; Unit characteristics Contains physical units Location and strain field data within the historical time window, and current work step identifier; Step 3: Establish a structural dependency graph based on the topological relationships between entity units, and then combine it with unit features. Construct graph structure features; Step 4: Geometric-Strain Dual-Field Coupling Prediction: Strain field prediction employs a graph neural network, based on graph structure features for prediction. The system continuously adapts to the changing field and updates the evolutionary dataset. Geometric field prediction employs a learned prediction network, based on location field data within historical time windows, current step identifiers, and predictions. Respond to changing situations at all times and make predictions The position field at each time point is updated in the evolutionary dataset; Step 5: When switching work steps, determine whether the mesh division of the entity unit before and after the switch is consistent; if so, the next work step inherits the evolution dataset of the previous work step as the historical time window data; otherwise, map the evolution dataset of the previous work step onto the entity unit of the next work step to construct the historical time window data; repeat steps 2 to 4.

[0013] Preferably, in step 1, the collected data at least covers the field variable simulation data of each entity unit in the initial stage of the first work step: The collected data includes field variable simulation data corresponding to the initial stage of each step; the initial stage of the first step contains M time points, where M is a pre-set historical time window length used to predict subsequent field variables; the initial stages of non-first steps contain less than or equal to M time points.

[0014] Preferably, in step 2, the constructed unit features Contains physical units The corresponding M consecutive historical moments include the position, position increment, equivalent plastic strain, and equivalent plastic strain increment, as well as the current step identifier; M is the length of the historical time window.

[0015] Preferably, in step 3, when constructing the graph structure features, node weights and edge weights are calculated based on node states and differences between adjacent nodes; the node states are the unit features of entity units. ; The node weights are determined as follows:

[0016] The edge weights are determined as follows:

[0017] in, For physical units The node weight of the corresponding node, For the Sigmoid function, and The learnable parameter matrix for calculating node weights; For physical units and physical unit Edge weights between them For physical units The unit characteristics, For smooth activation function, and This is the learnable parameter matrix for calculating edge weights.

[0018] Preferably, step 4, the geometry-strain dual-field coupling prediction, specifically includes: Strain field prediction employs a graph neural network, taking the graph structure features as input and outputting... Predicted equivalent plastic strain increment at time t And combined with the current evolutionary dataset The equivalent plastic strain at time t is calculated. Predicted equivalent plastic strain at time 1 ;Will Predicted equivalent plastic strain at time 1 Update to the evolutionary dataset; Geometric field prediction is achieved using a learned prediction network. The input features are obtained from the location and location increment of M consecutive historical time points, the current step identifier, and strain field prediction. Predicted equivalent plastic strain at time 1 Composition, output Predicted position increment at time and combined with evolutionary data Position at time, calculation Predicted location value at time ;Will Predicted location value at time Update the evolutionary dataset.

[0019] Preferably, the geometric field prediction further includes a heuristic extrapolation prediction step and a fusion step: The heuristic extrapolation prediction step includes: Based on historical location, kinematic extrapolation is used to obtain the extrapolated value of location increment. ; Based on strain field prediction Predicted equivalent plastic strain at time 1 and in the current evolutionary dataset Equivalent plastic strain at time ,calculate Equivalent plastic strain increment at time ; Obtain the 95th percentile of the equivalent plastic strain of all solid elements at the current moment, and use it as the global reference strain benchmark. ; Calculate the position increment scaling correction factor To correct for positional changes caused by deformation:

[0020] in, Represents the hyperbolic tangent function; Using position increment scaling correction factor Correct the extrapolated value of the position increment. Obtain the heuristic extrapolation of the position increment prediction value ; The fusion steps are as follows: Heuristic extrapolation of position increment prediction values Location increment predictions output by the learning prediction network Weighted summaries are used to obtain the fused location increment predictions; these fused location increment predictions are then added to the evolutionary dataset. Location at any given time, obtain Predicted location value at time .

[0021] Preferably, in the heuristic extrapolation step, the predicted position increment value obtained by heuristic extrapolation is obtained. Then, further position cropping is performed: Extract data from the two most recent historical moments from the evolutionary dataset; Determine the maximum position value of all entity elements. and position minimum value ; Determine the 99th quantile of the absolute value of the position increment ; Using the 99th percentile The multiple determines the boundary margin ; The clipping boundary is defined as: the lowest boundary. highest boundary ; When the heuristic extrapolation of the position increment prediction value When the value exceeds the corresponding clipping boundary, the predicted position increment value of the heuristic extrapolation beyond the clipping boundary will be used. Clip to the corresponding boundary value.

[0022] Preferably, in the strain field prediction process, the predicted value of the equivalent plastic strain increment is obtained. Afterwards, if If the value is less than or equal to 0, then the value in the current evolutionary dataset is used. The equivalent plastic strain at time t is used as Predicted equivalent plastic strain at time 1 To achieve the constraint that the equivalent plastic strain prediction value increases monotonically; if If it is greater than 0, then Superimposed on the current evolutionary dataset On the equivalent plastic strain at time, obtain Predicted equivalent plastic strain at time 1 .

[0023] Preferably, in step 5, mapping the evolution dataset data from the previous step to the entity unit of the next step involves: At the moment of the next step Extract field variable simulation data for historical time windows from the evolution dataset of this step; if some or all of the historical time windows are missing from the current evolution dataset, extract the field variable simulation data with missing historical moments and heterogeneous grids from the tail of the evolution dataset of the previous step. The extracted field variable simulation data with missing historical moments and heterogeneous meshes are mapped onto the solid element mesh of this step. The mapping method is as follows: for the missing historical moment m, search the field variable simulation data of the extracted historical moment m for distances from the solid elements in this step. The K nearest heterogeneous meshes are selected, where K is a set positive integer. Mapping weights are determined based on mesh distance; the closer the meshes, the greater the mapping weight. The field variable simulation data from these K heterogeneous meshes are weighted and fused using these mapping weights to form the solid element corresponding to time m in this step's historical data. The field variable simulation data is used to complete the cross-step data mapping.

[0024] Preferably, when training the graph neural network, the total loss is a weighted Huber loss; The weighted Huber loss uses the Huber loss as the base loss, and for high-strain regions where the strain field is greater than a threshold, an importance weight is introduced for each entity unit within the region; entity unit Importance weight for: , For adjustment coefficients, This represents the true value of the strain field; The strain field is the equivalent plastic strain or the equivalent plastic strain increment.

[0025] Beneficial effects: (1) This invention designs a cross-step variable mesh historical state migration scheme: Most existing proxy models are limited to single-step, fixed mesh modeling, which cannot adapt to the common working conditions of multi-step, variable mesh in engineering simulation. After the step switching, the historical damage time sequence information is directly broken, which is a common bottleneck in the industry that has not been solved for a long time. This invention designs a multi-step time sequence state continuation mechanism. For the two scenarios of consistent mesh and different mesh after the step switching, the direct state inheritance and cross-step similar mesh mapping strategies are adopted respectively to completely fill in the cross-step historical time window data, truly realize multi-stage continuous damage evolution prediction, and break through the inherent limitation of traditional proxy models that can only predict independently in a single segment.

[0026] (2) This invention provides a geometry-strain dual-field coupling prediction scheme: Existing data-driven prediction methods generally use a single-field independent fitting method for prediction, which completely decouples geometric deformation and damage strain, ignoring the physical nature of their linkage, resulting in long-term prediction distortion. This invention innovatively constructs a geometry-strain dual-field coupling prediction scheme, adding a geometric evolution branch for geometric field prediction. The data basis for this geometric branch prediction incorporates the strain field prediction results, that is, it considers the position change caused by deformation, thus increasing the accuracy of position prediction.

[0027] (3) In a preferred scheme, the geometric field prediction step adds a heuristic extrapolation prediction branch; the position increment extrapolation value is solved based on the kinematic laws, and the position increment scaling correction coefficient is constructed in combination with the strain prediction. The extrapolation result is adaptively corrected, and finally the corrected heuristic extrapolation value and the prediction value output by the learning prediction network are merged.

[0028] The position increment scaling correction factor is based on the plastic strain increment and can adaptively adjust the position increment amplitude according to the degree of local deformation of the structure. When the plastic deformation of the element intensifies, the displacement increment is automatically amplified to match the physical characteristic of increased node offset under large deformation; when the deformation tends to level off or springback occurs, the displacement increment is appropriately narrowed to suppress prediction deviation. This effectively compensates for the shortcomings of pure kinematic extrapolation, which only considers rigid body motion and ignores plastic deformation, so that the position prediction results are highly consistent with the actual deformation law of the structure.

[0029] Simultaneously, the extrapolated and predicted values ​​are integrated, taking into account both the physical continuity of structural motion and the nonlinear fitting capability of the data-driven model, effectively improving the accuracy and stability of geometric position prediction under large deformation conditions.

[0030] (4) In one preferred scheme, an adaptive weighted key area enhancement modeling scheme was designed: Traditional graph neural networks use uniform weighting and undifferentiated information propagation, which cannot adapt to the characteristics of localized damage concentration and gradient abrupt changes under impact and penetration conditions, resulting in poor accuracy in key areas and serious interference from redundant information. This invention designs a learnable adaptive weighting mechanism, which dynamically adjusts node weights and edge weights based on node states and differences between adjacent nodes, accurately enhancing the features of high-damage and high-gradient core areas; combined with the weighted loss function in another preferred scheme, it focuses on learning key failure areas, realizing differentiated modeling and effectively solving the problem of insufficient prediction accuracy in key areas.

[0031] (5) In a preferred embodiment, a multi-dimensional physical constraint mechanism is designed: Pure data-driven models have no physical rule constraints, which are prone to abnormal results, resulting in unusable prediction results in engineering. This invention adds mechanisms such as strain non-negative correction and position clipping to ensure the physical acceptability and temporal stability of the prediction results, and solves the problems of missing pure data fitting mechanisms and result distortion.

[0032] (6) In the process of strain field prediction, the present invention does not directly predict the strain field. Instead of calculating the absolute strain at any given moment, we predict the strain increment and then combine it with historical absolute strain data. The absolute strain at any given moment perfectly matches the physical nature of the gradual accumulation of plastic deformation; and it is easy to control the non-negative constraint of the strain increment, ensuring the monotonic increase of the predicted absolute strain. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall process of the multi-step continuous prediction method based on cross-step state transition and dual-field coupling of the present invention.

[0034] Figure 2 This is a schematic diagram of the geometric field dual-channel prediction of the present invention.

[0035] Figure 3 This is a schematic diagram of the direct inheritance of historical states across work steps and the inverse distance weighted mapping in this invention. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0037] This invention is applicable to the rapid prediction of complex nonlinear responses such as impact, explosion, collision, penetration, and collapse. Taking a two-stage penetration and explosion process as an example, the multiple steps refer to penetration, explosion, secondary penetration, and secondary explosion.

[0038] Figure 1This is a multi-step continuous prediction method based on cross-step state transition and dual-field coupling in the embodiments of the present invention, which is used for structural damage evolution analysis of multiple steps; the method includes the following steps: Step 1: Collect structural damage evolution process data through dynamic simulation, the collected data at least covers the field variable simulation data of solid elements in the initial stage of the first step, and store the data into an evolution dataset.

[0039] In this step, LS-DYNA can be used to implement dynamic simulation, construct an explicit kinematics model of structural damage evolution, simulate the multi-step structural damage evolution process, and output the d3plot file of each step. The d3plot file records the structural damage evolution process data corresponding to the current step. The structural damage evolution process data needs to cover the position field data and strain field data of each solid element. The position field data includes the position of the solid element (centroid coordinates are used in this embodiment) and the position increment (the position variation between two adjacent moments), and the strain field data includes the equivalent plastic strain and equivalent plastic strain increment of the solid element (the plastic strain variation between two adjacent moments). If there is no increment data in the d3plot file, it can also be calculated and obtained using the position data sequence / strain data sequence recorded in the d3plot file.

[0040] In order to reduce the data calculation amount of dynamic simulation, only the field variable simulation data in the initial stage of the step is collected. The field variable evolution data at the subsequent time of the step is obtained through prediction.

[0041] The determination method of the initial stage is: For the first step: in order to construct historical time window data for prediction, it is necessary to collect a sufficient number of field variable simulation data according to the historical time window length M. Therefore, the initial stage of the first step includes M moments.

[0042] For subsequent steps: fewer (<M) field variable simulation data can be collected, and the missing historical moment data is provided by the prediction result of the previous step to complete the historical time window length. Therefore, the number of moments included in the initial stage of non-first steps can be less than or equal to M. Of course, it is also possible to not simulate and collect any field variable simulation data of subsequent steps, and the historical time window data required for prediction completely depends on cross-step data.

[0043] The field variable simulation data collected in this Step 1 is stored in the evolution dataset. This evolution dataset is continuously updated as the prediction time advances. After the dynamic simulation of Step 1 is executed, the evolution dataset only contains the collected data; when the prediction process of Steps 2 to 4 is started to obtain predicted field variables, the predicted field variables are added to this evolution dataset. The data of the historical time window is extracted from this evolution dataset.

[0044] This step may also include filtering out non-structural domain elements, retaining only structural entity elements that participate in load-bearing or damage evolution. To balance accuracy and efficiency, the target entity element set can be obtained preferentially based on part identification, part title, or automatic structural domain identification rules.

[0045] Step 2: Based on the current evolutionary dataset, construct the unit features of each entity unit. .

[0046] In this step, at time Constructing unit features for each entity based on the current evolutionary dataset. This unit feature Contains physical units The historical time window contains location and strain field data, as well as the current step identifier. In one example, the historical time window length can be selected as 4.

[0047] In this preferred embodiment, the position field and strain field include not only absolute quantities but also increments, thus constructing the element features. Contains physical units The corresponding M consecutive historical moments, including their positions, position increments, equivalent plastic strains, and equivalent plastic strain increments, as well as the current step identifier, are concatenated to form the element features. , is represented as:

[0048] in, Indicates splicing, and Representing solid units respectively At any moment Position and position increment, and Representing solid units respectively At any moment The equivalent plastic strain and the equivalent plastic strain increment, For a moment The process step identification. This unit feature construction method can effectively characterize local geometric changes and material state accumulation.

[0049] The unit features are normalized for each dimension component using mean-variance standardization.

[0050] Step 3: Establish a structural dependency graph based on the topological relationships between entity units, and then combine it with unit features. Construct graph structure features.

[0051] In this step, a unit-level structural dependency graph is first constructed based on the geometric / physical relationships of the structural object to be predicted. Each node in the diagram For a given entity unit, the connections between nodes form edges. Edges can be generated by identity mappings, spatial nearest neighbor relationships, or topological adjacency relationships. For example, for large-scale grids, identity graphs can be used for block-based reasoning; for scenarios requiring explicit neighborhood interactions, a K-nearest neighbor-based graph construction method can be used. In specific implementations, a neighborhood search algorithm can be used to select a specified number (e.g., 10) of the closest neighbors for each node, and the adjacency relationships between nodes can be represented as edge index tensors. Used to describe nodes i With nodes j The connection relationship between them.

[0052] Next, the element features of the solid element are used. As the node states in the structural dependency graph, the node weights and edge weights are calculated to form the graph structure features.

[0053] In this embodiment, node weights and edge weights are calculated based on node states and differences between adjacent nodes: The node weights are determined as follows:

[0054] The edge weights are determined as follows:

[0055] in, For physical units The node weight of the corresponding node, For the Sigmoid function, and The learnable parameter matrix for calculating node weights; For physical units and physical unit Edge weights between them For physical units The unit characteristics, Smooth activation function , As the independent variable; and This is the learnable parameter matrix for calculating edge weights.

[0056] As can be seen, this embodiment uses the unit state features and inter-unit feature differences as inputs, and solves for node weights and edge weights through a learnable linear transformation combined with a nonlinear activation function. For key units with high strain, high gradient, and local damage, their feature inputs can obtain larger node weights after mapping, strengthening the unit's own feature expression; for key neighborhood combinations with significant state differences, the corresponding edge weights are increased simultaneously, enhancing the information transmission strength between neighborhoods. Meanwhile, the weights of conventional units and neighborhoods with gentle deformation and small state differences are correspondingly suppressed. By calculating node weights and edge weights in this way, key units and key neighborhoods receive higher weights and are given priority consideration in feature propagation and aggregation, thus highlighting the contributions of key units and key neighborhoods.

[0057] Step 4: Perform geometry-strain dual-field coupling prediction: including strain field prediction and geometry field prediction.

[0058] (1) Strain field prediction Strain field prediction employs a graph neural network, based on graph structure features for prediction. The strain field is constantly adjusted and updated in the evolutionary dataset.

[0059] In a specific implementation, the graph neural network can be a residual graph neural network. For example, a residual graph neural network can be composed of multiple stacked graph convolutional modules, configured with layer normalization, SiLU activation, random deactivation, and residual connections to improve the stability of deep propagation.

[0060] The graph structure features enhanced by the key units and key neighborhoods in step 3 are input into the residual graph neural network, and the output is... Predicted equivalent plastic strain increment at time t And combined with the current evolutionary dataset The equivalent plastic strain at time t is calculated. Predicted equivalent plastic strain at time 1 .here, The equivalent plastic strain at time point is obtained from historical data of the previous time point, derived from the evolution dataset. Depending on the time point, it may be dynamic simulation data or predicted data from historical time points. The predicted equivalent plastic strain value is then... Update the evolutionary dataset.

[0061] In this invention, there is no direct prediction. Instead of calculating the absolute strain at any given moment, we predict the strain increment and then combine it with historical absolute strain data. The absolute strain at any moment perfectly matches the physical nature of the gradual accumulation of plastic deformation.

[0062] Furthermore, predicting strain increments makes it easier to impose non-negativity constraints on the strain increments, ensuring a monotonically increasing predicted absolute strain and avoiding spurious growth or oscillations. Specifically, this applies to the prediction results of graph neural networks. Apply physical constraints to construct the predicted absolute strain value for the next time step. for:

[0063] in, This indicates taking the larger value, when When less than or equal to 0, through calculate, That is, to maintain Equivalent plastic strain value at time ;when When greater than 0, through calculate, for Superimposed .therefore This is used to maintain the monotonically increasing absolute strain prediction. Furthermore, considering... This is network output data, which is a normalized value. Here, we perform scale recovery to restore it to a physical strain value. This represents the scale recovery mapping.

[0064] Finally Add it to the evolutionary dataset and use it as part of the above formula in the prediction of the next time step. .

[0065] (2) Geometric field prediction Geometric field prediction employs a learned prediction network, based on location field data within historical time windows, current work step identifiers, and strain field predictions, to predict the strain field. The position field at time t is then updated in the evolutionary dataset.

[0066] In one specific implementation scheme, the input features of the learning-based prediction network are obtained from the position and position increment of M consecutive historical time points, the current step identifier, and strain field prediction. Predicted equivalent plastic strain at time 1 Composition; Learning-based prediction network output Predicted position increment at time and combined with evolutionary data Position at time, calculation Predicted location value at time ; predict the location value Update the evolutionary dataset.

[0067] In such Figure 2In the preferred embodiment shown, the geometric field prediction includes a data-driven prediction channel and a heuristic extrapolation prediction channel. The two channels obtain position predictions separately, and their fusion yields a more stable geometric update result. The working process and fusion method of the two channels are described in detail below.

[0068] (2.1) Data-driven prediction channel A data-driven prediction channel implements the prediction scheme of the aforementioned learning-based prediction network. In one specific implementation, the learning-based prediction network can employ a lightweight MLP, which uses training data to learn the evolution law of the displacement field during the training phase. Specifically, this MLP has 4 hidden layers, 256 hidden units, and the activation function is... The loss function used is Huber loss. In the dual-prediction-channel scheme, the learned prediction network outputs the position increment prediction value. After that, do not calculate directly. Predicted location value at time Instead, it merges the prediction results with the heuristic extrapolation prediction channel to form a new position increment prediction value, and then calculates the position prediction value.

[0069] (2.2) Heuristic extrapolation prediction channel The heuristic extrapolation prediction channel, based on physical intuition and simple geometric rules, calculates the velocity and acceleration at historical moments to predict the position increment caused by motion. Based on the predicted strain increment, the position increment is adjusted to incorporate the position change caused by predicted deformation. By combining the minimum / maximum boundaries calculated from historical positions, the predicted new position is clipped within these boundaries to prevent nodes from flying out of the reasonable area.

[0070] The specific implementation process of heuristic extrapolation prediction includes the following steps: ① Kinematic extrapolation: Based on historical location, kinematic extrapolation is used to obtain the extrapolated value of the location increment. ; First, based on three adjacent times Location , and Calculate the first velocity characterization quantity Second velocity characterization quantity For simplicity, subscripts are not used here. .

[0071]

[0072]

[0073] Next, the position increment is extrapolated to obtain the extrapolated position increment value. :

[0074] The first term represents velocity, the second term represents average velocity, and the third term represents acceleration. (Solid element) of This is denoted as the extrapolated value of the position increment. .

[0075] ② Calculate the position increment scaling correction factor to correct the position change caused by deformation.

[0076] First, based on strain field prediction... Predicted equivalent plastic strain at time 1 and in the current evolutionary dataset Equivalent plastic strain at time ,calculate Equivalent plastic strain increment at time .

[0077]

[0078] Next, the 95th percentile of the equivalent plastic strain of all solid elements at the current moment is obtained as the global reference strain benchmark. .

[0079]

[0080] in, This indicates taking the larger value; This represents the 95th percentile of the equivalent plastic strain of all physical elements at the current moment, reflecting the strain level of the vast majority of physical elements in the data at the current moment; To prevent the occurrence of a minimum value.

[0081] Then, calculate the position increment scaling correction factor. :

[0082] in, This represents the hyperbolic tangent function.

[0083] ③ Utilize position increment scaling correction factor Correcting the extrapolated value of the position increment Obtain the heuristic extrapolation of the position increment prediction value :

[0084] ④ Position trimming: The purpose of position clipping is to prevent physical quantities from going out of bounds, and it specifically includes the following steps: Extract the two most recent historical time points from the evolutionary dataset; based on the extracted historical time points, determine the maximum position values ​​of all entity units. and position minimum value And determine the 99th quantile of the absolute value of the position increment. ; Using the 99th percentile The multiple determines the boundary margin For example, if 2.5 is chosen as the multiple, then the boundary margin... ; The final clipping boundary is determined as: the lowest boundary. highest boundary ; When the heuristic extrapolation of the position increment prediction value When located outside the final clipping boundary, the clipping boundary is taken as... The value of is used to trim the predicted position increments that exceed the boundary to the corresponding boundary value.

[0085] (2.3) Fusion of two prediction channels The data-driven prediction channel position increment prediction value is expressed as: The predicted position increment of the channel is expressed as follows: The two values ​​are weighted to obtain the fused position increment prediction value. Different weight values ​​can be used for different work steps to make the prediction of different work steps more targeted.

[0086] Add the fused location increment predictions to the evolutionary dataset. Location at any given time, obtain Predicted location value at time .

[0087] Finally, the predicted strain field and geometric field are fed back to the evolution dataset to achieve continuous time-series response extrapolation across multiple work steps.

[0088] Step 5: Cross-step state transition.

[0089] like Figure 3 As shown, when switching work steps, it is determined whether the mesh division of the entity unit before and after the switch is consistent; if so, the next work step inherits the evolution dataset of the previous work step as the historical time window data; otherwise, the evolution dataset of the previous work step is mapped onto the entity unit of the next work step to construct the historical time window data; repeat steps 2 to 4.

[0090] Specifically, at the step switching time, it is determined whether the grid at the end of the previous step and the initial grid of the next step are consistent. The method to determine whether the grid is consistent can be to determine whether the position is consistent or within the tolerance range. If consistent, the historical state can be directly copied. If inconsistent, inverse distance weighted (IDW) mapping is used to map the data in the historical time window of the previous step to the entity unit of the next step, so as to complete the historical window.

[0091] Specifically, at the moment of the next step Extract field variable simulation data for historical time windows from the evolution dataset of this step; if some or all of the historical time windows are missing from the current evolution dataset, extract the field variable simulation data with missing historical moments and heterogeneous grids from the tail of the evolution dataset of the previous step. The extracted field variable simulation data with missing historical moments and heterogeneous meshes are mapped onto the solid element mesh of this step. The mapping method is as follows: for the missing historical moment m, search the field variable simulation data of the extracted historical moment m for distances from the solid elements in this step. The K nearest heterogeneous meshes of the grid, where K is a set value, for example, K=8.

[0092] The mapping weight is determined based on the grid distance; the closer the distance, the greater the mapping weight. The mapping weight is expressed as follows:

[0093] in, For the solid unit in this step The grid position, For the K heterogeneous meshes, the th The location of the nearest neighboring entity cell grid; The distance decay exponent, To prevent tiny constants with a denominator of zero.

[0094] The field variable simulation data of K heterogeneous meshes are weighted and fused using mapping weights to serve as the solid element corresponding to the historical time m of this step. The field variable simulation data is used to complete the cross-step data mapping.

[0095] The weighted fusion is expressed as follows:

[0096] In the above formula, Represents the Kth heterogeneous mesh. Equivalent plastic strain of each grid. To predict the solid unit of this step using cross-step data The equivalent plastic strain.

[0097] Using the above method, the strain history memory learned in the previous step can be smoothly transferred to the next step, thereby avoiding the temporal fragmentation problem caused by completely independent modeling of each step. For multi-step continuous prediction scenarios, this invention allows multiple steps, such as step 1, step 2, and step 3, to be chained together to gradually propagate historical states.

[0098] Training and Deployment: During the training phase, the loss function of the graph neural network model is designed to balance the model's fitting accuracy to most nodes with its sensitivity to high-strain regions, while also enhancing its robustness to outliers. The entire loss function is calculated on the normalized target value space, using a weighted Huber loss as the total loss. This weighted Huber loss uses the Huber loss as the base loss, introduces importance weights to each entity unit within the high-strain region, and guides the model to prioritize high-strain regions through sample weights.

[0099] Specifically, adopt This represents the predicted value of the equivalent plastic strain or equivalent plastic strain increment output by the graph neural network model. express The corresponding true values ​​are derived from the complete structural damage evolution data from LS-DYNA simulations.

[0100] First, calculate the basic Huber loss. Its definition is:

[0101] in, This controls the threshold for transitioning from squared loss to linear loss. This loss form uses a squared penalty for small errors and a linear penalty for large errors, thereby reducing the impact of outliers on the gradient.

[0102] To highlight the importance of the high-strain region, a weight is introduced for each node. ,in To adjust the coefficients, nodes with larger target values ​​receive higher loss weights. Ultimately, for nodes containing... The total loss is defined as the average of the weighted losses of all nodes.

[0103] This loss function combines the robustness of Huber loss with the physical focusing characteristics of the weighted strategy, enabling the model to stably learn the evolution of the strain field in LS-DYNA simulations.

[0104] During the prediction phase, the time length (number of frames) of the model's autoregressive prediction is determined by the total number of frames in the training data for the corresponding work step. Specifically, the training data for each work step comes from the d3plot file sequence output by LS-DYNA. Since the training data is generated with the physical parameters (such as the equivalent plastic strain of concrete) ensuring that the time series covers the complete process from initial loading to a relatively stable state, the time window range of the prediction phase naturally inherits this length, i.e., starting from a known frame, predicting all subsequent frames until the last frame. This design allows the model to perform frame-by-frame extrapolation without needing to specify an additional prediction end time, ensuring consistency between the prediction results and the temporal logic of the simulation process. Simultaneously, the stable nature of the physical parameters in the training data provides reasonable boundary conditions for the model, preventing non-physical divergence behavior during long-term extrapolation.

[0105] The specific embodiments described above only illustrate the design principles of the present invention. The shapes and names of the components in this description may differ and are not limited. Therefore, those skilled in the art can modify or make equivalent substitutions to the technical solutions described in the foregoing embodiments; and such modifications and substitutions do not depart from the inventive spirit and technical solutions of the present invention, and should all fall within the protection scope of the present invention.

Claims

1. A multi-step continuous prediction method based on cross-step state transition and dual-field coupling, characterized in that, This method is used for multi-step structural damage evolution analysis and includes: Step 1: Collect data on the structural damage evolution process through dynamic simulation. The collected data should at least cover the field variable simulation data of each entity unit in the initial stage of the first step and be stored in the evolution dataset. Step 2: At time Constructing unit features for each entity based on the current evolutionary dataset. ; Unit characteristics Contains physical units Location and strain field data within the historical time window, and current work step identifier; Step 3: Establishing structure dependency graph based on topological relationship between entity units, and then combining unit characteristics Constructing graph structure features; Step 4: Geometric-Strain Dual-Field Coupling Prediction: Strain field prediction adopts graph neural network, based on graph structure features prediction Strain field at the moment, and update to the evolution dataset; The geometric field prediction adopts a learning prediction network, and is based on position field data in a historical time window, a current work step identifier and a predicted moment strain field to predict a position field at a moment and update to an evolution data set; Step 5: When switching work steps, determine whether the mesh division of the entity unit before and after the switch is consistent; if so, the next work step inherits the evolution dataset of the previous work step as the historical time window data; otherwise, map the evolution dataset of the previous work step onto the entity unit of the next work step to construct the historical time window data; repeat steps 2 to 4.

2. The multi-pass continuous prediction method based on cross-pass state migration and dual field coupling of claim 1, wherein, In step 1, the collected data at least covers the field variable simulation data of each entity unit in the initial stage of the first work step: The collected data includes field variable simulation data corresponding to the initial stage of each step; the initial stage of the first step contains M time points, where M is a pre-set historical time window length used to predict subsequent field variables; the initial stages of non-first steps contain less than or equal to M time points.

3. The multi-pass continuous prediction method based on cross-pass state migration and dual field coupling of claim 1, wherein, In step 2, the unit features constructed comprising physical units the positions, position increments, equivalent plastic strains and equivalent plastic strain increments of the M consecutive historical time instants, and the current operation step identification; M is the length of the historical time window.

4. The multi-pass continuous prediction method based on cross-pass state migration and dual field coupling of claim 1, wherein, In the step 3, when constructing the graph structure feature, the node weight and the edge weight are calculated according to the node state and the difference between adjacent nodes; the node state is the cell feature of the entity cell ​ The node weights are determined as follows: The edge weights are determined as follows: in, For physical units The node weight of the corresponding node, For the Sigmoid function, and The learnable parameter matrix for calculating node weights; For physical units and physical unit Edge weights between them For physical units The unit characteristics, For smooth activation function, and This is the learnable parameter matrix for calculating edge weights.

5. The method of claim 1, wherein the method is based on cross-step state migration and double field coupling. Step 4, the geometry-strain dual-field coupling prediction, specifically includes: Strain field prediction employs a graph neural network, taking the graph structure features as input and outputting... Predicted equivalent plastic strain increment at time t And combined with the current evolutionary dataset The equivalent plastic strain at time t is calculated. Predicted equivalent plastic strain at time 1 ;Will Predicted equivalent plastic strain at time 1 Update to the evolutionary dataset; Geometric field prediction is achieved using a learned prediction network. The input features are obtained from the location and location increment of M consecutive historical time points, the current step identifier, and strain field prediction. Predicted equivalent plastic strain at time 1 Composition, output Predicted position increment at time and combined with evolutionary data Position at time, calculation Predicted location value at time ;Will Predicted location value at time Update the evolutionary dataset.

6. The multi-pass continuous prediction method based on cross-pass state migration and dual field coupling of claim 5, wherein, The geometric field prediction further includes a heuristic extrapolation prediction step and a fusion step: The heuristic extrapolation prediction step includes: Based on the historical position, a position delta extrapolation value is obtained using kinematic extrapolation ; based on the strain field prediction the equivalent plastic strain prediction value at the time instant and the current evolution dataset the equivalent plastic strain at the time instant , the increment of the equivalent plastic strain at the time instant is calculated ​ Obtain the 95% quantile of the equivalent plastic strain of all entity elements at the current time as the global reference strain benchmark ; Computing position increment scaling correction coefficients to correct for position changes due to deformation: wherein denotes the hyperbolic tangent function; Utilizing position delta scaling correction factor , correcting the position delta extrapolation value , obtaining a heuristic extrapolated position delta prediction value ; The fusion steps are as follows: The position delta prediction value of the heuristic extrapolation The position delta prediction value output by the learning prediction network The fused position delta prediction value is obtained by weighting The position at the time instant is obtained by adding the fused position delta prediction value to the position in the evolution data set The position prediction value at the time instant .

7. The multi-pass continuous prediction method based on cross-pass state migration and dual field coupling of claim 6, wherein, In the heuristic extrapolation step, a position increment prediction value for the heuristic extrapolation is obtained After, further position pruning is performed: Extract data from the two most recent historical moments from the evolutionary dataset; determining a position maximum value for all entity units and a position minimum value ; Determining a 99th percentile of absolute values of position increments ; Utilizing the 99th percentile of the multiples to determine the margin of error ; determining the cropping boundary as: the lowest boundary , the highest boundary ; when the heuristic extrapolated position delta prediction value exceeds the corresponding clipping boundary, clip the heuristic extrapolated position delta prediction value to the corresponding boundary value.

8. The multi-pass continuous prediction method based on cross-pass state migration and dual field coupling of claim 5, wherein, In the strain field prediction process, the equivalent plastic strain increment prediction value is obtained. Afterwards, if If the value is less than or equal to 0, then the value in the current evolutionary dataset is used. The equivalent plastic strain at time t is used as Predicted equivalent plastic strain at time 1 To achieve the constraint that the equivalent plastic strain prediction value increases monotonically; if If it is greater than 0, then Superimposed on the current evolutionary dataset On the equivalent plastic strain at time, obtain Predicted equivalent plastic strain at time 1 .

9. The multi-pass continuous prediction method based on cross-pass state migration and dual field coupling of claim 1, wherein, In step 5, mapping the evolution dataset data from the previous step to the entity unit of the next step involves: at the time of the latter step extracting the simulation data of the field variables of the historical time window from the evolution dataset of the present step; if the data of the historical time window is missing in part or in whole from the evolution dataset of the present step, extracting the simulation data of the field variables of the missing historical time from the end of the evolution dataset of the previous step; The extracted grid heterogeneous field variable simulation data at the missing historical moment is mapped to the entity unit grid of the present step, and the mapping manner is that: for the missing historical moment m, in the extracted field variable simulation data at the historical moment m, K heterogeneous grids closest to the grid of the entity unit in the present step are searched, and K is a set positive integer. ​ The mapping weights are determined based on the grid distance; the closer the grid, the greater the mapping weight. The simulation data of field variables from K heterogeneous grids are then weighted and fused using these mapping weights to form the solid element corresponding to time m in this step of the process. The field variable simulation data is used to complete the cross-step data mapping.

10. The multi-pass continuous prediction method based on cross-pass state migration and dual field coupling of claim 1, wherein, When training the graph neural network, the total loss is a weighted Huber loss. The weighted Huber loss uses the Huber loss as the base loss; for high-strain regions where the strain field is greater than the threshold, an importance weight is introduced for each entity unit within the region; entity unit importance weight is: , is a modulation coefficient, is the true value of the strain field; The strain field is the equivalent plastic strain or the equivalent plastic strain increment.

Images