A deep learning-based structural multi-working-condition collapse response prediction method and system

Abstract

The application discloses a structure multi-working condition collapse response prediction method and system based on deep learning, and relates to the technical field of civil engineering structure continuous collapse resistance analysis and intelligent prediction. The method comprises the following steps: obtaining structure parameters and failure working conditions; calculating a physical prior response curve based on an analytical model; discretizing the structure into a graph structure and dynamically constructing a failure perception graph according to a failure position; extracting topological features by using a graph neural network to obtain a graph-level representation; constructing displacement sampling points into a sequence vector and fusing the graph-level representation into each sequence position through a gating mechanism; and finally, outputting a complete response curve through a Transformer encoder. The application significantly improves the accuracy and generalization ability of structure response prediction under different column failure working conditions through the collaborative fusion of physical prior, graph topology modeling and sequence modeling, and is suitable for engineering rapid evaluation.

Description

Technical Field

[0001] This application relates to the field of progressive collapse analysis and intelligent prediction technology for civil engineering structures, and in particular to a method and system for predicting the collapse response of structures under multiple working conditions based on deep learning. Background Technology

[0002] Progressive collapse of a structure refers to the failure of a local load-bearing component due to a random event, forcing the load above the failed area to redistribute through the remaining structure, leading to the sequential failure of adjacent components and ultimately causing a large-scale collapse or even complete collapse of the entire structure. Analysis of progressive collapse induced by column failure is an important research topic in structural engineering. In practical engineering assessments, it is usually necessary to repeatedly conduct analyses under multiple failure locations, boundary conditions, and floor participation modes. If each sub-condition relies on high-cost finite element analysis, the efficiency of assessing the structure's resistance to progressive collapse will be severely reduced.

[0003] Existing analytical model analysis methods are typically established for specific failure locations or force transmission mechanisms. Based on structural geometric parameters, material constitutive models, and boundary conditions, and through membrane effects or cable action, they derive the functional relationship between resistance and deformation, obtaining the resistance curve after local failure, which has good mechanical interpretability. However, the force transmission paths and boundary constraints corresponding to different failure locations vary significantly. When facing complex floor participation modes and nonlinear material degradation, the accuracy and applicability of existing analytical model analysis methods are greatly limited, making it difficult to meet the needs of rapid assessment under multiple working conditions in engineering. Summary of the Invention

[0004] Therefore, it is necessary to provide a deep learning-based method for predicting structural collapse response under multiple operating conditions to address the aforementioned technical problems.

[0005] The following technical solution is adopted in this specification: This specification provides a deep learning-based method for predicting structural collapse response under multiple operating conditions, including: Obtain the physical parameters and failure condition parameters of the target structure; the failure condition parameters include the failure location category and floor participation mode; based on the physical parameters and failure condition parameters, obtain the physical prior resistance values ​​of each preset displacement sampling point of the target structure; Each floor of the target structure is represented by a floor subgraph. The nodes in each floor subgraph include column nodes and plate element nodes within that floor. The edges in each floor subgraph include horizontal adjacent connections between column nodes within that floor and connections between plate element nodes and their corner column nodes. Through the inter-floor connection relationships between adjacent floors, the floor subgraphs are stitched together into a global graph structure and represented by an adjacency matrix. The failure column node is determined according to the failure location category. The failure column node and the edge associated with the failure column node are shielded from the adjacency matrix. The floor subgraph that does not participate in message passing is shielded from the adjacency matrix according to the floor participation mode, so as to construct a failure-aware dynamic graph. Based on the aforementioned failure-aware dynamic graph, graph-level pooling is performed using a graph neural network to obtain a graph-level representation; Based on the physical parameters, the prior physical resistance values, and the graphical representation, the resistance prediction curve of the target structure is obtained.

[0006] Furthermore, based on the physical parameters, the prior physical resistance values, and the graphical representation, the resistance prediction curve of the target structure is obtained, including: The displacement values, the corresponding prior physical resistance values, and the physical parameters of each displacement sampling point are concatenated to form the original vector of the position of each displacement sampling point. The original vector is then mapped into a sequence representation through a shared linear mapping. The graph-level representation and the sequence representation are fused using a gating mechanism to form a joint input sequence; The joint input sequence is input into the Transformer encoder, and the long-range dependency between different displacement sampling points is modeled through the multi-head self-attention mechanism of the Transformer encoder, and the hidden state corresponding to each displacement position is output. The hidden states at each location are mapped to predicted response values ​​through the linear output layer of the Transformer encoder.

[0007] Furthermore, the graph-level representation and the sequence representation are fused through a gating mechanism to form a joint input sequence, including: The graph-level representation is linearly mapped to obtain the mapped graph vector. For each sequence position, a position-related gate vector is obtained through a gate function based on the sequence representation of that position and the mapping vector. Based on the gate vector, the mapping vector is superimposed onto the sequence representation of the sequence position by element-wise multiplication to obtain the fused sequence vector; The mapping vector is used as an independent global vector, which together with each fused sequence vector and its position code constitutes a joint input sequence.

[0008] Furthermore, based on the physical parameters and failure condition parameters, the prior physical resistance values ​​at each preset displacement sampling point of the target structure are obtained, including: Based on the physical parameters and failure condition parameters, and according to the membrane effect decomposition theory, the compressive membrane effect resistance curves in the small deformation stage and the tensile membrane effect resistance curves in the large deformation stage are obtained; wherein, the compressive membrane effect resistance is described by an expression with a degradation coefficient, and the tensile membrane effect resistance is described by an expression with a shape factor. The pressure membrane effect resistance and the tension membrane effect resistance are superimposed to generate the initial prior response trend corresponding to the failure location category. The trend is then discretized according to the preset number of displacement sampling points to obtain the sequence of physical prior resistance values ​​for each displacement sampling point of the target structure.

[0009] Furthermore, the step of performing graph-level pooling through a graph neural network to obtain a graph-level representation includes: The adjacency matrix of the failure perception graph is normalized to obtain a normalized adjacency matrix. Based on the normalized adjacency matrix, a multi-layer graph neural network is used to pass and update the node features, and during the update process, the information propagation of inactive nodes is shielded by an active node mask. Perform masked maximum pooling on the active nodes after message passing to obtain the graph-level representation.

[0010] Furthermore, the physical parameters of the target structure include: span in the X direction, span in the Y direction, beam height in the X direction, beam height in the Y direction, beam width in the X direction, beam width in the Y direction, slab thickness, top reinforcement ratio of beam in the X direction, bottom reinforcement ratio of beam in the X direction, top reinforcement ratio of beam in the Y direction, bottom reinforcement ratio of beam in the Y direction, slab reinforcement ratio, steel bar yield strength, steel bar ultimate strength, steel bar ultimate strain, and concrete compressive strength. The failure location categories include inner column failure, side column failure, and corner column failure.

[0011] Furthermore, the resistance prediction curve of the target structure is a complete resistance displacement response curve containing 30 displacement sampling points, which are distributed at equal intervals with a step size of 0.005 starting from displacement 0.

[0012] This specification provides a deep learning-based structural multi-condition collapse response prediction system, including: The parameter acquisition module is used to acquire the physical parameters and failure condition parameters of the target structure; the failure condition parameters include the failure location category and the floor participation mode; based on the physical parameters and failure condition parameters, the physical prior resistance values ​​of each preset displacement sampling point of the target structure are obtained; The graph structure construction module is used to create a floor subgraph for each floor of the target structure. The nodes in each floor subgraph include column nodes and plate element nodes within that floor. The edges in each floor subgraph include horizontal adjacent connections between column nodes within that floor and connections between plate element nodes and their corner column nodes. Through the inter-floor connection relationships between adjacent floors, the floor subgraphs are spliced ​​into a global graph structure and represented by an adjacency matrix. The failure graph construction module is used to determine the failure column node according to the failure location category, shield the failure column node and the edge related to the failure column node from the adjacency matrix, and shield the floor subgraph that communicates with the failure column node from the adjacency matrix according to the floor participation mode, so as to construct a failure-aware dynamic graph. The graph-level representation construction module is used to obtain a graph-level representation based on the failure-aware dynamic graph by performing graph-level pooling through a graph neural network. The prediction module is used to obtain the resistance prediction curve of the target structure based on the physical parameters, the prior physical resistance value and the graphical representation.

[0013] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the multi-condition collapse response of a frame structure.

[0014] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described method for predicting the multi-condition collapse response of a frame structure.

[0015] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This presents a challenge for existing analytical models in meeting the needs of assessing various column failure conditions. The multi-condition collapse response prediction method for frame structures provided in this specification utilizes a failure-aware dynamic graph constructed based on failure location categories and floor participation modes. This method achieves unified representation and modeling of different failure location conditions. It can dynamically shield failed column nodes and their associated edges, and adaptively adjust the subgraph message passing range according to the floor participation mode. This allows the same depth model to naturally adapt to various topological change conditions without requiring separate training or model switching for each failure location, significantly reducing model management costs in engineering assessment. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0017] Figure 1This document provides a flowchart illustrating a multi-condition collapse response prediction method for frame structures. Figure 2 This document provides a schematic diagram of the experimental objects and the nine types of column failure sub-conditions defined in this manual. Figure 3 This document provides a schematic diagram of the failure perception graph representation, failure node masking, and graph message passing. Figure 4 This is a schematic diagram of the gating fusion sequence modeling provided in this specification; Figure 5 This is a schematic diagram of the hierarchical results based on MAPE provided in this specification; Figure 6 This is a schematic diagram comparing the MAPE of the five models provided in this manual in the displacement dimension. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0019] Existing machine learning methods can input structural parameters into regression models to predict structural responses. While this can improve prediction speed, these methods only use physical parameters as tabular features and the response values ​​at different displacement positions as independent output points. They cannot explicitly express the relationship between the location of the failed column and the structural connection, and it is difficult to maintain the continuous shape of the entire response curve. Furthermore, they do not make full use of the prior trends provided by existing structural mechanics analytical models, and their stability is insufficient when the sample size is limited or the working conditions are significantly different.

[0020] To address the shortcomings of existing technologies, this invention proposes a fast prediction method and system that combines physical prior curves, failure perception graph structure representation, and Transformer sequence modeling, which can solve the following technical problems: (1) Solve the problem that existing fast prediction methods cannot simultaneously utilize structural physical parameters, failure location topology information and the sequential correlation of complete response curves.

[0021] (2) Solve the problem that pure data-driven models lack structural mechanics priors and have insufficient stability in predicting the response evolution curves of complex working conditions.

[0022] (3) Solve the problem that traditional finite element or complex analytical calculations have low computational efficiency and are not convenient for rapid engineering evaluation under batch working conditions.

[0023] (4) Provide a system solution that can be integrated into engineering software, so that structural parameter input, physical prior calculation, failure diagram construction, resistance response prediction and result output can be executed automatically.

[0024] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0025] Figure 1 This is a flowchart illustrating a multi-condition collapse response prediction method for a frame structure as described in this specification, which specifically includes the following steps: S101: Obtain the physical parameters, failure location categories, boundary conditions, floor participation modes, and target response data of the target frame structure, and perform data preprocessing.

[0026] S102: Calculate the analytical response prior curve based on structural physical parameters and operating condition information.

[0027] S103: Discretize the frame structure into a graph structure containing column nodes, plate element nodes and inter-story connections, and construct a failure perception graph based on the location of the failed column and the floor participation mode.

[0028] S104: Use graph neural networks to perform message passing and graph-level pooling on the dynamic failure graph to obtain graph terms.

[0029] S105: Construct sequence vectors from the 30 displacement sampling points and fuse graph words into each sequence position through a gating mechanism.

[0030] S106: The fused sequence input is a sequence encoder based on a self-attention mechanism, and the output is a 30-point structural response curve.

[0031] S107: Performs rapid prediction based on the trained model and outputs structural resistance evolution curves and visualization results under different failure conditions.

[0032] Example 1: Two-layer 5×5 column grid frame structure.

[0033] 1. Structure Object and Input / Output Definition The target structure is a two-story 5×5 column grid frame, with each story including 25 column nodes and 16 plate element nodes, for a total of 82 drawing nodes across the two stories. The task is to obtain the physical parameters, failure location categories, boundary conditions, and floor participation modes of the target structure.

[0034] The physical parameters of the target structure include: the X-direction span of the target structure. L x Span in the Y direction L y Beam height in the X / Y directions H x / H y Beam width in the X / Y directions B x / B y , plate thickness H s Reinforcement ratio of beam top in X / Y directions tx / ty Reinforcement ratio at the bottom of beams in the X / Y directions bx / by Slab reinforcement ratio s , yield strength of steel bars fy f y Ultimate strength of steel bars f u Ultimate strain of steel reinforcement u and concrete compressive strength f c .

[0035] Based on the physical parameters, resistance response vectors at 30 displacement positions can be output.

[0036] This invention defines the prediction target for the i-th sample as the intact resistance-displacement response curve: in, For the k-th displacement sampling point, Let be the predicted response of the i-th sample at that displacement.

[0037] 2. Physics Prior Module According to the membrane effect decomposition theory, the internal work of a beam-slab structure includes two parts: the work done by the yield line bending moment and the work done by the additional bending moment caused by the axial force. Therefore, the structural resistance curve can be fully decomposed. F =[ p 1,…, p 30 Preliminary decomposition into small deformation compression film effect resistance F M_CMA Resistance to large deformation membrane effect F N_TMA .

[0038] Pressure film effect resistance F M_CMA The resistance initially increases with increasing displacement until it reaches its peak value, after which it gradually decreases with further increases in displacement. The resistance to the pressure film effect can be described by the following expression:

[0039] In the formula, F cma and Δ c These represent the bearing capacity and corresponding displacement due to the pressure film effect, and the degradation coefficient, respectively. c Used to control the rate at which the resistance to the pressure film effect decreases.

[0040] In comparison, the resistance to membrane effect F N_TMA Under small deformations, the impact on structural resistance is relatively small, but as the displacement increases, F N_TMA The contribution gradually increases. Therefore, the following expression can be used to describe the resistance to the membrane effect:

[0041] In the formula, Δ t Take 0.15 min { L x , L y}, ; This indicates the load-bearing capacity due to the membrane stretching effect.

[0042] 3. Construction of Failure Awareness Map Structure This invention maps the target structure as a graph. In the two-layer 5×5 column grid embodiment, each floor consists of 25 column nodes and 16 slab unit nodes; the column nodes are connected according to horizontal and vertical adjacency, the slab unit nodes are connected to their four corner column nodes, and the two floors are connected by corresponding column nodes.

[0043] For a given failure condition, the system identifies the failed column node based on boundary conditions and removes that node and its associated edges from the adjacency matrix; it then determines whether the second-level subgraph participates in message passing based on the floor participation mode. This yields the dynamic failure adjacency matrix. The normalized adjacency matrix is:

[0044] in, for The degree matrix, This is a unit item that is only valid for the active node. This acts as a mask for active nodes. This process prevents failed nodes and non-participating floors from propagating invalid information to subsequent network nodes.

[0045] 4. Message Transmission Suppose the graph neural network's first... Layer nodes are represented as , No. The trainable weights of the layer are The nonlinear activation function is The node update process is as follows: in, This represents element-wise multiplication; after two layers of message passing, the system performs a max-pooling layer on the active nodes to obtain a graph-level representation: This level represents Graph branches are used as graph terms in subsequent sequence modeling. Compared with using only one-hot encoding of load case labels, graph branches can express the spatial location of the failed node in the structure, the adjacent connection relationship, and the floor participation range, thus making them more suitable for load cases with complex topological relationships such as edge columns and corner columns.

[0046] 5. Gated fusion and sequence construction This invention treats the 30 displacement positions as an ordered input sequence. For the first... k Each displacement position, the original sequence consists of displacements. Physical prior values and structural physical variables The sequence is formed by splicing and then obtained through a shared linear mapping:

[0047] in, W x and b x These represent the weights and biases in the hidden layer of the neural network.

[0048] To avoid applying image information with a fixed intensity to all displacement positions, this invention employs a position-dependent and channel-dependent gating fusion mechanism. Let... Given the mapped graph words, the gated vector and the fused sequence are respectively:

[0049] in To perform element-wise multiplication, the system then retains the independent graph terms, which, together with the 30 fused sequences, form the joint input sequence: in, This is a trainable position encoding. The design allows graph structure information to participate in self-attention computation as a global structural context, and can also be injected into different displacement positions through gating.

[0050] 5.7 Transformer Response Curve Prediction The Transformer encoder models long-range dependencies between different displacement positions through multi-head self-attention. For any attention head, the Query, Key, and Value matrices are denoted as follows: , , Then the self-attention output is:

[0051] After passing through multiple encoders, the system retrieves the hidden states corresponding to 30 displacement positions. The predicted value is obtained through a linear output layer: In one embodiment, the model employs a two-layer graph neural network, a three-layer Transformer encoder, four attention heads, and hidden dimensions. The feedforward layer has a dimension of 512, the optimizer is Adam, and the learning rate is... The number of samples selected for each training session is 32, and the number of training rounds is 100.

[0052] The present invention also provides a system for rapid prediction of structural response curves, comprising: (1) Parameter input module: used to receive structural geometric parameters, material parameters, reinforcement parameters, failure location category, boundary conditions and floor participation mode.

[0053] (2) Data preprocessing module: used to uniformly perform z-score standardization on input and output.

[0054] (3) Physical Prior Calculation Module: Used to calculate analytical prior curves for failure of inner columns, side columns and corner columns respectively.

[0055] (4) Failure Graph Construction Module: Used to build the basic structure graph and output the dynamic adjacency matrix based on the position of the failure column and the mask of the failure node.

[0056] (5) Fusion prediction module: including graph neural network module, physical prior module and Transformer module, used to output complete response curve.

[0057] (6) Results output module: used to output structural resistance evolution curves and related visualization charts under different failure conditions.

[0058] (7) Model management module: used to save normalizers, training history, running configurations and evaluation results, so as to facilitate the reproduction of experiments and engineering deployment.

[0059] Compared with the prior art, the beneficial effects of the present invention include: (1) It can uniformly handle the failure conditions of inner columns, side columns and corner columns. By unifying data organization and failure perception map construction, the present invention incorporates the failure conditions of inner columns, side columns and corner columns into the same model framework, avoiding the need to build independent models for different locations.

[0060] (2) It can explicitly express the structural topology and failure location. The graph neural network module dynamically adjusts the adjacency matrix according to the failure node and the floor participation mode, so that the model not only utilizes the structural physical parameters, but also the node connection relationship and failure propagation path.

[0061] (3) It can maintain the sequential consistency of the entire response curve. The Transformer encoder is used to capture the interdependence of the structural collapse response under different displacements, which is more conducive to maintaining the overall shape of the curve than point-by-point regression.

[0062] (4) It can improve stability by utilizing structural mechanics priors. The physical prior branch provides analytical response trends, and the gating fusion mechanism enables the model to adaptively absorb graphical structural information and physical prior information according to different displacement positions.

[0063] (5) It can significantly improve the efficiency of batch prediction. After the model is trained, the complete response curve can be quickly output by inputting structural parameters and failure conditions, which is suitable for the rapid screening of a large number of parameter combinations and conditions.

[0064] (6) High prediction accuracy. On the test set, the established model (GNN-PHY-TF) achieved MAPE (mean absolute percentage error) = 1.72%, RMSE (root mean square error) = 35.83, MAE (mean absolute error) = 22.97, and R² (coefficient of determination) = 0.9986; the MAPE for the three failure conditions of inner columns, side columns, and corner columns were 1.49%, 1.41%, and 2.26%, respectively. (Meaning: mean absolute percentage error, root mean square error, mean absolute error, and coefficient of determination.)

[0065] (7) Ablation results show that each module contributes. Compared with w / o GNN (no graph neural network module), w / o Physics (no physical prior module), w / o Transformer (no Transformer module) and MLP baseline (multilayer perceptron model), the established model (GNN-PHY-TF) maintains better performance in overall indicators and most sub-conditions.

[0066] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A deep learning-based method for predicting structural collapse response under multiple operating conditions, characterized in that, include: Obtain the physical parameters and failure condition parameters of the target structure; The failure condition parameters include failure location category and floor participation mode; Based on the physical parameters and failure condition parameters, the physical prior resistance values ​​of each preset displacement sampling point of the target structure are obtained; Each floor of the target structure is represented by a floor subgraph. The nodes in each floor subgraph include column nodes and plate element nodes within that floor. The edges in each floor subgraph include horizontal adjacent connections between column nodes within that floor and connections between plate element nodes and their corner column nodes. Through the inter-floor connection relationships between adjacent floors, the floor subgraphs are stitched together into a global graph structure and represented by an adjacency matrix. The failure column node is determined according to the failure location category. The failure column node and the edge associated with the failure column node are shielded from the adjacency matrix. The floor subgraph that does not participate in message passing is shielded from the adjacency matrix according to the floor participation mode, so as to construct a failure-aware dynamic graph. Based on the aforementioned failure-aware dynamic graph, graph-level pooling is performed using a graph neural network to obtain a graph-level representation; Based on the physical parameters, the prior physical resistance values, and the graphical representation, the resistance prediction curve of the target structure is obtained.

2. The deep learning-based structural multi-condition collapse response prediction method as described in claim 1, characterized in that, Based on the physical parameters, prior physical resistance values, and the graphical representation, the resistance prediction curve of the target structure is obtained, including: The displacement values, the corresponding prior physical resistance values, and the physical parameters of each displacement sampling point are concatenated to form the original vector of the position of each displacement sampling point. The original vector is then mapped into a sequence representation through a shared linear mapping. The graph-level representation and the sequence representation are fused using a gating mechanism to form a joint input sequence; The joint input sequence is input into the Transformer encoder, and the long-range dependency between different displacement sampling points is modeled through the multi-head self-attention mechanism of the Transformer encoder, and the hidden state corresponding to each displacement position is output. The hidden states at each location are mapped to predicted response values ​​through the linear output layer of the Transformer encoder.

3. The deep learning-based structural multi-condition collapse response prediction method as described in claim 2, characterized in that, The graph-level representation and the sequence representation are fused using a gating mechanism to form a joint input sequence, including: The graph-level representation is linearly mapped to obtain the mapped graph vector. For each sequence position, a position-related gate vector is obtained through a gate function based on the sequence representation of that position and the mapping vector. Based on the gate vector, the mapping vector is superimposed onto the sequence representation of the sequence position by element-wise multiplication to obtain the fused sequence vector; The mapping vector is used as an independent global vector, which together with each fused sequence vector and its position code constitutes a joint input sequence.

4. The deep learning-based structural multi-condition collapse response prediction method as described in claim 1, characterized in that, Based on the physical parameters and failure condition parameters, the prior physical resistance values ​​of each preset displacement sampling point of the target structure are obtained, including: Based on the physical parameters and failure condition parameters, and according to the membrane effect decomposition theory, the compressive membrane effect resistance curves in the small deformation stage and the tensile membrane effect resistance curves in the large deformation stage are obtained; wherein, the compressive membrane effect resistance is described by an expression with a degradation coefficient, and the tensile membrane effect resistance is described by an expression with a shape factor. The pressure membrane effect resistance and the tension membrane effect resistance are superimposed to generate the initial prior response trend corresponding to the failure location category. The trend is then discretized according to the preset number of displacement sampling points to obtain the sequence of physical prior resistance values ​​for each displacement sampling point of the target structure.

5. The deep learning-based structural multi-condition collapse response prediction method as described in claim 1, characterized in that, The graph-level pooling obtained through a graph neural network to obtain a graph-level representation includes: The adjacency matrix of the failure perception graph is normalized to obtain a normalized adjacency matrix. Based on the normalized adjacency matrix, a multi-layer graph neural network is used to pass and update the node features, and during the update process, the information propagation of inactive nodes is shielded by an active node mask. Perform masked maximum pooling on the active nodes after message passing to obtain the graph-level representation.

6. The deep learning-based structural multi-condition collapse response prediction method as described in claim 1, characterized in that, The physical parameters of the target structure include: X-direction span, Y-direction span, X-direction beam height, Y-direction beam height, X-direction beam width, Y-direction beam width, slab thickness, X-direction beam top reinforcement ratio, X-direction beam bottom reinforcement ratio, Y-direction beam top reinforcement ratio, Y-direction beam bottom reinforcement ratio, slab reinforcement ratio, steel bar yield strength, steel bar ultimate strength, steel bar ultimate strain, and concrete compressive strength. The failure location categories include inner column failure, side column failure, and corner column failure.

7. The deep learning-based structural multi-condition collapse response prediction method as described in claim 1, characterized in that, The resistance prediction curve of the target structure is a complete resistance displacement response curve containing 30 displacement sampling points, which are distributed at equal intervals with a step size of 0.005, starting from displacement 0.

8. A deep learning-based structural multi-condition collapse response prediction system, characterized in that, include: The parameter acquisition module is used to acquire the physical parameters and failure condition parameters of the target structure. The failure condition parameters include failure location category and floor participation mode; Based on the physical parameters and failure condition parameters, the physical prior resistance values ​​of each preset displacement sampling point of the target structure are obtained; The graph structure construction module is used to create a floor subgraph for each floor of the target structure. The nodes in each floor subgraph include column nodes and plate element nodes within that floor. The edges in each floor subgraph include horizontal adjacent connections between column nodes within that floor and connections between plate element nodes and their corner column nodes. Through the inter-floor connection relationships between adjacent floors, the floor subgraphs are spliced ​​into a global graph structure and represented by an adjacency matrix. The failure graph construction module is used to determine the failure column node according to the failure location category, shield the failure column node and the edge related to the failure column node from the adjacency matrix, and shield the floor subgraph that communicates with the failure column node from the adjacency matrix according to the floor participation mode, so as to construct a failure-aware dynamic graph. The graph-level representation construction module is used to obtain a graph-level representation based on the failure-aware dynamic graph by performing graph-level pooling through a graph neural network. The prediction module is used to obtain the resistance prediction curve of the target structure based on the physical parameters, the prior physical resistance value and the graphical representation.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1 to 7.

Images