A method and system for predicting the finite element response of foundation pits based on spatiotemporal diagrams

By constructing a node-unit heterogeneous graph and a stage-conditional heterogeneous spatiotemporal graph neural network, the problems of low computational efficiency and poor prediction consistency in three-dimensional finite element simulation of foundation pit excavation are solved, and fast and accurate prediction of soil displacement and stress field is achieved, which is suitable for real-time response assessment of large-scale foundation pit projects.

CN122797248BActive Publication Date: 2026-11-10CHINA CONSTR FIFTH ENG DIV CORP LTD +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611268701.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-10
Estimated Expiration
2046-08-20

AI Technical Summary

Technical Problem

Existing three-dimensional finite element simulation methods for foundation pit excavation are computationally inefficient and cannot meet the needs of rapid comparison and real-time prediction during the construction phase. Furthermore, traditional graph neural networks are insufficient in distinguishing the physical roles of soil nodes and soil elements, resulting in poor consistency between displacement field and stress field predictions.

Method used

A spatiotemporal graph-based finite element response prediction method for foundation pits is adopted. A node-element heterogeneous graph is constructed to generate three-dimensional geometric-mechanical edge features. The stage-conditional heterogeneous spatiotemporal graph neural network model is used for message passing. Combined with the Transformer time refinement module, the joint prediction of soil node displacement and soil element stress is realized.

Benefits of technology

It enables rapid and accurate prediction of soil displacement and stress fields in foundation pits, improves computational efficiency, enhances the consistency of joint field prediction, adapts to dynamic changes during the construction phase, and supports efficient training and inference of large-scale three-dimensional models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122797248B_ABST
    Figure CN122797248B_ABST
Patent Text Reader

Abstract

This invention provides a method and system for predicting the finite element response of foundation pits based on spatiotemporal graphs. It is applied to predicting the stress and displacement fields of soil during foundation pit excavation. The method includes acquiring three-dimensional finite element sample data; constructing a node-element heterogeneous graph; generating three-dimensional geometric-mechanical edge features; conditionally encoding time information and construction stage states; and inputting soil node feature vectors, soil element feature vectors, three-dimensional geometric-mechanical edge features, and time-construction stage embedded feature vectors into a stage-conditional heterogeneous spatiotemporal graph neural network model to jointly predict the three-dimensional displacement of nodes and the three-dimensional stress of elements. This invention innovatively designs a dual-entity heterogeneous graph and stage-conditional message vector transmission, enabling changes in construction stages to directly affect the effective graph structure and information flow, effectively reducing the time required for repeated finite element solutions under multiple working conditions and improving the efficiency of foundation pit construction response prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of geotechnical engineering, foundation pit engineering, finite element simulation modeling, computational mechanics and graph neural network technology, and in particular to a spatiotemporal graph-based finite element response prediction method and system for foundation pits, applied to the prediction of soil stress field and displacement field during foundation pit excavation. Background Technology

[0002] Finite Element Analysis (FEA) is an important technique for analyzing soil deformation and stress in foundation pit engineering, geotechnical engineering, and underground engineering. In foundation pit excavation projects, by establishing a three-dimensional finite element model that includes the soil region, retaining and supporting structures, an unstructured three-dimensional mesh, soil material parameters, boundary conditions, surface loads or construction loads, and construction stage information, engineering responses such as soil nodal displacements, soil element stresses, equivalent stresses, and potential plastic zones can be obtained. This provides a basis for foundation pit support design, construction control, and safety assessment.

[0003] Existing three-dimensional finite element simulation methods for foundation pit excavation still face the problem of insufficient computational efficiency in engineering applications. For foundation pit models that include a large number of soil nodes, soil elements, retaining and supporting structures, and construction stages, whenever soil parameters, support stiffness, excavation sequence, surface loads, or boundary conditions change, it is often necessary to recalculate, resulting in long calculation times. This makes it difficult to meet the needs of rapid comparison and selection of foundation pit schemes, real-time prediction during construction stages, and multi-condition safety assessment. Therefore, relying solely on repeated calculations using traditional finite element methods is insufficient to balance accuracy and efficiency in complex foundation pit engineering scenarios.

[0004] Currently, research has attempted to approximate finite element solvers using deep learning surrogate models, mainly including convolutional neural networks, fully connected networks, and graph neural networks. Convolutional neural networks are more suitable for regular meshes or image-based field data, but struggle to adapt to unstructured 3D finite element meshes of foundation pit soil. Fully connected networks have limited adaptability to variations in the number of soil nodes, soil elements, and excavation topology. While ordinary graph neural networks can describe topological relationships, most methods only use nodes as graph vertices, failing to adequately distinguish the different physical roles of foundation pit soil nodes and soil elements. Furthermore, they suffer from insufficient representation of 3D adjacency relationships of soil elements, inadequate utilization of dynamic information during construction, and insufficient consistency between the joint prediction of soil displacement and stress fields. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for predicting the finite element response of foundation pits based on spatiotemporal diagrams, which can realize rapid and accurate prediction of the displacement field and stress field of the foundation pit soil.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is: In a first aspect, this application provides a finite element response prediction method for foundation pits based on spatiotemporal diagrams, applied to the prediction of soil stress and displacement fields during foundation pit excavation, comprising the following steps: S1: Obtain three-dimensional finite element sample data for the foundation pit excavation project; S2: Based on the three-dimensional finite element sample data, construct a node-element heterogeneous diagram to characterize the topological relationship of the finite element mesh of the foundation pit soil. The node-element heterogeneous diagram includes soil node entities, soil element entities, node-node edges, element-element edges, node-to-element edges, and element-to-node edges. S3: Based on the three-dimensional finite element sample data and the node-element heterogeneous diagram, generate three-dimensional geometric-mechanical edge features; the three-dimensional geometric-mechanical edge features are used to characterize the displacement coordination relationship between soil nodes, the stress transfer relationship between soil elements, and the displacement-stress coupling relationship between soil nodes and soil elements; S4: Conditionally encode the construction stage information and construction state vector to obtain the time-construction stage embedded feature vector; The construction state vector is composed of the soil excavation state, the activation state information of the retaining and supporting structure, the activation state of the material region, the boundary conditions of the foundation pit, and the load state information in the three-dimensional finite element sample data. S5: The three-dimensional finite element sample data, three-dimensional geometric-mechanical edge features, and time-construction stage embedded feature vectors are input into the stage-conditional heterogeneous spatiotemporal graph neural network model. The stage-conditional heterogeneous spatiotemporal graph neural network model performs heterogeneous message passing through stage gating, attention aggregation, and FiLM modulation (Feature-wise Linear Modulation). It also performs time refinement based on the Transformer time refinement module that introduces the construction stage attention bias to obtain node representation vectors and unit representation vectors. S6: Based on node representation vectors and element representation vectors, jointly predict the three-dimensional displacement vectors of soil nodes and the three-dimensional stress tensors of soil elements during the foundation pit excavation process.

[0007] In one possible implementation, in step S2, for the finite element mesh of the foundation pit soil, the formula is used... Constructing a node-cell heterogeneous graph , where the set of nodes , For the set of soil node entities, For a collection of soil unit entities, Let be a set of edge types, where , , , These represent node-to-node edge, cell-to-cell edge, node-to-cell edge, and cell-to-node edge, respectively. For a set of node-node edges, For cell-cell edge set, For the set of edges from nodes to cells, This is the set of element-to-node edges. If two nodes belong to the same soil element (soil finite element element) or the endpoint of a real mesh edge, then a node-to-node edge is created between these two nodes. If two soil elements share a surface, share an edge, or share a point, then an element-to-element edge is created between these two soil elements. If a soil node belongs to the soil element connection table, then a node-to-element edge and an element-to-node edge are created.

[0008] Through this heterogeneous graph definition, the soil nodes and soil elements in the foundation pit are no longer mixed into the same type of graph vertices, but instead carry the displacement information of the soil nodes and the stress, soil material, and excavation activation status information of the soil elements respectively.

[0009] The soil element connection table is a topological data structure describing the relationship between nodes and elements in the soil mesh. It records which specific nodes constitute each soil element and which elements each soil node participates in. The soil element connection table originates from the finite element mesh file of the foundation pit soil.

[0010] The node-unit heterogeneous diagram uses dynamic activation states to represent the changes in soil unit excavation and removal, activation of retaining and supporting structures, soil material state, foundation pit boundary switching, and construction load state during the construction stage. It also distinguishes between undisturbed soil, excavated soil areas, areas affected by retaining and supporting structures, and boundary constraint areas through entity type coding.

[0011] In one possible implementation, step S3 utilizes the general formula Construct three-dimensional geometric-mechanical edge features, where, Indicates the type is edge eigenvectors; and Let them represent the source entity and the target entity of the edge, respectively. This represents a directed edge from source entity s to target entity q. Indicates the type of edge relationship; This represents the normalized geometric quantities in the finite element mesh of the foundation pit soil. This indicates the influencing properties such as soil material, region, and retaining and supporting structures. This indicates information about the foundation pit boundary, load status, and construction stage. For the type The edge-specific feature constructor is used to concatenate and normalize the input parameters. This general expression can be further specified for different relation types, as follows: Node-to-node edge features are adopted Generate, where Indicates the type is edge eigenvectors, Indicates the type is The feature constructor function specific to the edge. Indicates from soil node Pointing to soil nodes The directed edge, For reference length, the span in the x, y, and z directions of the bounding box of the finite element sample of the foundation pit can be used. The maximum value is determined, that is Sure; For soil nodes and soil nodes Normalized distance between them and Soil nodes and soil nodes coordinates and Soil nodes and soil nodes The boundary load characteristics. This formula is used to express the coordination of local displacements and the propagation of boundary effects in the foundation pit soil, enabling node-to-node messages to distinguish between ordinary soil geometric neighbors, nodes with common boundary constraints, nodes in the common load zone, and nodes near the soil material interface.

[0012] Cell-to-cell edge features are adopted Generate, where Indicates the type is edge eigenvectors, Indicates the type is The feature constructor function specific to the edge. and These represent the soil element and the soil element number of the edge, respectively. Indicates from soil element Pointing to soil element The directed edge, and Soil units and soil unit The center of mass, Indicates the adjacency type of a soil element sharing a face, edge, or point; and Soil units and soil unit Shared surface normal and shared area; soil unit and soil unit Adjacency strength Can be written as ,in The weight functions correspond to different shared adjacency types; For target soil unit Total surface area; This is a numerically stable term. This formula is used to explain how soil element-to-soil element message passing distinguishes different 3D topological adjacencies and their significance in foundation pit mechanics: shared surfaces have more explicit directions of normal and shear force transmission, while shared edges or shared points mainly provide weaker spatial context relationships.

[0013] Node-to-cell edge features are adopted Generate, where Indicates the type is edge eigenvectors, Indicates the type is The feature constructor function specific to the edge. Indicates from soil node Pointing to soil element The directed edge, For local coordinates, The weights are the shape function weights.

[0014] Cell-to-node edge features are adopted Generate, where Indicates the type is edge eigenvectors, Indicates the type is The feature constructor function specific to the edge. Indicates from soil element Pointing to soil nodes The directed edge, For soil units Material characteristics. For element-to-node edges Soil element material features are added to the node-element edge features. This is used to express the feedback relationship between soil element stress and soil material properties on the prediction of soil node displacement.

[0015] In one possible implementation, in step S4, for the foundation pit excavation construction process, the formula is used... Generate construction phase Time-construction stage embedded feature vector ,in For stage coding functions, , and These represent the construction stages. The category code, normalized time value (normalized time value of the current excavation stage throughout the entire construction process), and construction state vector.

[0016] In one possible implementation, in step S5, before the phase-conditional heterogeneous message passing, the original input feature vectors of each type of entity are first encoded and initialized. (Soil node) The initial hidden state vector is generated by the soil node encoder. The mapping yields: ,in For soil nodes The original input feature vector, superscript Indicates node channels, with the superscript 0 indicating layer 0 (initial layer); soil element. The initial hidden state vector is generated by the soil element encoder. The mapping yields: ,in For soil units The original input feature vector, superscript Indicates the element channel. Soil node encoder. Soil element encoder Each contains independent learnable parameters, and the output dimensions are the same to ensure numerical compatibility in subsequent cross-type message passing.

[0017] For the four types of edge relationships, the original three-dimensional geometric-mechanical edge characteristics of each type of edge. Each is encoded by its own dedicated edge encoder. Mapped to edge implicit representation vector: , Indicates the type is edge The implicit representation vector of the side encoder. Different relation types are set independently, so that the edge features of different topological relations are encoded in their respective independent semantic spaces.

[0018] The initial hidden state vectors of soil nodes, soil elements, and edge hidden representation vectors together constitute the initial state input for conditional heterogeneous message passing in subsequent stages.

[0019] In one possible implementation, in step S5, during phase-conditional heterogeneous message passing, the formula is used... Conditional message vectors are generated during the generation phase, where Indicates the type is edge The phase conditional message vector on, Represents the source entity The hidden state vector. Represents the target entity The hidden state vector. Indicates the type of membership. The message generation function for the edges. This formula shows that the message is not determined solely by the hidden state vectors of adjacent entities, but also depends on the three-dimensional geometric-mechanical properties of the foundation pit soil edges and the embedded feature vectors of the time-construction stage, thus causing the same foundation pit soil topology to produce different propagation behaviors at different excavation stages.

[0020] Using phase gating to control message propagation in dynamically active states, and using formulas Controlling message propagation, among which Represents the edges after gating. The message vector on; Indicates the construction stage side The gating coefficient on, , , and Representing the source entity respectively Target Entity and edge During the construction phase The dynamic activation state (effective state) value is determined based on the soil node activation state, soil element activation state, and edge activation state in the construction state vector, corresponding to the soil excavation state, retaining and support structure activation state, material region activation state, foundation pit boundary condition state, and load state. Soil element activation state: determined based on the soil excavation state in the construction state vector. If the element is in the excavation and removal area during the current construction stage, the dynamic activation state value is 0; otherwise, it is 1. Soil node activation state: determined based on the activation states of all soil elements to which the node belongs. If the node does not belong to any currently effective soil element, the dynamic activation state value is 0; otherwise, it is 1. Edge activation state: determined by the logical AND of the activation states of the source entity and target entity connected to the edge. Additionally, for contact edges or constraint edges related to the retaining and support structures, the activation state information of the corresponding structure must also be considered. The dynamic activation state value of the corresponding edge is 1 only when the structure is in the current construction stage's activation state; otherwise, it is 0. For trainable gated weight matrix, This is the Sigmoid activation function. The product term of the dynamic activation states is used to filter out invalid propagation paths, and the feature vector is embedded based on the time-phase construction stage. The generated gating factor is used to learnably weight valid messages. This formula expresses the removal of soil elements, activation of retaining and support structures, changes in soil material state, and switching of boundary conditions during the excavation phase. When a certain type of soil entity or edge is invalid in the current excavation phase, its message is suppressed, thus preventing the model from propagating non-physical information in non-existent soil areas, excavated areas, or invalid boundary relationships. The dynamic activation state is used to form a hard constraint filter before message aggregation, ensuring that invalid entities and invalid relationships do not participate in message propagation in the current construction phase, and the corresponding messages do not enter subsequent attention aggregation.

[0021] Using formula To achieve attention convergence, among which Represents the target entity In the category of The aggregated message vector on the edge, For the type edge The attention weight coefficients on the surface can be determined by the entity latent state vector, edge features, and time-construction stage embedding feature vector. For target entity The relationship type is Neighbor set; soil node passage uses formula Aggregate node—aggregate message vector on node edge and aggregate message vector from element to node edge; soil element channel uses formula. The aggregation unit—the aggregation message vector on the unit edge and the aggregation message vector from the node to the unit edge—forms the target entity. Aggregated messages .

[0022] Embedded feature vectors based on time-construction stage Generate FiLM modulation scale parameters and bias parameters Using the formula FiLM modulation of aggregated messages ( (This represents element-wise multiplication), resulting in the aggregated message vector after FiLM modulation. .

[0023] Using formula Update the hidden state vector of the node or the hidden state vector of the cell; where... and They represent the first Layer and first The hidden state vector of the target entity in the layer; Number the network layers. For state update functions, This is the layer normalization function.

[0024] Therefore, dynamic activation state and stage gating first control the effective message propagation under the current construction stage, attention aggregation further determines the importance of messages in different neighborhoods, and FiLM modulation then adjusts the scale and bias of the aggregated messages according to the construction stage, and updates the node hidden state vector and the unit hidden state vector accordingly.

[0025] After completing several layers of heterogeneous message passing, the node hidden state vectors of the same soil node and the element hidden state vectors of the same soil element under continuous construction stages are respectively formed into node hidden state vector sequences and element hidden state vector sequences according to the construction stage time sequence, and then input into the Transformer time refinement module. For any of the hidden state vector sequences, query mapping, key mapping, and value mapping are performed on the hidden state vector sequence through different trainable mapping matrices to obtain the query matrix. Key matrix Sum matrix According to the current construction stage The corresponding time-phase embedding feature vectors determine the attention bias matrix for the construction phase. Each matrix element represents the bias effect of the stage association between two corresponding construction stages on the time attention score. Using the formula... Perform time attention calculations, where The dimension is the key vector; the result of the temporal attention calculation. After processing with feedforward network, residual connection and layer normalization, the time-refined node representation vector and element representation vector corresponding to each construction stage are obtained.

[0026] The Transformer temporal refinement module comprises multiple stacked temporal refinement modules. The process is as follows: the node hidden state sequence or unit hidden state sequence is used as input to the first-layer temporal refinement module. This module performs temporal self-attention calculation, residual connections, layer normalization, and feedforward network mapping, outputting the first-layer temporal refined features. These first-layer features are then used as input to the second-layer temporal refinement module, and the same operation is repeated. This process continues until a predetermined number of stacked temporal refinement modules are processed. Finally, the features output by the last layer serve as the temporally refined node representation. Each temporal refinement module is independently computed based on a multi-head self-attention mechanism, and the learnable parameters between layers are independent. This hierarchical abstraction of temporal features through inter-layer progression enables the model to capture multi-scale temporal dependencies from local mutations to global evolution during foundation pit construction.

[0027] This section enables the model to learn the long-term soil response evolution during the excavation process, rather than being limited to spatial propagation in a single time frame. This construction-phase attention bias allows the temporal refinement to focus on state changes significant to the excavation project, rather than simply performing a general fit on the time series.

[0028] In one possible implementation, in step S6, the formula is used. Predict the three-dimensional displacement vector of soil nodes. Indicates the construction stage Lower soil node The predicted three-dimensional displacement vector, where Indicates the construction stage Lower soil node The time-refined node representation vector, For displacement prediction function (nodal prediction head); using the formula Predict the six-dimensional stress vector of the soil element, where, Indicates the construction stage Lower soil unit The predicted six-dimensional stress vector, Indicates the construction stage Lower soil unit Time-refined unit representation vector, This represents the stress prediction function (element prediction head).

[0029] in, The three-dimensional symmetric stress tensor of the foundation pit soil is reconstructed from six stress components. This is used for subsequent calculations of von Mises stress, principal stresses, and stress contour maps. This reconstruction method ensures that the soil stress prediction results satisfy the three-dimensional symmetric stress tensor structure, avoiding contradictions between the six components and derived quantities.

[0030] Using formula Deterministic calculation of von Mises stress; where, This represents the predicted von Mises stress value. To predict the six-dimensional stress vector The function that calculates the von Mises stress square term, representing the equivalent stress invariant. , To prevent abnormally positive square roots, this formula directly links the six-component stress of the soil element output by the neural network with the equivalent stress commonly used in foundation pit engineering, ensuring that the engineering derived quantity remains consistent with the soil stress tensor.

[0031] The principal stresses are determined by their eigenvalues: ,in This represents the predicted three-dimensional symmetric stress tensor of the foundation pit soil. The three principal stress prediction values ​​obtained through eigenvalue decomposition This is a function for solving the eigenvalues ​​of a symmetric matrix, used to calculate the three eigenvalues ​​of a three-dimensional symmetric stress tensor. The principal stresses can be used as post-processing results in foundation pit engineering, or as auxiliary monitoring or evaluation indicators.

[0032] The total training loss used for predicting the displacement and stress fields of foundation pit soil is obtained by weighting multiple losses, as shown in the formula: ,in and These are, respectively, the soil nodal displacement monitoring loss, the soil element stress monitoring loss, the von Mises stress loss, and the physical consistency loss. and All are non-negative loss weights; the physical consistency loss is calculated using the formula... ,in, and These are, respectively, boundary condition loss, constitutive consistency loss, and equilibrium residual loss; and These are the activation coefficients for boundary condition loss, constitutive consistency loss, and equilibrium residual loss, respectively, and their values ​​are determined based on whether the corresponding physical constraints are calculable. These losses correspond to data fitting, engineering index consistency, and soil mechanical constraint consistency in the three-dimensional finite element prediction of the foundation pit, respectively.

[0033] In one possible implementation, the displacement monitoring loss and stress monitoring loss are calculated using the soil node displacement labels and soil element stress labels in step S1, respectively, and the stage-conditional heterogeneous spatiotemporal graph neural network model is trained by combining the engineering derived quantity loss and physical consistency loss.

[0034] In one possible implementation, in step S6, for large-scale foundation pit 3D finite element samples, a joint strategy of spatial subgraph sampling, k-hop halo expansion, and time window sampling is adopted during the training process: The spatial sampling strategy is as follows: A central node set is sampled from the set of soil nodes in the foundation pit, and a central soil element set is determined; these two sets constitute a central subgraph. Starting from the central subgraph, a neighborhood of k layers is expanded outward along the edges of each type of the node-element heterogeneous graph to obtain the halo region. The supervised loss is only applied to the set of soil nodes within the central subgraph. and soil element set In the above calculation, entities within the halo region are used to provide the spatial context for message passing and are not included in the supervision loss. The subgraph supervision loss is written as... , For time windows.

[0035] The time window sampling strategy is as follows: time windows are sampled from the complete construction phase sequence. ,in The time window length is defined, and temporal attention is calculated for consecutive construction stages within the window. This sampling strategy reduces memory usage while improving the model's ability to learn from frames that switch between foundation pit excavation stages, peak soil response frames, and frames that show abrupt changes in construction loads.

[0036] The inference and prediction results of the stage-conditional heterogeneous spatiotemporal graph neural network model can be written back to the original three-dimensional finite element mesh of the foundation pit according to the soil node number and soil element number.

[0037] The soil nodal displacement monitoring loss and the soil element stress monitoring loss are respectively written as: , ,in and The labels are the predicted displacement vector and the actual displacement vector, respectively. and These are the labels for the predicted stress vector and the actual stress vector, respectively. The Huber function is used to reduce the instability of the overall training caused by local abnormal elements or peak response noise in the foundation pit, while maintaining sensitivity to small error regions. These two supervised losses constrain the nodal-level displacement prediction and the element-level stress prediction of the foundation pit soil, respectively, and are fundamental terms for model training.

[0038] von Mises stress loss is written as ,in This loss is calculated from the true stress vector label (the six-dimensional stress vector of the true soil). It is used to enhance the predictive ability of engineering parameters in high-stress areas, near retaining and supporting structures, and near soil material interfaces during foundation pit excavation.

[0039] Boundary condition loss can be written as Where B is the set of boundary nodes in the finite element model of the foundation pit. Represents boundary nodes The predicted three-dimensional displacement vector, Represents boundary nodes For a given displacement boundary value, the zero vector is taken for a fixed constraint, and the corresponding non-zero displacement value is taken for a given displacement boundary. Boundary nodes The degree-of-freedom mask is set to 1 for constrained directions and 0 for free directions; The L1 norm is represented. The boundary condition loss is used to constrain the fixed boundary, far-field boundary, and given displacement boundary of the foundation pit, so that the predicted soil displacement field does not violate the basic boundary settings of the foundation pit finite element model.

[0040] The total physical consistency loss can be written as ,in , and This is an availability mask. It is used to comprehensively constrain the foundation pit boundary conditions, soil constitutive consistency, and equilibrium residuals under available physical operators, ensuring that the predicted soil displacement and stress fields simultaneously meet the requirements of basic boundary settings, soil material relationships, and local force equilibrium. The availability mask allows for the application of mechanical constraints adapted to the current foundation pit sample under different soil material types, excavation stages, and physical operator reliability conditions, avoiding the forced application of unreliable constraints to all soil elements.

[0041] When the soil material model and the soil element strain-displacement matrix are available, the constitutive equation can be written as: , , For soil units The predicted strain vector, For soil units The strain-displacement matrix is ​​determined by the element type and nodal coordinates; For soil units The predicted displacement vectors of all nodes (stacked in node order); For soil units The constitutive consistent stress prediction value; For soil units The soil material matrix. The constitutive consistency loss can be written as... , The formula represents the total number of soil elements, where n is the stress directly predicted by the neural network. The stress derived from the predicted displacement using constitutive equations If the two are consistent, it indicates that the prediction results satisfy both the geometric equations and the physical equations, and conform to the basic laws of mechanics. As a constitutive consistency availability mask, when Use 1 if reliable, otherwise use 0. For elastoplastic path-related soil materials, contact problems, or black-box materials, Unreliable, this loss item passed Close to prevent incorrect physical constraints from harming training.

[0042] When the area, surface normal, body forces, and boundary forces of the soil element are reliable, the equilibrium residual can be written as: ,in, Represents soil element The balance residual, For soil units No. The area of ​​the surface, For soil units No. The normal unit vector of the surface, For soil units volume, For soil units Physical activity, For soil units The equivalent boundary force term. The equilibrium residual loss is taken as... This loss constrains the prediction of soil stress field from the perspective of local force balance. To balance the residual availability mask, a value of 1 is used when the above geometric and mechanical data are reliable, and a value of 0 is used otherwise. To balance residual availability mask.

[0043] Secondly, this application provides a spatiotemporal diagram-based finite element response prediction system for foundation pits, used to implement the aforementioned spatiotemporal diagram-based finite element response prediction method for foundation pits, including: The data acquisition module is used to acquire three-dimensional finite element sample data for foundation pit excavation projects; The graph construction module is used to construct a node-element heterogeneous graph to characterize the topological relationship of the finite element mesh of the foundation pit soil based on the three-dimensional finite element sample data. The node-element heterogeneous graph includes soil node entities, soil element entities, node-node edges, element-element edges, node-to-element edges, and element-to-node edges. The feature generation module is used to generate three-dimensional geometric-mechanical edge features based on the three-dimensional finite element sample data and the node-element heterogeneous diagram; the three-dimensional geometric-mechanical edge features are used to characterize the displacement coordination relationship between soil nodes, the stress transfer relationship between soil elements, and the displacement-stress coupling relationship between soil nodes and soil elements. The time-construction stage embedding feature coding module performs conditional coding on construction stage information and construction state vector to obtain the time-construction stage embedding feature vector. The graph neural network model calculation module is used to input the feature vectors of soil nodes and soil elements, three-dimensional geometric-mechanical edge features, and time-construction stage into the stage-conditional heterogeneous spatiotemporal graph neural network model. The stage-conditional heterogeneous spatiotemporal graph neural network model performs heterogeneous message passing through stage gating, attention aggregation, and FiLM modulation, and performs time refinement based on the Transformer time refinement module that introduces construction stage attention bias to obtain node representation vectors and element representation vectors. The joint prediction module is used to jointly predict the three-dimensional displacement vector of soil nodes and the three-dimensional stress tensor of soil elements during the foundation pit excavation process, based on the node representation vector and the element representation vector.

[0044] Thirdly, this application provides an electronic device, including: a memory and a processor; The memory is used to store computer programs; The processor is used to invoke the computer program to execute the method described above.

[0045] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed on an electronic device, causes the electronic device to perform the method described above.

[0046] Fifthly, this application provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described above.

[0047] The specific implementation methods of the second to fifth aspects of this application can refer to the implementation methods of the first aspect, and will not be elaborated here.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] (1) Significantly improves computational efficiency and enables rapid reasoning. By constructing a high-precision deep learning proxy model, this invention can quickly infer and predict the displacement field and stress field of the foundation pit soil under different construction stages after training. This effectively solves the technical problem of high cost of repeated solving of large-scale foundation pit three-dimensional finite element models in multi-condition and multi-stage analysis, and provides an efficient tool for rapid comparison of foundation pit schemes and real-time response evaluation during construction stages.

[0050] (2) Distinguishing physical roles enhances the consistency of joint field prediction. This invention constructs a graph structure containing dual entity types to clearly distinguish the physical roles of soil nodes and soil elements—nodes bear displacement information, while elements bear stress and material information. Based on this, a two-way coupling relationship is established between entities, enabling the prediction of displacement and stress fields to be mutually constrained and optimized collaboratively. This fundamentally overcomes the problem of inconsistent field prediction caused by the confusion of physical roles in existing methods, and improves the physical reliability of the prediction results.

[0051] (3) Refined expression of three-dimensional spatial adjacency and mechanical transfer relationships. This invention, targeting the characteristics of three-dimensional finite element meshes for foundation pit soil, constructs multiple types of spatial adjacency relationship expressions, capable of distinguishing different levels of three-dimensional topological adjacency between soil elements, such as shared surfaces, shared edges, and shared points. Simultaneously, key geometric-mechanical information such as shared area, normal direction, and shape function weights are incorporated into the relationship features, enabling the model to precisely characterize the stress transfer and deformation coordination differences corresponding to different adjacency types, overcoming the shortcomings of existing methods that result in coarse adjacency relationship expressions and loss of mechanical information.

[0052] (4) Deep integration of dynamic changes in construction stages to adapt to physical topological changes. This invention encodes construction stage information into conditional embedding vectors and allows them to directly participate in the control of message propagation paths within the model. This enables changes in construction states, such as soil excavation and removal, step-by-step activation of retaining structures, and switching of boundary conditions, to dynamically affect the effective information flow. The model can adaptively adjust the inference path according to the current construction stage, accurately capture the soil response changes caused by key construction procedures, and overcome the limitations of traditional time-series models that only use construction stages as static inputs and cannot adapt to dynamic changes in physical topology.

[0053] (5) Supports efficient training and inference of large-scale 3D models. This invention designs a space-time joint sampling strategy for large-scale 3D finite element models of foundation pits. By using spatial subgraph sampling and halo region expansion, it effectively reduces the memory usage and focuses on key construction stages by using time window sampling. Under the condition of limited computing resources, it retains local spatial correlation and construction stage evolution information, making this invention applicable to large-scale 3D foundation pit models in engineering practice and having good engineering practical value. Attached Figure Description

[0054] Figure 1 This is an overall flowchart of the method in one embodiment of the present invention; Figure 2 This is a schematic diagram of the construction of the node-unit heterogeneous diagram and edge features of the foundation pit soil in one embodiment of the present invention; Figure 3 This is a schematic diagram of the conditional heterogeneous spatiotemporal graph network structure during the foundation pit excavation stage in one embodiment of the present invention; Figure 4 This is a schematic diagram of the combined prediction of soil displacement and stress in a foundation pit according to one embodiment of the present invention; Figure 5 This is a schematic diagram of soil mechanics consistency training and space-time sampling in one embodiment of the present invention. Detailed Implementation

[0055] To enable those skilled in the art to better understand the present application, the technical solution of the present application will be further described in detail below with reference to the embodiments and accompanying drawings.

[0056] See Figure 1 , Figure 1 This is a flowchart of the present invention. This embodiment uses a phased excavation project for foundation pits as the application object. The foundation pit project includes the soil region, retaining and supporting structures, foundation pit boundary conditions, and surface loads or construction loads. The foundation pit finite element response prediction method of the present invention is used to predict the soil nodal displacement field and soil element stress field at different excavation stages. Its overall process includes: S1: Obtain three-dimensional finite element sample data for the foundation pit excavation project.

[0057] Specifically, it includes: S1-1: Sample data obtained from the three-dimensional finite element model (coupled finite element model of foundation pit soil-retaining and supporting structure) or finite element calculation results of the foundation pit excavation project. This data may include soil node numbers, three-dimensional coordinates of soil nodes, soil element numbers, soil element connection relationships, soil element types, soil material parameters, foundation pit boundary conditions, surface or load state information at each construction stage, construction stage information, contact and constraint relationships between the retaining and supporting structure and the soil, activation state information of the retaining and supporting structure, and soil finite element response results (including displacement labels for some soil nodes and stress labels for some soil elements). Among these, soil nodes... The three-dimensional coordinates are as follows The centroid of the soil element is calculated based on the three-dimensional coordinates of the soil nodes and the connection relationship of the soil elements. During the construction phase The displacement label value is denoted as Soil unit During the construction phase The stress label value is denoted as ,in and These represent the soil node and element number, respectively. and These represent displacement and stress, respectively. This is a construction phase number.

[0058] S1-2: Perform a consistency check on soil node coordinates, soil element connection relationships, and the result field to ensure that the same soil node number, soil element number, and construction stage number remain consistent across geometric information, material information, boundary load information, and response labels. For the coordinate scales of different foundation pit engineering samples, calculate the reference length using the formula: The coordinates of soil nodes and the centroids of soil elements are then scaled. , .in, This represents the center of the bounding box of the finite element sample space for the foundation pit. This process ensures that the model inputs for different foundation pit engineering dimensions are within a similar numerical range, while preserving the three-dimensional relative geometric relationships of the foundation pit soil.

[0059] S1-3: Convert the boundary constraints, nodal loads, or equivalent surface loads of the soil nodes in the foundation pit into structural features. The soil node input features can be written as follows: .in, This indicates the fixed constraint direction, the given displacement boundary direction, and the displacement boundary value. Indicates the nodal force or equivalent load and its direction. Representing soil nodes During the construction phase The effective state is 0 or 1. This feature makes boundary conditions and load paths no longer implicitly present in the data file, but rather enter the model as learnable inputs.

[0060] S1-4: The three-dimensional finite element response results of the foundation pit excavation are organized into soil nodal displacement labels and soil element stress labels. The soil nodal displacement labels are as follows: The soil element stress label is This labeling method is consistent with the physical quantity definitions in the three-dimensional finite element method for foundation pits: displacement belongs to the soil nodal field, and stress belongs to the soil element field, providing a foundation for subsequent node-element heterogeneous diagram modeling from the data organization level.

[0061] S2: Based on the three-dimensional finite element sample data, construct a node-element heterogeneous diagram to characterize the topological relationship of the finite element mesh of the foundation pit soil; the node-element heterogeneous diagram includes soil node entities, soil element entities, node-node edges, element-element edges, node-to-element edges, and element-to-node edges.

[0062] Specifically, it includes: See Figure 2 The nodes in the three-dimensional finite element mesh of the foundation pit are abstracted as Abstracting the soil unit as and using the formula Forming a heterogeneous graph, in which , The enclosure and support structures are entered into the node-element heterogeneous diagram through the activation state of the support structure, material area properties, contact or constraint edge characteristics, and boundary condition characteristics.

[0063] Node-node edge construction is performed based on the finite element mesh connection relationship of the foundation pit soil. If soil node i and soil node j belong to the same soil element, or are the two endpoints of the actual finite element mesh edge, then node-node edge construction is established. and Nodes—node edges are used to express the relationships between local soil displacement coordination, boundary constraint transmission, and the diffusion of surface loads or construction loads.

[0064] Construct element-element edges based on the adjacency relationships of soil elements. If soil element a and soil element b share a face, an edge, or a point, then establish element-element edges. and The adjacency type is represented using a one-hot vector: ,in and These represent shared surfaces, shared edges, and shared points, respectively. When multiple sharing relationships exist simultaneously, shared surfaces are prioritized, followed by shared edges, and finally shared points. This priority aligns with the engineering common sense that in three-dimensional finite element methods for foundation pit soil, surface adjacency provides more direct stress transfer than point adjacency.

[0065] Construct node-to-element edges and element-to-node edges based on the soil element connection table. If soil node i belongs to the connection table of soil element a, then construct... and The node-to-element edge is used to learn the influence of soil nodal displacement, boundary conditions, and construction loads on the strain and stress state of soil elements; the element-to-node edge is used to learn the feedback of soil element material characteristics, regional properties, and the element implicit state vector formed by message passing on the prediction of soil nodal displacement.

[0066] A dynamic activation diagram is constructed based on the foundation pit excavation construction stage. For soil nodes, soil elements, and edges, separate settings are configured. , and This represents the dynamic activation state value during construction phase t. If a soil element is in the excavation and removal phase, the dynamic activation state value of that soil element and its related edges is 0 in the current phase. If the retaining and support structures are completed or begin to bear load in a certain phase, the dynamic activation state values ​​of their related constraints, adjacent soil elements, and edges change from 0 to 1. This mechanism allows the graph structure to dynamically change with the construction phase of the foundation pit. Before performing message aggregation within the relation, the effective message propagation path for the current construction phase is determined based on the activation states of the source entity, target entity, and edges. This ensures that changes in the dynamic graph are reflected not only in changes in input features but also in changes in the soil entities and mechanical relationships involved in the propagation.

[0067] S3: Based on the three-dimensional finite element sample data and the node-element heterogeneous diagram, generate three-dimensional geometric-mechanical edge features; the three-dimensional geometric-mechanical edge features are used to characterize the displacement coordination relationship between soil nodes, the stress transfer relationship between soil elements, and the displacement-stress coupling relationship between soil nodes and soil elements.

[0068] Specifically, it includes: See Figure 2 For node-to-node edges, calculate the relative positions, distances, and boundary load relationships of the soil nodes, and then... Organization is an edge feature. Among them, For normalized distance, and This is a boundary load feature for soil nodes; this edge is used to express the coordination of local displacements and the propagation of boundary effects in the foundation pit soil. This feature enables node-to-node messages to distinguish between ordinary soil geometric neighbors, nodes with common boundary constraints, nodes in the common load zone, and nodes near the soil material interface.

[0069] For element-element edges, calculate the relative position vector of the soil element centroid, sharing type, sharing surface normal, and sharing area. The sharing area ratio or adjacency strength can be calculated based on the shared area and the target soil element surface area. Organization is used as an edge feature to enhance the ability of cell-to-cell messages to identify soil material interfaces, interfaces affecting retaining and supporting structures, and areas near the excavation face of the foundation pit.

[0070] For nodes to cell edges Edge features are constructed using the relative positions of soil nodes to the centroids of soil elements, local numbering, local coordinates, and shape function weights. The core idea is that the displacement field of soil nodes influences the strain and stress of soil elements through local interpolation relationships. Therefore, the edge features record not only distance but also... , and The local information of the finite element allows the network to learn local correlations similar to those in finite element shape function interpolation.

[0071] For cell to node edge By adding soil element material parameters to the node-element edge features, the soil element material characteristics and regional properties can be fed back into the soil node displacement prediction. Thus, the node channel and the element channel form a closed loop through bidirectional coupling edges, rather than predicting independently.

[0072] The soil nodal feature vectors and soil element feature vectors are organized as follows: , .in, Used for soil node coding Used for soil element encoding. The above feature vector design enables the model input to simultaneously include the soil geometry of the foundation pit, soil material, boundaries, loads, and activation information of the excavation stage.

[0073] S4: Perform conditional encoding on the construction stage information and construction state vector to obtain the time-construction stage embedded feature vector.

[0074] Specifically, it includes: Based on construction stage information (including construction stage category and normalized time) and construction state vector, using the formula Obtain the time-construction stage embedded feature vector.

[0075] Will This serves as a conditional input for subsequent heterogeneous message passing and time refinement modules. In this embodiment, Used to generate stage gating and FiLM modulation parameters; the time-construction stage embedded feature vectors corresponding to continuous construction stages are also used to determine the construction stage attention bias matrix in the Transformer time refinement module. Phase gating is used to control the propagation strength of valid messages in the current construction phase, FiLM modulation is used to scale and translate aggregated messages, and the construction phase attention bias matrix is ​​used to adjust the temporal attention scores between different construction phases.

[0076] A dynamic activation mask is generated based on the construction state vector (composed of soil excavation state, retaining and support structure activation state information, material region activation state, foundation pit boundary conditions, and load state information from the 3D finite element sample data). The dynamic activation mask, in conjunction with the activation states of soil nodes, soil elements, and edges, ensures that information propagation in the network at the current excavation stage occurs only on valid entities and relationships. Compared to simply appending stage numbers to the input, this step directly incorporates changes in the foundation pit construction stage into the information flow within the model.

[0077] S5: Input the soil node feature vector, soil element feature vector, three-dimensional geometric-mechanical edge features, and time-construction stage embedded feature vector into the stage-conditional heterogeneous spatiotemporal graph neural network model. The stage-conditional heterogeneous spatiotemporal graph neural network model performs heterogeneous message passing through stage gating, attention aggregation, and FiLM modulation, and performs time refinement based on the Transformer time refinement module that introduces construction stage attention bias to obtain node representation vector and element representation vector.

[0078] Specifically, it includes: See Figure 3 The soil node features (which can be formed by the three-dimensional coordinates of the soil nodes, the boundary conditions of the foundation pit, and the load state information), soil element features (which can be formed by the centroid of the soil element and the material parameters of the soil layer), and edge features are respectively input into the node encoder, the element encoder, and the relation-specific edge encoder to obtain... , and .

[0079] Calculate messages for each of the four types of relations. Using the edges on relation r... For example, the message calculation formula is: This formula simultaneously incorporates source entity, target entity, edge features, and stage embedding. That is, the message in this embodiment is not determined solely by the graph topology, but jointly by the three-dimensional geometric-mechanical relationship of the foundation pit soil and the excavation construction stage.

[0080] This step adjusts the aggregated messages according to their stage conditions using stage gating and FiLM modulation. Dynamic activation masks first eliminate invalid propagation paths, and stage gating then applies learnable weighting to valid messages.

[0081] Using formula Controlling message propagation, among which Represents the edges after gating. The message vector on; Indicates the construction stage side The gating coefficient on, , , and Representing the source entity respectively Target Entity and edge During the construction phase The dynamic activation state value; For trainable gated weight matrix, Use the Sigmoid activation function; Attention aggregation is performed on the similar neighbors of the target entity q, and attention weights are applied. Determined jointly by entity hidden state vectors, edge features, and construction stage embeddings, this feature characterizes the contribution of different adjacent soil nodes, adjacent soil elements, and node-element coupling relationships to soil displacement and stress prediction at the current excavation stage. The relation-level aggregation message is written as... This aggregation process can adaptively distinguish the importance of different adjacent edges. For example, in soil element-soil element messages, shared surface adjacencies may receive higher weights; in soil node-soil element messages, nodes near critical foundation pit boundaries, support structure influence zones, or construction load areas may contribute more to soil element stress prediction.

[0082] Soil node channel adopts formula Aggregation node—aggregation message vector on the node edge and aggregation message vector from the cell to the node edge; Soil unit channel uses formula Aggregation cell—aggregation message vector on cell edge and aggregation message vector from node to cell edge; Time-construction phase embedded feature vector Generate FiLM scale parameters and bias parameters and utilize FiLM modulation is applied to the aggregated messages. This step allows aggregated messages of the same edge type to have different scales and offsets at different excavation stages. For example, soil element-soil element messages related to the excavation face can be enhanced during the excavation stage, while messages related to removed soil elements can be suppressed.

[0083] use Update the hidden state vectors of nodes or elements. After completing multi-level updates, the corresponding hidden state vectors of nodes and elements are obtained at each construction stage, which are used in the next step to construct the sequence of hidden state vectors of nodes and elements for the continuous construction stages. This update method maintains the stability of the residuals while enabling bidirectional message coupling between soil node channels and soil element channels.

[0084] After completing the multi-layer heterogeneous message passing, the node hidden state vectors of the same soil node and the element hidden state vectors of the same soil element under the continuous construction stage are respectively formed into node hidden state vector sequences and element hidden state vector sequences according to the construction stage time sequence, and then input into the Transformer time refinement module. For any of the hidden state vector sequences, query mapping, key mapping, and value mapping are performed according to the hidden state vector sequence to obtain the query matrix. Key matrix Sum matrix The attention bias matrix for each construction stage is determined based on the time-stage embedded feature vectors corresponding to the continuous construction stages. and using the formula Temporal attention calculations are performed. The temporal attention results are processed through a feedforward network, residual connections, and layer normalization, and then refined temporally through multi-layer stacking to obtain the node representations corresponding to each construction stage. and unit representation .

[0085] S6: Based on node representation vectors and element representation vectors, jointly predict the three-dimensional displacement vectors of soil nodes and the three-dimensional stress tensors of soil elements during foundation pit excavation. Specifically: See Figure 4 By inputting the time-refined soil node representation into the displacement prediction head, the three-dimensional displacement vector of the soil nodes during the foundation pit excavation stage is obtained. The predicted displacement vectors of all soil nodes are recombined to obtain the overall soil node displacement field. This displacement field can be used for post-processing in foundation pit engineering, as well as for calculating the strain required to determine constitutive consistency loss.

[0086] Inputting the time-refined soil element representation into the stress prediction head yields the six-dimensional stress vector of the soil element during the foundation pit excavation stage. .

[0087] Reconstruct the six components into a three-dimensional symmetric stress tensor. Then, all soil elements are reorganized to obtain the stress field of the soil elements.

[0088] This step may also include: Engineering derivatives, including von Mises stress and principal stresses, are calculated from the predicted soil stress tensor. Von Mises stress is calculated using the formula... ,in The equivalent stress invariant is used. Principal stresses are calculated using the eigenvalues ​​of the soil stress tensor. .

[0089] Both of the above-mentioned derived quantities are determined by the soil stress tensor, thus avoiding inconsistencies between the predicted stress components and the evaluation indicators of the foundation pit project.

[0090] The prediction results are used to characterize the soil displacement field, soil stress field, and response results required for engineering safety assessment during the foundation pit excavation process.

[0091] Using the soil node displacement labels and soil element stress labels from step S1, displacement monitoring loss and stress monitoring loss can be calculated respectively, and combined with engineering derived quantity loss and physical consistency loss, the stage-conditional heterogeneous spatiotemporal graph neural network model is trained. Specifically: See Figure 5 Construct a joint training loss. The total loss adopts... The Huber loss can reduce the unstable impact of abnormally high response points in the pit on training, while retaining the ability to penalize high-error samples.

[0092] The physical consistency loss is expressed by the formula ,in, and These are, respectively, boundary condition loss, constitutive consistency loss, and equilibrium residual loss; and These are the activation coefficients for boundary condition loss, constitutive consistency loss, and equilibrium residual loss, respectively.

[0093] Boundary condition losses are introduced for the fixed boundary and given displacement boundary of the foundation pit, and are uniformly written as... Where B is the set of boundary nodes of the foundation pit. For the boundary degree of freedom mask, Given a displacement value, the boundary condition loss ensures that the predicted soil displacement is close to zero in the fixed constraint direction and close to the known boundary value in the given displacement direction.

[0094] When the soil element strain-displacement matrix and soil material matrix are reliable, constitutive consistency loss is introduced, and its core relationship is as follows: Constitutive uniform stress is calculated from the soil material matrix: Constitutive consistency loss is denoted as .

[0095] in, The determination is jointly made by the soil material type, the excavation stage of the foundation pit, and the availability of the operator. For strongly nonlinear, contact, or path-dependent soil materials, if the equivalent material matrix... If unreliable, then let Thus in The constitutive consistency loss term is closed in the middle.

[0096] When the surface normal, area, body forces, and boundary forces of the soil element are reliable, and equilibrium residual loss is introduced, the local equilibrium residual of the soil element is: .

[0097] The loss of the balanced residual is denoted as And by balancing the residual availability mask Controls are enabled or disabled. This loss represents the discrete form of the local force equilibrium of the soil element, used to constrain the predicted soil stress field to meet the basic force equilibrium requirements; when the relevant physical operators are unreliable, it is used through... Turn this option off.

[0098] The total loss is obtained by weighted summation of the above losses: During training, the network parameters are updated using a gradient descent-like optimization algorithm to gradually reduce the total loss. (The weights are listed below.) , , and It can be set according to the key tasks of the foundation pit project, or it can be activated in stages.

[0099] For large-scale foundation pit 3D finite element samples, a space-time sampling training strategy is adopted. The central soil region participates in the supervised loss, while the halo region only provides message passing context; the subgraph loss can be written as... .

[0100] At the same time, a time window is adopted. Training was conducted to increase the sampling probability of frames switching during the foundation pit excavation stage, peak soil response frames, and construction load abrupt change frames. Central region supervision, halo region context supplementation, and priority sampling of key time frames worked together to preserve the local spatial correlation of the foundation pit soil and the evolution information of the construction stages under limited memory conditions.

[0101] Therefore, this embodiment completes the three-dimensional finite element response proxy modeling of the foundation pit in the following order: “distinguishing the heterogeneous physical roles of soil nodes and units in the foundation pit → screening effective paths in the excavation stage → conditional propagation of effective messages → time refinement in the continuous excavation stage → joint prediction of soil displacement and stress fields → conditional mechanical constraints”. Each step together serves the consistent prediction of soil displacement and stress fields during the foundation pit excavation process.

[0102] After model training, the geometry, soil material, foundation pit boundary conditions, surface load or construction load status information at each construction stage, construction stage information, contact and constraint relationships between the retaining and supporting structures and the soil, and activation status information of the retaining and supporting structures of the foundation pit engineering sample to be predicted are processed according to steps S1 to S4, and input into the trained stage-conditional heterogeneous spatiotemporal graph neural network model to obtain output results such as soil nodal displacement field, soil element stress field, von Mises stress, and principal stress. The output results can be further exported to CSV, VTK, or VTU formats for error statistics, 3D cloud map display of foundation pit soil, and engineering safety assessment.

[0103] Those skilled in the art should understand that, without departing from the core technical concept of predicting soil stress and displacement fields during foundation pit excavation, the specific message passing network can be replaced with GAT, MPNN, GraphSAGE or other heterogeneous graph neural network forms, the Transformer temporal refinement can be replaced with GRU, temporal convolution or other sequence modeling modules, and the edge features, loss weights, sampling strategies and derivation calculation methods can be adjusted according to the foundation pit engineering scenario.

[0104] This invention focuses on a coupled finite element model of soil, retaining structure, and support structure during foundation pit excavation. It utilizes a node-element heterogeneous graph and four types of edge relationships with different mechanical meanings to represent soil node displacement, soil element stress, soil material state, and bidirectional coupling information. Based on the foundation pit construction stage, an effective message propagation path is formed, suppressing non-physical propagation caused by excavated soil elements, inactive retaining and support structure entities, and invalid edge relationships before message aggregation. Further, stage gating, attention bias, and FiLM modulation are used to modify the propagation patterns under different excavation stages along the effective path, and time-refined modeling is employed to model the foundation pit construction process response evolution. By jointly predicting the three-dimensional displacement of soil nodes and the six-dimensional stress vector of soil elements, deterministically calculating engineering derivatives, and using openable / closeable mechanical consistency constraints, the consistency and reliability of the predicted soil stress and displacement fields are improved. Spatial subgraphs, k-hophalo, and time window sampling support the training and rapid inference of large-scale three-dimensional finite element samples of foundation pits.

[0105] Example 2: This embodiment provides a spatiotemporal diagram-based finite element response prediction system for foundation pits, used to implement the spatiotemporal diagram-based finite element response prediction method for foundation pits described in Embodiment 1, including: The data acquisition module is used to acquire three-dimensional finite element sample data for foundation pit excavation projects; The graph construction module is used to construct a node-element heterogeneous graph to characterize the topological relationship of the finite element mesh of the foundation pit soil based on the three-dimensional finite element sample data. The node-element heterogeneous graph includes soil node entities, soil element entities, node-node edges, element-element edges, node-to-element edges, and element-to-node edges. The feature generation module is used to generate three-dimensional geometric-mechanical edge features based on the three-dimensional finite element sample data and the node-element heterogeneous diagram; the three-dimensional geometric-mechanical edge features are used to characterize the displacement coordination relationship between soil nodes, the stress transfer relationship between soil elements, and the displacement-stress coupling relationship between soil nodes and soil elements. The time-construction stage embedding feature coding module performs conditional coding on construction stage information and construction state vector to obtain the time-construction stage embedding feature vector. The graph neural network model calculation module is used to input the feature vectors of soil nodes and soil elements, three-dimensional geometric-mechanical edge features, and time-construction stage into the stage-conditional heterogeneous spatiotemporal graph neural network model. The stage-conditional heterogeneous spatiotemporal graph neural network model performs heterogeneous message passing through stage gating, attention aggregation, and FiLM modulation, and performs time refinement based on the Transformer time refinement module that introduces construction stage attention bias to obtain node representation vectors and element representation vectors. The joint prediction module is used to jointly predict the three-dimensional displacement vector of soil nodes and the three-dimensional stress tensor of soil elements during the foundation pit excavation process, based on the node representation vector and the element representation vector.

[0106] Example 3: This embodiment provides an electronic device, including: a memory and a processor; The memory is used to store computer programs; The processor is configured to invoke the computer program to execute the method as described in Embodiment 1.

[0107] Example 4: This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is run on an electronic device, it causes the electronic device to perform the method described in Embodiment 1.

[0108] Example 5: This embodiment provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described in Embodiment 1.

[0109] The specific implementation of the system, electronic device, computer-readable storage medium, and computer program product provided in this application can be referred to the specific embodiments of the above methods, and will not be repeated here.

[0110] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0111] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A finite element method for predicting the response of a foundation pit based on a spatiotemporal diagram, applied to the prediction of soil stress and displacement fields during foundation pit excavation, characterized in that... Includes the following steps: S1: Obtain three-dimensional finite element sample data for the foundation pit excavation project; S2: Based on the three-dimensional finite element sample data, construct a node-element heterogeneous diagram to characterize the topological relationship of the finite element mesh of the foundation pit soil; the node-element heterogeneous diagram includes soil node entities, soil element entities, node-node edges, element-element edges, node-to-element edges, and element-to-node edges; S3: Generate three-dimensional geometric-mechanical edge features based on the three-dimensional finite element sample data and the node-element heterogeneous diagram; The three-dimensional geometric-mechanical edge features are used to characterize the displacement coordination relationship between soil nodes, the stress transfer relationship between soil elements, and the displacement-stress coupling relationship between soil nodes and soil elements. S4: Conditionally encode the construction stage information and construction state vector to obtain the time-construction stage embedded feature vector; S5: The feature vectors of soil nodes and soil elements, 3D geometric-mechanical edge features, and time-construction stage embedded feature vectors are input into a stage-conditional heterogeneous spatiotemporal graph neural network model. This model uses stage gating, attention aggregation, and FiLM modulation for heterogeneous message passing, and performs time refinement based on a Transformer time refinement module that introduces construction stage attention bias, resulting in node representation vectors and element representation vectors. Specifically, this includes: In step S5, the formula is used. Conditional message vectors are generated during the generation phase, where Indicates the type is edge The phase conditional message vector on, Represents the source entity The hidden state vector. Represents the target entity The hidden state vectors of soil nodes and soil elements are obtained by mapping the feature vectors of soil nodes and soil elements through the soil node encoder and soil element encoder, respectively. Indicates the type is edge The implicit representation vector, by Generated via edge encoder mapping; Indicates the type of membership. The message generation function for the edge; Using formula Controlling message propagation, among which Represents the edges after gating. The message vector on; Indicates the construction stage side The gating coefficient on, , , and Representing the source entity respectively Target Entity and edge During the construction phase The dynamic activation state value; For trainable gated weight matrix, Use the Sigmoid activation function; Using formula To achieve attention convergence, among which Represents the target entity In the category of The aggregated message vector on the edge, For the type edge Attention weight coefficients on For target entity The relationship type is The set of neighbors; Soil node channel adopts formula Aggregation node—aggregation message vector on the node edge and aggregation message vector from the cell to the node edge; Soil unit channel uses formula Aggregation cell—aggregation message vector on cell edge and aggregation message vector from node to cell edge; Embedded feature vectors based on time-construction stage Generate FiLM modulation scale parameters and bias parameters Using the formula The aggregated message is subjected to FiLM modulation to obtain the FiLM-modulated aggregated message vector. ; Using formula Update the hidden state vector of the node or the hidden state vector of the cell; where... and They represent the first Layer and first The hidden state vector of the target entity in the layer; Number the network layers. For state update functions, For layer normalization function; After completing heterogeneous message passing, the node hidden state vectors of the same soil node and the element hidden state vectors of the same soil element under continuous construction stages are respectively formed into node hidden state vector sequences and element hidden state vector sequences according to the construction stage time sequence, and then input into the Transformer time refinement module. For any of the hidden state vector sequences, the Transformer time refinement module performs query mapping, key mapping, and value mapping on the hidden state vector sequence through different trainable mapping matrices to obtain the query matrix. Key matrix Sum matrix According to the current construction stage The corresponding time-phase embedding feature vectors determine the attention bias matrix for the construction phase. Each matrix element represents the bias effect of the stage association between two corresponding construction stages on the time attention score; using the formula Perform time attention calculations, where The dimension is the key vector; the result of the temporal attention calculation. After processing with feedforward network, residual connection and layer normalization, the time-refined node representation vector and element representation vector corresponding to each construction stage are obtained; S6: Based on node representation vectors and element representation vectors, jointly predict the three-dimensional displacement vectors of soil nodes and the three-dimensional stress tensors of soil elements during the foundation pit excavation process.

2. The finite element response prediction method for foundation pits according to claim 1, characterized in that, In step S2, the formula is used. Construct a node-cell heterogeneous graph, in which , For the set of soil node entities, For a collection of soil unit entities, Let be a set of edge types, where , , , These represent node-to-node edge, element-to-element edge, node-to-element edge, and element-to-node edge, respectively. If two nodes belong to the same soil element or the endpoint of the real mesh, a node-to-node edge is established between these two nodes. If two soil elements share a face, an edge, or a point, an element-to-element edge is established between these two soil elements. If a soil node belongs to the soil element connection table, a node-to-element edge and an element-to-node edge are established.

3. The finite element response prediction method for foundation pits according to claim 2, characterized in that, In step S3, the general formula is used. Construct three-dimensional geometric-mechanical edge features, where, Indicates the type is edge eigenvectors; and Let them represent the source entity and the target entity of the edge, respectively. This represents a directed edge from source entity s to target entity q. Indicates the type of edge relationship; This represents the normalized geometric quantities in the finite element mesh of the foundation pit soil. This indicates the influencing properties such as soil material, region, and retaining and supporting structures. This indicates information about the foundation pit boundary, load status, and construction stage. For the type The edge-specific feature constructor is used to concatenate and normalize the input parameters.

4. The finite element response prediction method for foundation pits according to claim 3, characterized in that, In step S4, the formula is used. Generate construction phase Time-construction stage embedded feature vector ,in For stage coding functions, , and These represent the construction stages. The category code, normalized time value, and construction state vector.

5. The finite element response prediction method for foundation pits according to claim 4, characterized in that, In step S6, the formula is used. Predict the three-dimensional displacement vector of soil nodes. Indicates the construction stage Lower soil node The predicted three-dimensional displacement vector, where Indicates the construction stage Lower soil node The time-refined node representation vector, For displacement prediction function; using the formula Predict the six-dimensional stress vector of the soil element, where, Indicates the construction stage Lower soil unit The predicted six-dimensional stress vector, Indicates the construction stage Lower soil unit Time-refined unit representation vector, Represents the stress prediction function; The three-dimensional symmetric stress tensor of the soil element is reconstructed from the six stress components of the predicted six-dimensional stress vector of the soil element. And using the formula Calculate the von Mises stress; where, This represents the predicted von Mises stress value. To predict the six-dimensional stress vector A function for calculating the von Mises stress square term. To prevent abnormally positive numbers from being extracted.

6. The finite element response prediction method for foundation pits according to claim 5, characterized in that, The total training loss function of the stage-conditional heterogeneous spatiotemporal graph neural network model is: ,in and These are, respectively, the soil nodal displacement monitoring loss, the soil element stress monitoring loss, the von Mises stress loss, and the physical consistency loss. and All are non-negative loss weights; the physical consistency loss is calculated using the formula... ,in, and These are the activation coefficients for boundary condition loss, constitutive consistency loss, and equilibrium residual loss, respectively. Their values ​​are determined based on whether the corresponding physical constraints are calculable. and These are the boundary condition loss, constitutive consistency loss, and equilibrium residual loss, respectively.

7. The finite element response prediction method for foundation pits according to any one of claims 1 to 6, characterized in that, The training process of the stage-conditional heterogeneous spatiotemporal graph neural network model adopts a joint strategy of spatial subgraph sampling, k-hop halo expansion, and temporal window sampling. The spatial subgraph sampling strategy is as follows: A central node set is sampled from the set of soil nodes in the foundation pit, and a central soil element set is determined; these two constitute a central subgraph. Starting from the central subgraph, a neighborhood of k layers is expanded outward along the edges of each type of the node-element heterogeneous graph to obtain the halo region. The supervision loss is applied only to the set of soil nodes within the central subgraph. and soil element set In the above calculation, entities within the halo region are used to provide the spatial context for message passing. The time window sampling strategy is as follows: sample time windows from the complete construction phase sequence, and perform time attention calculations on the continuous construction phases within the time window.

8. A finite element response prediction system for foundation pits based on spatiotemporal diagrams, characterized in that, The method for predicting the finite element response of a foundation pit based on a spatiotemporal diagram, as described in any one of claims 1 to 7, comprises: The data acquisition module is used to acquire three-dimensional finite element sample data for foundation pit excavation projects; The graph construction module is used to construct a node-element heterogeneous graph to characterize the topological relationship of the finite element mesh of the foundation pit soil based on the three-dimensional finite element sample data. The node-element heterogeneous graph includes soil node entities, soil element entities, node-node edges, element-element edges, node-to-element edges, and element-to-node edges. The feature generation module is used to generate three-dimensional geometric-mechanical edge features based on the three-dimensional finite element sample data and the node-element heterogeneous diagram; the three-dimensional geometric-mechanical edge features are used to characterize the displacement coordination relationship between soil nodes, the stress transfer relationship between soil elements, and the displacement-stress coupling relationship between soil nodes and soil elements. The time-construction stage embedding feature coding module performs conditional coding on construction stage information and construction state vector to obtain the time-construction stage embedding feature vector. The graph neural network model calculation module is used to input the feature vectors of soil nodes and soil elements, three-dimensional geometric-mechanical edge features, and time-construction stage into the stage-conditional heterogeneous spatiotemporal graph neural network model. The stage-conditional heterogeneous spatiotemporal graph neural network model performs heterogeneous message passing through stage gating, attention aggregation, and FiLM modulation, and performs time refinement based on the Transformer time refinement module that introduces construction stage attention bias to obtain node representation vectors and element representation vectors. The joint prediction module is used to jointly predict the three-dimensional displacement vector of soil nodes and the three-dimensional stress tensor of soil elements during the foundation pit excavation process, based on the node representation vector and the element representation vector.

Citation Information

Patent Citations

  • Pipeline structure stress proxy model construction method based on deep learning

    CN119558135A

  • Rail transit foundation pit support modeling method based on multi-source data fusion

    CN120562009A