A shear wall design method based on graph node classification graph neural network

Abstract

The application discloses a kind of based on graph node classification graph neural network shear wall design method, specifically related to graph node classification technical field;Symmetrical node pair in shear wall is identified by geometric coordinates and connection relationship, and it is marked in graph structure, then the shear force response data of these nodes under horizontal load and the material creep data caused by stress difference under long-term load are obtained, which are converted into feature vectors, input machine learning model for training, judge the accuracy of stress distribution prediction, according to the prediction result, the symmetric nodes are divided into accurate, possibly accurate and inaccurate prediction results, for the possibly accurate node, extract local subgraph and dynamically adjust edge weight according to the degree of abnormal prediction, to further optimize the prediction accuracy of model, through the accurate identification and dynamic optimization of symmetric node, effectively solve the problem that the mechanical performance of symmetric node in shear wall is not accurately predicted in the prior art, improve the prediction accuracy and design reliability of model.

Description

Technical Field

[0001] The present invention relates to the technical field of graph node classification, and in particular to a shear wall design method based on a graph node classification graph neural network. Background Art

[0002] Graph neural networks based on graph node classification for shear wall design model the mechanical properties of building structures (such as shear walls) as a graph structure, where nodes represent key points or components of the shear wall, and edges represent the connections or mechanical interactions between them. By applying graph neural networks (GNNs), nodes can be classified and their mechanical properties, stress distribution, or other structural behaviors can be predicted, thereby optimizing the design of shear walls.

[0003] Prior art applications of graph neural networks based on graph node classification in shear wall design begin by modeling the shear wall structure as a graph, with nodes representing key points or components of the shear wall and edges representing the mechanical relationships between components. Next, data on the structural mechanical properties is collected and fed into the graph neural network model as node features. Subsequently, the graph neural network is used to classify the nodes and predict their mechanical performance or stress distribution. Finally, based on the classification results, the shear wall is optimized to identify structural weaknesses and enhance overall performance.

[0004] The existing technology has the following shortcomings:

[0005] In existing technologies, when using graph neural network models to classify nodes and predict their stress distribution, GNNs may not be able to effectively distinguish the mechanical performance of certain symmetrical parts of shear walls (such as bilaterally symmetrical wall sections). Even if these nodes actually carry different loads or stress conditions, the model may treat them as the same node, thereby producing the same prediction results. And because symmetrical structures are often related to critical mechanical equilibrium, if the symmetry problem is not handled properly, it may lead to misjudgment of key nodes, which in turn may cause weak links in the design to be overlooked. Especially in the case of complex boundary conditions or loads, the error may be amplified. Summary of the Invention

[0006] The purpose of the present invention is to provide a shear wall design method based on graph node classification graph neural network to address the shortcomings of the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a shear wall design method based on a graph node classification graph neural network, comprising the following steps:

[0008] S1: Determine whether the node pairs in the shear wall structure are symmetrical by comparing the geometric coordinates and connection relationships, mark the symmetrical node pairs identified in the shear wall structure, and construct a shear wall graph structure to keep the symmetrical node pairs in the same graph;

[0009] S2: After identifying the symmetrical nodes, obtain the horizontal shear response data of the symmetrical nodes of the wall under horizontal loads, as well as the material creep data generated based on the stress difference of the symmetrical nodes under long-term loads;

[0010] S3: Convert the acquired horizontal shear response data and material creep data into feature vectors, use the feature vectors as input to the machine learning model, train the machine learning model, and determine the accuracy of the predicted results of the symmetrical node stress distribution based on the model output results;

[0011] S4: The accuracy of the prediction results of the stress distribution of the symmetrical nodes is divided into different levels, namely accurate prediction results, possibly accurate prediction results and inaccurate prediction results;

[0012] S5: For accurate prediction results, the stress distribution of the node is directly used as the reference benchmark for shear wall design; for symmetrical nodes with inaccurate prediction results, the local graph structure is retrained;

[0013] S6: From the overall shear wall graph, extract each node that may accurately predict the result and its direct neighbor nodes, construct a local subgraph, and dynamically adjust the edge weights according to the abnormal degree of the accuracy of the prediction results of the symmetric node stress distribution.

[0014] Preferably, in S2, a shear response anomaly index is generated based on the shear force data at the extracted symmetrical nodes, wherein the shear response deviation index is obtained by:

[0015] Extract the shear response data of all symmetrical nodes in the shear wall and construct a feature matrix. The feature matrix of the shear response data of each symmetrical node i is composed of: the shear response τ of node i in the horizontal shear direction i,xy and the shear response τ of node i in the vertical shear direction i,xz , the shear response eigenvector f of each node i i is: f i =[τ i,xy ,τ i,xz ]; For N nodes, form a feature matrix X, X=[f1,f2,…,f N ] T ; Where X is an N×d-dimensional matrix, d is the dimension of the feature, T is the matrix transpose, and it is standardized. For the standardized feature matrix X′, its covariance matrix C is calculated, and the expression is: Where C is a d×d dimensional matrix, which represents the covariance between the shear response characteristics. The covariance matrix C is subjected to eigenvalue decomposition to obtain the eigenvalues ​​λ1,λ2,…,λ d and the corresponding eigenvectors v1,v2,…,v d ; The eigenvalue represents the variance contribution of each principal component direction, while the eigenvector represents the direction of the principal component. The expression is: Cv i =λ i v i ; Among them, the eigenvalue λ i Indicates the importance of the i-th principal component, v i is the corresponding eigenvector, and the principal components v1, v2, ..., v corresponding to the first g largest eigenvalues ​​are selected. g To reduce the dimension, the original feature vector f i Project it into the principal component space, and reconstruct the approximate value Sf of the original eigenvector through the principal component vector. i , calculate the difference between the original eigenvector and the reconstructed eigenvector, that is, calculate the shear response deviation index, the expression is: HK=||f i -Sf i ||2; where HK is the shear response deviation index.

[0016] Preferably, the creep rate anomaly index is generated after analyzing the creep rate of each node under long-term load, and the method for obtaining the creep rate anomaly index is:

[0017] Construct the creep rate characteristic vector of each node. For each node, its creep rate vector Q i Contains creep rates at multiple moments: in: Node i at time t n The creep rate of n is the number of time points. For N nodes, an N×n creep rate matrix Y is constructed. The Euclidean distance between each node is calculated. For node i and node j, the Euclidean distance d(i, j) is: Where d(i,j) represents the Euclidean distance between the creep rate feature vectors of nodes i and j. For each node i, a parameter k is selected to represent the number of nearest neighbor nodes to be considered; the k-distance d of node i is k (i refers to the distance from node i to its kth nearest neighbor node: Among them, N i Represents the k nearest neighbor nodes of node i, d(i,j) is the Euclidean distance from node i to node j, and Reach-dist is the reachable distance of node i to node j. k (i,j), the expression is: Reach-dist k (i,j)=max{dk (j), d(i, j)}; Calculate the local reachability density. The local reachability density of node i is the average LRD of the reciprocal of its reachability distance to its nearest neighbor node. k (i) The creep rate anomaly index is calculated based on the average value of the inverse of the calculated reachable distance, and the expression is: Where, LRD k (j) is the local reachability density of node j, LRD k (i) is the local reachability density of node i, and KM is the creep rate anomaly index.

[0018] Preferably, in S3, the acquired horizontal shear response data and material creep data are converted into feature vectors, the feature vectors are used as inputs of a machine learning model, the machine learning model is trained, and the accuracy of the prediction results of the symmetrical node stress distribution is determined based on the model output results, specifically:

[0019] The shear response deviation index and the creep rate anomaly index are converted into the first eigenvector, and the first eigenvector is used as the input of the machine learning model. The machine learning model uses the accuracy value label of the symmetric node stress distribution prediction result predicted by each group of first eigenvectors as the prediction target, and minimizes the sum of the prediction errors of the accuracy value labels of all symmetric node stress distribution prediction results as the training target. The machine learning model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The accuracy value of the symmetric node stress distribution prediction result is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

[0020] Preferably, in S4, the accuracy of the symmetric node stress distribution prediction results is divided into different levels, namely, accurate prediction results, possibly accurate prediction results, and inaccurate prediction results, specifically:

[0021] Comparing the obtained accuracy value of the symmetric node stress distribution prediction result with the gradient standard threshold, the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and comparing the accuracy value of the symmetric node stress distribution prediction result with the first standard threshold and the second standard threshold respectively;

[0022] If the accuracy value of the symmetric node stress distribution prediction result is greater than the second standard threshold, it indicates that the accuracy of the symmetric node stress distribution prediction result is high, and a high accuracy signal is generated at this time, and the symmetric node stress distribution prediction result under which the high accuracy signal is generated is classified as an accurate prediction result;

[0023] If the accuracy value of the symmetric node stress distribution prediction result is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it means that the accuracy of the symmetric node stress distribution prediction result is average, and a medium accuracy signal is generated at this time, and the symmetric node stress distribution prediction result under the medium accuracy signal is classified as a possible accurate prediction result;

[0024] If the accuracy value of the symmetric node stress distribution prediction result is less than the first standard threshold, it means that the accuracy of the symmetric node stress distribution prediction result is low. At this time, a low accuracy signal is generated, and the symmetric node stress distribution prediction result under the low accuracy signal is classified as an inaccurate prediction result.

[0025] Preferably, in S6, each node that may accurately predict the result and its direct neighboring nodes are extracted from the overall shear wall graph to construct a local subgraph. The edge weights are dynamically adjusted according to the abnormality of the accuracy of the prediction results of the symmetric node stress distribution, specifically as follows:

[0026] Symmetric node i is a node that may accurately predict the result. The node and its direct neighbor nodes are extracted to form a local subgraph. The node set V of the local subgraph is i Includes: V i ={i}∪{j| ( i,j ) ∈E}; where E is the edge set of the graph, and j is the neighbor node directly connected to node i;

[0027] In the local subgraph, the initial weight of each edge (i, j) is defined according to the geometric distance. is a fixed value, and the expression is: Where d(i,j) is the measure of the geometric distance between nodes i and j;

[0028] If the accuracy value of the symmetric node stress distribution prediction result is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, the symmetric node stress distribution prediction result generated under the medium accuracy signal is divided into a possible accurate prediction result, and the accuracy values ​​of the symmetric node stress distribution prediction results generated in a subsequent fixed time period that are greater than or equal to the first standard threshold and less than or equal to the second standard threshold are collected and a corresponding data set is established. The standard deviation of the data set is calculated as the abnormality degree of the stress distribution prediction, marked as A(i); the prediction accuracy adjustment factor α(i) of node i is defined according to A(i), which is negatively correlated with A(i), and the expression is: α ( i ) =1-A ( i ) ; Among them, α(i)∈[0,1], represents the accuracy of node prediction.

[0029] Preferably, in a local subgraph, the edge weight According to the dynamic adjustment of the abnormality of the stress distribution prediction results of the node, the edge weight is dynamically adjusted as follows by combining the prediction accuracy adjustment factor α(i) of node i and its neighbor node j: is the adjusted edge weight; after dynamically adjusting the edge weight, the weight matrix of the local subgraph is updated.

[0030] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0031] 1. This method compares the geometric coordinates and connection relationships of nodes, marks symmetrical nodes, and obtains detailed shear response and creep data. It then constructs feature vectors and inputs them into a machine learning model for training and accuracy prediction. The model output not only improves the accuracy of node stress distribution predictions but also, by grading the prediction results, provides a more accurate reference for shear wall structural design. This reduces misjudgments and improves the reliability of structural design, especially in situations with complex boundary conditions.

[0032] 2. By dynamically adjusting edge weights within the local graph structure, this method effectively improves the processing accuracy of nodes with potentially accurate predictions, ensuring that the model performs refined predictions and optimization around nodes with greater uncertainty. This approach dynamically adjusts edge weights by analyzing anomalies in creep rate and shear response. This not only improves the retraining effect on inaccurate nodes, but also reduces the risk of potential weak links in the overall shear wall structural design, thereby significantly enhancing the shear wall's overall mechanical resistance and long-term structural stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0034] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0036] For examples, see Figure 1 As shown, the shear wall design method based on the graph node classification graph neural network described in this embodiment includes the following steps:

[0037] S1: Determine whether the node pairs in the shear wall structure are symmetrical by comparing the geometric coordinates and connection relationships, mark the symmetrical node pairs identified in the shear wall structure, and construct a shear wall graph structure to keep the symmetrical node pairs in the same graph;

[0038] S2: After identifying the symmetrical nodes, obtain the horizontal shear response data of the symmetrical nodes of the wall under horizontal loads, as well as the material creep data generated based on the stress difference of the symmetrical nodes under long-term loads;

[0039] S3: Convert the acquired horizontal shear response data and material creep data into feature vectors, use the feature vectors as input to the machine learning model, train the machine learning model, and determine the accuracy of the predicted results of the symmetrical node stress distribution based on the model output results;

[0040] S4: The accuracy of the prediction results of the stress distribution of the symmetrical nodes is divided into different levels, namely accurate prediction results, possibly accurate prediction results and inaccurate prediction results;

[0041] S5: For accurate prediction results, the stress distribution of the node is directly used as the reference benchmark for shear wall design; for symmetrical nodes with inaccurate prediction results, the local graph structure is retrained;

[0042] S6: From the overall shear wall graph, extract each node that may accurately predict the result and its direct neighbor nodes, construct a local subgraph, and dynamically adjust the edge weights according to the abnormal degree of the accuracy of the prediction results of the symmetric node stress distribution.

[0043] Among them, in S1, the symmetry of the node pairs in the shear wall structure is determined by comparing the geometric coordinates and the connection relationship. The symmetrical node pairs identified in the shear wall structure are marked, and the graph structure of the shear wall is constructed to retain the symmetrical node pairs in the same graph.

[0044] First, it is necessary to obtain the geometric coordinates of all nodes in the shear wall structure. Geometric coordinates usually include the three-dimensional spatial position of each node, such as (x, y, z) coordinates. For shear wall design, common node coordinate information is generally two-dimensional plane coordinates (x, y) because shear walls are usually located in the building plane. Extract the geometric coordinate information of each node from the design model (such as a CAD model or a finite element model). If a finite element model is used, software such as ABAQUS, ANSYS, or ETABS can export the coordinate data of the structural nodes. The nodes and their geometric coordinate information are stored in a matrix form [N×2], where N is the number of nodes and the coordinates are the x and y values, respectively.

[0045] In graph neural networks, nodes are connected by edges, which represent mechanical associations or physical connections between components. Therefore, it is necessary to clarify the connection relationship between the nodes in the shear wall structure. Node connections can usually reflect the force transmission path. Construct the node adjacency matrix A, where A[i,j]=1 indicates that there is a connection between node i and node j, and A[i,j]=0 indicates that there is no connection. The connection relationship can be determined based on the topology of the structural diagram, which is usually exported by the design software or generated based on the meshing of the finite element method. The adjacency matrix can be generated and processed using a graph processing library such as NetworkX or DGL to ensure that the node connection relationship correctly reflects the mechanical connection.

[0046] Use the node's geometric coordinates and connection relationships to identify symmetrical node pairs in shear walls. Symmetrical node pairs typically have identical geometric characteristics (such as the axis of symmetry relative to the wall) and similar connection relationships. For symmetrical nodes, the geometric coordinates typically reflect their mirror image relationship about the axis of symmetry.

[0047] For example, assuming that the shear wall is symmetrical along the x-axis, the following conditions can be used to determine whether nodes i and j are symmetrical:

[0048] xi=-xj and yi=yj;

[0049] By comparing the geometric coordinates, find the node pairs that satisfy the above symmetry relationship.

[0050] After the geometric symmetry is determined, the similarity of these symmetric nodes is further verified through the connection relationship. Ideally, symmetric nodes should have similar connection topologies (the distribution and connection strength of adjacent nodes should be similar). For each pair of geometrically symmetric nodes, check the connection status in their adjacency matrix, that is, check whether their direct neighbors maintain a symmetric relationship with the neighbors of their geometrically symmetric points. For example, if node i and node j are identified as geometrically symmetric nodes, check whether the neighboring nodes ik and jk of the symmetric node pair also meet the symmetry. If not, it can be considered that the node pair does not meet the symmetry requirement.

[0051] The identified symmetric node pairs are retained in the overall graph structure of the shear wall. The constructed graph needs to reflect the geometric position, connection relationship and symmetry relationship of each node in the shear wall. Use graph processing tools such as NetworkX or DGL to import all nodes and the connection relationships between them into the graph. Add special marks to symmetric node pairs, such as assigning labels to symmetric node pairs through node attributes, so as to distinguish symmetric nodes from asymmetric nodes in subsequent model training. In the constructed graph, the feature vector of each node includes not only basic geometric information and connection information, but also symmetry relationships (such as "symmetric" or "asymmetric" labels). Visualize the constructed graph, check the distribution of symmetric node pairs in the graph, and verify whether the geometric symmetry and connection relationships are accurately reflected in the graph structure. Visualization tools such as Matplotlib and NetworkX's drawing function can be used to display nodes and connections. The verification process includes manually checking several important node pairs to ensure that their symmetry is correctly identified.

[0052] In this application, the symmetry of node pairs in a shear wall structure is determined by geometric coordinates and connection relationships, and symmetric nodes are marked to construct a shear wall structure containing symmetric nodes. This process ensures that the graph neural network can correctly identify and process symmetric nodes, thus taking symmetry into account when predicting the mechanical performance of nodes, avoiding misjudgment of key nodes and improving prediction accuracy.

[0053] S2: After identifying the symmetrical nodes, obtain the horizontal shear response data of the symmetrical nodes of the wall under horizontal loads, as well as the material creep data generated based on the stress difference of the symmetrical nodes under long-term loads.

[0054] In building structures (such as shear walls), horizontal loads usually include wind loads, earthquake effects, external impacts, etc. These loads will exert lateral forces on the wall, generating horizontal shear forces. Horizontal shear force refers to the force parallel to the plane of the wall, which will induce different force responses at symmetrical nodes. Determine the source and intensity of the horizontal load. For example, for seismic design, horizontal loads can be defined based on seismic parameters (such as peak acceleration, peak velocity, etc.); for wind loads, horizontal forces can be defined based on wind speed and wind pressure. Determine the direction and location of the load. Horizontal loads are usually applied in a direction parallel to the wall and can act on the top of the wall, around the opening, or on the entire structure.

[0055] Use a finite element model (FEM) to simulate the mechanical response of shear walls under horizontal loads. FEM can accurately calculate the stress, strain, and shear force at each node under external loads. Use finite element analysis software (such as ABAQUS, ANSYS, or ETABS) to create a three-dimensional or two-dimensional model of the shear wall. Define the material properties and boundary conditions of the nodes (such as the fixed relationship between the wall and the foundation), and set the load conditions. At the symmetrical node positions, define the displacement boundary conditions of the structure to ensure that the model reasonably reflects the actual engineering conditions. Apply horizontal loads to the wall structure. The applied loads can be static loads with linear growth or loads based on dynamic conditions, such as time-varying loads to simulate earthquakes.

[0056] Run a finite element simulation to calculate the horizontal shear response of each symmetric node under the applied load. Shear forces can be obtained from the node's force matrix or stress components, which represent the shear stress at the node in the horizontal plane. Use software processing tools (such as "Field Output" in ABAQUS or "Postprocessing" in ANSYS) to directly extract shear data at the symmetric nodes. The focus here is on extracting the shear response at the symmetric nodes.

[0057] The shear response anomaly index is generated based on the shear force data at the extracted symmetrical nodes, where the shear response deviation index is obtained as follows:

[0058] Extract the shear response data of all symmetrical nodes in the shear wall and construct a feature matrix. The feature matrix of the shear response data of each symmetrical node i is composed of: the shear response τ of node i in the horizontal shear direction i,xy and the shear response τ of node i in the vertical shear direction i,xz , the shear response eigenvector f of each node i i is: f i =[τ i,xy ,τ i,xz ]; For N nodes, form a feature matrix X, X=[f1,f2,…,f N ] T ; Where X is an N×d-dimensional matrix, d is the dimension of the feature, T is the matrix transpose, and it is standardized. For the standardized feature matrix X′, its covariance matrix C is calculated, and the expression is: Where C is a d×d dimensional matrix, which represents the covariance between the shear response characteristics. The covariance matrix C is subjected to eigenvalue decomposition to obtain the eigenvalues ​​λ1,λ2,…,λ d and the corresponding eigenvectors (i.e. principal components) v1, v2,…, v d ; The eigenvalue represents the variance contribution of each principal component direction, while the eigenvector represents the direction of the principal component. The expression is: Cv i=λ i v i ; Among them, the eigenvalue λ i Indicates the importance of the i-th principal component, v i is the corresponding eigenvector. In order to capture the main shear response characteristics, the principal components v1, v2, ..., v corresponding to the first g largest eigenvalues ​​are selected. g To reduce the dimension, the original feature vector f i Project it into the principal component space, and reconstruct the approximate value Sf of the original eigenvector through the principal component vector. i , calculate the difference between the original eigenvector and the reconstructed eigenvector, that is, calculate the shear response deviation index, the expression is: HK=||f i -Sf i ||2; where HK is the shear response deviation index.

[0059] A larger shear response deviation index indicates that the stress distribution at symmetrical nodes deviates significantly from the predicted results, indicating that the model's stress distribution predictions for these nodes are inaccurate. This deviation may indicate that the node exhibits unusual behavior under actual load conditions, such as concentrated stress or differences in material properties, which the model fails to capture. Therefore, a larger shear response deviation index indicates that these nodes warrant further attention, potentially requiring reinforcement design or model recalibration to improve prediction accuracy.

[0060] Conversely, when the shear response deviation index is smaller, the actual stress distribution at the symmetrical nodes closely matches the model prediction, indicating that the model's predictions at these nodes are highly accurate. This means that the model effectively captures the mechanical differences and stress states between symmetrical nodes, and the predictions are relatively reliable. In this case, the stress distribution at these nodes can be considered reasonable and can be directly used in the design and assessment of shear wall structures without additional modification.

[0061] In building structures, shear walls are often subjected to long-term loads such as deadweight, dead loads, and service loads. These loads exert continuous stress on the shear wall materials (such as concrete and steel), causing them to creep over time. Creep is the gradual, permanent deformation of a material under constant stress over time and typically manifests itself in three stages: an initial stage (rapid deformation), a secondary stage (slower, more stable deformation), and a final stage (accelerated deformation with potential material failure). Symmetrical nodes subjected to different stress states may experience different creep behaviors.

[0062] Use finite element analysis (such as ABAQUS, ANSYS, etc.) to simulate creep behavior under long-term loads. Input the creep characteristics of the material into the simulation model, and based on creep theory and empirical formulas (such as Norton-Bailey and Findley creep models), the creep data of the symmetrical nodes of the shear wall under long-term loads can be calculated. Common creep equations can describe the law of creep strain change over time. For example, the Norton-Bailey creep model can be expressed as: c( t ) =∈0+A·t n ; where ∈ c( t ) is the creep strain, ∈0 is the initial strain, A is the material constant, n is the creep exponent, and t is time.

[0063] By extracting the stress state of each symmetrical node (such as axial force, bending moment, shear force, etc.), the stress difference is calculated. Assume that the stress of symmetrical nodes i and j is σ i and σ j , stress difference Δσ ij It can be defined as: Δσ ij =|σ i -σ j |; Combined with the stress difference, the creep model of the material is used to calculate the creep rate of each node under long-term load.

[0064] After analyzing the creep rate of each node under long-term load, the creep rate anomaly index is generated. The method for obtaining the creep rate anomaly index is as follows:

[0065] Construct the creep rate characteristic vector of each node. For each node, its creep rate vector Q i Contains creep rates at multiple moments: in: Node i at time t n The creep rate of n is the number of time points. For N nodes, construct an N×n creep rate matrix Y; calculate the Euclidean distance between each node, which is used to measure the difference in creep rate between different nodes. For node i and node j, their Euclidean distance d(i,j) is: Where d(i,j) represents the Euclidean distance between the creep rate feature vectors of nodes i and j. For each node i, a parameter k is selected (usually 5 or 10), which represents the number of nearest neighbor nodes to be considered. The k-distance d of node i is k (i) refers to the distance from node i to its kth nearest neighbor node: Among them, N iRepresents the k nearest neighbor nodes of node i, d(i,j) is the Euclidean distance from node i to node j, and Reach-dist is the reachable distance of node i relative to node j. k (i,j), the expression is: Reach-dist k (i,j)=max{d k (j), d(i, j)}; the reachable distance is the larger value between the k-distance of node j and the distance from node i to node j. The purpose is to prevent some data points from being misjudged due to the small distance between their neighbors. The local reachable density of node i is the average LRD of the reciprocal of the reachable distance between it and its nearest neighbor nodes. k (i) The creep rate anomaly index is calculated based on the average value of the inverse of the calculated reachable distance, and the expression is: Where, LRD k ( j) is the local reachability density of node j, LRD k (i) is the local reachability density of node i, and KM is the creep rate anomaly index.

[0066] A larger creep rate anomaly index indicates that the creep rate of a node differs significantly from that of its neighboring nodes under long-term loading. This generally indicates that the model's stress distribution predictions at these nodes are significantly biased. Nodes with large creep rate anomaly indices may be experiencing excessive stress concentration, material aging, or increased local deformation, and the model may not adequately capture these complex mechanical effects. Therefore, a large anomaly index indicates low prediction accuracy, necessitating further analysis or model adjustments for these nodes.

[0067] Conversely, when the creep rate anomaly index is smaller, the creep rate of the node is very close to that of its neighboring nodes, consistent with the model's prediction of the stress distribution under long-term loading. This means that the model effectively captures the stress changes and creep behavior of the node, resulting in relatively accurate predictions. In this case, nodes with a small anomaly index are considered normal, indicating that the model's stress distribution predictions for these nodes are reliable and require no further adjustment.

[0068] S3: Convert the acquired horizontal shear response data and material creep data into feature vectors, use the feature vectors as input to the machine learning model, train the machine learning model, and determine the accuracy of the symmetric node stress distribution prediction results based on the model output results.

[0069] The shear response deviation index and the creep rate anomaly index are converted into the first eigenvector, and the first eigenvector is used as the input of the machine learning model. The machine learning model uses the accuracy value label of the symmetric node stress distribution prediction result predicted by each group of first eigenvectors as the prediction target, and minimizes the sum of the prediction errors of the accuracy value labels of all symmetric node stress distribution prediction results as the training target. The machine learning model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The accuracy value of the symmetric node stress distribution prediction result is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

[0070] The accuracy value of the symmetric node stress distribution prediction result is obtained by obtaining the corresponding function expression KR=F(HK, KM) from the first eigenvector training data of the trained machine learning model; where F is the output function of the model, HK is the shear response deviation index, KM is the creep rate anomaly index, and KR is the accuracy value of the symmetric node stress distribution prediction result.

[0071] S4: The accuracy of the prediction results of the stress distribution of the symmetrical nodes is divided into different levels, namely accurate prediction results, possibly accurate prediction results and inaccurate prediction results.

[0072] Comparing the obtained accuracy value of the symmetric node stress distribution prediction result with the gradient standard threshold, the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and comparing the accuracy value of the symmetric node stress distribution prediction result with the first standard threshold and the second standard threshold respectively;

[0073] If the accuracy value of the symmetric node stress distribution prediction result is greater than the second standard threshold, it indicates that the accuracy of the symmetric node stress distribution prediction result is high, and a high accuracy signal is generated at this time, and the symmetric node stress distribution prediction result under which the high accuracy signal is generated is classified as an accurate prediction result;

[0074] If the accuracy value of the symmetric node stress distribution prediction result is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it means that the accuracy of the symmetric node stress distribution prediction result is average, and a medium accuracy signal is generated at this time, and the symmetric node stress distribution prediction result under the medium accuracy signal is classified as a possible accurate prediction result;

[0075] If the accuracy value of the symmetric node stress distribution prediction result is less than the first standard threshold, it means that the accuracy of the symmetric node stress distribution prediction result is low. At this time, a low accuracy signal is generated, and the symmetric node stress distribution prediction result under the low accuracy signal is classified as an inaccurate prediction result.

[0076] S5: For accurate prediction results, the stress distribution of the node is directly used as the reference benchmark for shear wall design; for inaccurate prediction results, the local graph structure of the symmetrical node is retrained.

[0077] Nodes with accurate predictions have stress distributions that closely align with the actual mechanical behavior of the structure, making them a direct reference for shear wall design. This means the model's stress and deformation predictions at these nodes are reliable, allowing engineers to rely on these results for structural assessment and design without the need for additional corrections or verification. These accurately predicted nodes serve as a stable benchmark for the entire design, ensuring that the shear wall's mechanical performance meets design requirements under long-term loads and complex operating conditions, thereby optimizing overall safety and efficiency.

[0078] Symmetrical nodes with inaccurate predictions require focused training of the local graph structure of these nodes, as the model fails to correctly capture the stress state or creep effects at these nodes. Specifically, the local mechanical features of these nodes and their neighbors are extracted, and the feature inputs are adjusted or symmetry constraints are added, and the local graph is further refined and trained. This retraining helps the model better understand the complex mechanical behavior of these nodes under long-term loads, especially capturing stress concentration issues caused by asymmetric loads, boundary conditions, or material nonlinearity, thereby improving prediction accuracy and optimizing structural design.

[0079] S6: From the overall shear wall graph, extract each node that may accurately predict the result and its direct neighbor nodes, construct a local subgraph, and dynamically adjust the edge weights according to the abnormal degree of the accuracy of the prediction results of the symmetric node stress distribution.

[0080] Symmetric node i is a node that may accurately predict the result. This node and its direct neighbor nodes (i.e., nodes with edge connections to i) are extracted to form a local subgraph. The node set V of the local subgraph is i Includes: V i ={i}∪{j|(i,j)∈E}; where E is the edge set of the graph and j is the neighbor node directly connected to node i;

[0081] In the local subgraph, the initial weight of each edge (i, j) is defined according to the geometric distance or the material properties of the node. Usually the initial weight is a fixed value, and the expression is: Among them, d(i,j) is the geometric distance or structural correlation measure between nodes i and j. The larger the initial weight, the closer the connection between the nodes.

[0082] If the accuracy value of the symmetric node stress distribution prediction result is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, the symmetric node stress distribution prediction result under the generated accuracy signal is divided into a possible accurate prediction result, and the accuracy values ​​of the symmetric node stress distribution prediction results generated in the subsequent fixed time period that are greater than or equal to the first standard threshold and less than or equal to the second standard threshold are collected and a corresponding data set is established. The standard deviation of the data set is calculated as the abnormality degree of stress distribution prediction, marked as A(i); the abnormality degree A(i) of stress distribution prediction reflects the deviation of the prediction result of node i. The larger its value, the greater the prediction deviation. According to A(i), the prediction accuracy adjustment factor α(i) of node i is defined, which is negatively correlated with A(i) and the expression is: α ( i ) =1-A ( i ) ; where α(i)∈[0,1] represents the accuracy of the node prediction. The closer α(i) is to 1, the more accurate the prediction; the smaller α(i) is, the less accurate the prediction.

[0083] In a local subgraph, edge weights According to the dynamic adjustment of the abnormality of the stress distribution prediction results of the node, the edge weight is dynamically adjusted as follows by combining the prediction accuracy adjustment factor α(i) of node i and its neighbor node j: is the adjusted edge weight. The smaller the edge weight, the greater the uncertainty or error in the prediction of the node pair connected by the edge; the larger the weight, the more accurate the prediction of the connected node pair.

[0084] When the stress distribution prediction result of node i or neighboring node j is abnormal (i.e. A(i) or A(j) is large, resulting in α(i) or α(j) being small), the adjusted edge weight This means that the strength of the association between these nodes in the local graph structure has weakened, reflecting that the model is not accurately capturing the mechanical interactions between these nodes, prompting subsequent models to correct these local relationships. Conversely, if the stress distribution predictions for nodes i and j are more accurate, then α(i) and α(j) are higher, and the adjusted edge weights remain high, indicating that the model is able to better capture the mechanical relationships between these nodes.

[0085] By dynamically adjusting the edge weights, the weight matrix Wi of the local subgraph is updated. The new weight matrix Wi(new) reflects the node association strength, corrected based on the degree of abnormality in the prediction results. Retraining the updated local subgraph in the graph neural network can further improve the stress distribution prediction accuracy for nodes that are likely to have accurate prediction results.

[0086] In this embodiment, symmetrical node pairs in the shear wall structure are identified by comparing geometric coordinates and connection relationships, and these nodes are marked in the graph structure. Next, the shear response data of the symmetrical nodes under horizontal loads and the material creep data caused by stress differences under long-term loads are obtained, and these data are converted into feature vectors and input into the machine learning model for training to determine the stress distribution prediction accuracy of the symmetrical nodes. Then, the nodes are classified according to the accuracy of the prediction results and divided into accurate, possibly accurate and inaccurate prediction results. For accurate prediction results, the stress distribution of the node is directly used for shear wall design, while for nodes with inaccurate prediction results, their local graph structure is extracted and retrained. For nodes that are likely to be accurate, these nodes and their neighboring nodes are extracted from the overall shear wall graph, a local subgraph is constructed, and the edge weights are dynamically adjusted according to the degree of abnormality of the prediction results to optimize the model prediction.

[0087] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0088] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0089] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A shear wall design method based on graph node classification graph neural network, characterized by: The following steps are included: S1: Determine whether the node pairs in the shear wall structure are symmetrical by comparing the geometric coordinates and connection relationships, mark the symmetrical node pairs identified in the shear wall structure, and construct a shear wall graph structure to keep the symmetrical node pairs in the same graph; Specifically, for each pair of geometrically symmetric nodes, check the connectivity in their adjacency matrix, that is, check whether their direct neighbors maintain a symmetric relationship with the neighbors of their geometrically symmetric points; if nodes i and j are identified as geometrically symmetric nodes, check whether the neighboring nodes ik and jk of the symmetric node pair also satisfy the symmetry; if not, the node pair is considered to not meet the symmetry requirement; S2: After identifying the symmetrical nodes, obtain the horizontal shear response data of the symmetrical nodes of the wall under horizontal loads, as well as the material creep data generated based on the stress difference of the symmetrical nodes under long-term loads; S3: Convert the acquired horizontal shear response data and material creep data into feature vectors, use the feature vectors as input to the machine learning model, train the machine learning model, and determine the accuracy of the predicted results of the symmetrical node stress distribution based on the model output results; S4: The accuracy of the prediction results of the stress distribution of the symmetrical nodes is divided into different levels, namely accurate prediction results, possibly accurate prediction results and inaccurate prediction results; S5: For accurate prediction results, the stress distribution of the node is directly used as the reference benchmark for shear wall design; for symmetrical nodes with inaccurate prediction results, the local graph structure is retrained; S6: From the overall shear wall graph, extract each node that may accurately predict the result and its direct neighbor nodes, construct a local subgraph, and dynamically adjust the edge weights according to the abnormal degree of the accuracy of the prediction results of the symmetric node stress distribution.

2. The shear wall design method based on graph node classification graph neural network according to claim 1 is characterized by: In S2, the shear response anomaly index is generated based on the shear force data at the extracted symmetrical nodes, wherein the shear response deviation index is obtained as follows: Extract the shear response data of all symmetrical nodes in the shear wall and construct a feature matrix. The feature matrix of the shear response data of each symmetrical node i is as follows: the shear response of node i in the horizontal shear direction and the shear response of node i in the vertical shear direction , the shear response eigenvector of each node i for: ; For N nodes, form the feature matrix X, ; Where X is an N×d-dimensional matrix, d is the dimension of the feature, T is the matrix transpose, and it is standardized. For the standardized feature matrix X′, its covariance matrix C is calculated, and the expression is: ; Where C is a d×d dimensional matrix, which represents the covariance between the shear response characteristics. The eigenvalue decomposition of the covariance matrix C is performed to obtain the eigenvalue and the corresponding eigenvector ; The eigenvalue represents the variance contribution in the direction of each principal component, while the eigenvector represents the direction of the principal component. The expression is: ; Among them, the eigenvalue represents the importance of the i-th principal component, is the corresponding eigenvector, and the principal components corresponding to the first g largest eigenvalues ​​are selected To reduce the dimension, the original feature vector Project it into the principal component space and reconstruct the approximate value of the original eigenvector through the principal component vector. , calculate the difference between the original eigenvector and the reconstructed eigenvector, that is, calculate the shear response deviation index, which is expressed as: ; Where HK is the shear response deviation index.

3. The shear wall design method based on graph node classification graph neural network according to claim 2 is characterized by: After analyzing the creep rate of each node under long-term load, the creep rate anomaly index is generated. The method for obtaining the creep rate anomaly index is as follows: Construct the creep rate characteristic vector of each node. For each node, its creep rate vector Contains creep rates at multiple moments: ;in: Node i at time The creep rate of n is the number of time points. For N nodes, an N×n creep rate matrix Y is constructed. The Euclidean distance between each node is calculated. For node i and node j, the Euclidean distance d(i, j) is: ; Where d(i,j) represents the Euclidean distance between the creep rate feature vectors of nodes i and j. For each node i, a parameter k is selected to represent the number of nearest neighbor nodes to be considered. The k-distance of node i is It refers to the distance from node i to its kth nearest neighbor node: ;in, Represents the k nearest neighbor nodes of node i, d(i,j) is the Euclidean distance from node i to node j, and calculates the reachable distance of node i to node j , the expression is: ; Calculate the local reachability density. The local reachability density of node i is the average of the reciprocal of its reachability distance to its nearest neighbor node , the creep rate anomaly index is calculated based on the average value of the inverse of the calculated reachable distance, and the expression is: Where, is the local reachability density of node j, is the local reachability density of node i, and KM is the creep rate anomaly index.

4. The shear wall design method based on graph node classification graph neural network according to claim 3 is characterized by: In S3, the acquired horizontal shear response data and material creep data are converted into feature vectors, which are used as inputs to the machine learning model. The machine learning model is trained and the accuracy of the prediction results of the symmetrical node stress distribution is determined based on the model output results. Specifically: The shear response deviation index and the creep rate anomaly index are converted into the first eigenvector, and the first eigenvector is used as the input of the machine learning model. The machine learning model uses the accuracy value label of the symmetric node stress distribution prediction result predicted by each group of first eigenvectors as the prediction target, and minimizes the sum of the prediction errors of the accuracy value labels of all symmetric node stress distribution prediction results as the training target. The machine learning model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The accuracy value of the symmetric node stress distribution prediction result is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

5. The shear wall design method based on graph node classification graph neural network according to claim 4 is characterized by: In S4, the accuracy of the prediction results of the stress distribution of the symmetrical nodes is divided into different levels, namely accurate prediction results, possibly accurate prediction results and inaccurate prediction results, specifically: Comparing the obtained accuracy value of the symmetric node stress distribution prediction result with the gradient standard threshold, the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and comparing the accuracy value of the symmetric node stress distribution prediction result with the first standard threshold and the second standard threshold respectively; If the accuracy value of the symmetric node stress distribution prediction result is greater than the second standard threshold, it indicates that the accuracy of the symmetric node stress distribution prediction result is high, and a high accuracy signal is generated at this time, and the symmetric node stress distribution prediction result under which the high accuracy signal is generated is classified as an accurate prediction result; If the accuracy value of the symmetric node stress distribution prediction result is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it means that the accuracy of the symmetric node stress distribution prediction result is average, and a medium accuracy signal is generated at this time, and the symmetric node stress distribution prediction result under the medium accuracy signal is classified as a possible accurate prediction result; If the accuracy value of the symmetric node stress distribution prediction result is less than the first standard threshold, it means that the accuracy of the symmetric node stress distribution prediction result is low. At this time, a low accuracy signal is generated, and the symmetric node stress distribution prediction result under the low accuracy signal is classified as an inaccurate prediction result.

6. The shear wall design method based on graph node classification graph neural network according to claim 1 is characterized by: In S6, each node that can accurately predict the result and its direct neighbor nodes are extracted from the overall shear wall diagram to construct a local subgraph. The edge weights are dynamically adjusted according to the abnormality of the accuracy of the stress distribution prediction results of the symmetrical nodes. Specifically, Symmetric node i is a node that may accurately predict the result. The node and its direct neighbor nodes are extracted to form a local subgraph. The node set of the local subgraph is include: ; Where E is the edge set of the graph, and j is the neighbor node directly connected to node i; In the local subgraph, the initial weight of each edge (i, j) is defined according to the geometric distance. is a fixed value, and the expression is: ; Where d(i,j) is the measure of the geometric distance between nodes i and j; If the accuracy value of the symmetric node stress distribution prediction result is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, the symmetric node stress distribution prediction result generated under the medium accuracy signal is classified as a possible accurate prediction result. The accuracy values ​​of the symmetric node stress distribution prediction results generated in a subsequent fixed time period that are greater than or equal to the first standard threshold and less than or equal to the second standard threshold are collected and a corresponding data set is established. The standard deviation of the data set is calculated as the abnormality degree of the stress distribution prediction, which is marked as A(i). The prediction accuracy adjustment factor α(i) of node i is defined based on A(i), which is negatively correlated with A(i) and is expressed as: ; Among them, α(i)∈[0,1], represents the accuracy of node prediction.

7. The shear wall design method based on graph node classification graph neural network according to claim 6 is characterized by: In a local subgraph, edge weights Dynamically adjust the abnormality of the stress distribution prediction results of the node by combining the prediction accuracy adjustment factor of node i and its neighbor node j , the edge weights are dynamically adjusted as follows: is the adjusted edge weight; after dynamically adjusting the edge weight, the weight matrix of the local subgraph is updated.

Images