Method for fast prediction of structural response based on physical coding graph network

Abstract

The application discloses a kind of structure response fast prediction methods based on physical coding graph network, comprising: constructing physical coding graph network;The physical parameters of target structure are input into physical coding graph network coding;Message passing is carried out to the physical parameters after coding, message passing includes successively carried out node mapping and edge mapping, and the data flow of message passing meets mathematical logic;The result of node mapping and edge mapping is decoded, and displacement prediction result and internal force prediction result are obtained.The application, by constructing physical coding graph network, realizes the fast prediction of structure response in multiple working conditions;Wherein, the coding, message passing, decoding process adapts complex structure parameters, node mapping, edge mapping division completes the accurate prediction of displacement and internal force, solves the problem of repeated calculation of traditional method, meets the real-time interaction demand of digital twin, efficiently supports fast iterative design and real-time performance evaluation, provides better solution for structure engineering calculation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of engineering data analysis, and in particular to a structure response rapid prediction method based on a physical encoding graph network. BACKGROUND

[0002] In the whole process of modern engineering construction, numerical simulation is an important means of calculation analysis and design.

[0003] Traditional finite element analysis can only perform single-condition calculation of a single structure model, and repeated operations are required in the face of similar load conditions. In the complex multi-condition scene of structure operation and maintenance, the calculation is complicated and time-consuming, and it is difficult to meet the requirements of rapid iterative design and real-time performance evaluation. Further, with the application of digital twin technology in structural engineering, it is required to obtain load data in real time and quickly feedback the structure response. The traditional calculation method cannot meet this timeliness requirement.

[0004] The existing structure response prediction technology has many problems, which are embodied in:

[0005] (1) Data-driven models rely on a large amount of high-quality data, which is difficult to meet in the engineering field;

[0006] (2) Physical-driven models (such as physical information neural networks PINN) are difficult to adapt to complex structures because the mapping network is only a multilayer perceptron, and the training is difficult due to the physical constraints integrated into the loss function, resulting in convergence and stability problems;

[0007] (3) Existing researches focus on the forward and inverse problems of a single condition, and lack of research on condition generalization. The prediction accuracy and range cannot meet the actual structural calculation requirements.

[0008] A Chinese invention patent with the authorized announcement number CN116432274B discloses a rod system structure optimization method and device, electronic equipment and storage medium, but the invention is only applicable to a single rod system structure, and has low transferability.

[0009] In summary, there is an urgent need for a new efficient and rapid structure response rapid prediction technical solution. SUMMARY

[0010] The main purpose of the present application is to provide a structure response rapid prediction method based on a physical encoding graph network, which aims to solve the technical problems that the existing technology requires repeated operations in the face of similar load conditions in structure analysis, and the calculation is complicated and time-consuming in the complex multi-condition scene of structure operation and maintenance, which is difficult to meet the requirements of rapid iterative design and real-time performance evaluation.

[0011] To achieve the above purpose, the present application provides a structure response rapid prediction method based on a physical encoding graph network, comprising the following steps:

[0012] S1: constructing a physical encoding graph network, which is used for encoding, message passing and decoding of physical parameters of a target structure to obtain a structural response, the structural response at least including displacement prediction and internal force prediction;

[0013] S2: inputting the physical parameters of the target structure into the physical encoding graph network for encoding;

[0014] S3: performing the message passing on the encoded physical parameters, the message passing including node mapping and edge mapping in sequence, wherein the data flow of the mapping network conforms to mathematical logic, the node mapping is used for the displacement prediction, and the edge mapping is used for the internal force prediction;

[0015] S4: decoding the results of the node mapping and the edge mapping to obtain displacement prediction results and internal force prediction results.

[0016] Preferably, the construction of the physical encoding graph network includes the following steps:

[0017] A1: obtaining geometric relations and the physical parameters of the target structure; wherein the physical parameters include node data and member data of physical properties of the target structure constrained by the geometric relations; node features are obtained and encoded based on the node data, and edge features are obtained and encoded based on the member data;

[0018] A2: obtaining a small amount of response data set of the target structure; training the physical encoding graph network based on A1 and the response data set, so that the physical encoding graph network encodes the node features and the edge features, performs the message passing based on different load cases input, and decodes and outputs different results of the node mapping and the edge mapping; wherein the response data set stores real responses of the target structure under different load cases, the load cases act on nodes corresponding to the node data, and the node data is taken as input;

[0019] A3: outputting the trained physical encoding graph network.

[0020] Preferably, the nodes include free nodes and constraint nodes;

[0021] The node data of the free nodes is defined as load node data;

[0022] The node data of the constraint nodes is defined as constraint node data;

[0023] The node data includes the load node data and the constraint node data.

[0024] As preferred, the obtaining node features and encoding based on the node data comprises:

[0025] Based on the load node data, load node features are obtained.

[0026] The load node features are encoded as:

[0027]

[0028] wherein, represents the stiffness of a load node , which is defined as the load node data, is the load node number, represents a set of surrounding neighbor nodes represents the stiffness influence on the surrounding nodes to the node .

[0029] As preferred, the obtaining node features and encoding based on the node data further comprises:

[0030] Based on the constraint node data, constraint node features are obtained.

[0031] The constraint node features are encoded as:

[0032]

[0033] wherein, represents the stiffness of a constraint node , and the stiffness of the constraint node is set to infinity to express the constraint attribute.

[0034] As preferred, the obtaining edge features and encoding based on the member data comprises:

[0035] Obtaining member data corresponding to the load node data to obtain load node edge features.

[0036] Based on the encoding of the load node features and the load node edge features, the load node edge features are encoded as:

[0037]

[0038] wherein, is the edge feature of the first adjacent edge and the load node after encoding, is the stiffness between the load node and its first adjacent edge corresponding to the surrounding neighbor nodes , The second adjacent edge and load node after encoding edge features, For load nodes The surrounding neighbor nodes corresponding to its second adjacent edge Stiffness between The third adjacent edge and load node after encoding edge features, For load nodes The surrounding neighbor nodes corresponding to its third adjacent edge The stiffness between them.

[0039] Preferably, the step of obtaining and encoding edge features based on the component data further includes:

[0040] Obtain the component data corresponding to the constraint node data to obtain the constraint node edge features;

[0041] Based on the encoding of the constraint node features and the constraint node edge features, the constraint node edge features are encoded as follows:

[0042]

[0043] in, The first adjacent edge and constraint node after encoding Edge features, The second adjacent edge and constraint node after encoding Edge features.

[0044] Preferably, the displacement prediction includes:

[0045] The encoded node features and edge features are input into the physical coding graph network, and displacement prediction is performed using a node mapping formula. The result of this displacement prediction is defined as the result of the node mapping; wherein, the node mapping formula is:

[0046]

[0047] in, The result of the node mapping, It is a node The initial eigenvectors, It is a node neighboring nodes The set, It is a continuously updated neighbor node Node characteristics, It is an unupdated node. and its neighboring nodes Edge features between them representing an aggregation of a node and its neighbor nodes, and is a multi-layer perceptron.

[0048] As preferred, the internal force prediction comprises:

[0049] The encoded node features and the edge features are input into the physical encoded graph network, an internal force prediction is performed through an edge mapping formula, and the result of the internal force prediction is defined as the result of the edge mapping; wherein the edge mapping formula is:

[0050]

[0051] wherein, is a node and its neighbor nodes between the node and the neighbor nodes, is the node feature of the node being updated, is the node feature of the neighbor node being updated, is the edge feature between the node and the neighbor node not being updated, is a multi-layer perceptron. As preferred, the target structure at least comprises a truss structure. When the target structure is the truss structure, the node is a hinged point of the truss structure, and the component is a rod of the truss structure; wherein the structural calculation follows the mathematical logic, and the mathematical logic at least comprises a structural mechanics calculation logic.

[0052] As preferred, the target structure at least comprises a truss structure.

[0053] When the target structure is the truss structure, the node is a hinged point of the truss structure, and the component is a rod of the truss structure; wherein the structural calculation follows the mathematical logic, and the mathematical logic at least comprises a structural mechanics calculation logic.

[0054] Beneficial effects: the structure response rapid prediction method based on the physical encoded graph network of the present application retains the advantages of data flow and training mode of the data-driven model, and the efficient structure attribute encoding mode and the physical guided architecture design greatly enhance the physical interpretability and generalization performance of the model, realize maintaining high generalization prediction accuracy through small samples, thereby solving the technical problems that the prior art needs to repeat operations in the face of similar load working conditions in structure analysis, and in the complex multi-working condition scene of structure operation and maintenance, the calculation is complicated and time-consuming, and it is difficult to meet the requirements of rapid iterative design and real-time performance evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0055] ​​In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0056] Figure 1 The flow chart of the structural response fast prediction method based on the physical encoding graph network provided by the embodiment of the present application is shown in the figure.

[0057] Figure 2 The flow chart of the construction of the physical encoding graph network in the structural response fast prediction method based on the physical encoding graph network provided by the embodiment of the present application is shown in the figure.

[0058] Figure 3 The schematic diagram of the node in the structural response fast prediction method based on the physical encoding graph network provided by the embodiment of the present application is shown in the figure.

[0059] Figure 4 The prediction condition based on the GCN provided by the embodiment of the present application is shown in the figure.

[0060] Figure 5 The prediction condition based on the GIN provided by the embodiment of the present application is shown in the figure.

[0061] Figure 6 The prediction condition based on the GAT provided by the embodiment of the present application is shown in the figure.

[0062] Figure 7 The prediction condition of the structural response fast prediction method based on the physical encoding graph network provided by the embodiment of the present application is shown in the figure.

[0063] Figure 8 The error bar contrast chart of the prediction condition based on the GCN, the prediction condition based on the GIN, the prediction condition based on the GAT and the prediction condition of the structural response fast prediction method based on the physical encoding graph network provided by the embodiment of the present application is shown in the figure.

[0064] Figure 9 The truss structure and load schematic diagram for displacement prediction and internal force prediction provided by the embodiment of the present application is shown in the figure.

[0065] Figure 10 The real displacement condition based on the truss structure and load for displacement prediction provided by the embodiment of the present application is shown in the figure.

[0066] Figure 11 The predicted displacement condition based on the truss structure and load for displacement prediction provided by the embodiment of the present application is shown in the figure.

[0067] Figure 12The truss structure and load based on the real internal force situation for internal force prediction are provided for the embodiments of the present application.

[0068] Figure 13 The prediction position situation based on the truss structure and load for internal force prediction is provided for the embodiments of the present application.

[0069] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0070] It should be understood that the specific embodiments described herein are intended to explain the present application, but not to limit the present application.

[0071] Traditional methods such as finite element analysis, although play a key role in solving structural mechanics problems, can only realize single working condition calculation of a structure model, and even if facing similar load working conditions, still need to carry out repeated calculation. In practical application, engineering structures are affected by various factors in the operation and maintenance stage, and the structures need to bear various complex load working conditions, and each working condition needs to be re-calculated with tedious finite element calculation, resulting in serious repeated calculation problem, time-consuming and laborious, and it is difficult to meet the requirements of rapid iterative design and real-time performance evaluation. Further, with the rise of digital twinning and the application of digital twinning system in structure, structural digital twinning emphasizes real-time interaction between physical entity and virtual model, and puts forward higher requirements for structure calculation, that is, after real-time acquisition of load data, the structure response is quickly fed back. However, the traditional calculation method is difficult to meet this timeliness, so it is necessary to break through the repeated calculation dilemma and establish an efficient and rapid structure response prediction method.

[0072] In view of the above technical difficulties, with reference to Figure 1 The embodiment discloses a structure response rapid prediction method based on a physical coding graph network, comprising the following steps:

[0073] S1: constructing a physical coding graph network, the physical coding graph network is used for encoding, message passing and decoding of physical parameters of a target structure to obtain a structure response, the structure response at least includes displacement prediction and internal force prediction;

[0074] S2: inputting the physical parameters of the target structure into the physical coding graph network for encoding;

[0075] S3: performing message passing on the encoded physical parameters, the message passing includes node mapping and edge mapping performed in sequence, wherein the data flow of the mapping network conforms to mathematical logic, the node mapping is used for displacement prediction, and the edge mapping is used for internal force prediction;

[0076] S4: decoding the results of the node mapping and the edge mapping to obtain displacement prediction results and internal force prediction results.

[0077] By the above, by constructing a physical coding graph network, multi-working condition structural response fast prediction is realized; wherein, the coding, message passing, and decoding process adapt to complex structural parameters, and the node mapping and edge mapping complete the accurate prediction of displacement and internal force, solve the problem of repeated calculation of traditional methods, meet the real-time interaction demand of digital twin, efficiently support rapid iterative design and real-time performance evaluation, and provide a better solution for structural engineering calculation.

[0078] With the innovation of artificial intelligence technology, agent models based on deep learning methods have been successfully applied to the prediction tasks of complex engineering systems. Among them, graph neural networks have the ability to express the characteristics of graph structures composed of nodes and edges, and the data form of the graph structure is similar to the structure system. Based on the similar data form and structure system of the graph structure, using graph neural networks as the mapping module for structural working condition generalization prediction tasks is very suitable. However, if only based on simple structural similarity for mapping, the data-driven model often needs a large amount of high-quality data as the training data set, and the data set in the engineering field, especially for structural calculation, cannot meet the data requirements. Further, in existing physical-driven models, the physical information neural network (PINN) uses a multilayer perceptron as a mapping network, which is difficult to adapt to the complex characteristics of engineering structures, and the physical-driven model makes the model training difficult by using the physical constraints as the loss function, and new convergence and stability problems arise. It is particularly pointed out that the existing research on physical-driven models mainly focuses on the positive problem and inverse problem of a single working condition of the structure, and represents sample-free solution and parameter inversion. There is a lack of research on working condition generalization of structural calculation or the prediction accuracy and range are difficult to support structural calculation. In order to solve the current limitations of constructing structural generalization calculation and realize structural response fast prediction, a new physical coding graph network is constructed in the embodiment.

[0079] Specifically, referring to Figure 2 , the construction of the physical coding graph network includes the following steps:

[0080] A1: Obtain the geometric relationship and physical parameters of the target structure; wherein, the physical parameters include node data and member data of the physical properties of the target structure constrained by the geometric relationship; based on the node data, the node features are obtained and coded, and based on the member data, the edge features are obtained and coded;

[0081] A2: Obtain a small amount of response data set of the target structure; based on A1 and the response data set, train the physical coding graph network, so that the physical coding graph network encodes the node features and the edge features, and based on the input of different load working conditions, performs message passing, and decodes the results of different node mappings and edge mappings; wherein, the response data set stores the true response of the target structure under different load working conditions, the load working condition acts on the node corresponding to the node data, and the node data is taken as the input;

[0082] A3: output the trained physical encoding graph network.

[0083] In the specific application of the embodiment, the target structure at least includes a truss structure;

[0084] When the target structure is a truss structure, the nodes are the hinged points of the truss structure, and the members are the rod members of the truss structure.

[0085] The truss structure and the graph structure of the graph neural network have similarities, so the hinged points and the rod members of the truss structure can naturally be expressed by the node features and the edge features of the graph neural network, respectively.

[0086] Specifically, referring to Figure 3 , the nodes include free nodes and constraint nodes to which loads are applied;

[0087] The node data of the free nodes is defined as load node data;

[0088] The node data of the constraint nodes is defined as constraint node data;

[0089] The node data includes the load node data and the constraint node data.

[0090] Generally, the rod system attribute tensile stiffness of the truss structure is expressed by edge features. The load working condition applied to the nodes can be input by node features, but the constraint condition of the nodes as node feature input affects the generalization of the load.

[0091] The problem faced by the static force calculation of the structure in the finite element can be expressed as an algebraic equation of , wherein is the overall stiffness matrix of the structure, and if the load vector is known, the displacement vector is solved, the overall stiffness matrix needs to be processed according to the boundary constraint condition, then the inverse matrix of is calculated, and finally the value of the vector is obtained. Taking the displacement calculation of the truss structure as an example, each node involves two degrees of freedom of direction and direction, and if the number of nodes is , then is a matrix of rows and columns. The overall processing mode of directly inverting the overall stiffness matrix does not conform to the node discrete processing mode of the graph neural network, so a new structure feature encoding method is established in combination with the numerical iteration theory. The iteration formula of the numerical iteration theory is:

[0092]

[0093] The iterative formula is used to calculate the algebraic equation , is the stiffness matrix at node , is the stiffness relationship matrix between node and node , is the approximate solution of node at the first iteration, is the approximate solution of a certain node around node at the first iteration.

[0094] In order to simultaneously express the unit stiffness and constraint nodes of the truss structure in the neural network, we take the stiffness influence relationship matrix between nodes as the edge feature to represent the unit stiffness.

[0095] Specifically, referring to Figure 3 , the node features are obtained based on the node data and are coded, including:

[0096] Based on the load node data, the load node features are obtained;

[0097] The load node features are coded as:

[0098]

[0099] wherein, represents the stiffness of the load node , and the stiffness is defined as the load node data, is the load node number, represents the set of neighbor nodes around the load node represents the stiffness influence of the surrounding nodes on the node .

[0100] Based on the coding of the load node features, specifically, referring to Figure 3 , the edge features are obtained based on the component data and are coded, including:

[0101] The component data corresponding to the load node data are obtained, and the load node edge features are obtained;

[0102] Based on the coding of the load node features and the load node edge features, the load node edge features are coded as:

[0103]

[0104] wherein, the edge feature of the first adjacent edge and the load node after encoding, the stiffness between the load node and the surrounding neighbor node corresponding to the first adjacent edge thereof after encoding, the edge feature of the second adjacent edge and the load node after encoding, the stiffness between the load node and the surrounding neighbor node corresponding to the second adjacent edge thereof after encoding, the edge feature of the third adjacent edge and the load node after encoding, the stiffness between the load node and the surrounding neighbor node corresponding to the third adjacent edge thereof after encoding.

[0105] In addition, for the constraint node, the constraint attribute is expressed by directly setting the stiffness of the constraint node to infinity.

[0106] Specifically, referring to Figure 3 , the node features are obtained based on the node data and encoded, and further comprising:

[0107] obtaining constraint node features based on constraint node data;

[0108] encoding the constraint node features into:

[0109]

[0110] wherein, the stiffness of the constraint node is set to infinity to express the constraint attribute.

[0111] Based on the constraint node feature encoding, specifically, referring to Figure 3 , the edge features are obtained based on the member data and encoded, and further comprising:

[0112] obtaining constraint node edge features by obtaining member data corresponding to the constraint node data;

[0113] encoding the constraint node edge features into:

[0114]

[0115] wherein, the edge feature of the first adjacent edge and the constraint node after encoding, the edge feature of the second adjacent edge and the constraint node Edge features.

[0116] By using the above structural encoding, all truss structural attributes (especially constraint nodes) are taken as input in the form of edge features, and separated from the generalized load node features; the data normalization is also directly achieved based on the encoding method of stiffness influence relationship matrix.

[0117] It should be noted that the reason why graph neural networks are based on numerical iteration theory is that: (1) Both neural network optimization theory and numerical iteration theory are based on optimization as the underlying theory, so they are highly similar; (2) Graph neural networks are discrete calculations with nodes as objects, similar to the processing method of sparse matrices, which corresponds to iterative calculations; (3) Numerical iteration theory can clearly express the strict mathematical logical relationship between physical variables, and the computational logical relationship can provide a clear design idea for the data flow of the network model.

[0118] However, in general, both matrix calculations and iterative calculations are methods for calculating single working conditions. The significance of neural network models lies in overcoming the problem of repetitive calculations of working conditions, achieving generalized calculations of structural working conditions, and establishing a meta-model for proxy calculations. Therefore, we introduce graph neural networks. Graph neural networks mainly establish the mapping relationship between node features and edge features, which is called the message passing mechanism, represented by a propagation function and an aggregation function. The propagation function corresponds to message propagation, updating the features of neighboring nodes based on the master node. The aggregation function corresponds to message aggregation, updating the features of the master node based on the master node and the updated neighboring nodes. The message passing mechanism can be expressed as:

[0119]

[0120] in, and It is a fully connected layer in a neural network; and They are Layer and the next layer Feature vectors of data points in the layer; It is the edge-connected feature vector. It is the feature vector of the corresponding neighboring data points.

[0121] It is clear that the structural load case generalization calculation method based on the physical coding graph network in this embodiment is based on the mathematical logic relationship of physical quantity updates in numerical theory, and combines the message passing mechanism of graph neural networks to carry out the network architecture design. Based on this design, the physical coding graph network can perform node prediction and edge prediction tasks for displacement prediction and internal force prediction of truss structures.

[0122] Specifically, displacement prediction includes:

[0123] The encoded node features and edge features are input into the physical encoding graph network, displacement prediction is performed through a node mapping formula, and the result of the displacement prediction is defined as the result of node mapping; wherein the node mapping formula is:

[0124]

[0125] wherein, is the result of node mapping, is the initial feature vector of the node , is the set of neighbor nodes of the node , is the node feature of the always-updated neighbor node , is the edge feature between the node and its neighbor nodes, denotes aggregation of the node and its neighbor nodes, and are multilayer perceptrons. It should be noted that the design of node mapping is based on numerical value iteration theory, i.e. iteration formula. According to the iteration formula, there are two node features in the iteration process, i.e. the un-updated load feature and the always-updated displacement feature, and the displacement feature is aggregated in the updating process, while the load feature is in a superimposed manner.

[0126] Specifically, internal force prediction includes:

[0127] The encoded node features and edge features are input into the physical encoding graph network, internal force prediction is performed through an edge mapping formula, and the result of the internal force prediction is defined as the result of edge mapping; wherein the edge mapping formula is:

[0128]

[0129]

[0130] wherein, is the result of edge mapping between the node and its neighbor nodes, is the node feature of the always-updated node , is the node feature of the always-updated neighbor node , is the edge feature between the node and its neighbor nodes, and are multilayer perceptrons.

[0131] ​​​It should be noted that in a specific application, the node mapping formula is embedded in the node mapping module, the edge mapping formula is embedded in the edge mapping module, and the node mapping module is connected to the edge mapping module in the physical encoding graph network, so as to realize the node mapping and edge mapping in sequence.

[0132] The design principle of the physical encoding graph network is to focus on the causal relationship of the physical quantity calculation of the physical formula, and combine the causal relationship to design the update order of the node features and the edge features in the network, such as message aggregation and vector superposition. In addition, it should be specially pointed out that the advantage of the physical guide network design is that for the vectors appearing in the iterative formula such as , if not as input of the network, the features can be expressed by a mapping function such as a multi-layer perceptron, which has better flexibility.

[0133] To analyze the application effect of the embodiment, the same training data, test data and optimal value are used to compare the GCN (graph convolutional neural network), GIN (graph isomorphism neural network), GAT (graph attention network) and the structural response fast prediction method based on the physical encoding graph network of the embodiment. Specifically, Figure 4 the prediction based on the GCN is shown, Figure 5 the prediction of the GIN is shown, Figure 6 the prediction of the GAT is shown, Figure 7 the prediction of the embodiment is shown. The Figures 4 to 7 comparison is made, and in combination with Figure 8 the error bar contrast shown, it is found that the prediction of the structural response fast prediction method based on the physical encoding graph network of the embodiment is significantly better than the prediction of the GCN, GIN and GAT.

[0134] In a specific application of the embodiment, as shown in Figure 9 , in a truss structure, the black circle solid circle represents a free node, the red solid triangle represents a constraint node, the red arrow represents the load applied on the corresponding node, and the length of the red arrow represents the size of the applied load. Figure 10 the real displacement obtained based on the load condition of Figure 9 is shown, Figure 11 the predicted displacement obtained based on the displacement prediction of the embodiment is shown. By Figure 10 and Figure 11 , it can be clearly seen that the displacement prediction of the embodiment still has a high-precision prediction result under the condition of a small amount of response data set, and at least provides a more optimal technical solution for the structural response in the field of civil engineering.

[0135] Still based on the application shown in Figure 9 , in the truss structure, Figure 12 the real displacement obtained based on the load condition of Figure 9the true internal force obtained under the load condition, Figure 13 The predicted internal force based on the internal force prediction of the embodiment is shown. By Figure 12 and Figure 13 Through comparison, we can clearly see that the internal force prediction of the embodiment still has a high-precision prediction result under the condition of a small amount of response data set, and at least provides a more optimal technical solution for structural response in the field of civil engineering.

[0136] Based on the above application, we can deduce that in the field of civil engineering, when analyzing truss structures as representative of the truss structure, the mechanical calculation logic followed by the truss structure corresponds to the mathematical logic of the embodiment. Based on this correspondence, the embodiment is applicable to the analysis of engineering components in the field of civil engineering to achieve rapid prediction of structural response.

[0137] Based on the above description, the structure response rapid prediction method based on the physical coding graph network of the embodiment realizes the following technical innovations:

[0138] (1) The numerical simulation theory is used as the theoretical support to design the physical coding graph network, which greatly enhances the physical interpretability and generalization performance of the model, and has obvious advantages compared with conventional graph networks.

[0139] (2) The stiffness influence matrix is used as the normalized coding of the structure attribute. This method not only realizes efficient extraction of the structure attribute features, but also promotes the generalization of the load point features as an edge feature processing method.

[0140] Based on the above technical innovations, the embodiment realizes the following technical effects:

[0141] The physical coding graph network retains the data flow and training method of data-driven models, and the efficient structure attribute coding method and physically guided architecture design greatly enhance the physical interpretability and generalization performance of the model, achieving high generalization prediction accuracy with small samples, thereby solving the technical problems that the existing technology needs to be repeated in the face of similar load conditions in structural analysis, and in the complex multi-condition scene of structure operation and maintenance, the calculation is complicated and time-consuming, and it is difficult to meet the needs of rapid iterative design and real-time performance evaluation, especially suitable for the analysis of engineering components in the field of civil engineering that meet the mechanical calculation logic.

[0142] It should be understood that the above is only for illustration, and does not constitute any limitation on the technical solutions of the present application. In specific applications, those skilled in the art can set it up as needed, and the present application does not limit it.

[0143] It should be noted that the above-described workflow is merely illustrative and does not limit the protection scope of the present application, and in actual applications, a person skilled in the art can select part or all of the above-described workflow to achieve the purpose of the embodiment scheme according to actual needs, which is not limited herein.

[0144] It should be noted that in this paper, the term "including", "containing" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or system. Without more limitations, the element defined by the sentence "including a" does not exclude the presence of other identical elements in the process, method, article or system including the element.

[0145] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, it can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, an optical disk) and includes a number of instructions for making a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the method described in each embodiment of the present application.

[0146] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for fast prediction of structural response based on physical encoded graph network, characterized in that, The method comprises the following steps: S1: constructing a physical encoding graph network for encoding, message passing and decoding physical parameters of a target structure to obtain a structural response; wherein the physical parameters comprise node data and member data of physical properties of the target structure constrained by geometric relationships, and the structural response comprises at least displacement prediction and internal force prediction; S2: inputting the physical parameters of the target structure into the physical encoding graph network for encoding; S3: performing the message passing on the encoded physical parameters, wherein the message passing comprises node mapping and edge mapping performed in sequence, and data flow of the mapping network conforms to mathematical logic, the node mapping is used for performing the displacement prediction, and the edge mapping is used for performing the internal force prediction; S4: decoding results of the node mapping and the edge mapping to obtain displacement prediction results and internal force prediction results; the nodes comprise free nodes and constraint nodes to which loads are applied; the node data of the free nodes is defined as load node data; the node data of the constraint nodes is defined as constraint node data; the node data comprises the load node data and the constraint node data; the node data is used to obtain node features and encoding, comprising: the load node data is used to obtain load node features; the load node features are encoded as: wherein, represents the stiffness of a load node , defined as the load node data, is a load node number, represents the set of surrounding neighbor nodes of a node represents the stiffness influence of the surrounding nodes on the node on the node . the node data is used to obtain node features and encoding, further comprising: the constraint node data is used to obtain constraint node features; the constraint node features are encoded as: wherein, represents the stiffness of a constraint node and sets the stiffness of the constraint node to infinity to express the constraint property; the member data is used to obtain edge features and encoding, comprising: member data corresponding to the load node data is obtained to obtain load node edge features; based on the encoding of the load node features and the load node edge features, the load node edge features are encoded as: wherein, is the edge feature of the first adjacent edge and the load node after encoding, is the load node and the stiffness between its first adjacent edge corresponding surrounding neighbor nodes , is the edge feature of the second adjacent edge and the load node after encoding, is the load node and the stiffness between its second adjacent edge corresponding surrounding neighbor nodes , is the edge feature of the third adjacent edge and the load node after encoding, is the load node and the stiffness between its third adjacent edge corresponding surrounding neighbor nodes .

2. The method for fast prediction of structural response based on physical encoded graph network of claim 1, wherein, the construction of the physical encoding graph network comprises the following steps: A1: obtaining geometric relationships and the physical parameters of the target structure; based on the node data, node features are obtained and encoded, and based on the member data, edge features are obtained and encoded; A2: obtaining a small amount of response data sets of the target structure; based on A1 and the response data sets, the physical encoding graph network is trained, so that the physical encoding graph network encodes the node features and the edge features, performs the message passing based on different load cases input, and decodes and outputs different results of the node mapping and the edge mapping; wherein the response data sets store real responses of the target structure under different load cases, the load cases act on nodes corresponding to the node data, and the node data is used as input; A3: outputting the trained physical encoding graph network. 3.The physical encoded graph network based structural response fast prediction method of claim 1, wherein, the member data is used to obtain edge features and encoding, further comprising: member data corresponding to the constraint node data is obtained to obtain constraint node edge features; based on the encoding of the constraint node features and the constraint node edge features, the constraint node edge features are encoded as: wherein, is an edge feature of the first adjacent edge and the constrained node after encoding, is an edge feature of the second adjacent edge and the constrained node after encoding, is an edge feature of the first adjacent edge and the constrained node after encoding,​ 4. The method of fast prediction of structural response based on physical encoded graph network according to any one of claims 1 or 3, wherein, the displacement prediction comprises: Input the encoded node features and edge features into the physical coding graph network, perform displacement prediction through a node mapping formula, and define the result of the displacement prediction as the result of the node mapping; wherein the node mapping formula is: wherein, is the result of the mapping of the node, is the initial eigenvector of the node is the initial eigenvector of the node is the set of neighbor nodes of the node is the set of neighbor nodes of the node is the set of neighbor nodes of the node is the node feature of the always-updated neighbor node is the node feature of the always-updated neighbor node is the edge feature between the node and its neighbor node is the edge feature between the node and its neighbor node is the edge feature between the node and its neighbor node denotes the aggregation of the node and its neighbor node, and is a multilayer perceptron.

5. The method of fast prediction of structural response based on physical encoded graph network according to any one of claims 1 or 3, wherein, The internal force prediction comprises: Input the encoded node features and edge features into the physical coding graph network, perform internal force prediction through an edge mapping formula, and define the result of the internal force prediction as the result of the edge mapping; wherein the edge mapping formula is: wherein is the node and its neighbor nodes between which the edge mapping is performed, is the always-updated node node feature, is the always-updated neighbor node node feature, is the non-updated node and its neighbor nodes between which the edge feature is performed, is a multi-layer perceptron. 6.The physical encoded graph network based structural response fast prediction method of claim 2, wherein, The target structure at least comprises a truss structure; When the target structure is the truss structure, the nodes are connection points of the truss structure, and the components are members of the truss structure; The structure calculation follows the mathematical logic, and the mathematical logic at least comprises a structural mechanics calculation logic.

Images