Universal building structure tensorization feature extraction method and system and related equipment

Abstract

The invention discloses a general building structure tensorization feature extraction method and system and related equipment, and the method comprises the steps: S1, selecting the node, edge and overall features in a building structure, and selectively selecting a label for an end application; s2, establishing a random sample library of the building structures of the same type based on the features and the labels selected in the step S1; s3, performing XYZ three-direction scanning on the building structure based on the building structure random sample library established in the step S2 to form a three-direction series diagram, and storing a three-dimensional diagram at the same time; s4, establishing a feature extraction and tensor data representation network; and S5, feature extraction, tensor data representation network training and model solidification are carried out. According to the scheme, the problems existing in the prior art can be effectively solved, the universal data feature representation content of the building structure can be formed, and the method can be applied to various tail ends such as subsequent access design algorithms, optimization algorithms and response prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent design and optimization of civil structures, and particularly to a general tensor feature extraction scheme for building structures. Background Art

[0002] At present, the development of computer and artificial intelligence technologies is rapid, and a large number of cutting-edge algorithms and models are applied in various fields, becoming a bridge for the transformation of many traditional fields to intelligence. The informatization and intelligent development of the construction industry also rely on the empowerment of cutting-edge technologies. Conducting cross-research on artificial intelligence technologies and construction technologies is the core strategy for serving the national strategic transformation and sustainable development, accelerating the innovation of key cutting-edge technologies, and empowering the higher-level development of the large construction industry. However, the popularity and integration depth of such cutting-edge technologies in construction-related fields lag far behind other fields. The key reason is that building structures contain a large amount of complex information such as nodes, components, and spatial configurations, which have characteristics such as multi-dimensionality and difficulty in serialization, making it difficult to be represented as unified tensor data, and there is a significant gap with the current intelligent algorithm and model technologies centered on tensor data flow and probability, making it difficult to deeply integrate and apply.

[0003] At the same time, with the proposal and popularization of a large number of complex structural systems and new components, traditional design methods and concepts cannot adapt to structures or components with various characteristics, and relevant architectural and structural design methods and concepts also need to keep up with the times. Establishing design methods for different new structures or components respectively has great difficulties in practice, consumes a lot of manpower and resources, and cannot be synchronized to engineering applications in real time, thus hindering the implementation of advanced systems.

[0004] If, based on the background of the current intelligent era, through an abstract method at the structural level, the building structure is digitalized, and subsequent design and optimization are carried out based on general format data streams, the problem of different design methods required due to inconsistent characteristics of different structures and components can be effectively solved. Extensive prediction work can also be carried out based on building data streams in earthquake prediction, vulnerability analysis, economic analysis, and regional disaster analysis, introducing advanced algorithm models in multiple fields, and empowering the intelligent development of the civil engineering and construction fields.

[0005] However, for feature extraction and digitalization of building structures, the current digitalization schemes for building structures have the following limitations in actual application processes:

[0006] 1) Existing digitalization schemes can only represent information for local parts. Due to limited data input volume, general building structure feature extraction and digitalization methods are difficult to organize reasonable data formats when facing a large amount of building structure data, so they can only represent information for local parts.

[0007] 2) Existing data conversion solutions cannot effectively consider the spatial characteristics of building structures. The spatial characteristics of building structures are multi-dimensional data, and it is difficult to express them in a reasonable way in a flattened tensor or other data formats. Moreover, directly constructing a spatial format will pose great obstacles to subsequent data processing and algorithm application.

[0008] 3) The data representation forms output by existing data conversion solutions are limited or fixed. Traditional methods will vectorize and combine data or directly input it into an algorithm to obtain output data. Its form is extremely dependent on the process formulator, and the final output data form will be affected by the quantity, dimension, framework type, etc. of the original data, resulting in poor generality.

[0009] 4) Existing data conversion solutions do not have self-adaptability for different building structure types, forms, and parameters. When the building structure type, form, parameters, etc. change, it means that the feature extraction and data conversion methods to be applied need to process input parameters of different quantities, types, and scales, as well as different intermediate processes and different output parameters. Traditional methods do not have similar adaptability. Summary of the Invention

[0010] Aiming at the problems and limitations existing in the existing feature extraction and data conversion solutions for building structures, the purpose of the present invention is to provide a general tensor feature extraction method for building structures, which can effectively overcome the problems existing in the prior art, can form general data feature representation content of building structures, and can be connected to various end applications such as design algorithms, optimization algorithms, and response prediction in the future; the present invention further provides a system and related device solutions that can implement this method.

[0011] To achieve the above purpose, the general tensor feature extraction method for building structures provided by the present invention includes:

[0012] Step S1: Select nodes, edges, and overall features in the building structure, and selectively select labels for end applications;

[0013] Step S2: Establish a random sample library of the same type of building structures based on the features and labels selected in Step S1;

[0014] Step S3: Based on the random sample library of building structures established in Step S2, scan the building structure in three directions of X, Y, and Z to form a three-direction series of graphs and save the three-dimensional graph at the same time;

[0015] Step S4: Build a feature extraction and tensor data representation network;

[0016] Step S5: Train the feature extraction and tensor data representation network and solidify the model.

[0017] Further, when performing the selection of nodes, edges, and overall features in step S1 and selectively selecting labels for the end applications, the following steps are included:

[0018] Step (S11): Analyze the component types, node types, and special attributes in the target building structure;

[0019] Step (S12): Based on the node types parsed in step (S11), perform node feature selection;

[0020] Step (S13): Based on the component types parsed in step (S11), perform component feature selection;

[0021] Step (S14): Select the features that can represent the overall characteristics of the building structure as the overall features;

[0022] Step (S15): Select the parameters required for the end application as the labels;

[0023] Step (S16): Calculate the dimensions of the nodes, edges, overall features, and labels based on the features and labels selected in steps (S12) to (S15) for subsequent network construction.

[0024] Further, when step S2 establishes a random sample library of the same type of building structure, the following steps are included:

[0025] Step (S21): Set the generation ranges of the spatial parameters and construction parameters of the building structure;

[0026] Step (S22): Establish a selection database for component cross-sections;

[0027] Step (S23): Establish an analysis library such as a seismic wave record selection database or other analysis libraries required for model analysis;

[0028] Step (S24): Develop random generation rules for the building structure;

[0029] Step (S25): Based on the parameters set in steps (S21) - (S23), the constructed databases, and the random generation rules for the building structure developed in step (S24), construct a sample random generation model;

[0030] Step (S26): Randomly generate samples based on the sample random generation model in step (S24) and collect the target number of samples.

[0031] Further, when step S3 performs XYZ three-directional scanning on the building structure, the following steps are included:

[0032] Step (S31): Set the scanning direction for the samples generated in step (S2);

[0033] Step (S32): Scan the building structure according to the scanning direction set in step (S31) and obtain the two-dimensional building structure framework during the scanning process;

[0034] Step (S33): Extract the nodes, edges, and overall features selected in step (S1) in the building structure to form a tensor format;

[0035] Step (S34): Establish a node mapping mechanism and an edge connection representation tensor of an undirected graph for the two-dimensional framework and the three-dimensional structure obtained in step (S32) respectively;

[0036] Step (S35): Independently save the graphs constructed by the two-dimensional framework and the three-dimensional structure.

[0037] Furthermore, when performing feature extraction and tensor data representation network construction in step (S4) based on step (S3), the following steps are included:

[0038] Step (S41): Select the architecture of the graph neural network and the node feature update mechanism;

[0039] Step (S42): Set the neural network architecture and related hyperparameters for the output of edge features and overall features;

[0040] Step (S43): Update the node features of the target graph through the graph neural network in step (S41);

[0041] Step (S44): Output the edge features updated to a specific dimension after passing the edge features through the neural network specified in step (S42);

[0042] Step (S45): Aggregate the edge features and adjacent node features through data aggregation to form new edge features;

[0043] Step (S46): For the new edge features formed in step (S45), accumulate and aggregate the edge features of the entire graph and splice the overall features of the aggregated graph to form new graph overall features;

[0044] Step (S47): Select a method of directional feature aggregation or non-discriminatory aggregation to aggregate the overall features of all two-dimensional framework graphs to form the final tensorized features of the building structure;

[0045] Step (S48): Obtain the features of the three-dimensional framework graph according to the same steps and splice them with the overall features of the two-dimensional framework graph in step (S47) to form the final tensorized features of the building structure.

[0046] Furthermore, when performing feature extraction and tensor data representation network training and model solidification in step (S5), it is first necessary to determine whether to train for end applications:

[0047] If so, train with the network and method determined in step (S4), using the parameters required by the end application as labels. After achieving the desired effect, solidify and save the model parameters to form a dedicated model.

[0048] If not, train with the network and method determined in step (S4), using the parameters of the structure itself as labels. After achieving the desired effect, discard the final output layer of the model and only solidify and save the other model parameters.

[0049] To achieve the above object, the general building structure tensorized feature extraction system provided by the present invention includes: a selection module, a sample library establishment module, a three-way scanning module, a feature extraction and tensor data representation network module, and a training and model solidification module.

[0050] The selection module is configured to be able to select nodes, edges, and overall features, and selectively select labels for end applications.

[0051] The sample library establishment module is configured to interact with the selection module in terms of data, and be able to establish a random sample library of the same type of building structure.

[0052] The three-way scanning module is configured to interact with the sample library establishment module in terms of data, and be able to perform three-way scanning of the building structure in the XYZ directions to form a three-way series of images and simultaneously save the three-dimensional image.

[0053] The feature extraction and tensor data representation network module is configured to interact with the three-way scanning module in terms of data, and be able to build a feature extraction and tensor data representation network.

[0054] The training and model solidification module is configured to interact with the feature extraction and tensor data representation network module in terms of data, and be able to train the feature extraction and tensor data representation network and solidify the model.

[0055] Further, the building structure tensorized feature extraction system further includes a transfer learning module. The transfer learning module is configured to interact with the training and model solidification module in terms of the dataset, and be able to connect to the end application and add labels for transfer learning.

[0056] To achieve the above object, the present invention also provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, the steps of the above general building structure tensorized feature extraction method are implemented.

[0057] To achieve the above object, the present invention also provides a processor. The processor is used to run a program, and when the program runs, the steps of the above general building structure tensorized feature extraction method are executed.

[0058] To achieve the above object, the present invention further provides a terminal device, which includes a processor, a memory, and a program stored on the memory and executable on the processor. The program code is loaded and executed by the processor to implement the steps of the above general method for extracting tensor features of building structures.

[0059] To achieve the above object, the present invention further provides a computer program product, which is adapted to execute the steps of the above general method for extracting tensor features of building structures when executed on a data processing device.

[0060] Compared with the prior art, the general solution for extracting tensor features of building structures provided by the present invention has the following advantages:

[0061] 1) The present invention can effectively capture and process the global information of building structures.

[0062] 2) The present invention incorporates the spatial features of building structures and represents them in multiple dimensions such as nodes and the entire graph.

[0063] 3) The present invention can flexibly control the extraction process and the form of the final output tensor data stream.

[0064] 4) The present invention can adapt to different types and forms of building structures.

[0065] When the general solution for extracting tensor features of building structures provided by the present invention is specifically applied, it has the following advantages:

[0066] 1) It has strong information extraction ability and a wide information view, and can upgrade the expression of local information to the expression of global information.

[0067] 2) It has the ability to process the spatial features of building structures and comprehensively consider them through multiple dimensions.

[0068] 3) It has a high degree of controllability over the process and the final output tensor data stream, improving the freedom of end applications and intermediate model extraction.

[0069] 4) It can flexibly adapt to different types and forms of building structures with the same process and method. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The present invention will be further described below in conjunction with the drawings and specific embodiments.

[0071] Figure 1 is a flowchart of the general method for extracting tensor features of building structures in the present invention;

[0072] Figure 2 is a flowchart of selecting node, edge, and overall features and selectively selecting labels for end applications in the present invention;

[0073] Figure 3 Flow chart for establishing a random sample library of the same type of building structure in the present invention;

[0074] Figure 4 Flow chart for three-way scanning of a building structure in the present invention to form a three-way series of diagrams and simultaneously save a three-dimensional diagram;

[0075] Figure 5 Flow chart for building a feature extraction and tensor data characterization network in the present invention;

[0076] Figure 6 Flow chart for training and model solidification of the feature extraction and tensor data characterization network in the present invention;

[0077] Figure 7 Example diagram of a two-dimensional framework scanned in an example of the present invention;

[0078] Figure 8 Example diagram of feature extraction of a graph neural network in an example of the present invention. Detailed implementation manners

[0079] In order to make the technical means, creative features, achieved purposes and effects realized by the present invention easy to understand, the present invention will be further described below with reference to specific diagrams.

[0080] The solution of the present invention provides a general tensorized feature extraction solution for building structures, which can perform abstraction processing on the structural level of building structures and complete feature extraction and dataization of building structures.

[0081] In this general tensorized feature extraction solution for building structures, first, based on graph theory, the building structure is abstracted into multiple planar diagrams in three dimensions in the way of "XYZ three-dimensional depth scanning", and the parameters related to nodes, components, etc. are converted into node features and edge features in standard formats and added to each diagram; then, the node features and edge features are updated through corresponding neural network algorithms such as graph convolution and self-attention graph neural networks; on this basis, the graph features in multiple dimensions are further aggregated or spliced with additional overall supplementary features to obtain corresponding tensor data streams, and these tensor data streams are used as the general data feature characterization content of the building structure. Thus, the tensorized feature extraction of the building structure is completed, and the extracted feature data can be subsequently connected to various end applications such as design algorithms, optimization algorithms, and response prediction.

[0082] Based on the above scheme mechanism, the present invention specifically provides a general tensorized feature extraction method for building structures. As shown in Figure 1 This general tensorized feature extraction method for building structures is mainly composed of the following steps in cooperation:

[0083] Step S1: Select the nodes, edges, and overall features of the building structure, and selectively select labels for the end applications.

[0084] Step S2: Based on the features and labels selected in Step S1, establish a random sample library of the same type of building structure.

[0085] Step S3: Based on the random sample library of the building structure established in Step S2, perform three-dimensional scans of the building structure in the X, Y, and Z directions to form a three-way series of diagrams and save the three-dimensional diagram simultaneously.

[0086] Step S4: Build a feature extraction and tensor data representation network.

[0087] Step S5: Train the feature extraction and tensor data representation network and solidify the model.

[0088] Step S6: Connect to the end applications and add labels for transfer learning.

[0089] In some embodiments of the present invention, when performing the selection of nodes, edges, and overall features and selectively selecting labels for the end applications in Step S1 of the present extraction method, as Figure 2 shown, it is specifically implemented through the following steps:

[0090] Step S11: Analyze the component types, node types, and special attributes in the target building structure.

[0091] Step S12: Based on the node types parsed in Step S11, select node features.

[0092] Step S13: Based on the component types parsed in Step S11, select component features.

[0093] Step S14: Select the features that can represent the overall characteristics of the building structure as the overall features.

[0094] Step S15: Select the parameters required for the end applications as labels.

[0095] Step S16: Based on the features and labels selected in Steps S12 to S15, calculate the dimensions of the nodes, edges, overall features, and labels for subsequent network construction.

[0096] As a further explanation, when analyzing the component types, node types, and special attributes in the target building structure in Step (11) of the present extraction method, it is preferable to analyze the target building structure, sort out the types of components therein, such as whether there are supports in addition to common beam and column components, and clarify the node types and their physical properties, such as whether there are both rigid and hinged nodes at the same time.

[0097] On this basis, this step further analyzes whether there are other special attributes related to the building structure. If there are other special attributes related to the building structure, they are characterized in the form of tensor data and can be added to the tensor data of the overall features subsequently.

[0098] As a further illustration, when selecting node features in step (S12) of this extraction method, the selected node features are specifically related to the attributes involved at the structural node, such as the spatial coordinates of the node, the binary representation of whether it is fixed, the nodal inertial mass, the node type, or other physical parameters, etc., and can ultimately form a one-dimensional node feature tensor with a unified format.

[0099] As a further illustration, when selecting edge features in step (S13) of this extraction method, the selected edge features are specifically related to the attributes involved in the component, such as the length of the component, the cross-sectional dimension parameters, the local coordinate direction, and other parameters related to the structural performance, etc., and can ultimately form a one-dimensional edge feature tensor with a unified format.

[0100] As a further illustration, when selecting features that can represent the overall characteristics of the building structure as the overall features in step (S14) of this extraction method, a series of features that can represent the overall characteristics of the building structure are selected as its overall features. The overall features can be direct features related to the structure, such as the storey height, the number of spans, etc., or indirect features obtained through finite element analysis, such as the Nth-order period, etc., and ultimately a one-dimensional edge feature tensor with a unified format.

[0101] As a further illustration, step (S15) in this extraction method selects the parameters required for the end application, which is an optional step compared to the solution in step S1. Specifically, if training a general structural feature model, this step is temporarily not required; if training a model for a specific end application, the parameters required for it need to be selected as the labels of the subsequent dataset. For example, if the end application is the prediction of the maximum displacement of a layer under seismic excitation, the maximum displacement of the layer will be used as the label.

[0102] As a further illustration, when calculating the dimensions of the nodes, edges, overall features, and labels in step (S16) of this extraction method, specifically based on the nodes, edges, overall features, and labels selected in the aforementioned steps (S12 - 15), the dimensions corresponding to them as tensors are determined, which are used for the construction of the network architecture in the subsequent graph neural network and neural network and the determination of dimension conversion. As an example, if a node feature is [n1, n2, n3], its dimension is 3, and the node input dimension of the subsequent graph neural network will also be 3.

[0103] In some embodiments of the present invention, step S2 in this extraction method cooperates with the following 6 sub-steps to establish a random sample library of the same type of building structure, as Figure 3 shown, and the specific implementation steps are as follows:

[0104] Step (S21): Set the generation ranges of the spatial parameters and construction parameters of the building structure.

[0105] Step (S22): Establish a selection database for component cross-sections.

[0106] Step (S23): Establish an analysis database such as a seismic wave record selection database or other analysis databases required for model analysis.

[0107] Step (S24): Formulate the random generation rules for the building structure.

[0108] Step (S25): Based on the parameters set in steps (S21)-(S23), the established databases, and the random generation rules for the building structure formulated in step (S24), construct a sample random generation model.

[0109] Step (S26): Randomly generate samples based on the sample random generation model in step (S24) and collect the target number of samples.

[0110] As a further illustration, since a random sample library needs to be established in step S2 of this extraction method, when setting the generation ranges of the spatial parameters and construction parameters of the building structure in step (S21), first determine the generation ranges of the spatial parameters and construction parameters. Among them, the spatial parameters such as the number of spans in two directions, the length of the spans, the number of structural layers, the storey height, etc. are limited, which can ensure that the randomly generated building structure samples are controlled within a certain range and will not form a structure that is geometrically divorced from reality; the construction parameters such as the nodal inertial mass, etc. are limited, which can control the general properties of the building structure and ensure that there will be no obvious design deviations.

[0111] As a further illustration, when establishing a selection database for component cross-sections in step (S22) of this extraction method, it is specifically configured to establish a database containing the cross-section parameter information of beams, columns, and other components, where the cross-section parameter information of a single cross-section is stored as item information. Based on this cross-section database, when generating subsequent building structure random samples, the cross-section will be selected and assigned for each generated component from the cross-section database.

[0112] As a further illustration, step (S23) of this extraction method to establish an analysis database such as a seismic wave record selection database or other analysis databases required for model analysis is an optional step relative to the solution in step S2.

[0113] Specifically, if there is a need for end applications, the corresponding database or analysis database can be established as needed. For example, if the subsequent model needs to predict the maximum storey displacement of the building structure under seismic excitation, its label is the maximum storey displacement.

[0114] On this basis, when collecting samples, each sample will undergo a finite element analysis of seismic excitation to obtain its maximum story displacement. The seismic wave records used in this process can be randomly selected from the established seismic wave record selection database and scaled to the required intensity level as required.

[0115] As a further illustration, when formulating the random generation rules for building structures in step (S24) of this extraction method, it is preferred to formulate a set of rules for guiding and constraining the generation of random samples of building structures. This set of rules includes the generation methods of spatial parameters and construction parameters of building structures, the distribution patterns of frame sections in different directions, the sectional change patterns along the building height, etc. It is required that the building structures generated according to the rules do not deviate from the actual design specifications and local weaknesses and unevenness should be avoided as much as possible.

[0116] As a further illustration, when constructing the sample random generation model in step (S25) of this extraction method, specifically, the parameters and data specified in steps (S21) to (S23) are used to construct the corresponding sample random generation model according to the rules formulated in step (S24).

[0117] As an example, the sample random generation model here is configured to be able to first randomly generate the spatial parameters and construction parameters of the sample, form the distribution information of nodes and members, select the member sections according to the sample geometry and dimensions, generate the model file and information record file, conduct modal analysis and analysis of end application requirements (such as dynamic analysis), and save the analysis results.

[0118] As a further example, the sample random generation model here can be constructed by a corresponding software program.

[0119] As a further illustration, when randomly generating samples based on the configuration of the sample random generation model constructed in step (S25) and collecting a target number of samples in step (S26) of this extraction method, first, a target number is determined to determine the final required number of samples. This number should meet the requirements of the dataset required for subsequent model training. A larger number of samples is beneficial to the model effect, but it also means higher resource consumption. Therefore, a balance needs to be struck between cost and model effect. The number of samples is usually at least in the order of thousands or even higher.

[0120] In some embodiments of the present invention, step S3 in this extraction method is performed on each randomly generated building structure sample in the random sample library of the same type of building structures established in step S2, that is, three-dimensional scans are performed on each randomly generated building structure sample in the X, Y, and Z directions to form three series of diagrams and the three-dimensional diagrams are saved simultaneously.

[0121] Specifically, as Figure 4 shown, step S3 in this extraction method is specifically composed of the following 5 sub-steps in cooperation:

[0122] Step (S31): Set the scanning direction for the samples generated in step (S2).

[0123] Step (S32): Scan the building structure according to the scanning direction set in step (S31) and obtain the two-dimensional building structure framework during the scanning process.

[0124] Step (S33): Extract the nodes, edges, and overall features selected in step (S1) from the building structure to form a tensor format.

[0125] Step (S34): Establish a node mapping mechanism and an edge connection representation tensor of an undirected graph for the two-dimensional framework and the three-dimensional structure obtained in step (S32) respectively.

[0126] Step (S35): Independently save the graphs constructed by the two-dimensional framework and the three-dimensional structure.

[0127] As a further explanation, when this extraction method sets the scanning direction for the samples generated in step (S2) in step (S31), for the scanning direction set for the building structure, if it is a regular building structure, the three directions of x, y, and z can be set, and the three directions are orthogonal to each other; if there is an irregular-shaped building structure, multiple scanning directions can be set to ensure that the building structure can be completely decomposed into a two-dimensional plane framework in each scanning direction.

[0128] As a further explanation, when this extraction method scans the building structure and obtains the two-dimensional building structure framework during the scanning process in step (S32), specifically according to each direction determined in step (S31), a virtual plane is used to scan along the direction respectively, and the framework in the building structure framework that coincides with the virtual plane is the two-dimensional building structure framework obtained by scanning in this direction; and the two-dimensional frameworks are classified and collected according to the direction.

[0129] As a further explanation, when this extraction method extracts the nodes and edge features selected in step (S1) from the building structure to form a tensor format in step (S33), the features of the two-dimensional framework obtained by scanning in step (S32) are extracted, which are the nodes and edge features selected in step (S1), and the source of the features is the original information saved when generating random samples in step (S2).

[0130] On this basis, further extract the nodes and edge features for the three-dimensional framework of the building structure itself. The extracted features are processed and constructed into a tensor format.

[0131] As a further explanation, when this extraction method establishes a node mapping mechanism and an edge connection representation tensor of an undirected graph for the two-dimensional framework and the three-dimensional structure respectively in step (S34),

[0132] First, it is necessary to establish a mapping mechanism for the nodes extracted from the two-dimensional framework and the three-dimensional structure. Regarding the problem that nodes in a graph neural network usually do not have a numbering mechanism but use sequential encoding, while node numbering is required to construct a finite element model in sample generation and finite element analysis, in this step, by mapping the original node numbers to fixed sequential encodings, sequential encodings can be established in ascending order of the original node numbers and saved and solidified in a specific data format such as an array.

[0133] Secondly, since the graph formed by the building structure framework is an undirected graph, that is, there is no directionality between the two nodes of a single edge. In this step, an edge connection tensor in the unique format of an undirected graph is established, and its format needs to clearly indicate that it belongs to an undirected graph; depending on the different network training frameworks used, the representation methods are different. For example, in pyg of torch, the two nodes of a single edge are stored in reverse twice to represent the undirected nature of the edge. Usually, after completing the node mapping mechanism, the node numbers at both ends of the components in the building structure are converted into sequential encodings and represented and stored in tensor format.

[0134] As a further illustration, in step (S34) of this extraction method, when independently saving the graphs constructed from the two-dimensional framework and the three-dimensional structure, it is necessary to independently save the graph corresponding to the scanned two-dimensional framework and the graph formed by the three-dimensional structure itself. At the same time, each graph will contain node features, edge connection tensors, edge features, and global features, thereby meeting the requirement that each graph needs to be processed by a graph neural network and feature aggregation is performed in subsequent model construction.

[0135] In some embodiments of the present invention, in step (S4) of this extraction method, based on step (S3), when building a feature extraction and tensor data representation network, as Figure 5 shown, it is specifically implemented through the following 8 sub-steps:

[0136] Step (S41): Select the architecture of the graph neural network and the node feature update mechanism;

[0137] Step (S42): Set the neural network architecture and related hyperparameters for outputting edge features and global features;

[0138] Step (S43): Update the node features of the target graph after passing through the graph neural network in step (S41);

[0139] Step (S44): Output the edge features as updated edge features in a specific dimension after passing the edge features through the neural network specified in step (S42);

[0140] Step (S45): Aggregate the edge features and adjacent node features through a data aggregation method to form new edge features;

[0141] Step (S46): Form new edge features for step (S45), accumulate and aggregate the edge features of the whole graph, and splice the overall features of the aggregated graph to form new overall graph features;

[0142] Step (S47): Select the method of aggregating features in different directions or aggregating without discrimination to aggregate the overall features of all two-dimensional frame diagrams, and form the final tensorized features of the building structure;

[0143] Step (S48): Obtain the features of the three-dimensional frame diagram according to the same steps and splice them with the overall features of the two-dimensional frame diagram in step (S47) to form the final tensorized features of the building structure.

[0144] As a further explanation, when selecting the architecture of the graph neural network and the node feature update mechanism in step (S41) of this extraction method, first select a graph neural network architecture for subsequent node feature update.

[0145] As an example, common network types include graph convolutional network (GCN), GraphSAGE, GAT, etc. Different network types represent different node update methods and propagation methods;

[0146] Then, after completing the selection of the network type, further determine the node feature update mechanism. Different networks involve the number of propagation layers, the way of aggregating node feature information, attention mechanism parameters, etc.

[0147] As a further explanation, when setting the neural network architecture and related hyperparameters for edge feature and overall feature output in step (S42) of this extraction method, it is necessary to determine the architecture and hyperparameters for the edge feature processing neural network and the overall feature output neural network, including the number of layers, the number of hidden layer nodes, the type of activation function, the application ratio of dropout, etc.

[0148] As a further explanation, when updating the node features by passing the target graph through the graph neural network in step (S41) in step (S43) of this extraction method, and when outputting the edge features as updated edge features of a specific dimension after passing the edge features through the neural network specified in step (S42) in step (S44), the node and edge features are updated respectively. Only by inputting the original graph information with node and edge features into the network can the update be completed.

[0149] As a further explanation, when aggregating the edge features and adjacent node features to form new edge features based on the corresponding data aggregation method in step (S45) of this extraction method, specifically, it can be summation, averaging, splicing, or more complex output through a neural network. Among them, the summation and averaging methods require the updated node features and edge features to have the same dimension; while splicing or output through a neural network does not require the same dimension. After data aggregation, new edge features will be obtained.

[0150] As a further illustration, when the extraction method accumulates and aggregates the edge features of the whole graph and splices the overall features of the aggregated graph to form new overall features of the graph in step (S46), first, all the new edge features in step (S45) should be accumulated and aggregated to obtain a one-dimensional feature, and then this one-dimensional feature is spliced with the overall features of the graph to obtain new overall features of the graph, which are the final overall features of a single two-dimensional or three-dimensional framework.

[0151] As a further illustration, when the extraction method aggregates the overall features of all two-dimensional framework graphs in step (S47) by selecting the method of directional feature aggregation or non-discriminatory aggregation to form the final tensorized features of the building structure, two methods of aggregating the features of two-dimensional framework graphs are preferably adopted:

[0152] One is to aggregate the features of graphs in the same direction, and finally form one-dimensional features at multiple direction levels; on this basis, they are spliced to form the features of the final building structure;

[0153] The other is to ignore the directionality of the two-dimensional framework, but to accumulate and aggregate its overall features without discrimination, and finally form the features of the building structure.

[0154] As a further illustration, step (S48) in the extraction method is an optional step compared with the solution in step S4. Based on step (S47), the final features obtained by processing the two-dimensional framework are spliced with the overall features of the three-dimensional framework to form new features, which are used as the final features of a single building structure.

[0155] In some embodiments of the present invention, when step (S5) in the extraction method performs feature extraction, tensor data representation network training, and model solidification, as Figure 6 shown, it is specifically implemented through the following steps:

[0156] First, it is necessary to determine whether to train for the end application (S51):

[0157] If so, first enter step (S521), and use the network and method determined in step (S4) to train with the parameters required for the end application as labels; then enter step (S522), and after achieving the effect, solidify and save the model parameters to form a dedicated model;

[0158] If not, first enter step (S531), and train with the network and method determined in step (S4), using the parameters of the structure itself as labels. The parameters of the structure itself should be high-dimensional parameters obtained through analysis, which can, to a certain extent, vividly represent the characteristics of the structure, such as multi-order natural vibration periods, displacement ratios of pushover, etc.; then enter step (S532). After achieving the effect, discard the final output layer of the model and only solidify and save other model parameters. Since the model network architecture needs to pass through a neural network when obtaining the final output result, and the general feature extraction model will ultimately be used to extract features in the series of end applications, in this step, it is necessary to discard the output layer at the very end after model training and only retain the parameters of other layers. When applying, the output of the second-to-last hidden layer can be combined with the network architecture of the end application to form a new network model.

[0159] In some embodiments of the present invention, step S6 in this extraction method for docking with end applications and adding labels for transfer learning is an optional step, and it is determined whether to further adopt step 6 according to the actual situation.

[0160] The general method for extracting tensorized features of building structures formed based on the above solution has the following advantages compared with the prior art:

[0161] 1) It can achieve the capture and processing of global information of building structures. Through the graph neural network, the transmission of information in the graph and the accumulation of multiple graph information can be realized, so as to extract global information as close as possible.

[0162] 2) Incorporate spatial features and represent them in multiple dimensions such as nodes and the whole graph. By integrating spatial features into node features and edge features, the original spatial features are completely retained; and the spatial features of the building structure are concatenated through the way of two-dimensional frame scanning and three-dimensional frame diagrams.

[0163] 3) It can flexibly control the extraction process and the form of the final output tensor data stream to meet the requirements of end applications. This method combines two basic models, the graph neural network and the neural network, and controls the form of the tensor data stream during the process and the final output by flexibly adjusting data aggregation, data splicing, and the flow direction of the tensor stream to adapt to different end application requirements or the extraction of intermediate models.

[0164] 4) The native concept adapts to different types and forms of building structures. By means of two-dimensional frame scanning and inputting into the graph neural network for processing, it is inherently adaptable to any structural form. It can flexibly process the acquisition of original data of various structural forms and subsequent high-dimensional feature extraction by increasing or decreasing the scanned virtual planes and changing the network hyperparameters.

[0165] For the general method for extracting tensorized features of building structures given in the solution of the present invention, in specific applications, a corresponding software program can be formed to create a corresponding general system for extracting tensorized features of building structures. When this software program runs, it can execute the above-mentioned general method for extracting tensorized features of building structures and is stored in a corresponding storage medium for the processor to retrieve and execute.

[0166] The thus-formed general system for extracting tensorized features of building structures mainly includes, in terms of function: a selection module, a sample library establishment module, a three-way scanning module, a feature extraction and tensor data representation network module, a training and model solidification module, and a transfer learning module.

[0167] Among them, the selection module is configured to be able to select nodes, edges, and overall features, and selectively select labels for end applications.

[0168] The sample library establishment module is configured to interact with the selection module in terms of data and be able to establish a random sample library of building structures of the same type.

[0169] The three-way scanning module is configured to interact with the sample library establishment module in terms of data and be able to perform three-way scanning of the building structure in the XYZ directions to form a series of three-way diagrams and save the three-dimensional diagram at the same time.

[0170] The feature extraction and tensor data representation network module is configured to interact with the three-way scanning module in terms of data and be able to build a feature extraction and tensor data representation network.

[0171] The training and model solidification module is configured to interact with the feature extraction and tensor data representation network module in terms of data and be able to train and solidify the feature extraction and tensor data representation network;

[0172] The transfer learning module is configured to interact with the training and model solidification module in terms of data sets, be able to interface with end applications and add labels for transfer learning.

[0173] In some embodiments of the present invention, the selection module in this system is configured to be implemented based on the solution of steps (S11) to (S16).

[0174] In some embodiments of the present invention, the sample library establishment module in this system is configured to be implemented based on the solution of steps (S21) to (S26).

[0175] In some embodiments of the present invention, as Figure 6 shown, the three-way scanning module in this system is configured to be implemented based on the solution of steps (S31) to (S35).

[0176] In some embodiments of the present invention, the feature extraction and tensor data representation network module in the system is configured to be implemented based on the solution of steps (S41) to (S48).

[0177] In some embodiments of the present invention, the training and model solidification module in the system is configured to be implemented based on the solution of steps (S51) to (S522) and / or the solution of step (S532).

[0178] For the general building structure tensorized feature extraction solution given by the solution of the present invention, the following will be further illustrated by specific application examples.

[0179] Here, a three-story frame building structure with a stepped appearance is taken as an example for illustration.

[0180] It is assumed that after the generation of random samples, a three-story frame building structure as shown in Figure 7 is obtained, and three directions are specified for it, namely the X, Y, and Z directions.

[0181] Virtual planes are added in the three directions, and the virtual planes pass through the building structure along this direction, and the two-dimensional frames that coincide and intersect with it are separated and displayed, so that all the two-dimensional frame planes in the three directions can be obtained. It can be seen from the implementation case that four planes are separated in the X direction, namely the first-floor two-dimensional frame X-1, the second-floor two-dimensional frame X-2, the third-floor two-dimensional frame X-3, and the third-floor two-dimensional frame X-4. Similarly, 4 and 3 two-dimensional frames are obtained in the Y and Z directions respectively. The two-dimensional frames obtained in the three directions can respectively form a single graph for subsequent training of the graph neural network model.

[0182] As shown in Figure 8 , this example takes the Z-3 two-dimensional frame in the three-story frame building structure as an example to illustrate the operation mechanism of the graph neural network. There are a total of 8 nodes and 10 members (edges) in Z-3. It is assumed that the node features are tensors in the format of [n1, n2, n3...], and the edge features are tensors in the format of [e1, e2, e3...].

[0183] First, the node features need to be updated through a specified graph neural network. The core is information transmission, that is, the information of surrounding nodes is concentrated on the target node through a certain propagation method and aggregation method. When the same processing is performed on each node, the global transmission and uniform diffusion of information can be realized. The edge features are extracted in high dimensions by using a neural network. The input features are the original [e1, e2, e3...], and the output features are [fe1, fe2, fe3...]. The output features depend on the architecture of the neural network.

[0184] Secondly, the edge features after high-dimensional extraction are concatenated with the features of adjacent nodes after update, thereby obtaining updated edge features in the format of [n1, n2, n3... fe1, fe2, fe3... e1, e2, e3...], where the order of the tensors does not affect the result. All updated edge features have the same dimension, and they are all accumulated to obtain the overall feature of the graph. The original overall feature can also be considered to be added to expand the feature dimension.

[0185] Finally, the overall features of all two-dimensional framework diagrams are accumulated to obtain the overall feature of the building structure, and the final result is output via a neural network. The overall feature of the three-dimensional framework diagram can also be considered to be aggregated in the overall feature of the building structure.

[0186] As can be seen from the above examples, the general building structure tensorized feature extraction scheme provided by the present invention captures and processes the effective global information of the building structure.

[0187] The general building structure tensorized feature extraction scheme provided by the present invention incorporates the spatial features of the building structure and is characterized in multiple dimensions such as nodes and the entire graph.

[0188] The general building structure tensorized feature extraction scheme provided by the present invention can flexibly control the extraction process and the form of the final output tensor data stream.

[0189] The general building structure tensorized feature extraction scheme provided by the present invention can adapt to processing different types and forms of building structures.

[0190] An embodiment of the present invention also provides a processor, and the processor is used to run a program. When the program runs, it executes the steps of the above general building structure tensorized feature extraction method.

[0191] An embodiment of the present invention also provides a terminal device. The device includes a processor, a memory, and a program stored on the memory and executable on the processor. The program code is loaded and executed by the processor to implement the steps of the above general building structure tensorized feature extraction method.

[0192] The present invention also provides a computer program product, which is suitable for executing the steps of the above general building structure tensorized feature extraction method when executed on a data processing device.

[0193] In the above embodiments, the descriptions of the various embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0194] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the systems, devices, and modules described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0195] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0196] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, apparatuses (systems), and computer program products of the embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0197] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0198] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0199] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and a memory.

[0200] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0201] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0202] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0203] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0204] The above method of the present invention, or a specific system unit, or a part thereof, is a pure software architecture and can be deployed on a physical medium through program code, such as a hard disk, an optical disc, or any electronic device (such as a smart phone, a computer-readable storage medium). When the machine loads and executes the program code (such as a smart phone loads and executes), the machine becomes a device for implementing the present invention. The method and device of the present invention can also be transmitted in the form of program code through some transmission media, such as cables, optical fibers, or any transmission type. When the program code is received, loaded, and executed by a machine (such as a smart phone), the machine becomes a device for implementing the present invention.

[0205] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification is only to illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and all these changes and improvements fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A general tensor feature extraction method for building structures, characterized in that: include: Step S1: Select nodes, edges, and overall features in the building structure, and selectively select labels for terminal applications; Step S2: establishing a random sample library of the same type of building structures based on the features and labels selected in step S1; Step S3: Based on the random sample library of building structures established in step S2, the building structures are scanned in three directions, XYZ, to form a three-dimensional series of images and a three-dimensional image is saved at the same time; Step S4: Building a feature extraction and tensor data representation network; Step S5: Feature extraction and tensor data representation network training and model solidification.

2. The general building structure tensor feature extraction method according to claim 1, characterized in that: The step S1 includes the following steps when selecting nodes, edges, overall features, and selectively selecting labels for terminal applications: Step (S11): parsing component types, node types, and special attributes in the target building structure; Step (S12): Based on the node type analyzed in step (S11), node feature selection is performed; Step (S13): Based on the component type analyzed in step (S11), component feature selection is performed; Step (S14): selecting a feature that can represent the overall characteristics of the building structure as the overall feature; Step (S15): Select the parameters required by the terminal application as labels; Step (S16): Based on the features and labels selected in steps (S12) to (S15), the dimensions of nodes, edges, overall features and labels are calculated for subsequent network construction.

3. The general building structure tensor feature extraction method according to claim 1, characterized in that: The step S2 of establishing a random sample library of the same type of building structures comprises the following steps: Step (S21): setting the generation range of building structure space parameters and construction parameters; Step (S22): establishing a selection database for component cross sections; Step (S23): establishing a seismic wave record selection database or other analysis database that needs to be coordinated with the model analysis; Step (S24): formulating random generation rules for building structures; Step (S25): constructing a sample random generation model based on the parameters set in steps (S21) to (S23), the constructed database, and the building structure random generation rules formulated in step (S24); Step (S26): Randomly generate samples based on the sample random generation model of step (S24) and collect the target number of samples.

4. The general building structure tensor feature extraction method according to claim 1, characterized in that: When the step S3 performs XYZ three-dimensional scanning on the building structure, the following steps are included: Step (S31): setting a scanning direction for the sample generated in step (S2); Step (S32): Scanning the building structure according to the scanning direction set in step (S31) and obtaining a two-dimensional building structure framework during the scanning process; Step (S33): extracting the nodes, edges, and overall features of the building structure selected in step (S1) to form a tensor format; Step (S34): establishing a node mapping mechanism and an edge connection representation tensor of an undirected graph for the two-dimensional framework and the three-dimensional structure obtained in step (S32); Step (S35): independently save the graph constructed by the two-dimensional framework and the three-dimensional structure.

5. The general building structure tensor feature extraction method according to claim 1, characterized in that: When the step (S4) is performed on the basis of the step (S3) to build a feature extraction and tensor data representation network, it includes the following steps: Step (S41): Selecting the architecture of the graph neural network and the node feature update mechanism; Step (S42): setting the neural network architecture and related hyperparameters for edge feature and overall feature output; Step (S43): updating node features after the target graph passes through the graph neural network in step (S41); Step (S44): outputting the edge features as edge features updated with a specific dimension after passing through the neural network specified in step (S42); Step (S45): Aggregating edge features with adjacent node features by data aggregation to form new edge features; Step (S46): forming a new edge feature in step (S45), accumulating and aggregating the edge features of the entire graph and concatenating the overall features of the aggregated graph to form a new overall feature of the graph; Step (S47): selecting a directional feature aggregation method or an indifferent aggregation method to aggregate the overall features of all two-dimensional frame graphs to form a final tensor feature of the building structure; Step (S48): Follow the same steps to obtain the features of the three-dimensional framework diagram and combine them with the overall features of the two-dimensional framework diagram in step (S47) to form the final building structure tensor features.

6. The general building structure tensor feature extraction method according to claim 1, characterized in that: When the step (S5) performs feature extraction and tensor data representation network training and model solidification, it is first necessary to determine whether to perform training for the terminal application: If yes, the network and method determined in step (S4) are used for training with the parameters required by the terminal application as labels. After the effect is achieved, the model parameters are solidified and saved to form a dedicated model; If not, the network and method determined in step (S4) are used for training with the parameters of the structure itself as labels. After the effect is achieved, the final output layer of the model is discarded and only other model parameters are solidified and saved.

7. A general building structure tensor feature extraction system, characterized in that: include: Selection module, sample library establishment module, three-way scanning module, feature extraction and tensor data representation network module, training and model solidification module; The selection module is configured to select nodes, edges, and overall features, and selectively select labels for terminal applications; The sample library establishment module is configured to interact with the selection module data and is capable of establishing a random sample library of the same type of building structures; The three-dimensional scanning module is configured to interact with the sample library building module data, and can perform XYZ three-dimensional scanning on the building structure to form a three-dimensional series of images and simultaneously save a three-dimensional image; The feature extraction and tensor data representation network module is configured to interact with the three-way scanning module data, and can build a feature extraction and tensor data representation network; The training and model solidification module is configured to interact with the feature extraction and tensor data representation network module data, and can train and solidify the model of the feature extraction and tensor data representation network.

8. The universal building structure tensor feature extraction system according to claim 7, characterized in that: The building structure tensor feature extraction system also includes a transfer learning module, which is configured to interact with the training and model solidification module for data sets, and can connect to terminal applications and add labels for transfer learning.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the general building structure tensor feature extraction method described in any one of claims 1-6 are implemented.

10. A computer program product, characterized in that When executed on a data processing device, it is suitable for executing the steps of the general building structure tensor feature extraction method described in any one of claims 1-6.

Images