A method, device and medium for extracting spatiotemporal features of steel frame fire collapse warning

Abstract

The present invention relates to a method, device, and medium for extracting spatiotemporal features of steel frame fire collapse warning, comprising: establishing a finite element numerical analysis model of a multi-layer steel frame structure in combination with actual engineering cases; obtaining fire response data of the multi-layer steel frame structure for a deep learning model training process based on the numerical analysis model; determining the displacement of easy-to-measure nodes and difficult-to-measure nodes of the multi-layer steel frame structure; constructing a deep learning model based on a graph neural network and a recurrent neural network; training the deep learning model until the accuracy requirements are met; and applying the deep learning model obtained through training. Compared with the prior art, the present invention can accurately and real-time deduce the displacements of the top and internal key nodes required in the fire collapse warning theory and method, and is applicable to multi-layer steel frame structures with diverse topological forms, saving the computational cost of repeated training of the deep learning model for different structures, and facilitating rapid emergency response in actual fire rescue.

Description

Technical Field

[0001] The present invention relates to the fields of public safety rapid emergency response and artificial intelligence deep learning, and in particular to a method, device, and medium for extracting spatiotemporal features of steel frame fire collapse warning. Background Art

[0002] Multi-story steel frame structures are widely used in commercial buildings, such as large supermarkets, commercial complexes, and residential plazas. These buildings are large and prone to fire. Once a fire breaks out, it can easily spread rapidly horizontally and vertically through atria, escalators, and other facilities, resulting in a large, three-dimensional fire in a short period of time. Given the significant degradation of steel's mechanical properties at high temperatures, a collapse of a multi-story steel frame structure poses a serious threat to the lives of firefighters and those trapped within.

[0003] In recent years, relevant scholars have proposed a fire-induced collapse warning method for steel structures based on the evolution of key physical quantities (displacement and displacement rate). This method can achieve graded warning of the entire collapse process of steel structures under fire and quantitatively provide a reliable remaining time for structural collapse. Compared with the original collapse warning method based on existing experience, this fire-induced collapse warning method is more reliable and accurate. However, there are still some unresolved issues in the actual application of this method. The main problem is that the displacements of some key nodes required in the warning theory and method are difficult to directly obtain through external measurement means. For multi-story steel frames, the difficult-to-measure nodes are located at the top and inside of the structure. Summary of the Invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and provide a method, equipment and medium for extracting spatiotemporal features of steel frame fire collapse warning, which can accurately and in real time calculate the displacement of the top and internal key nodes required in the fire collapse warning theory and method, and is suitable for multi-layer steel frame structures with various topological forms, saving the computational cost generated by the repeated training of deep learning models for different structures, and is conducive to rapid emergency response in actual fire rescue.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] The present invention provides a method for extracting spatiotemporal features of steel frame fire collapse warning. Based on the structural fire response data obtained by a high-precision finite element numerical analysis model, a deep learning model combining a graph neural network and a recurrent neural network is constructed. The method can realize real-time estimation of the difficult-to-predict displacement of multi-story steel frame structures according to the displacement and temperature data that are easily obtained at the actual fire rescue scene, thereby facilitating the implementation of the theory and method of fire-induced collapse warning for multi-story steel frame structures.

[0007] The following steps are involved:

[0008] S1: Based on actual engineering case information, determine the geometric characteristics of the finite element numerical analysis model of the multi-story steel frame structure and establish the finite element numerical analysis model of the multi-story steel frame structure;

[0009] S2: Obtaining fire response data of multi-story steel frame structures for deep learning model training based on finite element numerical analysis models;

[0010] S3: Determine the displacement of easily measurable nodes and difficult-to-measurable nodes in multi-story steel frame structures;

[0011] S4: Abstract the multi-layer steel frame structure into a graph structure and build a deep learning model based on graph neural network and recurrent neural network. The deep learning model includes a feature extraction module, a graph convolution module, and a graph structure dynamic feature prediction module.

[0012] S5: The displacements of all nodes, the temperatures of the rods, and the geometric parameters of the multi-layer steel frame structure are used as the input layer, and the real unpredictable node displacements are used as the output layer. The deep learning model is trained until the accuracy requirements are met. The displacements of all nodes include the real easy-to-measure node displacements and the artificially randomly set difficult-to-measure node displacements. The geometric parameters include node coordinates and rod lengths. After the input layer is imported into the deep learning model, it first passes through the feature extraction module to enable the model to maximize the extraction of more useful information from the initial input features. Then, it passes through the graph convolution module, using global convolution, edge convolution, and node convolution operations to achieve information interaction between the measured easy-to-measure node displacements and the unknown unpredictable node displacements, enhance the model's ability to predict the overall behavior of the structure and capture the dependency relationship between local adjacent nodes. Finally, it passes through the graph structure dynamic feature prediction module to achieve real-time inference and prediction of all difficult-to-measure node displacements. The model training stage should be completed before the fire of the multi-layer steel frame structure occurs. The purpose is to pre-determine the relevant model parameters of the feature extraction module, graph convolution module, and graph structure dynamic feature prediction module in the deep learning model through a large amount of reliable numerical analysis data;

[0013] S6: The trained deep learning model is used as the final model for real-time estimation and prediction of the displacement of difficult-to-predict nodes in fires of multi-story steel frame structures with various topological forms in actual situations. The model usage stage refers to the stage after a fire occurs in an actual multi-story steel frame structure and the fire rescue personnel arrive at the scene. Based on the trained deep learning model and the easy-to-measure node displacement and temperature data measured at the fire scene, the real-time estimation and prediction of the displacement of difficult nodes is realized, which helps in the subsequent early warning of fire-induced collapse of multi-story steel frame structures.

[0014] Furthermore, in S1, the geometric characteristics of the finite element numerical analysis model of the multi-story steel frame structure are determined based on typical and representative actual engineering cases. The random variables include the temperature rise curve, fire location, load conditions, component material properties and cross-sectional dimensions. The probability distribution of the random variables is determined based on relevant statistical data in actual conditions. The random variables are randomly sampled based on the probability distribution to establish the final finite element numerical analysis model.

[0015] Furthermore, in S1, after all parameters are determined, a finite element numerical analysis model is established using ABAQUS, a common finite element software in the field of structural engineering. The explicit dynamic analysis step in ABAQUS is used to perform a thermal-mechanical coupling analysis on the multi-story steel frame structure exposed to fire, and the displacement time history curves of all nodes and the temperature time history curves of all rods are obtained.

[0016] Furthermore, in S2, the fire response data of the multi-story steel frame structure includes the displacement time history curves of all nodes and the temperature time history curves of all rods throughout the fire process.

[0017] Furthermore, in S3 and S5, the easily measurable node displacements include the horizontal and vertical displacements of the tops of the side columns and corner columns on each floor of the multi-story steel frame structure. The easily measurable node displacements are obtained by microwave radar.

[0018] Difficult-to-measure node displacements include the horizontal and vertical displacements of the top and internal nodes of multi-story steel frame structures. Difficult-to-measure node displacements are displacement data that is difficult to obtain directly through external measurement methods at the fire scene.

[0019] The temperature of the rods is measured by embedding thermocouples in the multi-story steel frame structure.

[0020] Furthermore, in S4, the multi-story steel frame structure can be abstracted as a graph structure, and the connection nodes and beam and column components in the multi-story steel frame structure are represented by nodes and edges in the graph structure respectively.

[0021] Furthermore, in S4, the feature extraction module includes node feature extraction and edge feature extraction. Node feature extraction is based on the static features (initial coordinates of key nodes in a multi-story steel frame structure) and dynamic features (node ​​displacement) of the nodes in the graph structure. By means of dimensional expansion, as much information as possible is extracted from the original features to generate a node comprehensive feature matrix containing static and dynamic information. Edge feature extraction refers to the static features (length of each rod in a multi-story steel frame structure) and dynamic features (rod temperature) of the edges in the graph structure. By means of dimensional expansion, as much information as possible is extracted from the original features to generate an edge comprehensive feature matrix containing static and dynamic information.

[0022] The graph convolution module includes three graph-based convolution operations: global convolution, edge convolution, and node convolution. The global convolution operation is to establish a virtual main node, transfer the features of the easy-to-measure node (corresponding to the easy-to-measure node in the multi-layer steel frame structure) to the virtual main node, and then transfer the features of the virtual main node to the features of the difficult-to-measure node (corresponding to the difficult-to-measure node in the multi-layer steel frame structure), which is equivalent to indirectly embedding the features of the easy-to-measure node into the features of the difficult-to-measure node; the edge convolution operation refers to the process of embedding the comprehensive features of the edge into the comprehensive features of the nodes connected at both ends in the graph structure; and the node convolution operation refers to the information transfer process between a node and its adjacent nodes in the graph structure.

[0023] The graph structure dynamic feature prediction module is based on an optimized recurrent neural network that integrates graph structure node convolution operations. It is used to predict the displacement of difficult-to-predict nodes in multi-story steel frame structures in real time. The input of this module is the comprehensive feature matrix of all nodes obtained after processing by the above-mentioned feature extraction module and graph convolution module. The output is the displacement time history curve of all difficult-to-predict nodes in the multi-story steel frame structure.

[0024] Furthermore, in S5, the deep learning model uses the mean square error (MSE) as the loss function during the training process. At the same time, the update of various hyperparameters in the deep learning model is achieved through the error back propagation algorithm;

[0025] The final model is determined according to the following criteria: the performance of the trained deep learning model on the test set is evaluated by the mean absolute error (MAE) and correlation coefficient (r), and MAE < 10mm or r > 0.9 is defined as the judgment criterion for high-precision samples. The test set is divided into a training framework test set (based on the framework structure used for model training, but the data set is not involved in the training at all) and an untrained framework test set (based on the framework structure that does not participate in the model training). The training framework test set is used to test the prediction performance of the deep learning model in potential unknown real fire scenarios, while the non-training framework test set is used to test the universality of the deep learning model in multi-story steel frames with other topological forms. Finally, the deep learning model with the highest cumulative proportion of high-precision test samples on the above two test sets will be used as the final deep learning model for actual fire rescue.

[0026] The present invention also provides an electronic device, including a memory and a processor, wherein the processor is used to execute a program in the memory, thereby realizing the spatiotemporal feature extraction method for steel frame fire collapse warning as described above.

[0027] The present invention also provides a storage medium containing computer executable instructions. When the storage medium containing computer executable instructions is executed by a computer processor, it is used to execute the spatiotemporal feature extraction method for steel frame fire collapse warning as described above.

[0028] Based on the theories and methods of artificial intelligence deep learning, this invention proposes a real-time estimation method for the displacement of difficult-to-predict nodes of multi-story steel frame structures under fire, thereby facilitating the implementation of fire-induced collapse early warning methods for multi-story steel frame structures, reducing losses and casualties caused by structural collapse under fire, and ensuring the safety of people's lives and property, which has great public safety significance.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] (1) The present invention uses the displacement of key nodes, the temperature of rods, and the geometric parameters of the structure (node ​​coordinates, rod length) as data sources, and combines graph neural networks and recurrent neural networks to build a deep learning model. It can accurately calculate the displacement of top and internal key nodes required in fire collapse warning theories and methods, and is suitable for multi-story steel frame structures with diverse topological forms.

[0031] (2) It greatly saves the computational cost of repeated training of deep learning models for different structures. The deep learning model can perform real-time inference and prediction of the difficult-to-predict displacements of multi-story steel frame structures with various topological forms in real fire scenarios without repeated training, which is conducive to rapid emergency response in actual fire rescue. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Flowchart of the spatiotemporal feature extraction method for steel frame fire collapse warning.

[0033] Figure 2 Schematic diagram of the structure of a multi-story steel frame.

[0034] Figure 3 This is a structural diagram of the deep learning model.

[0035] Figure 4 Schematic diagram of deep learning model training method.

[0036] Figure 5 Schematic diagram of the method applied to the deep learning model.

[0037] Figure 6 A real-time acquisition method for rod temperature when applying deep learning models. DETAILED DESCRIPTION

[0038] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. Any features such as component models, material names, connection structures, control methods, algorithms, etc. that are not clearly stated in this technical solution shall be deemed to be common technical features disclosed in the prior art. It should be noted in particular that the key technology of the present invention is to adopt a deep learning model based on graph neural networks and recurrent neural networks to achieve real-time extrapolation and prediction of difficult-to-predict node displacements in real fire scenarios of multi-story steel frame structures of various topological forms without the need for repeated training of the model. Any changes or modifications that are made to the main design ideas and spirit of the present invention that have no substantive significance, and whose technical methods are still consistent with the present invention, should be included in the scope of protection of the present invention.

[0039] Example 1

[0040] This embodiment provides a method for extracting spatiotemporal features for steel frame fire collapse warning. Based on the structural fire response data obtained by a high-precision finite element numerical analysis model, by building a deep learning model combining a graph neural network and a recurrent neural network, it can realize real-time estimation of the difficult-to-predict displacements of multi-story steel frame structures based on the displacement and temperature data that are easily obtained at the actual fire rescue scene, thereby facilitating the implementation of the theory and method of fire-induced collapse warning for multi-story steel frame structures.

[0041] like Figure 1 As shown, the following steps are included:

[0042] S1: Based on actual engineering cases, determine the geometric characteristics of the finite element numerical analysis model of multi-story steel frame structure and establish the finite element numerical analysis model of multi-story steel frame structure;

[0043] S2: Obtaining fire response data of multi-story steel frame structures for deep learning model training based on finite element numerical analysis models;

[0044] S3: Determine the displacement of easily measurable nodes and difficult-to-measurable nodes in multi-story steel frame structures;

[0045] S4: Abstract the multi-layer steel frame structure into a graph structure and build a deep learning model based on graph neural network and recurrent neural network, such as Figure 3 As shown in the figure, the deep learning model includes a feature extraction module, a graph convolution module, and a graph structure dynamic feature prediction module;

[0046] S5: If Figure 3 、 Figure 4As shown, the displacements of all nodes, the temperatures of the rods, and the geometric parameters of the multi-layer steel frame structure are used as the input layer, and the real unpredictable node displacements are used as the output layer. The deep learning model is trained until the accuracy requirements are met. The displacements of all nodes include the real easy-to-measure node displacements and the artificially randomly set difficult-to-measure node displacements. The geometric parameters include node coordinates and rod lengths. After the input layer is imported into the deep learning model, it first passes through the feature extraction module to enable the model to maximize the extraction of more useful information from the initial input features. Then, it passes through the graph convolution module, using global convolution, edge convolution, and node convolution operations to achieve information interaction between the measured easy-to-measure node displacements and the unknown unpredictable node displacements, enhance the model's ability to predict the overall behavior of the structure and capture the dependency between local adjacent nodes. Finally, it passes through the graph structure dynamic feature prediction module to achieve real-time inference and prediction of all difficult-to-measure node displacements. The model training stage should be completed before the fire of the multi-layer steel frame structure occurs. The purpose is to predetermine the relevant model parameters of the feature extraction module, graph convolution module, and graph structure dynamic feature prediction module in the deep learning model through a large amount of reliable numerical analysis data.

[0047] S6: As Figure 5 As shown, the trained deep learning model is used as the final model for real-time estimation and prediction of difficult-to-measure node displacements in fire-affected multi-story steel frame structures of various topological forms. In the application phase, the displacements of all nodes (including easily measurable node displacements measured on-site and artificially set difficult-to-measure node displacements), measured component temperatures, and the structure's geometric parameters (node ​​coordinates and member lengths) are first input into the pre-trained deep learning model. The deep learning model then estimates and outputs the difficult-to-measure node displacements in real time based on the input data. Finally, the measured easy-to-measure node displacement time history data and the difficult-to-measure node displacement time history data calculated by the deep learning model, which are essential for the early warning theory and method of fire-induced collapse of multi-story steel frame structures, are introduced into the early warning algorithm. A graded early warning is issued for fire-induced collapse of the affected steel frame, and a reliable remaining time to collapse is given for each level of warning, thereby guiding fire rescue efforts at the fire scene.

[0048] Furthermore, in S1, the geometric characteristics of the finite element numerical analysis model of the multi-story steel frame structure are determined based on typical and representative actual engineering cases. The random variables include the fire location, load conditions, component material properties, and cross-sectional dimensions. The probability distribution of the random variables is determined based on relevant statistical data in actual conditions. The random variables are randomly sampled based on the probability distribution, thereby determining the fire conditions of the multi-story steel frame structure under this parameter combination, and establishing the final finite element numerical analysis model.

[0049] Furthermore, in S1, after the parameters are determined, the finite element numerical analysis model is established using ABAQUS, a common finite element software in the field of structural engineering. The explicit dynamic analysis step in ABAQUS is used to perform a thermal-mechanical coupling analysis on the multi-story steel frame structure exposed to fire, and the displacement time history curves of all nodes and the temperature time history curves of all rods are obtained as the training set for the deep learning model.

[0050] In a specific embodiment, in S2, the fire response data for the multi-story steel frame structure includes displacement time-history curves for all nodes and temperature time-history curves for all rods throughout the fire process. In S3 and S5, easily measurable node displacements include the horizontal and vertical displacements of the tops of the side and corner columns on each floor outside the multi-story steel frame structure, acquired via microwave radar. Difficult-to-measurable node displacements include the horizontal and vertical displacements of the top and internal nodes of the multi-story steel frame structure, which are difficult to directly obtain through external measurement methods at the fire scene. Rod temperatures are measured using thermocouples embedded in the multi-story steel frame structure.

[0051] In a specific implementation, in S4, the multi-story steel frame structure may be abstracted as a graph structure, and the connection nodes and beam and column components in the multi-story steel frame structure are represented by nodes and edges in the graph structure, respectively.

[0052] In a specific embodiment, in S4, the feature extraction module includes node feature extraction and edge feature extraction. Node feature extraction is based on the static features (initial coordinates of each key node in the multi-story steel frame structure) and dynamic features (node ​​displacement) of the nodes in the graph structure. By means of dimensional expansion, as much information as possible is extracted from the original features to generate a node comprehensive feature matrix containing static and dynamic information. Edge feature extraction refers to the static features (length of each rod in the multi-story steel frame structure) and dynamic features (rod temperature) of the edges in the graph structure. By means of dimensional expansion, as much information as possible is extracted from the original features to generate an edge comprehensive feature matrix containing static and dynamic information.

[0053] The graph convolution module includes three graph-based convolution operations: global convolution, edge convolution, and node convolution. The global convolution operation is to establish a virtual main node, transfer the features of the easy-to-measure node (corresponding to the easy-to-measure node in the multi-layer steel frame structure) to the virtual main node, and then transfer the features of the virtual main node to the features of the difficult-to-measure node (corresponding to the difficult-to-measure node in the multi-layer steel frame structure), which is equivalent to indirectly embedding the features of the easy-to-measure node into the features of the difficult-to-measure node; the edge convolution operation refers to the process of embedding the comprehensive features of the edge into the comprehensive features of the nodes connected at both ends in the graph structure; and the node convolution operation refers to the information transfer process between a node and its adjacent nodes in the graph structure.

[0054] The graph structure dynamic feature prediction module is based on an optimized recurrent neural network that integrates graph structure node convolution operations. It is used to predict the displacement of difficult-to-predict nodes in multi-story steel frame structures in real time. The input of this module is the comprehensive feature matrix of all nodes obtained after processing by the above-mentioned feature extraction module and graph convolution module. The output is the displacement time history curve of all difficult-to-predict nodes in the multi-story steel frame structure.

[0055] In a specific embodiment, in S5, the deep learning model uses the root mean square error (MSE) as the loss function during the training process. At the same time, the update of various hyperparameters in the deep learning model is achieved by the error back propagation algorithm;

[0056] The final model is determined according to the following criteria: the performance of the trained deep learning model on the test set is evaluated by the mean absolute error (MAE) and correlation coefficient (r). The training samples will be divided into training set, validation set and test set in a ratio of 3:1:1. Only the training set and validation set participate in model training. The test set is divided into a training framework test set (based on the framework structure used for model training, but the data set does not participate in training at all) and an untrained framework test set (based on the framework structure that does not participate in model training). The training framework test set is used to test the prediction performance of the deep learning model in potential unknown real fire scenarios, while the untrained framework test set is used to test the universality of the deep learning model in multi-story steel frames with other topological forms. When the model's prediction of a sample in the test set meets MAE < 10mm or r > 0.9, it is considered that the model's prediction of the test sample has reached a high level of accuracy. When the model achieves the above-mentioned high-precision level in predictions for more than 85% of the samples in the test set, it is considered that the model has basically met the needs of practical applications. Finally, the deep learning model with the highest cumulative proportion of high-precision test samples on the above two test sets will be used as the final deep learning model for actual fire rescue, and the training will be completed.

[0057] In a specific embodiment, Figure 6 As shown, for a multi-story steel frame structure, the easily measurable node displacement refers to the horizontal and vertical displacements u of the tops of the side columns and corner columns on each floor located outside the structure. E1,i 、u E2,i (i=1,2,3,…,n E ), the unpredictable node displacement refers to the horizontal and vertical displacement u of the top of the structure and each internal node D1,j 、u D2,j (j=1,2,3,…,n D ), where n E represents the total number of testable nodes, and n D represents the total number of difficult-to-measure nodes; component temperature refers to the temperature T of beams and columns in multi-story steel frame structures. k(k=1,2,3,…,n e ), where n e The above-mentioned easily measurable node displacement data can be obtained by setting up a microwave radar at the fire scene, while the component temperature data can be obtained by embedding thermocouples in the multi-story steel frame structure.

[0058] This embodiment also provides an electronic device, including a memory and a processor, wherein the processor is used to execute a program in the memory, thereby implementing the spatiotemporal feature extraction method for steel frame fire collapse warning as described above. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components; the memory may include a random access memory (RAM) and may also include a non-volatile memory, such as at least one disk storage. The memory can be an internal memory of the random access memory (RAM) type, and the processor and memory can be integrated into one or more independent circuits or hardware, such as an application-specific integrated circuit (ASIC). It should be noted that the computer program in the above-mentioned memory can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods of each embodiment of the present invention.

[0059] This embodiment also provides a storage medium containing computer-executable instructions. When the storage medium of the computer-executable instructions is executed by a computer processor, it is used to perform the spatiotemporal feature extraction method for steel frame fire collapse warning as described above. The storage medium can be an electronic medium, a magnetic medium, an optical medium, an electromagnetic medium, an infrared medium, or a semiconductor system or a propagation medium. The storage medium can also include semiconductor or solid-state memory, a magnetic tape, a removable computer disk, a random access memory (RAM), a read-only memory (ROM), a hard disk, and an optical disk. The optical disk can include a compact disk - read-only memory (CD-ROM), a compact disk - read / write (CD-RW), and a DVD.

[0060] The above description of the embodiments is intended to facilitate understanding and use of the invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to these embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above-described embodiments. Improvements and modifications made by those skilled in the art based on the disclosure of the present invention, without departing from the scope of the present invention, should be within the scope of protection of the present invention.

Claims

1. A method for extracting spatiotemporal features of steel frame fire collapse warning, characterized in that: The following steps are involved: S1: Based on actual engineering case information, determine the geometric characteristics of the finite element numerical analysis model of the multi-story steel frame structure and establish the finite element numerical analysis model of the multi-story steel frame structure; S2: Obtaining fire response data of multi-story steel frame structures for deep learning model training based on finite element numerical analysis models; S3: Determine the displacements of easily measurable nodes and difficult-to-measurable nodes of the multi-story steel frame structure; the easily measurable node displacements include the horizontal and vertical displacements of the tops of the side columns and corner columns of each floor outside the multi-story steel frame structure, and the easily measurable node displacements are acquired by microwave radar; The difficult-to-measure node displacements include the horizontal and vertical displacements of the top and internal nodes of the multi-story steel frame structure. The difficult-to-measure node displacements are displacement data that are difficult to obtain directly through external measurement methods at the fire scene. S4: Abstract the multi-layer steel frame structure into a graph structure and build a deep learning model based on graph neural network and recurrent neural network. The deep learning model includes a feature extraction module, a graph convolution module, and a graph structure dynamic feature prediction module. In the process of abstracting the multi-layer steel frame structure into a graph structure, the connection nodes and beam and column components in the multi-layer steel frame structure are represented by nodes and edges in the graph structure respectively; the feature extraction module includes two parts: node feature extraction and edge feature extraction; the graph convolution module includes three graph-based convolution operations: global convolution, edge convolution, and node convolution; the graph structure dynamic feature prediction module is based on an optimized recurrent neural network that integrates the graph structure node convolution operation; S5: Using the displacements of all nodes, the temperatures of rods, and the geometric parameters of the multi-layer steel frame structure as the input layer and the actual hard-to-measure node displacements as the output layer, the deep learning model is trained until the accuracy requirements are met. The rod temperatures are measured by embedding thermocouples in the multi-layer steel frame structure. The displacements of all nodes include the actual easy-to-measure node displacements and the artificially randomly set hard-to-measure node displacements. The geometric parameters include the node coordinates and the rod lengths. S6: The trained deep learning model is used as the final model for real-time prediction of unpredictable node displacements in multi-story steel frame structures with various topological forms under fire in actual situations.

2. A method for extracting spatiotemporal features of steel frame fire collapse warning according to claim 1, characterized in that: In S1, the geometric characteristics of the finite element numerical analysis model of the multi-story steel frame structure are determined based on typical and representative actual engineering cases. The random variables include the temperature rise curve, the fire location, the load condition, the component material properties and the cross-sectional dimensions. The probability distribution of the random variables is determined based on the relevant statistical data in the actual situation. The random variables are randomly sampled based on the probability distribution to establish the final finite element numerical analysis model.

3. A method for extracting spatiotemporal features of steel frame fire collapse warning according to claim 1, characterized in that: In S1, the finite element numerical analysis model is established using the finite element software ABAQUS, and the explicit dynamic analysis step in ABAQUS is used to perform a thermal-mechanical coupling analysis on the multi-story steel frame structure exposed to fire.

4. A method for extracting spatiotemporal features of steel frame fire collapse warning according to claim 1, characterized in that: In S2, the fire response data of the multi-story steel frame structure includes displacement time history curves of nodes and temperature time history curves of rods during the entire fire process.

5. The method for extracting spatiotemporal features of steel frame fire collapse warning according to claim 1, characterized in that: In S5, the deep learning model uses the root mean square error during training. MSE As a loss function, at the same time, the update of various hyperparameters in the deep learning model is achieved through the error back propagation algorithm; The final model was determined according to the following criteria: MAE and correlation coefficient r To evaluate the performance of the trained deep learning model on the test set, define MAE < 10 mm or r > 0.9 is the criterion for judging high-precision samples. The test set is divided into a trained framework test set and an untrained framework test set. The deep learning model with the highest cumulative proportion of high-precision test samples on the two test sets is used as the final deep learning model.

6. An electronic device comprising a memory and a processor, characterized in that: The processor is used to execute the program in the memory, so as to implement the spatiotemporal feature extraction method for steel frame fire collapse warning as described in any one of claims 1 to 5.

7. A storage medium containing computer-executable instructions, characterized in that: The storage medium of the computer executable instructions, when executed by a computer processor, is used to execute the spatiotemporal feature extraction method for steel frame fire collapse warning as described in any one of claims 1 to 5.

Images