Machine Learning and Escape Planning Method for the Explosion Power Field of Buildings in Limited Open Spaces

Abstract

The present invention discloses a machine learning and escape planning method for the explosion power field of buildings in a limited open space. By using a database of the power field of buildings in a limited open space under explosion shock, which includes a training data set and a validation data set, the power field interaction collaborative graph neural network is trained and optimized to obtain a prediction model for the power field of buildings in a limited open space under explosion shock, realizing machine learning of the three-dimensional explosion power field of buildings in a limited open space. According to the prediction of the power field of buildings in a limited open space under explosion shock and comparison with historical test data, the personnel damage field of the corresponding underground space building is obtained, and the starting point and ending point of personnel escape are obtained. Using a path optimization method, the emergency escape plan is realized in a manner that satisfies the shortest escape time and the maximum survival rate. The method of the present invention has high accuracy, high efficiency, and strong generalization ability.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and particularly relates to a machine learning and escape planning method for the explosion power field of a limited open space building with high prediction efficiency, strong generalization ability, and high accuracy. Background Art

[0002] A limited open space refers to a building space that is neither completely open nor completely closed, such as a tunnel, a ship's cabin, an underground space, etc.

[0003] An explosion shock power field refers to an explosion effect field sufficient to cause damage or kill to a target, such as a peak overpressure field, a peak impulse field, etc.

[0004] Power field prediction refers to predicting the important parameters of the power field under explosion shock, including impact load, overpressure, impulse, arrival time.

[0005] 1. Impact load: Impact load refers to the change of the force acting on an object with time, describing the short-term effect of the force on the object. In the case of an explosion shock wave, the impact load usually refers to the short-term force exerted by the explosion-generated shock wave on a structure or object. The characteristics of this load are high energy and short-time action, which can cause serious damage to the structure.

[0006] 2. Overpressure: Overpressure refers to the difference between the pressure on the shock wave front of an explosion and the ambient atmospheric pressure. When an explosion occurs, the front of the shock wave will cause a sharp increase in local pressure, and the part higher than the atmospheric pressure is the overpressure. Overpressure is an important parameter for measuring the intensity of the shock wave, and it directly affects the damage degree of the explosion shock wave to buildings and biological targets.

[0007] 3. Impulse: Impulse is the cumulative effect of force over time, defined as the integral of force with respect to time. In the context of an explosion shock wave, impulse is usually used to describe the total effect produced by the change of the force exerted by the explosion shock wave on a structure or object over time, and is used to evaluate the damage potential of the explosion shock wave.

[0008] 4. Arrival time: Arrival time refers to the time when the explosion shock wave reaches a specific area or target. After an explosion occurs, the shock wave needs a certain time to propagate to the target area, and this time is the arrival time. Arrival time is very important for evaluating the impact of the explosion shock wave and formulating protective measures, because it determines when the target begins to be affected by the shock wave after the explosion.

[0009] Predicting the power field of a limited open space under explosion shock is the basis for safety protection design, and is also an important prerequisite for evaluating explosion consequences and formulating emergency response plans.

[0010] Currently, the prediction of the power field of a finite open - space building under explosion shock is as described in the paper "Research Status and Prospect of Leakage and Combustion - explosion Characteristics of Hydrogen - energy Vehicles in Confined Spaces" (Chen Jiayan, Yang Juntao, He Qize, et al. Journal of Wuhan University of Technology, Edition of Information & Management Engineering, 2024, 46(3): 382 - 386, 409. DOI: 10.3963 / j.issn.2095 - 3852.2024.03.006.). It discloses a determination method for a hydrogen - energy vehicle to form jet fire and leakage - ignition explosion in a confined space, including: using the computational fluid dynamics numerical simulation method to study the jet - fire behavior of hydrogen under different ventilation conditions, leakage angles, and leakage rates; using the results of the numerical simulation to further predict the surface explosion - shock power field in the confined space.

[0011] However, the above - mentioned method relies on a complex simulation system. The complex model operations will greatly increase the calculation time and consume a large amount of computing resources, which is very time - consuming and laborious and not efficient enough. On the other hand, due to the many assumptions in the numerical simulation, it can only predict the explosion - shock response under specific working conditions and has weak generalization ability.

[0012] Also, as described in the paper "Real - time Simulation Method of Explosion Flow Field Based on the Fusion of Physical Model Analysis and Deep Neural Network" (Zhou Shennan, Wang Zhongqi, Li Qizhong. Journal of Safety and Environment, 2024, 24(5): 1681 - 1690. DOI: 10.13637 / j.issn.1009 - 6094.2023.0758.), it discloses a machine - learning method for the three - dimensional explosion power field of a finite open - space building based on deep learning, including: using the computational fluid dynamics numerical simulation method to simulate and analyze a variety of typical explosion scenarios to construct a data set, using a convolutional variational auto - encoder model to map high - dimensional features to latent space variables and reduce the dimensionality of the data set; establishing a mapping relationship between explosion - related parameters, the observation time of the flow field, and the latent variables through a multi - layer feed - forward neural network model to achieve data regression; feeding the latent variables output by the multi - layer feed - forward neural network model back to the convolutional variational auto - encoder model to reconstruct or confirm the corresponding regional explosion pressure field and achieve data generation; using the trial - and - error method to optimize each hyper - parameter related to the model architecture one by one, and through a progressive training method, forming a basic deep - neural - network prediction model; analyzing the prediction results through the relative L2 - norm error and comparing them with the numerical simulation results to evaluate the prediction accuracy and real - time performance of the prediction model; combining the real - time data provided by the monitoring system and using the established model to efficiently obtain the regional explosion - risk analysis results and achieve real - time prediction.

[0013] However, the neural network prediction model adopted in the above method is limited by specific types of explosion scenarios and conditions, lacks a diverse and representative dataset, and the deep learning neural network adopted above only relies on a multi-layer feedforward neural network, resulting in low prediction accuracy.

[0014] In summary, the problems existing in the prior art are: the efficiency of predicting the power field of buildings in a limited open space under explosion shock is not high, the generalization ability is not strong, and the accuracy is not high. Summary of the Invention

[0015] The purpose of the present invention is to provide a machine learning method for the three-dimensional explosion power field of buildings in a limited open space, which has high efficiency, strong generalization ability and high accuracy.

[0016] The technical solution for achieving the purpose of the present invention is as follows:

[0017] In a first aspect, there is provided a machine learning method for the three-dimensional explosion power field of buildings in a limited open space, including:

[0018] Training and optimizing a power field interaction collaborative graph neural network through a power field database of buildings in a limited open space under explosion shock, including a training dataset and a validation dataset, to obtain a prediction model for the power field of buildings in a limited open space under explosion shock, so as to realize the machine learning of the three-dimensional explosion power field of buildings in a limited open space; wherein, the power field includes peak overpressure, peak impulse, and shock wave arrival time;

[0019] The power field interaction collaborative graph neural network includes an encoder, a processor, and a decoder, wherein:

[0020] The encoder includes a cascaded multi-head self-attention layer, a first feedforward neural network layer, a residual connection and a layer normalization component. Among them, the first feedforward neural network layer includes a first fully connected layer and N identical first sub-layers. Each first sub-layer includes a second fully connected layer with a front-end cascaded ReLU activation function. The input of each first sub-layer is the output of the multi-head self-attention layer, and the output of each sub-layer is summarized to the first fully connected layer through a ReLU activation function;

[0021] The processor includes a TransGNN processor, a first convolutional layer, a multi-path network structure layer, a third fully connected layer, a pooling layer, and a readout layer. Among them, the multi-path network structure layer includes a cascaded graph attention layer, a second feedforward neural network layer, and a RefineNet multi-path network structure. The readout layer includes a cascaded global pooling layer and a fourth fully connected layer; wherein, the second feedforward neural network layer has the same structure as the first feedforward neural network layer;

[0022] The decoder includes multiple layers of masked self-attention layers, N identical second sub-layers, a stacking and normalization component, and a generation network layer. Each second sub-layer includes a second convolutional layer and a fifth fully-connected layer connected by a ReLU activation function. The input of each second convolutional layer is the output of the multiple layers of masked self-attention layers. The stacking and normalization component includes a cascaded stacking layer, a softmax function, a normalization layer, and a residual connection layer. The output of the residual connection layer and the outputs of all the fifth fully-connected layers are used as the input of the stacking layer.

[0023] As a further optimization scheme of the present invention, physical information constraints are embedded in the power field interaction collaborative graph neural network, including:

[0024] Using the random number method, random sample data of three-dimensional explosions of finite open space buildings in the near field, mid-field, and far field are extracted according to the proportional distance, and the empirical relationships between the peak overpressure, peak impulse, shock wave arrival time, and proportional distance are fitted:

[0025] ,

[0026] ,

[0027] ,

[0028] In the formula, I + represents the true value of the peak impulse, P represents the true value of the peak overpressure, t + represents the true value of the shock wave arrival time, A1, B1, C1, D1 represent fitting parameters, Z represents the proportional distance, A2 represents the parameter of the peak impulse, A3 represents the parameter of the shock wave arrival time, and W represents the TNT equivalent;

[0029] According to the empirical relationships corresponding to different proportional distances in the near field, mid-field, and far field, the empirical relationships are used as constraint conditions and embedded into the loss function of the graph neural network to achieve the embedding of physical information constraints in the graph neural network.

[0030] As a further optimization scheme of the present invention, after embedding the physical information constraints, a physical information constraint loss term L physics is added to the loss function of the obtained power field interaction collaborative graph neural network:

[0031] ,

[0032] ,

[0033] ,

[0034] ,

[0035] Among them, M represents the number of nodes, , and respectively represent the peak overpressure constraint loss term, the peak impulse constraint loss term, and the shock wave arrival time constraint loss term. represents the predicted peak overpressure of node j, and the true value of the peak overpressure of node j , represents the scaled distance of node j, and A1, B1, C1, D1 represent fitting parameters. represents the predicted peak impulse of node j, and W represents the TNT equivalent. represents the predicted shock wave arrival time of node j, A3 represents the parameter of the shock wave arrival time, and dt represents the infinitesimal element.

[0036] As a further optimization scheme of the present invention, the steps for establishing the database of the explosion shock power field of a finite open space building include:

[0037] By digital twin of the real finite open space building and the environment, randomly generate the explosion equivalent and the explosion center to construct an explosion event;

[0038] According to the explosion event, use the numerical simulation method to establish a numerical model of the explosion event power field and obtain the explosion event power field data;

[0039] Perform data preprocessing on the explosion event power field data, establish a sample set, and combine the selected explosion characteristics to establish a database of the explosion shock power field of a finite open space building including a training data set and a validation data set.

[0040] As a further optimization scheme of the present invention, the steps for constructing an explosion event include:

[0041] Through data crawling, obtain the map data source of the real finite open space building and the surrounding environment, and construct a three-dimensional model of the finite open space building and the surrounding environment through the building three-dimensional digital twin method;

[0042] Preset the range of the explosion equivalent and the explosion center, generate the explosion equivalent and the explosion center by the random number method, and perform data combination according to the explosion event number to obtain the explosion equivalent and the explosion center under the explosion event number;

[0043] Match the explosion center and the explosion equivalent under the same explosion event number with the three-dimensional model of the finite open space building and the surrounding environment to construct an explosion event.

[0044] As a further optimization scheme of the present invention, the steps for establishing a database of the explosion shock power field of a finite open space building including a training data set and a validation data set include:

[0045] Through missing value completion, amplitude adjustment, and spatial correlation conversion, the blast event power field data is converted into a two-dimensional / three-dimensional matrix form to achieve preprocessing of the blast event power field data;

[0046] Through the reconstruction simulation of the blast shock pressure field distribution based on limited measurement point data, the digital reconstruction of the blast shock power field parameters is carried out to construct a sample set. The blast shock power field parameters include peak overpressure, peak impulse, and shock wave arrival time;

[0047] Through fluid field analysis and numerical simulation, the characteristic parameters of the blast shock power field parameters are obtained;

[0048] Combining the characteristic parameters of the blast shock power field parameters with the blast event, a database of the blast shock power field of buildings in a finite open space is constructed, and the database of the blast shock power field of buildings in a finite open space is divided into a training data set and a validation data set.

[0049] In a second aspect, a method for planning an explosion escape in a building with a finite open space is also provided, including:

[0050] Using the machine learning method described above, the prediction result of the blast shock power field of buildings in a finite open space is obtained;

[0051] According to the prediction result of the power field, the percentage of personnel damage is calculated;

[0052] A weighted directed graph is used to represent the three-dimensional explosion scene of the building with a finite open space, and the percentage of personnel damage is used as the attribute of the nodes in the weighted directed graph, that is, the degree of injury of personnel at that node;

[0053] Based on the Dijkstra algorithm, under the constraint conditions that the attribute of any node on the planned path does not exceed the preset moderate injury level and the survival rate is the largest and the escape time is the shortest, the optimal escape path is found to complete the escape plan.

[0054] As a further optimized solution of the present invention, the calculation steps of the percentage of personnel damage include:

[0055] Using the nonlinear relationship data fitting method, the calculation empirical relationship formula between the percentage of personnel damage and peak overpressure, peak impulse, and shock wave arrival time is obtained:

[0056] ,

[0057] Among them, is the percentage of personnel damage, and E1, F1, and G1 are fitting parameters.

[0058] In a third aspect, there is also provided a computer-readable storage medium storing one or more programs, the one or more programs including instructions which, when executed by a computing device, cause the computing device to execute the machine learning method as described above.

[0059] In a fourth aspect, there is also provided an electronic device, including one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, and the one or more programs include instructions for executing the machine learning method as described above.

[0060] Compared with the prior art, the remarkable advantages of the present invention are as follows:

[0061] 1. Higher analysis and prediction efficiency: By constructing a physics-informed embedded graph neural network and establishing a prediction model for the blast field of buildings in a finite open space under blast shock, the present invention directly processes the three-dimensional physical field calculation of the blast field of buildings in a finite open space under blast shock, avoiding the traditional numerical format derivation and solution process to improve the calculation efficiency, and realizing the efficient response prediction of the blast field under blast shock in a finite open space.

[0062] 2. Stronger generalization ability: By establishing a database of the blast field of buildings in a finite open space under blast shock with random region characteristics, random facility characteristics, and random explosion characteristics, the present invention ensures the diversity of the database and the robustness of the neural network, and realizes the efficient generalization prediction of the blast field of buildings in a finite open space under blast shock (such as peak overpressure, peak impulse, shock wave arrival time, etc.).

[0063] 3. Higher analysis accuracy: By combining the graph neural network and the physics-informed embedding methods, the present invention enhances the ability of three-dimensional blast field information transmission, realizes the interactive collaborative graph neural network processing of blast field prediction, and realizes the high-precision prediction of physical problems of the blast field of buildings in a finite open space under blast shock.

[0064] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is the main flowchart of the machine learning and escape planning method for the blast field of buildings in a finite open space of the present invention.

[0066] Figure 2 is a schematic diagram of the encoder structure.

[0067] Figure 3 is a schematic diagram of the processor structure.

[0068] Figure 4 is a schematic diagram of the decoder structure.

[0069] Figure 5 It is a flow chart of the escape plan. Specific implementation mode

[0070] Such as Figure 1 As shown in the figure, the machine learning and escape planning method for the explosion power field of a limited open space building in the present invention includes the following steps:

[0071] S10. Establishment of the explosion shock power field database: Construct an explosion event, obtain the power field data of the explosion event, establish a sample set, select explosion characteristics, and establish a database of the power field of a limited open space building under explosion shock including a training data set and a verification data set.

[0072] The steps for establishing the explosion shock power field database include:

[0073] S11. Construction of the explosion event: Through digital twin of the real limited open space building and the environment, randomly generate the explosion equivalent and the explosion center to construct the explosion event.

[0074] The steps for constructing the explosion event include:

[0075] S111. Digital twin of the real limited open space building and the environment: Through data crawling, obtain the map data source of the real limited open space building and the surrounding environment, and construct a three-dimensional model of the limited open space building and the surrounding environment through the building three-dimensional digital twin method.

[0076] S112. Generation of the explosion equivalent and the explosion center: Preset the range of the explosion equivalent and the explosion center, generate the explosion equivalent and the explosion center through the random number method, and perform data combination according to the explosion event number to obtain the explosion equivalent and the explosion center under the explosion event number.

[0077] S113. Construction of the explosion event: Match the three-dimensional model of the limited open space building and the surrounding environment through the explosion center and the explosion equivalent under the same explosion event number to construct the explosion event; define the nodes of the limited open space building, the explosion equivalent and the explosion center as input data.

[0078] S12. Obtaining the explosion event power field data: According to the constructed explosion event, use the numerical simulation method to establish a numerical model of the explosion event power field, simulate and divide to obtain grid nodes, and obtain the explosion event power field data; define the peak overpressure, peak impulse and shock wave arrival time of the limited open space as output data.

[0079] S13. Establishment of the power field database: Perform data preprocessing on the explosion event power field data, splice the input data and the output data, establish a sample set, and establish a database of the power field of a limited open space building under explosion shock including a training data set and a verification data set in combination with the selected explosion characteristics.

[0080] The steps for establishing the power field database include:

[0081] S131. Data preprocessing: Through missing value completion, amplitude adjustment, and spatial correlation conversion. For example, uniformly setting negative overpressure values to zero and converting the power field data of the explosion event into a two-dimensional / three-dimensional matrix form to achieve the preprocessing of the power field data of the explosion event.

[0082] S132. Sample set construction: Through the reconstruction simulation of the explosion shock pressure field distribution based on limited measurement point data, digitally reconstructing the parameters of the explosion shock power field to construct a sample set. The parameters of the explosion shock power field include peak overpressure, peak impulse, and shock wave arrival time.

[0083] S133. Explosion feature selection: Through flow field analysis and numerical simulation, obtaining the characteristic parameters of the parameters of the explosion shock power field.

[0084] S134. Database construction: Combining the characteristic parameters of the parameters of the explosion shock power field with the explosion event to construct a power field database of a limited open space building under explosion shock, and dividing the power field database of the limited open space building under explosion shock into a training data set and a validation data set.

[0085] S20. Construction of the power field model of a limited open space building under explosion shock: Extracting small sample data from the database, fitting the empirical relationship between the peak overpressure, shock wave arrival time of the limited open space, and the scaled distance, constructing a physics-informed embedding equation based on the empirical relationship of the scaled distance, and combining an interactive collaborative graph neural network to construct an interactive collaborative graph neural network for the power field of a limited open space building under explosion shock. Using the power field database of the limited open space building under explosion shock to train, evaluate, and optimize it to obtain a maturely trained power field model of a limited open space building under explosion shock.

[0086] The steps for constructing the power field model of a limited open space building under explosion shock include:

[0087] S21. Fitting of the empirical relationship: Using the random number method to extract random sample data from the database according to the scaled distance for explosions in the near field, mid-field, and far field respectively, to obtain the empirical relationship between the peak overpressure, peak impulse of the limited open space, and the shock wave arrival time and the scaled distance after fitting.

[0088] 1. Scaled distance: The scaled distance is obtained by the distance from the explosion center to the TNT equivalent.

[0089] The distance from the explosion center to the node refers to the distance between the explosion center and the node. Given the spatial coordinates of the known explosion center and the node, the distance from the explosion center to the node can be calculated, and further comparison can obtain the scaled distance.

[0090] 2. The empirical relationship between peak overpressure and scaled distance is given by the following formula (1):

[0091] (1)

[0092] Where P represents peak overpressure, Z represents scaled distance, and A1, B1, C1, D1 represent the fitted parameters.

[0093] 3. The empirical relationship between peak impulse and scaled distance is given by the following formula (2):

[0094] (2)

[0095] Where I + represents peak impulse, W represents TNT equivalent, Z represents scaled distance, and A2 represents the parameter of peak impulse.

[0096] 4. The empirical relationship between shock wave arrival time and scaled distance is given by the following formula (3):

[0097] (3)

[0098] Where t + represents shock wave arrival time, W represents TNT equivalent, Z represents scaled distance, and A3 represents the parameter of shock wave arrival time.

[0099] According to the empirical relationships of formulas (1)-(3), the relationships between peak overpressure, peak impulse and shock wave arrival time and scaled distance in a finite open space are fitted. Through data fitting, the parameters of the above empirical relationships are obtained, and the empirical relationships in the near field, mid field and far field are obtained.

[0100] S22. Construction of a graph neural network for interaction and cooperation in the power field: Construct a graph neural network composed of an encoder, a processor and a decoder, construct physical information constraints for explosion power field prediction, and according to the empirical relationships corresponding to different scaled distances in the near field, mid field and far field respectively, embed the empirical relationships as constraint conditions into the loss function of the graph neural network to achieve dynamic physical information embedding of the graph neural network, realize physical information integration, and construct a graph neural network for interaction and cooperation in the power field.

[0101] The construction steps of the graph neural network for interaction and cooperation in the power field include:

[0102] S221. Encoder

[0103] As Figure 2As shown in the figure, the encoder part consists of a multi-head self-attention layer, a feed-forward neural network layer, a residual connection, and a layer normalization component. The multi-head self-attention layer connects each graph data feature; the feed-forward neural network layer consists of a fully connected layer and N identical sub-layers. Each of the identical sub-layers interacts information on a fully connected layer through the ReLU activation function and then aggregates the information into a fully connected layer through the ReLU activation function to achieve information aggregation; using the residual connection and layer normalization, the processing of the feed-forward neural network is completed, helping the encoder capture the dependencies of all positions in the input sequence, obtaining a part of the potential interaction and collaborative graph data, and outputting the encoded representation of the interaction and collaborative neural network.

[0104] By defining the structure, assigning features, and initializing the multi-head self-attention mechanism for the input information, the preprocessing of the input information is completed; through the multi-head self-attention layer, each "head" is given a weight matrix of a set of linear transformations, and further calculate the attention weights between the self-attention "head" and the neighbor nodes. Each head independently calculates the attention weights and updates the data features. The results of the multi-heads are combined into one by concatenation or averaging to complete the combination of multi-head attention; using the ReLU activation function, the feature information processing is completed on N fully connected layers; the output information of the N fully connected layers is combined with the data processed by the interaction and collaborative convolution using the ReLU activation function to achieve the aggregation of neighbor information on a new fully connected layer; through a sub-layer including a residual connection that prevents the values of the node feature data in each layer from changing violently and a layer normalization component that realizes the integration of node feature data, the input and output of the data are completed on a sub-layer to improve the training speed of the model.

[0105] The above process gradually refines and integrates the information in the graph structure through the multi-head self-attention layer and the feed-forward network layer until the required encoding depth is reached. By coupling multiple layers of the multi-head self-attention layer and the feed-forward neural network, the potential graph data is obtained, and the encoded representation of each node is output to achieve the encoding of the interaction and collaborative neural network.

[0106] S222. Processor

[0107] As Figure 3 shown in the figure, the processor part consists of a TransGNN processor, a convolutional layer, a multi-path network structure layer, a fully connected layer, a pooling layer, and a readout layer. Using the TransGNN processor, through the ReLU activation function, feature aggregation is completed in the convolutional layer. The multi-path network structure layer consists of a graph attention layer, a feed-forward neural network layer, and a RefineNet multi-path network structure including a refinement module and a feature extraction module. The feed-forward neural network layer in the multi-path network structure layer of the processor has the same structure as the feed-forward neural network layer of the encoder. The readout layer consists of a global pooling operation and a fully connected layer.

[0108] For the encoding output by the encoder, using the TransGNN processor, the Transformer is used to aggregate more relevant data feature information, expand the receptive field and decouple the edge information aggregation, and improve the message passing of the GNN. Through positional encoding, the GNN layer is used to fuse the structure into node attributes, capture the graph structure information, and realize the improvement of the Transformer in the graph data. Using the ReLU activation function, the form of the interactive collaborative graph neural network is obtained; through feature aggregation in the convolutional layer, the retention of key graph feature information is realized; through the multi-path network structure layer, high-resolution semantic segmentation is carried out to accelerate the graph neural network and enhance the ability to transmit three-dimensional power field information; through the fully connected layer and the pooling layer respectively, the information exchange and feature transformation between nodes are completed; the explosion power field prediction is realized in the readout layer. By using global pooling operations (such as sum, average or max pooling), the aggregation of the embedding vectors of all nodes is realized, and then classification is carried out through a fully connected layer to obtain the graph-level prediction. By iterating the latent graph data multiple times, the processing of the interactive collaborative graph neural network is completed.

[0109] 1), TransGNN processor: TransGNN is a new type of graph neural network model that alternately integrates the Transformer and graph neural network (GNN) layers to enhance their capabilities mutually. In the alternating integration, the GNN layer first processes the features, aggregates the information of neighboring nodes through the message passing mechanism, updates the data representation, and obtains the local graph structure information; using the data representation output by the GNN layer, through the self-attention mechanism, the Transformer layer aggregates the information from other parts of the graph, realizes the capture of global dependencies and expands the receptive field of each node; using the GNN layer, the global information output from the Transformer layer back to the GNN layer is further used to update the data representation, and using the global context information provided by the Transformer layer, the local structure information fusion is further realized; by continuously iterating the alternating integration process, in each iteration, the GNN layer and the Transformer layer will work alternately, provide information to each other, and continuously update the data representation to ensure that the model can focus on the most important nodes. The way of alternating integration allows the model to switch between different layers, uses the Transformer layer to broaden the receptive field and decouple the information aggregation from the edges, so as to aggregate the information from more relevant nodes and enhance the message passing ability of the GNN; the GNN layer helps the Transformer layer perceive the graph structure information and obtain more relevant information of neighboring nodes. Through the TransGNN processor, the enhancement of the three-dimensional physical information learning ability is realized.

[0110] 2) Multi-path network structure layer: Through the graph attention layer, graph-structured data processing is achieved; through the feed-forward neural network layer, feature transformation is realized; by using the RefineNet multi-path network structure technology including a feature extraction module and a refinement module, the feature extraction and refinement processes are combined in parallel, preserving the information in the original image, improving the segmentation performance, performing high-resolution semantic segmentation, and accelerating the graph neural network.

[0111] 3) Readout layer: By using global pooling operations (such as sum, average, or max pooling), the aggregation of the embedding vectors of all nodes is realized. Thereafter, classification is performed through a fully connected layer to complete the graph-level prediction (such as graph classification) of complex graph data, and the interactive collaborative graph neural network processing is realized.

[0112] S223. Decoder

[0113] As Figure 4 shown, the decoder consists of multiple layers of masked self-attention layers, N identical layers, stacking and normalization components, and a generation network layer. And each identical layer contains a convolutional layer, an activation function, and a fully connected layer. Through the processing of these layers, it is ensured that when the decoder generates a sequence, it can take into account the previous outputs and avoid the influence of future information. Through the decoder, the final predicted part is obtained.

[0114] Through multiple layers of masked self-attention layers, using N identical sub-layers, the information features of neighbor nodes are aggregated through the convolutional layer, processed by the activation function, and transmitted to the fully connected layer. For the information of the processed fully connected layer, it is processed by the stacking and normalization components to achieve decoding, complete node update, restore the potential graph data to the explosion shock power field data structure, and recombine in the generation network layer to achieve data decoding.

[0115] 1) Multiple layers of masked attention layer: The graph autoencoder represents the vertices of the graph as low-dimensional vectors, learns the low-dimensional representation of graph data using the encoder and decoder structures, migrates the variational autoencoder to the graph domain, learns the distribution (mean and variance) of the low-dimensional vector representation of nodes through encoding (graph convolution), samples the vector representation of nodes from the distribution, realizes multiple layers of masked self-attention layers, and reconstructs the graph through link prediction decoding the attention layer.

[0116] 2) Convolutional layer: The information features of neighbor nodes are aggregated through convolution operations, and the methods of convolution based on spectral decomposition and convolution based on node spatial transformation are used to realize the update of node representation. By stacking N graph convolutional layers to extract node representation, node classification and graph classification tasks are realized.

[0117] 3) Generation network layer: Using the graph generation network as a building block, recombine the nodes and edges of the graph data according to certain rules to generate a target graph with specific attributes and requirements, and realize the generation of a graph based on a set of observable graphs.

[0118] S224. Embed physical information: Construct physical information constraints, fit through the empirical relationship, and dynamically add the corresponding physical information constraints after fitting to the loss of the interactive collaborative graph neural network for predicting the explosion power field, so as to obtain an interactive collaborative graph neural network for the explosion power field of a finite open space building under explosion shock embedded with dynamic physical information.

[0119] The loss term Loss of the model given by the following formula (4) is the data loss term L between the predicted value and the true value obtained by the MSE method data and the physical information loss term L physics constitute:

[0120] (4),

[0121] In the formula, λ data and λ physics respectively represent the constraint weights.

[0122] The step of embedding physical information includes:

[0123] S2241. Construction of physical information constraint loss term: By obtaining the parameters of the explosion power field of a finite open space building under explosion shock such as peak overpressure, peak impulse, and shock wave arrival time, construct a physical information constraint loss term to ensure that the model output conforms to physical laws;

[0124] As shown in formulas (5)-(8), the physical information constraint loss term L physics is composed of the peak overpressure constraint loss term L p , the peak impulse constraint loss term L I , and the shock wave arrival time constraint loss term L t , and the corresponding values are obtained through calculation:

[0125] (5),

[0126] (6),

[0127] (7)

[0128] (8),

[0129] where M represents the number of grid nodes in the numerical simulation of the power field, represents the predicted peak overpressure, Represents the scaled distance, Represents the predicted peak impulse, Represents the end time of the explosion, and W represents the TNT equivalent, Represents the predicted arrival time of the shock wave.

[0130] S2242. Physical information constraint embedding: Obtain the scaled distance through the explosion equivalent and the distance from the explosion center. Classify the data into near-field and far-field explosion types according to the magnitude of the scaled distance. When performing physical information constraint embedding, respectively embed the empirical relationships of the corresponding near-field, mid-field, and far-field explosions to construct a data loss term. By adding dynamic physical information constraints to the loss term of the model, achieve dynamic physical information constraint embedding classified by the scaled distance;

[0131] S2243. Physical information embedding: Divide near-field, mid-field, and far-field explosions according to the magnitude of the scaled distance. Add dynamic physical information constraints to the loss term of the model to achieve dynamic physical information embedding. The loss term of the model consists of a data loss term and a physical information constraint loss term. Since the final prediction focuses on the building power field in a finite open space under explosion shock, when performing physical information constraint embedding, respectively embed the empirical relationships of the corresponding near-field, mid-field, and far-field explosions to obtain a dynamic physical information constraint loss term and achieve dynamic embedding of physical information.

[0132] S23. Construction of the power field model: Use the training data set to train the interactive collaborative graph neural network for the building power field in a finite open space under explosion shock, and use the validation data set to optimize the interactive collaborative graph neural network for the building power field in a finite open space under explosion shock to obtain a model for the building power field in a finite open space under explosion shock.

[0133] S231. Model training and optimization: Adopt the Adam algorithm to update the model parameters. Train the model in a supervised manner and optimize the model by adjusting the hyperparameters.

[0134] The model parameters are randomly generated at the initial stage of training. It is necessary to use the backpropagation algorithm to backwardly transmit the value of the loss function and calculate the gradient of the model parameters with respect to the loss value to update the model parameters. The selected backpropagation algorithm is the Adam algorithm. Convert the pressure field in the three-dimensional explosion field into a gradient field, which combines the characteristics of gradient descent and momentum optimization, and set an appropriate transfer function to highlight that the shock wave has good performance and convergence speed.

[0135] When performing hyperparameter tuning, a series of key hyperparameters need to be concerned about, including the number of iterations, batch size, learning rate, and adjustment of the model structure, etc.

[0136] S232. Construction of the power field model of a building in a finite open space under explosion shock: Taking the peak overpressure, peak impulse, and shock wave arrival time of the power field parameters as the evaluation targets, and combining the relative error, the model is evaluated by substituting the verification data set to obtain the power field model of the building in the finite open space under explosion shock.

[0137] In the present invention, a physical information-embedded interactive collaborative graph neural network for the power field of a building in a finite open space under explosion shock is established, and the Transformer and GNN layers are alternately used to enhance the ability of three-dimensional power field information transmission, realizing the processing of the explosion power field prediction interactive collaborative graph neural network. It can not only ensure the learning ability of the model for the physical problems of the power field of the building in the finite open space under explosion shock, but also ensure the learning ability of the model for three-dimensional unstructured data, thus realizing the efficient and general prediction of the power field of the building in the finite open space under explosion shock.

[0138] S30. Prediction of the power field of a building in a finite open space under explosion shock: Through digital twin technology and explosion information collection, real explosion event data is obtained, and the real explosion event data is used as the input of the power field model of the building in the finite open space under explosion shock, and its output is the required power field of the building in the finite open space under explosion shock.

[0139] The steps of predicting the power field of the building in the finite open space under explosion shock include:

[0140] S31. Acquisition of real explosion event data: Through data collection, the explosion point and explosion information are obtained, and using digital twin technology, a three-dimensional model of the on-site building and environment is established to obtain a real explosion event, and through data processing and conversion, real explosion event data is obtained;

[0141] The steps of acquiring the real explosion event data include:

[0142] S311. Explosion information collection: By obtaining the explosion center and explosion equivalent, the explosion information in the finite open space is collected to obtain the input parameters of the explosion shock in the finite open space;

[0143] S312. Acquisition of the digital twin: By obtaining the explosion position information and combining with GIS technology, a three-dimensional model of the on-site building and environment is constructed to obtain the digital twin of the building in the finite open space;

[0144] S313. Data Processing and Conversion: Convert and combine the above input parameters of explosion shock in a limited open space and the digital twin of the building in the limited open space according to data characteristics to achieve data processing; format the above input parameters of explosion shock in a limited open space and the digital twin of the building in the limited open space according to data types, standardize the formatted data, and combine the standardized data according to the power field parameters of peak overpressure, peak impulse, and shock wave arrival time to obtain explosion input data.

[0145] S314. Acquisition of Real Explosion Event Data: Combine the explosion input data according to explosion cases to obtain real explosion event data.

[0146] S32. Prediction of Explosion Shock Power Field: Input the real explosion event data into the power field model of a building in a limited open space under explosion shock, and its output is the required power field of the building in a limited open space under explosion shock.

[0147] For example, the prediction result of the test set of the power field of a building in a limited open space under explosion shock corresponding to the grid nodes divided by a numerical model simulation is represented by [explosion case number, grid node number, peak overpressure, peak impulse, shock wave arrival time], which is [1, 2500, 5e6, 1e3, 2e-1]. Its content indicates that in the explosion event with the explosion case number of 1, the peak overpressure of the unit with the grid cell node number of 2500 in the three-dimensional physical field is 5e6 Pa, the impulse is 1e3 Pa•s, and the shock wave arrival time is 2e-1 s.

[0148] It should be noted here that in the present invention, the scene is saved in a three-dimensional data format such as stl for buildings, tunnels, etc., and can be meshed. Through meshing, grid nodes are obtained. The nodes in the graph data represent data nodes, which are based on the grid nodes, but are also the same nodes, except that the attached features in the nodes will change.

[0149] The present invention also provides an emergency escape planning method for an explosion in a building in a limited open space: According to the prediction of the power field of a building in a limited open space under explosion shock, and by comparing with historical test data, obtain the personnel damage field of the corresponding underground space building, obtain the starting and ending points of personnel escape, and use the path optimization method to achieve emergency escape planning in a manner that satisfies the shortest escape time and the maximum survival rate. The steps of the emergency escape planning include:

[0150] S41. Acquisition of personnel damage field: Compare the personnel damage levels under the peak overpressure, peak impulse, and arrival time of shock wave explosion power field parameters at different proportional distances in the existing limited open space, perform interpolation processing to obtain the personnel damage percentage; use the data fitting method to obtain the calculation empirical relationship between personnel damage and explosion power field parameters; according to the prediction of the explosion shock power field, obtain the explosion power field data, substitute it into the constructed calculation empirical relationship for application, and obtain the personnel damage field.

[0151] Currently, the research on the calculation empirical relationship between personnel damage and explosion power field parameters is as described in the paper "Simplified Calculation Research on the Overpressure-Impulse Curve of Structures under Explosion Conditions" (Chen Junjie, Gao Kanghua, Sun Ao. [J]. Journal of Vibration and Shock, 2016, 35(13): 224-232. DOI: 10.13465 / j.cnki.jvs.2016.13.036.), which discloses a dynamic fitting formula for a curve:

[0152] (9),

[0153] In the formula, represents the equivalent pressure, represents the equivalent impulse, A and B respectively represent the values of the quasi-static asymptote and the impulse asymptote, C is a constant affected by the load shape, and n is a parameter affected by the damping ratio ξ and the ductility ratio η. From this formula, it can be seen that this P-I curve realizes the rapid assessment of the damage effect of building structures under explosion loads by fitting the overpressure and impulse.

[0154] Without considering the damping energy dissipation, simplify this formula. Also, in order to better reflect the damage effect, use the relationship between the peak overpressure (P) and the peak impulse (I + ) to obtain the calculation empirical formula, which amplifies the influence on the calculation of the damage percentage. And introduce the arrival time of the shock wave this variable, amplify the influence of time on the calculation of the damage percentage, and thus obtain the new following empirical relationship:

[0155] (10),

[0156] The steps for obtaining the personnel damage field described above include:

[0157] S411. Classification of personnel damage field levels: Classify the loss field levels according to the peak overpressure, peak impulse, and arrival time of the shock wave at the time of explosion, perform interpolation processing to obtain the personnel damage percentage, and divide the personnel damage field into the following different levels according to the different degrees of personnel injury.

[0158] 1. Minor Injury Grade A (0% - 10%): The lowest degree of injury, which may only involve epidermal abrasions, minor burns, or minor impact injuries.

[0159] 2. Slight Injury Grade B (10% - 30%): The degree of injury increases, and medical treatment is required, such as suturing, dressing, or short-term treatment, but usually does not result in long-term disability.

[0160] 3. Moderate Injury Grade C (30% - 50%): A relatively high degree of injury, which may involve fractures, deep burns, or internal organ injuries. The person cannot walk for a short time and requires a long time for treatment and rehabilitation.

[0161] 4. Severe Injury Grade D (50% - 80%): A serious degree of injury, which may cause severe physical damage, requires emergency medical intervention and long-term treatment, and may carry the risk of long-term disability.

[0162] 5. Extremely Severe Injury Grade E (80% - 100%): An extremely high degree of injury, which may cause life-threatening injuries, such as severe internal organ injuries, extensive burns, or extreme physical traumas, and requires immediate medical rescue and possibly long-term care.

[0163] S412. Data fitting of personnel injury situation: From historical test data, it can be seen that the effects of peak overpressure, peak impulse, and shock wave arrival time on personnel injury are non-linear relationships. Using the non-linear relationship data fitting method. By subtracting the constant parameter E1 from the peak overpressure P, then multiplying by the peak impulse subtracting the constant parameter F1, and finally multiplying by the shock wave arrival time subtracting the constant parameter G1, an empirical formula is obtained for the personnel injury field . The result is a percentage value used to describe the possible degree of injury caused by the shock wave to personnel under specific conditions. Considering the relationship between the personnel injury percentage (Y), peak overpressure (P), peak impulse (I + ), and shock wave arrival time (t + ), the calculation empirical relationship formula between personnel injury and explosion power field parameters is shown in the following formula (11):

[0164] (11)

[0165] where Y is in the form of a percentage, the symbol is %, is a function of P, I, and t, and can be expressed by the following multiplicative formula (12):

[0166] (12)

[0167] Among them, E1, F1, and G1 are constants that need to be obtained through polynomial fitting using historical data on personnel damage and experimental data. P represents the peak overpressure, I + represents the peak impulse, and t represents the arrival time of the shock wave.

[0168] By comparing with historical experimental data, inputting the parameters of peak overpressure, peak impulse, and arrival time of the shock wave, and outputting the percentage of personnel damage, data fitting of the personnel damage situation is achieved.

[0169] S413. Acquisition of the personnel damage field: Through the prediction of the explosion power field, the parameters of the power field, namely peak overpressure (P), peak impulse (I), and arrival time of the shock wave (t), are obtained and substituted into the calculation empirical relationship to further obtain the personnel damage field.

[0170] S42. Optimization of the escape route: Define the problem as a shortest path problem with multi-objective constraints. According to the condition that the personnel damage along the path does not exceed the moderate injury level (which can be set to 50%), the shortest escape time, and the maximum survival rate, these are used as multiple constraint terms for the path optimization design.

[0171] The steps for optimizing the escape route include:

[0172] S421. Model establishment: Use a weighted directed graph in graph theory to represent the environment. Each node in the graph data represents a location, and each edge in the graph data represents a path from one location to another. The weight of each edge in the graph data represents the cost (such as time, distance, etc.) from one location to another. Each node and each edge in the graph data have corresponding attributes, such as the degree of injury, escape time, and survival rate. For example, each node has an attribute of the degree of injury, indicating the degree of injury of the personnel at that location.

[0173] S422. Setting of constraint conditions: Set the overall constraint condition that the degree of injury of any node on the path cannot exceed the moderate injury level, and use the rule that the total time to complete the path should be as short as possible and the maximum survival rate on the path as the permutation constraint conditions to complete the setting of constraint conditions.

[0174] Maximum survival rate: Through the described personnel damage field, substitute it into the established model for calculation to obtain the survival rate of a single node. If the calculated value Y is greater than the moderate damage level, the damage to this node is severe and it is not included in the calculation of the overall survival rate; by calculating the percentage of node damage on the entire path, obtain the corresponding node damage level, and obtain all nodes that meet the damage level not exceeding the moderate damage level. By calculating the proportion of all compliant characteristic nodes appearing on the entire path and through weighted calculation, obtain the maximum survival rate. The obtained survival rate is used as a factor in the final arrangement selection of the path. For example, the survival rate can be reflected by adjusting the weights of the edges in the graph data, so that the algorithm tends to select paths with higher survival rates.

[0175] S423. Path optimization algorithm: Use the Dijkstra algorithm to process the weighted graph. The flowchart of path optimization is as Figure 5 shown. Initialize the distance according to all constraint conditions to obtain the distance corresponding to each node; create a set of unvisited nodes, and add a check step to the algorithm to check whether the injury degree of the new node exceeds the moderate damage level. If the injury degree exceeds the moderate damage level, do not add this node to the path, realizing the constraint of the injury degree and ensuring that the injury degree of each node on the path does not exceed the moderate damage level; update according to the constraint conditions until all nodes are visited to obtain all paths that meet the overall constraint conditions; according to the arrangement rules, complete the arrangement of the paths that meet the arrangement constraint conditions based on the requirements of the shortest total time and the maximum survival rate of the path. After comprehensive consideration, find the best path from the starting point to the end point that meets all constraint conditions to realize the optimization of the escape path.

[0176] S43. Emergency escape plan: With the help of digital twin technology, realize the digitalization of the underground limited open space building. According to the described path optimization method, realize the rapid emergency escape plan.

[0177] Based on the same technical solution, the present invention also discloses a computer-readable storage medium storing one or more programs. The one or more programs include instructions that, when executed by a computing device, cause the computing device to execute the above-mentioned machine learning and escape planning method for the explosion power field of a limited open space building.

[0178] Based on the same technical solution, the present invention also discloses a computing device, including one or more processors, one or more memories, and one or more programs. One or more of the programs are stored in the one or more memories and are configured to be executed by the one or more processors. The one or more programs include instructions for executing the above-mentioned machine learning and escape planning method for the explosion power field of a limited open space building.

[0179] 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 a completely hardware embodiment, a completely 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 memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0180] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices produce a means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0181] 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 produce a manufactured article including an instruction means that implements the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0182] 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, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0183] The above embodiments are only used to illustrate the technical idea of the present invention, and the protection scope of the present invention cannot be limited thereby. Any modification made on the basis of the technical solution according to the technical idea proposed by the present invention falls within the protection scope of the present invention.

Claims

1. A machine learning method for the three-dimensional explosion power field of a limited open space building, characterized in that, Including: Training and optimizing the interaction and cooperation graph neural network of the power field under explosion shock through a database of the power field of limited open - space buildings under explosion shock including a training data set and a validation data set to obtain a prediction model of the power field of limited open - space buildings under explosion shock, and realizing machine learning of the three - dimensional explosion power field of limited open - space buildings; among them, the power field includes peak overpressure, peak impulse, and shock wave arrival time. The interaction and cooperation graph neural network of the power field includes an encoder, a processor, and a decoder, where: The encoder includes a cascaded multi - head self - attention layer, a first feed - forward neural network layer, a residual connection, and a layer normalization component. Among them, the first feed - forward neural network layer includes a first fully - connected layer and N identical first sub - layers. Each first sub - layer includes a second fully - connected layer with a front - end cascaded ReLU activation function. The input of each first sub - layer is the output of the multi - head self - attention layer, and the output of each sub - layer is summarized to the first fully - connected layer through a ReLU activation function. The processor includes a TransGNN processor, a first convolutional layer, a multi - path network structure layer, a third fully - connected layer, a pooling layer, and a read - out layer. Among them, the multi - path network structure layer includes a cascaded graph attention layer, a second feed - forward neural network layer, and a RefineNet multi - path network structure. The read - out layer includes a cascaded global pooling layer and a fourth fully - connected layer; among them, the second feed - forward neural network layer has the same structure as the first feed - forward neural network layer. The decoder includes a multi - layer masked self - attention layer, N identical second sub - layers, a stacking and normalization component, and a generation network layer. Among them, each second sub - layer includes a second convolutional layer and a fifth fully - connected layer connected by a ReLU activation function. The input of each second convolutional layer is the output of the multi - layer masked self - attention layer. The stacking and normalization component includes a cascaded stacking layer, a softmax function, a normalization layer, and a residual connection layer. The output of the residual connection layer and the outputs of all fifth fully - connected layers are used as the input of the stacking layer.

2. The machine learning method according to claim 1, wherein Embedding physical information constraints in the interaction and cooperation graph neural network of the power field, including: Using the random number method, randomly sampling data for the three - dimensional explosion of limited open - space buildings in the near - field, mid - field, and far - field according to the proportional distance, and fitting the empirical relationships between peak overpressure, peak impulse, shock wave arrival time, and proportional distance: ， ， ， Where, I + represents the true value of the peak impulse, P represents the true value of the peak overpressure, t + represents the true value of the shock wave arrival time, A1, B1, C1, D1 represent fitting parameters, Z represents the scaled distance, A2 represents the parameter of the peak impulse, A3 represents the parameter of the shock wave arrival time, and W represents the TNT equivalent; According to the empirical relationships corresponding to different proportional distances in the near - field, mid - field, and far - field, embedding the empirical relationships as constraint conditions into the loss function of the graph neural network to realize the embedding of physical information constraints of the graph neural network.

3. The machine learning method according to claim 1, characterized in that After embedding the physical information constraints, a physical information constraint loss term L is added to the loss function of the resulting power field interaction collaborative graph neural network physics : ， ， ， ， where M represents the number of nodes, , and represent the peak overpressure constraint loss term, the peak impulse constraint loss term, and the shock wave arrival time constraint loss term respectively, represents the predicted peak overpressure of node j, and the true value of the peak overpressure of node j , represents the scaled distance of node j, and A1, B1, C1, D1 represent fitting parameters, represents the predicted peak impulse of node j, and W represents the TNT equivalent, represents the predicted shock wave arrival time of node j, A3 represents the parameter of the shock wave arrival time, and dt represents the infinitesimal element.

4. The machine learning method according to claim 1, characterized in that The steps for establishing a database of the power field of limited open - space buildings under explosion shock include: Digital - twin the real limited open - space building and the environment, randomly generate the explosion equivalent and the explosion center, and construct an explosion event. According to the explosion event, use the numerical simulation method to establish a numerical model of the power field of the explosion event and obtain the power field data of the explosion event. Perform data pre - processing on the power field data of the explosion event, establish a sample set, and combine the selected explosion characteristics to establish a database of the power field of limited open - space buildings under explosion shock including a training data set and a validation data set.

5. The machine learning method according to claim 4, wherein The steps for constructing an explosion event include: Obtain the map data source of the real limited open space building and its surrounding environment through data crawling, and construct a three-dimensional model of the limited open space building and its surrounding environment through the method of building three-dimensional digital twin; Preset the explosion equivalent and the range of the explosion center, generate the explosion equivalent and the explosion center through the random number method, and perform data combination according to the explosion event number to obtain the explosion equivalent and the explosion center under this explosion event number; Match the three-dimensional model of the limited open space building and its surrounding environment with the explosion center and the explosion equivalent under the same explosion event number to construct an explosion event.

6. The machine learning method according to claim 4, wherein The steps for establishing a database of the power field of a limited open space building under explosion shock, including a training data set and a validation data set, include: Convert the power field data of the explosion event into a two-dimensional / three-dimensional matrix form through missing value completion, amplitude adjustment, and spatial correlation transformation to achieve preprocessing of the power field data of the explosion event; Reconstruct and simulate the distribution of the explosion shock pressure field based on the data of limited measuring points, digitally reconstruct the parameters of the explosion shock power field, and construct a sample set. The parameters of the explosion shock power field include peak overpressure, peak impulse, and shock wave arrival time; Obtain the characteristic parameters of the explosion shock power field parameters through flow field analysis and numerical simulation; Combine the characteristic parameters of the explosion shock power field parameters with the explosion event to construct a database of the power field of a limited open space building under explosion shock, and divide the database of the power field of a limited open space building under explosion shock into a training data set and a validation data set.

7. Method for explosion escape planning of buildings in limited open spaces, characterized in that, Include: Adopt the machine learning method described in any one of claims 1-6 to obtain the prediction result of the power field of a limited open space building under explosion shock; Calculate the percentage of personnel damage according to the prediction result of the power field; Use a weighted directed graph to represent the three-dimensional explosion scene of the limited open space building, and use the percentage of personnel damage as the attribute of the nodes in the weighted directed graph, that is, the degree of injury of personnel at this node; Based on the Dijkstra algorithm, find the optimal escape path under the constraint conditions that the attributes of any node on the planned path do not exceed the preset moderate injury level, the survival rate is the highest, and the escape time is the shortest, and complete the escape plan.

8. The escape planning method according to claim 7, characterized in that The calculation steps of the percentage of personnel damage include: Use the nonlinear relationship data fitting method to obtain the calculation empirical relationship between the percentage of personnel damage and peak overpressure, peak impulse, and shock wave arrival time: ， Among them, is the percentage of personnel damage, and E1, F1, G1 are fitting parameters. I + represents the true value of the peak impulse, P represents the true value of the peak overpressure, and t + represents the true value of the shock wave arrival time.

9. A computer-readable storage medium storing one or more programs, the one or more programs including instructions, characterized in that, When executed by a computing device, the instructions cause the computing device to execute the machine learning method described in any one of claims 1-6.

10. An electronic device, characterized in that, Include one or more processors, one or more memories, and one or more programs, where one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors. The one or more programs include instructions for executing the machine learning method described in any one of claims 1-6.

Images