Accurate analysis method of explosion shock wave distribution pattern based on machine learning

Abstract

The present invention discloses a method for accurately analyzing the distribution pattern of explosion shock waves based on machine learning, comprising the following steps: obtaining a multi-scenario explosion overpressure field database including a training data set and a verification data set; constructing a machine learning model, wherein the machine learning model comprises a data reading module, a feature recognition module and an output module; using the training data set and the verification data set to train and optimize the machine learning model to obtain an explosion overpressure field prediction model; using the explosion overpressure field prediction model to predict the on-site explosion overpressure field, and realizing accurate analysis of the distribution pattern of explosion shock waves. The method of the present invention has high accuracy, strong generalization ability and high efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular to a precise analysis method of explosion shock wave distribution patterns based on machine learning. Background Art

[0002] Explosion overpressure field refers to the instantaneous pressure generated by the explosion shock wave in space when an explosion occurs, which exceeds the ambient static pressure and forms a pressure distribution field that changes with distance, time and medium. Explosion overpressure field prediction refers to the process of using a large amount of explosion-related data (such as explosion equivalent, explosion center distance, environmental conditions, medium characteristics, etc.) to predict the overpressure distribution and its changing trend at different locations when an explosion occurs through data analysis, machine learning or deep learning and other technical means. Explosion overpressure field prediction is the basis for evaluating safety risks and the effectiveness of protective measures, and is also the prerequisite for formulating emergency response and damage control strategies.

[0003] At present, for the rapid prediction of explosion overpressure field, as described in the paper "Research on the distribution characteristics of explosive shock waves at different altitudes" (Wang L, Kong D. Defence Technology, 2023, 24: 340-348.), a shock wave overpressure prediction method considering the influence of altitude is proposed, including: testing shock wave overpressure data at different altitudes; obtaining the attenuation effect of different altitudes on the shock wave propagation process by comparing and analyzing the overpressure data; constructing a shock wave overpressure calculation formula considering the influence of altitude to achieve effective prediction of explosion shock wave overpressure in plateau environment. However, the above method relies on the empirical calculation formula constructed by traditional numerical simulation to predict the characteristic distribution of explosion overpressure field. Faced with the complexity of explosion wave propagation, it is impossible to accurately capture the overpressure distribution in complex scenarios, which affects the accuracy and universality of the method.

[0004] For example, regarding the prediction of explosion overpressure field, as described in the paper "A Machine Learning-Based Method for Fast Prediction of Explosion Wave Overpressure Distribution in Urban Areas" (Murong F, Zheng K, Xiong C, et al. / / 2023 2nd International Conference on Artificial Intelligence and Computer Information Technology (AICIT). IEEE, 2023: 1-3.), it discloses a method for fast prediction of explosion wave overpressure distribution in urban areas based on machine learning, including: generating an explosion scene slice model by numerically simulating explosions and explosion wave propagation in urban areas; establishing an explosion wave overpressure data set through batch simulation; and combining the explosion wave overpressure data set to construct an XGBoost-based explosion wave overpressure prediction model to achieve fast prediction of peak overpressure at the bottom and top of urban buildings. However, the above method predicts explosion overpressure by constructing an XGBoost-based explosion wave overpressure prediction model, which not only takes a lot of time to prepare labeled training data sets, but also takes up a lot of computing resources to train the model, which affects the efficiency and economy of the method to a certain extent.

[0005] In summary, the problems existing in the prior art are: the prediction accuracy of explosion overpressure field is not high enough, the versatility is poor, and the efficiency is relatively low. Summary of the invention

[0006] The purpose of the present invention is to provide a precise analysis method of explosion shock wave distribution pattern based on machine learning, which has high accuracy, strong generalization ability and high efficiency.

[0007] The technical solution to achieve the purpose of the present invention is:

[0008] In a first aspect, a method for accurately analyzing the distribution pattern of explosion shock waves based on machine learning is provided, comprising the following steps:

[0009] Obtain a multi-scenario explosion overpressure field database including a training data set and a verification data set;

[0010] Constructing a machine learning model, wherein the machine learning model includes a data reading module, a feature recognition module, and an output module;

[0011] The data reading module includes a fully connected layer and a wavelet decomposition and supervised attention layer; wherein the fully connected layer is composed of an input layer, an intermediate layer, a first ReLU activation layer and a non-local attention layer in series; the non-local attention layer includes the first to fourth convolutional layers, the first to third convolutional layers are parallel and the inputs are all the outputs of the first ReLU activation layer, the outputs of the first and second convolutional layers are point-multiplied and then Softmax operated, the output of the third convolutional layer is point-multiplied with the result of the Softmax operation and then used as the input of the fourth convolutional layer, the input of the fourth convolutional layer is superimposed with the output of the first ReLU activation layer as the output of the non-local attention layer; the wavelet decomposition and supervised attention layer includes a discrete wavelet transform DWT unit, the fifth to eleventh convolutional layers, and an integer wavelet transform IWT unit, the input of the DWT unit is the output of the non-local attention layer, and the output of the DWT unit is sequentially passed through the fifth to eighth convolutional layers and the IWT unit before outputting the feature , the inputs of the ninth and tenth convolutional layers are The output of the ninth convolutional layer is superimposed with the input of the data reading module as the input of the tenth convolutional layer. The output of the eleventh convolutional layer is subjected to a Sigmoid operation. The result of the Sigmoid operation is multiplied by the output of the tenth convolutional layer and then summed with Superimpose output attention feature map ;

[0012] The feature recognition module consists of the 12th to 16th convolutional layers and the channel attention layer in series; the channel attention layer includes the pooling layer, the 17th convolutional layer, the second ReLU activation layer and the 18th convolutional layer in series. The input of the pooling layer is , the output of the eighteenth convolutional layer is subjected to Sigmoid operation and The dot product is used as the output of the channel attention layer;

[0013] The output module includes a deconvolution layer and a third ReLU activation layer. The input of the deconvolution layer is the output of the channel attention layer.

[0014] The training data set and the validation data set are used to train and optimize the machine learning model to obtain the explosion overpressure field prediction model;

[0015] The explosion overpressure field prediction model is used to predict the on-site explosion overpressure field and realize accurate analysis of the explosion shock wave distribution pattern.

[0016] Preferably, the loss function of the machine learning model is The expression is:

[0017] ,

[0018] in, Represents the number of graph data nodes, Characterize the number of nodes in the fine area of ​​the graph data. The judgment standard of the fine area is: the edge length between two nodes is less than the average length of all edges in the graph data; Represents the entire area of ​​the graph data, Represents fine-grained regions in graph data, The prediction result of the machine learning model representing the graph data node n, Characterization The corresponding truth value, The prediction result of the machine learning model representing the graph data node f, Characterization The corresponding truth value, is the weight coefficient of weight loss.

[0019] Preferably, the steps of constructing a multi-scenario explosion overpressure field database include:

[0020] Acquisition of multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set: Based on multi-scenario and multi-dimensional actual working conditions, the overpressure distribution under explosion impact is solved through Abaqus numerical simulation, and the multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set is obtained;

[0021] Construction of a multi-scenario explosion overpressure field database: Data processing and data conversion are performed on the multi-scenario and multi-dimensional explosion overpressure distribution characteristic data sets to build a multi-scenario explosion overpressure field database containing training data sets and verification data sets.

[0022] Preferably, the step of acquiring the multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set includes:

[0023] By setting different explosion source characteristics, different building structures, different structural node types, and different proportional distances, the actual working conditions in multiple scenarios and dimensions can be obtained;

[0024] Based on multi-scenario and multi-dimensional actual working conditions, the overpressure distribution under explosion impact is solved through Abaqus numerical simulation, and the multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set is obtained.

[0025] Preferably, the step of acquiring multi-scenario and multi-dimensional actual working conditions includes:

[0026] Explosion source characteristic setting: Based on the principle of explosion mechanics, the equivalent, density, and energy release rate parameters of different types of explosives are determined. Combined with the physical and chemical properties of the explosives, the initial conditions of the explosion center are determined. The initial conditions of the explosion center include the initial pressure, temperature, and energy of the explosion center.

[0027] Building structure setting: according to the actual scene, set up a variety of building structures, and set the corresponding material properties and geometric shapes of each building structure;

[0028] Structural node type setting: set the structural node of each building structure;

[0029] Proportional distance setting: According to the actual scenario, set different distance ranges, and use the normalized distance formula to express the ratio of the explosion center distance to the explosion equivalent, which is the proportional distance:

[0030] ,

[0031] Among them, Z is the proportional distance, R is the distance from the center of the explosion, and W is the explosion equivalent.

[0032] Preferably, the construction of a multi-scenario explosion overpressure field database includes:

[0033] Data processing: Data preprocessing, feature extraction and data standardization are performed on the multi-scenario and multi-dimensional explosion overpressure field data set to obtain standardized data;

[0034] Data conversion: Use the adjacency matrix to represent the connection relationship between standardized data nodes, and combine the features of nodes and edges into a graph. Through random sampling, the graph is enhanced to obtain a standardized graph storage structure.

[0035] Construction of a multi-scenario explosion overpressure field database: Stratified sampling is performed on the standardized graph storage structure to construct a multi-scenario explosion overpressure field database containing training data sets and verification data sets.

[0036] Preferably, the data processing step comprises:

[0037] Data preprocessing: batch processing of data in the multi-scenario and multi-dimensional explosion overpressure field database and conversion into a graph structure that can be used for learning;

[0038] Feature extraction: vectorize the features of nodes and edges in the graph structure and construct feature vectors of multi-source data;

[0039] Data standardization: Normalize the feature vector to obtain standardized data.

[0040] Secondly, a precise analysis system for explosion shock wave distribution patterns based on machine learning is also provided, including:

[0041] A data acquisition unit, used for acquiring a multi-scenario explosion overpressure field database including a training data set and a verification data set;

[0042] Model unit, used to build machine learning models;

[0043] A training unit, used to train and optimize the machine learning model using a training data set and a validation data set to obtain an explosion overpressure field prediction model;

[0044] The prediction unit is used to predict the on-site explosion overpressure field using the explosion overpressure field prediction model, so as to realize the accurate analysis of the explosion shock wave distribution pattern.

[0045] According to a third aspect, a computer-readable storage medium storing one or more programs is provided. The one or more programs include instructions. When the instructions are executed by a computing device, the computing device executes the method as described above.

[0046] The fourth aspect also provides an electronic device, comprising 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 are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the method as described above.

[0047] Compared with the prior art, the present invention has the following significant advantages:

[0048] 1. More accurate prediction: Based on multi-scenario and multi-dimensional actual working conditions, the present invention constructs a multi-scenario explosion overpressure field database through numerical simulation to express the overpressure distribution characteristics under multi-scenario and multi-dimensional explosion impact. By establishing a digital intelligence simulation explosion overpressure prediction model based on a multi-module graph neural network, accurate overpressure distribution prediction of explosion impact in complex multi-dimensional and multi-scale scenarios is achieved;

[0049] 2. Strong generalization ability and high efficiency: The present invention establishes a multi-scenario explosion overpressure field database. By training the neural network, the digital simulation explosion overpressure prediction model based on the multi-module graph neural network is realized. The model has the ability to quickly learn and generalize the geometric structure characteristics of different equivalents, different proportional distances, and different facilities, thereby realizing efficient prediction of multi-scenario and multi-dimensional explosion overpressure distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is the main flow chart of the method for accurately analyzing the distribution pattern of explosion shock waves based on machine learning of the present invention.

[0051] Figure 2 It is a flowchart of the steps for constructing a multi-scenario explosion overpressure field database.

[0052] Figure 3 It is a flow chart of the steps for establishing the explosion overpressure field prediction model.

[0053] Figure 4 It is a structural diagram of the explosion overpressure field prediction model.

[0054] Figure 5 It is a schematic diagram of the non-local attention layer structure.

[0055] Figure 6 It is a schematic diagram of the wavelet decomposition and supervised attention layer structure.

[0056] Figure 7 It is a schematic diagram of the channel attention layer structure.

[0057] Figure 8 It is a flow chart of the optimization steps of the explosion overpressure field prediction model.

[0058] Fig. 9 It is a flow chart of the steps for predicting explosion overpressure field. DETAILED DESCRIPTION

[0059] like Figure 1 As shown, the method for accurately analyzing the distribution pattern of explosion shock waves based on machine learning of the present invention comprises the following steps:

[0060] S10. Construction of a multi-scenario explosion overpressure field database: Based on multi-scenario and multi-dimensional actual working conditions, a multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set is obtained through Abaqus numerical simulation, and data processing and data conversion are performed on the explosion overpressure distribution characteristic data set to construct a multi-scenario explosion overpressure field database including a training data set and a verification data set.

[0061] like Figure 2 As shown, the steps of constructing the S10 multi-scenario explosion overpressure field database include:

[0062] S11. Acquisition of multi-scenario and multi-dimensional explosion overpressure distribution characteristic dataset: Based on multi-scenario and multi-dimensional actual working conditions, the overpressure distribution under explosion impact is solved through Abaqus numerical simulation, and the multi-scenario and multi-dimensional explosion overpressure distribution characteristic dataset is obtained.

[0063] The acquisition of the S11 multi-scenario multi-dimensional explosion overpressure distribution characteristic data set includes:

[0064] S111. Acquisition of actual working conditions in multiple scenarios and dimensions: Acquisition of actual working conditions in multiple scenarios and dimensions by setting different explosion source characteristics, different building structures, different node types, and different proportional distances.

[0065] The S111 multi-scenario multi-dimensional actual working condition acquisition includes:

[0066] S1111. Explosion source characteristic setting: Based on the principle of explosion mechanics, determine the equivalent, density, and energy release rate parameters of different types of explosives. Combined with the physical and chemical properties of the explosives, determine the initial conditions of the explosion center, including the initial pressure, temperature, and energy of the explosion center.

[0067] S1112. Building structure type setting: according to the actual scene, set a variety of building structures (such as high-rise buildings, industrial facilities, residential buildings, underground structures, etc.), and set their corresponding material properties and geometric shapes;

[0068] S1113, node type setting: setting corresponding positions and structural nodes according to the different building structures;

[0069] S1114. Proportional distance acquisition: according to the actual scene, different distance ranges are set, and the ratio of the explosion center distance to the explosion equivalent, i.e., the proportional distance, is expressed by the normalized distance formula:

[0070] ,

[0071] Where Z is the proportional distance, R is the distance from the center of the explosion, and W is the explosion equivalent (TNT equivalent mass).

[0072] S112. Acquisition of multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set: Based on multi-scenario and multi-dimensional actual working conditions, through numerical simulation, the overpressure distribution under the explosion impact is solved, and the multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set is obtained.

[0073] S12. Construction of a multi-scenario explosion overpressure field database: performing data processing and data conversion on the multi-scenario multi-dimensional explosion overpressure distribution characteristic data set to construct a multi-scenario explosion overpressure field database including a training data set and a verification data set.

[0074] The construction of the S12 multi-scenario explosion overpressure field database includes:

[0075] S121, data processing: performing data preprocessing, feature extraction and data standardization on the multi-scenario and multi-dimensional explosion overpressure field data set to obtain standardized data.

[0076] The S121 data processing structure includes:

[0077] S1211, data preprocessing: batch processing the data set data in the multi-scenario multi-dimensional explosion overpressure field database, and converting it into a graph structure that can be learned;

[0078] ,

[0079] ,

[0080] ,

[0081] Where G represents the graph structure, V represents the node set, and node v n It represents a spatial location, building unit or measurement point, E represents a set of edges, and edges represent physical connections or spatial relationships;

[0082] S1212 Feature Extraction: Vectorize the features of nodes and edges in the graph structure and construct feature vectors of multi-source data;

[0083] S1213 Data standardization: normalize the feature vector to obtain standardized data;

[0084] The feature vector normalization process is as follows:

[0085] ,

[0086] Wherein, Z represents the normalized data, and X represents the feature vector;

[0087] The standardized data are as follows:

[0088] ,

[0089] Among them, M represents the standardized data, Z n Characterize the normalized data under the actual working conditions of multiple scenarios and multiple dimensions (for example, explosion equivalent, proportional distance, building structure geometric parameters, etc.).

[0090] S122, data conversion: using the adjacency matrix to represent the connection relationship between standardized data nodes, and combining the features of nodes and edges into a graph, achieving graph enhancement through random sampling, and obtaining a standardized graph storage structure;

[0091] The adjacency matrix is ​​expressed as follows:

[0092] ,

[0093] ,

[0094] Where A represents the adjacency matrix, ω ij v i and v j The weight between .

[0095] S123. Construction of a multi-scenario explosion overpressure field database: performing stratified sampling on the standardized graphic storage structure to construct a multi-scenario explosion overpressure field database including a training data set and a verification data set.

[0096] S20. Establishing an explosion overpressure field prediction model: constructing a machine learning model, and using the training data set to train the machine learning model to obtain an explosion overpressure field prediction model.

[0097] like Figure 3 As shown, the S20 explosion overpressure field prediction model is established including:

[0098] S21. Machine learning model construction: Figure 4 As shown, a machine learning model is obtained by constructing a data reading module, a feature recognition module and an output module.

[0099] like Figure 4As shown, the S21 machine learning model construction includes:

[0100] S211. Data reading module construction: Set up the fully connected layer, wavelet decomposition and supervised attention layer to build the data reading module.

[0101] Among them, the fully connected layer is composed of the input layer, the intermediate layer, the ReLU activation layer and the non-local attention layer in series, which enables the fully connected layer to identify the redundant features of the input data set and extract the deep features.

[0102] like Figure 5 As shown, the specific steps for setting the non-local attention layer are:

[0103] By setting three 1×1 convolutional layers to extract features, we get three corresponding output feature matrices: , by outputting the features Perform dot multiplication and calculate The output feature similarity is After the Softmax operation, the output feature similarity results are compared with the feature matrix Perform a dot multiplication, then pass the output through a 1×1 convolutional layer and superimpose it with the input to obtain a non-local result that is consistent with the input size.

[0104] like Figure 6 As shown in Figure 1, the specific steps of wavelet decomposition and supervised attention layer setting are:

[0105] By connecting four convolution kernels in series between discrete wavelet transform (DWT) and integer wavelet transform (IWT) to extract different frequency components, the high and low frequency feature information in the non-local results is processed to obtain the input feature , input features The features obtained through a 3×3 convolution are superimposed on the original input data to obtain intermediate data. The intermediate data is activated by a 1×1 convolution and a Sigmoid function and then combined with the input features. After a 3×3 convolutional layer, the feature point multiplication is the result of the input feature Output attention feature map after superposition and fusion , thereby achieving further extraction of feature information and fusion of wavelet features and non-local results.

[0106] S212, feature recognition module construction: by setting up layers, building a graph neural network message passing mechanism, and then constructing a feature recognition module.

[0107] The message passing mechanism of the graph neural network consists of five stacked convolutional layers connected in series with a channel attention layer to achieve message passing and update of node features.

[0108] The message transmission form is expressed as follows:

[0109] ,

[0110] Among them, h i (k+1) Characterizes the characteristics of node i at layer k+1; h i (k) Characterizes the characteristics of node i at layer k; M ij (k) Represents the message from node j to node i; Represents the set of neighbor nodes of node i; Aggregate(•) represents the aggregation function; Update(•) represents the node feature update function;

[0111] like Figure 7 As shown in the figure, the channel attention layer is composed of pooling compression and Sigmoid excitation network connected in sequence to enhance the deep features of the network and repair the detailed texture.

[0112] The graph neural network message passing mechanism is used to realize the feature aggregation of deep features in the feature recognition module.

[0113] S213, output module construction: By connecting the deconvolution layer and the ReLU activation layer in series, the output module is constructed to realize the decoding of deep features and visualization of calculation results of the output module.

[0114] S214. Loss function design: The loss function design is realized by using a composite loss function combining weighted mean square error loss. The basic form is as follows:

[0115] ,

[0116] ,

[0117] ,

[0118] in, Represents the number of graph data nodes, Characterize the number of nodes in the fine area of ​​the graph data. The criterion for determining the fine area is: the edge length between two nodes is less than the average length of all edges in the graph data; Represents the entire area of ​​the graph data, Represents fine-grained regions in graph data, Characterize the prediction results of machine learning models, Represents the corresponding truth value; is the weight coefficient of weight loss;

[0119] By designing a loss function combined with grid weights, the model can further identify and extract feature information and improve the model's predictive ability.

[0120] S22. Establishment of explosion overpressure field prediction model: Use the training data set to train the machine learning module to obtain the explosion overpressure field prediction model.

[0121] S30. Optimization of explosion overpressure field prediction model: using the validation data set to evaluate the explosion overpressure field prediction model, optimizing the model based on the evaluation results, and obtaining an optimized explosion overpressure field prediction model.

[0122] like Figure 8 As shown, the S30 explosion overpressure field prediction model optimization steps include:

[0123] S31. Error evaluation: The explosion overpressure field prediction model is evaluated using the validation data set, and the prediction results are evaluated using error metrics by comparing the explosion overpressure distribution results predicted by the prediction model with the simulation data or experimental data.

[0124] The S31 error evaluation step is specifically as follows:

[0125] The mean square error (MSE) is obtained by calculating the average of the square differences between the explosion overpressure distribution results predicted by the calculation model and the simulation data or experimental data, which measures the overall error. 2 To measure the prediction of the model and evaluate the overall fitting effect of the model.

[0126] S32. Optimization of explosion overpressure field prediction model: Based on the error evaluation results, an optimized explosion overpressure field prediction model is obtained through parameter tuning, structural optimization, data structure enhancement and balancing.

[0127] The S32 explosion overpressure field prediction model optimization steps include:

[0128] S321, Hyperparameter tuning: Combine the results of error evaluation to achieve parameter tuning;

[0129] S322, structural optimization: by adjusting the number of layers, node aggregation method and convolution method of the graph neural network, the prediction results of the structurally adjusted model are evaluated for errors;

[0130] S323, data structure enhancement and balance: balance the data with large errors under specific conditions through undersampling and oversampling methods, and achieve data enhancement and balance through scene data expansion, data perturbation, changing graph structure, etc.;

[0131] S324, model optimization: by repeating steps S321 to S323 until the error assessment standard is met, the digital intelligence model is optimized.

[0132] S40, explosion overpressure field prediction: construct an on-site explosion input data set according to the on-site environment and explosion data, input the on-site explosion input data set data into the optimized explosion overpressure field prediction model, and the output is the required on-site explosion overpressure field.

[0133] like Fig. 9 As shown, the S40 explosion overpressure field prediction step includes:

[0134] S41, constructing a prediction input data set: obtaining the required prediction site environment and explosion data, and constructing a site explosion input data set;

[0135] The steps of constructing the prediction input data set in S41 are specifically as follows:

[0136] According to the on-site environmental parameter conditions such as explosion source characteristics, building structure, node type, proportional distance, etc. in the required prediction scenario, a multi-source prediction input data set is obtained through data preprocessing and feature extraction, and a prediction input data set is constructed through data standardization and data conversion.

[0137] S42, explosion overpressure distribution prediction: input the data of the on-site explosion input data set into the optimized explosion overpressure field prediction model, and the output thereof is the required on-site explosion overpressure field.

[0138] The present invention is based on multi-scenario and multi-dimensional actual working conditions, obtains multi-scenario and multi-dimensional explosion overpressure distribution feature data sets through numerical simulation, constructs a multi-scenario explosion overpressure field database through data processing and conversion, establishes a machine learning module, and performs training based on the multi-scenario explosion overpressure field database, thereby obtaining an explosion overpressure field prediction model with fast learning ability and extremely strong generalization ability, and performs error evaluation and parameter tuning on the model to improve the prediction ability of the explosion overpressure field prediction model, and realizes real-time and accurate prediction of overpressure distribution under explosion impact in complex scenarios and at multiple scales.

[0139] The present invention also designs a precise analysis system of explosion shock wave distribution pattern based on machine learning, including:

[0140] A data acquisition unit, used for acquiring a multi-scenario explosion overpressure field database including a training data set and a verification data set;

[0141] Model unit, used to build machine learning models;

[0142] A training unit, used to train and optimize the machine learning model using a training data set and a validation data set to obtain an explosion overpressure field prediction model;

[0143] The prediction unit is used to predict the on-site explosion overpressure field using the explosion overpressure field prediction model, so as to realize the accurate analysis of the explosion shock wave distribution pattern.

[0144] The technical solution of the above-mentioned precise analysis system of the explosion shock wave distribution pattern is similar to the aforementioned precise analysis method of the explosion shock wave distribution pattern, which will not be repeated here.

[0145] Based on the same technical solution, the present invention also discloses a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, enable the computing device to perform the above-mentioned precise analysis method of the explosion shock wave distribution pattern.

[0146] 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, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the above-mentioned precise analysis method of the explosion shock wave distribution pattern.

[0147] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes 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 a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0149] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0150] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0151] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A precise analysis method of explosion shock wave distribution pattern based on machine learning, characterized in that: The steps include: Obtain a multi-scenario explosion overpressure field database including a training data set and a verification data set; Constructing a machine learning model, wherein the machine learning model includes a data reading module, a feature recognition module, and an output module; The data reading module includes a fully connected layer and a wavelet decomposition and supervised attention layer; wherein the fully connected layer is composed of an input layer, an intermediate layer, a first ReLU activation layer and a non-local attention layer in series; the non-local attention layer includes the first to fourth convolutional layers, the first to third convolutional layers are parallel and the inputs are all the outputs of the first ReLU activation layer, the outputs of the first and second convolutional layers are point-multiplied and then Softmax operated, the output of the third convolutional layer is point-multiplied with the result of the Softmax operation and then used as the input of the fourth convolutional layer, the input of the fourth convolutional layer is superimposed with the output of the first ReLU activation layer as the output of the non-local attention layer; the wavelet decomposition and supervised attention layer includes a discrete wavelet transform DWT unit, the fifth to eleventh convolutional layers, and an integer wavelet transform IWT unit, the input of the DWT unit is the output of the non-local attention layer, and the output of the DWT unit is sequentially passed through the fifth to eighth convolutional layers and the IWT unit before outputting the feature , the inputs of the ninth and tenth convolutional layers are The output of the ninth convolutional layer is superimposed with the input of the data reading module as the input of the tenth convolutional layer. The output of the eleventh convolutional layer is subjected to a Sigmoid operation. The result of the Sigmoid operation is multiplied by the output of the tenth convolutional layer and then summed with Superimpose output attention feature map ; The feature recognition module consists of the 12th to 16th convolutional layers and the channel attention layer in series; the channel attention layer includes the pooling layer, the 17th convolutional layer, the second ReLU activation layer and the 18th convolutional layer in series. The input of the pooling layer is , the output of the eighteenth convolutional layer is subjected to Sigmoid operation and The dot product is used as the output of the channel attention layer; The output module includes a deconvolution layer and a third ReLU activation layer. The input of the deconvolution layer is the output of the channel attention layer. The training data set and the validation data set are used to train and optimize the machine learning model to obtain the explosion overpressure field prediction model; The explosion overpressure field prediction model is used to predict the on-site explosion overpressure field and realize accurate analysis of the explosion shock wave distribution pattern.

2. The method according to claim 1, characterized in that Loss Functions for Machine Learning Models The expression is: ， in, Represents the number of graph data nodes, Characterize the number of nodes in the fine area of ​​the graph data. The criterion for determining the fine area is: the edge length between two nodes is less than the average length of all edges in the graph data; Represents the entire area of ​​the graph data, Represents fine-grained regions in graph data, The prediction result of the machine learning model representing the graph data node n, Characterization The corresponding truth value, The prediction result of the machine learning model representing the graph data node f, Characterization The corresponding truth value, is the weight coefficient of weight loss.

3. The method according to claim 1, characterized in that: The steps for constructing a multi-scenario explosion overpressure field database include: Acquisition of multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set: Based on multi-scenario and multi-dimensional actual working conditions, the overpressure distribution under explosion impact is solved through Abaqus numerical simulation, and the multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set is obtained; Construction of a multi-scenario explosion overpressure field database: Data processing and data conversion are performed on the multi-scenario and multi-dimensional explosion overpressure distribution characteristic data sets to build a multi-scenario explosion overpressure field database containing training data sets and verification data sets.

4. The method according to claim 3, characterized in that The steps for obtaining the multi-scenario and multi-dimensional explosion overpressure distribution characteristic dataset include: By setting different explosion source characteristics, different building structures, different structural node types, and different proportional distances, the actual working conditions in multiple scenarios and dimensions can be obtained; Based on multi-scenario and multi-dimensional actual working conditions, the overpressure distribution under explosion impact is solved through Abaqus numerical simulation, and the multi-scenario and multi-dimensional explosion overpressure distribution characteristic data set is obtained.

5. The method according to claim 4, characterized in that The steps for obtaining multi-scenario and multi-dimensional actual working conditions include: Explosion source characteristic setting: Based on the principle of explosion mechanics, the equivalent, density, and energy release rate parameters of different types of explosives are determined. Combined with the physical and chemical properties of the explosives, the initial conditions of the explosion center are determined. The initial conditions of the explosion center include the initial pressure, temperature, and energy of the explosion center. Building structure setting: according to the actual scene, set up a variety of building structures, and set the corresponding material properties and geometric shapes of each building structure; Structural node type setting: set the structural node of each building structure; Proportional distance setting: According to the actual scenario, set different distance ranges, and use the normalized distance formula to express the ratio of the explosion center distance to the explosion equivalent, which is the proportional distance: ， Among them, Z is the proportional distance, R is the distance from the center of the explosion, and W is the explosion equivalent.

6. The method according to claim 3, characterized in that The construction of multi-scenario explosion overpressure field database includes: Data processing: Data preprocessing, feature extraction and data standardization are performed on the multi-scenario and multi-dimensional explosion overpressure field data set to obtain standardized data; Data conversion: Use the adjacency matrix to represent the connection relationship between standardized data nodes, and combine the features of nodes and edges into a graph. Through random sampling, the graph is enhanced to obtain a standardized graph storage structure. Construction of a multi-scenario explosion overpressure field database: Stratified sampling is performed on the standardized graph storage structure to construct a multi-scenario explosion overpressure field database containing training data sets and verification data sets.

7. The method according to claim 6, characterized in that The data processing steps include: Data preprocessing: batch processing of data in the multi-scenario and multi-dimensional explosion overpressure field database and conversion into a graph structure that can be used for learning; Feature extraction: vectorize the features of nodes and edges in the graph structure and construct feature vectors of multi-source data; Data standardization: Normalize the feature vector to obtain standardized data.

8. A precise analysis system for explosion shock wave distribution pattern based on the method described in any one of claims 1 to 7, characterized in that: include: A data acquisition unit, used for acquiring a multi-scenario explosion overpressure field database including a training data set and a verification data set; Model unit, used to build machine learning models; A training unit, used to train and optimize the machine learning model using a training data set and a validation data set to obtain an explosion overpressure field prediction model; The prediction unit is used to predict the on-site explosion overpressure field using the explosion overpressure field prediction model, so as to realize the accurate analysis of the explosion shock wave distribution pattern.

9. A computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, characterized in that: When the instructions are executed by a computing device, the computing device is caused to perform the method as claimed in claim 7.

10. An electronic device, characterized in that: The method comprises 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 are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the method according to claim 7.

Images