A pressure effect field distribution prediction method based on deep learning

Abstract

The embodiment of the application discloses a pressure effect field distribution prediction method based on deep learning, and belongs to the field of deep learning. In order to accelerate the calculation speed of the pressure effect field distribution in the shock wave propagation process, a pressure effect field distribution prediction network based on deep learning is designed, and through the use of a coding-decoding network structure based on ConvGRU, the subsequent n frames of pressure effect field distribution are predicted from the first 4 frames of pressure effect field distribution. The method has the advantages that: the calculation speed is fast, and only 160 seconds are needed for predicting the first 50 frames of data; the accuracy is high, and the average relative error of the pressure distribution of the overpressure part and the average relative error of the maximum pressure are both about 10 %; and the generalization is good, and the method can be used in different scenes.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and specifically to a method for predicting the distribution of pressure effect fields based on deep learning. Background Technology

[0002] Simulation experiments and numerical simulations are important methods for analyzing the propagation process of explosion shock waves. Simulation experiments, especially large-scale tests, suffer from poor repeatability, high technical difficulty, and high cost. Numerical simulations offer high computational accuracy, but are computationally expensive, requiring recalculation when initial and boundary conditions change, making them unsuitable for rapid evaluation and real-time visualization. To address these issues, this invention proposes a deep learning-based method for predicting the pressure effect field distribution during shock wave propagation, achieving a runtime speed within minutes to meet the needs of rapid evaluation and real-time visualization. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning-based method for predicting the distribution of pressure effect fields during shock wave propagation, and to achieve a minute-level running speed to meet the needs of rapid evaluation and real-time display.

[0004] To achieve the above objectives, embodiments of the present invention provide a method for predicting the distribution of pressure effect fields based on deep learning.

[0005] The method includes the following steps:

[0006] S1: Data generation and data preprocessing;

[0007] S2: Sample the training data to generate a training set, which is then provided to the model for training;

[0008] S3: Use the improved ConvGRU to build a pressure effect field distribution prediction model, and train the model using the training set, using MSE Loss as the loss function;

[0009] S4: Based on the characteristics and needs of the data, customize the model evaluation index, input the test data into the trained pressure effect field distribution prediction model for testing, and evaluate the quality of the model based on the evaluation index;

[0010] S5: When using it, users only need to input the calculation parameters and terrain mask to predict the data for the next 50 frames.

[0011] Furthermore, in step S1, the calculation parameters and terrain information are input into the numerical simulation platform, and the calculation is performed using a uniform grid. The generated data is the pressure effect distribution data at each time point in VTK format and its corresponding time information. The time information is uniformly spaced, and there are differences between the data in the scene.

[0012] The VTK format data is converted into a four-dimensional array, with the first dimension being the number of frames and the following three dimensions representing the pressure distribution in each frame.

[0013] Treat each grid cell as a pixel and assign the grid cell value to the pixel cell.

[0014] Each raw data set contains N consecutive time-stamped VTK format data sets, which are then converted into N frames of 3D image data.

[0015] Furthermore, in step S2, the ratio of training data to test data is 2:1, and the scenarios of the training data and test data are significantly different to verify the generalization ability of the model.

[0016] Each training dataset contains N frames of 3D image data. m training samples are sampled from the midpoint. Each training sample consists of 8 consecutive frames of 3D image data, with the first 4 frames serving as model input and the last 4 frames serving as model output.

[0017] Furthermore, in step S3, a neural network is constructed based on the ConvGRU model commonly used in the field of video prediction. The two-dimensional network layer in ConvGRU is transformed into a three-dimensional network. Four frames of three-dimensional images are used as input, and the expected output is four subsequent frames of three-dimensional images.

[0018] Furthermore, in step S4, relative error is used as the standard for evaluating the quality of the prediction result. Since there are many background-irrelevant factors in the data, a threshold method is used to distinguish the key area from the background, and relative error is used to evaluate the key area, which is more in line with actual needs.

[0019] During testing, the test data is input from frames 1 to 4, and the subsequent 50 frames of data are predicted iteratively. Evaluation metrics are used to evaluate the prediction results to quantitatively judge the quality of the results.

[0020] Furthermore, in step S5, calculation parameters and terrain mask are input, and three-dimensional image data of the pressure effect field distribution of frames 1 to 4 are calculated using empirical formulas combined with existing data information.

[0021] Then, these 4 frames of data are input into the pressure effect field distribution prediction model to predict the 5th to 8th frames of data. Then, the 5th to 8th frames of data are used to predict the next 4 frames of data. The above process is repeated iteratively until the data for the next 50 frames is predicted. Attached Figure Description

[0022] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0023] The structures, proportions, and sizes illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0024] Figure 1 This diagram illustrates the data generation and preprocessing process provided in an embodiment of the present invention.

[0025] Figure 2 A diagram of a pressure effect field distribution prediction model based on ConvGRU provided for embodiments of the present invention.

[0026] Figure 3 The training flowchart is provided for an embodiment of the present invention.

[0027] Figure 4 The prediction flowchart is provided for an embodiment of the present invention. Detailed Implementation

[0028] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1

[0030] To make the above-mentioned objects, features and advantages of the present invention more readily understood, the present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0031] This application provides a method for predicting the distribution of pressure effect fields based on deep learning. The method includes the following steps:

[0032] S1: Data generation and data preprocessing;

[0033] S2: Sample the training data to generate a training set, which is then provided to the model for training;

[0034] S3: Use the improved ConvGRU to build a pressure effect field distribution prediction model, and train the model using the training set, using MSE Loss as the loss function;

[0035] S4: Based on the characteristics and needs of the data, customize the model evaluation index, input the test data into the trained pressure effect field distribution prediction model for testing, and evaluate the quality of the model based on the evaluation index;

[0036] S5: When using it, users only need to input the calculation parameters and terrain mask to predict the data for the next 50 frames.

[0037] Furthermore, in step S1, the calculation parameters and terrain information are input into the numerical simulation platform, and the calculation is performed using a uniform grid. The generated data is the pressure effect distribution data at each time point in VTK format and its corresponding time information. The time information is uniformly spaced, and there are differences between the data in the scene.

[0038] The VTK format data is converted into a four-dimensional array, with the first dimension being the number of frames and the following three dimensions representing the pressure distribution in each frame.

[0039] Treat each grid cell as a pixel and assign the grid cell value to the pixel cell.

[0040] Each raw data set contains N consecutive time-stamped VTK format data sets, which are then converted into N frames of 3D image data.

[0041] Furthermore, in step S2, the ratio of training data to test data is 2:1, and the scenarios of the training data and test data are significantly different to verify the generalization ability of the model.

[0042] Each training dataset contains N frames of 3D image data. m training samples are sampled from the midpoint. Each training sample consists of 8 consecutive frames of 3D image data, with the first 4 frames serving as model input and the last 4 frames serving as model output.

[0043] Furthermore, in step S3, a neural network is constructed based on the ConvGRU model commonly used in the field of video prediction. The two-dimensional network layer in ConvGRU is transformed into a three-dimensional network. Four frames of three-dimensional images are used as input, and the expected output is four subsequent frames of three-dimensional images.

[0044] Furthermore, in step S4, relative error is used as the standard for evaluating the quality of the prediction result. Since there are many background-irrelevant factors in the data, a threshold method is used to distinguish the key area from the background, and relative error is used to evaluate the key area, which is more in line with actual needs.

[0045] During testing, the test data is input from frames 1 to 4, and the subsequent 50 frames of data are predicted iteratively. Evaluation metrics are used to evaluate the prediction results to quantitatively judge the quality of the results.

[0046] Furthermore, in step S5, calculation parameters and terrain mask are input, and three-dimensional image data of the pressure effect field distribution of frames 1 to 4 are calculated using empirical formulas combined with existing data information.

[0047] Then, these 4 frames of data are input into the pressure effect field distribution prediction model to predict the 5th to 8th frames of data. Then, the 5th to 8th frames of data are used to predict the next 4 frames of data. The above process is repeated iteratively until the data for the next 50 frames is predicted.

[0048] Figure 1 This diagram illustrates the data generation and preprocessing process. In the data generation section, given initial parameters and terrain information are input into the numerical simulation platform. The platform outputs pressure effect field distribution data in VTK format along with corresponding time information. Since VTK format data cannot be directly processed by the model and uses a uniform grid, the data preprocessing module converts the VTK format data into a four-dimensional array of size (n, x, y, z), where n represents the total number of data frames, and x, y, and z represent the size of each frame's 3D image. In the experiment, x, y, and z were 200, 200, and 80 respectively, meaning the converted 3D image size was 200. 200 80.

[0049] Figure 2This diagram illustrates the neural network structure used in this invention, a ConvGRU-based encoder-decoder structure divided into an encoding section and a decoding section. The blue squares in the diagram represent the encoding section, and the brown squares represent the decoding section. In the encoding section, the pressure effect field distribution data is input into the Encoder ConvGRU_1 module, updating its state and passing the output to the next layer, the Encoder ConvGRU_2 module. The new state is retained and used for the next time step. When the input time series is input into the encoding section in the order of time steps, the three layers of Encoder ConvGRU in the Encoder section generate the final state and pass it to the decoding section as the initial state of the three layers of Decoder ConvGRU in the decoding section. The decoding section is the reverse process of the encoding section. The initial state of the Decoder ConvGRU in the decoding section is the output of the encoding section. At each time step, the Decoder ConvGRU module in the decoding section receives the output from the upper-layer Decoder ConvGRU module as input and continuously updates the state in its own memory unit. The output of the last Decoder ConvGRU in the decoding section at each time step is processed to obtain the predicted pressure effect field distribution data. In the figure, n represents the number of output frames, and m represents the number of output frames, that is, the first n frames are used to predict the next m frames. In this invention, both n and m are set to 4.

[0050] Figure 3 This demonstrates the training process of the present invention. Assuming the current frame is frame t, the input to the video prediction algorithm is the continuous pressure effect field distribution data of the previous n frames. The output is the predicted distribution data of the pressure effect field for the next m frames. This is an end-to-end process. In this invention, both n and m are set to 4. Constructing the training dataset is a crucial step in the training process. When constructing the training set, sampling of the training data is necessary. Each original dataset contains N frames of 3D image data. Analysis of the data shows that the shockwave from frames 50 to N becomes very weak and has low research value; therefore, only the first 50 frames are used when generating samples. Training samples are sampled from the first 50 frames at one-frame intervals. Each training sample consists of 8 consecutive frames of 3D image data, with the first 4 frames used as model input and the last 4 frames as model output. Each original dataset can yield 32 training samples. Simultaneously, due to the limitation of video memory size and the symmetrical propagation of the shockwave, the XZ and YZ planes are used to divide a 200... 200 The 80-sized sample was divided into four 100-sized samples. 100 80 samples were used, and these 4 samples were transformed to the same orientation using rotational symmetry.

[0051] Figure 4 This demonstrates the prediction process of the present invention, which is divided into a first 4-frame calculation module and an iterative prediction module. The process of the first 4-frame calculation module is as follows:

[0052] Input the initial calculation parameters, terrain mask, and initial peak radius r;

[0053] The existing pressure effect field distribution data are integrated according to empirical formulas to establish an index with the effective peak radius as the key and the effective radius-pressure relationship as the value. The index is denoted as refer_dict.

[0054] Based on the initial peak radius r, the peak radii of the first four frames are calculated to be r, r+2, r+4, and r+6, respectively.

[0055] Calculate the effective peak radius of the first 4 frames, use refer_dict to find the effective radius-pressure relationship corresponding to the nearest effective peak radius, and then restore the pressure effect field distribution data through a series of transformations.

[0056] After calculating the first four frames of data, these four frames are input into the iterative prediction module for subsequent predictions. The flow of the iterative prediction module is as follows:

[0057] The four frames of data obtained from the first four frames of the calculation module are used as the input to the deep learning model for the first to fourth frames, and the pressure effect field distribution data for the fifth to eighth frames is predicted.

[0058] The predicted data from frames 5 to 8 are input into the deep learning model to predict the pressure effect field distribution data from frames 9 to 12. The predicted data is then iteratively input into the deep learning model to predict subsequent data until frame 50.

[0059] Example 2

[0060] This application provides a method for predicting the distribution of pressure effect fields based on deep learning. The method includes the following steps:

[0061] Step 1: Data generation and preprocessing module.

[0062] In the data generation section, the calculation parameters and terrain file information are input into the numerical simulation platform, and the calculation grid is set to a uniform grid. The output is VTK format data of time information and pressure effect field distribution at each time point.

[0063] However, storing VTK format data requires a large amount of storage space, and neural networks cannot directly use VTK format data. Therefore, this invention converts VTK format data into a four-dimensional array of size (n, x, y, z), where n represents n frames, and each frame is a three-dimensional array of size (x, y, z). The specific conversion operation is as follows: each grid in the VTK format data is treated as a pixel, and the grid value is assigned to the pixel. Since the calculation grid is set to a uniform grid during simulation, each VTK data can be converted into a three-dimensional array. Since a sample contains VTK format data from multiple time points, time is taken as a dimension, and the data from each time point are concatenated to obtain the four-dimensional array.

[0064] To facilitate model training, the data is standardized using the following formula:

[0065]

[0066] Where X is the original data, Y is the standardized data, and 101300.0 is the size of one standard atmosphere. After standardization, the pixel values ​​in the data represent overpressure.

[0067] Step 2: Construction of the pressure effect field prediction model.

[0068] By preprocessing the raw data, this invention transforms N consecutive VTK format data moments into N frames of 3D image data, thus applying the problem to the field of video prediction. Assuming the current frame is frame t, the input to the video prediction algorithm is the continuous pressure effect field distribution of the previous n frames. The output is the predicted continuous pressure effect field distribution for the next m frames. It is an end-to-end process.

[0069] This invention uses a ConvGRU-based encoder-decoder structure to construct a pressure effect field distribution prediction model. The ConvGRU unit mainly comprises convolutional layers and recurrent gated units. The convolutional layers are used to extract spatial features from the 3D image, while the recurrent gated units are used to learn temporal features. The overall network structure is composed of ConvGRU units, divided into an encoding part and a decoding part. The encoding part extracts the temporal and spatial features of the input data, while the decoding part predicts subsequent frames based on the features extracted by the encoding part.

[0070] Step 3: Prediction process and evaluation indicators.

[0071] In the prediction phase, the first four frames of 3D image data from the test data are input to predict the data for the subsequent 40 frames. The specific process is as follows: input the first four frames of test data into the pressure effect field distribution prediction model to predict the data for the fifth to eighth frames. Then use the data for the fifth to eighth frames to predict the data for the subsequent four frames. Repeat the above process iteratively until the data for the subsequent 40 frames is predicted.

[0072] This invention uses relative error to evaluate the model, and the error calculation formula is as follows:

[0073]

[0074] Because the data contains many background-irrelevant factors, a threshold method is used to distinguish critical regions from the background, and relative error is used to evaluate the critical regions, which better reflects actual needs. In this invention, the threshold is set to 0, and the portion greater than 0 is the positive pressure portion, i.e., the critical region.

[0075] Step 4: Prediction process design.

[0076] The main function of this invention is to calculate the pressure effect field distribution during the propagation of a shock wave. Users only need to input the calculation parameters and terrain mask information to obtain the pressure effect field distribution at each moment during the propagation of the shock wave.

[0077] To achieve the above functions, this invention uses an empirical formula combined with known data to calculate the three-dimensional image data of the first four frames. The empirical formula is as follows:

[0078]

[0079] Where r is the shock wave propagation radius, Let w be the effective radius and w be the explosive equivalent. This formula normalizes data of different equivalents into a single dimension, achieving the purpose of integrating data of different equivalents together and establishing an index. When calculating the first four frames, the calculation parameters and terrain file information are input. After normalization, the effective radius-pressure relationship is obtained by searching on the previously established index. Through a series of transformations, the pressure effect field distribution data is restored.

[0080] After calculating the first 4 frames of data, these 4 frames of data are input into the pressure effect field distribution prediction model to iteratively predict the subsequent 50 frames of data.

[0081] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for predicting the distribution of pressure effect fields based on deep learning, characterized in that, The method includes the following steps: S1: Data generation and data preprocessing; S2: Sample the training data to generate a training set, which is then provided to the model for training; S3: Use the improved ConvGRU to build a pressure effect field distribution prediction model, and train the model using the training set, using MSE Loss as the loss function; S4: Based on the characteristics and needs of the data, customize the model evaluation index, input the test data into the trained pressure effect field distribution prediction model for testing, and evaluate the quality of the model based on the evaluation index; S5: When using it, users only need to input the calculation parameters and terrain mask to predict the data for the next 50 frames; In step S1, the calculation parameters and terrain information are input into the numerical simulation platform, and the calculation is performed using a uniform grid. The generated data is the pressure effect distribution data at each time point in VTK format and its corresponding time information. The time information is uniformly spaced, and there are differences between the data in the scene. The VTK format data is converted into a four-dimensional array, with the first dimension being the number of frames and the following three dimensions representing the pressure distribution in each frame. VTK format data is converted into a four-dimensional array of size (n, x, y, z), where n represents n frames, and each frame is a three-dimensional array of size (x, y, z). The specific conversion operation is as follows: each grid in the VTK format data is treated as a pixel, and the grid value is assigned to the pixel. Since the calculation grid is set to a uniform grid during simulation, each VTK data can be converted into a three-dimensional array. Since a sample contains VTK format data from multiple time points, time is taken as a dimension, and the data from each time point are concatenated to obtain the four-dimensional array. To facilitate model training, the data is standardized using the following formula: VTK format data is converted into a four-dimensional array of size (n, x, y, z), where n represents n frames, and each frame is a three-dimensional array of size (x, y, z). The specific conversion operation is as follows: each grid in the VTK format data is treated as a pixel, and the grid value is assigned to the pixel. Since the calculation grid is set to a uniform grid during simulation, each VTK data can be converted into a three-dimensional array. Since a sample contains VTK format data from multiple time points, time is taken as a dimension, and the data from each time point are concatenated to obtain the four-dimensional array. To facilitate model training, the data is standardized using the following formula: Where X represents the original data, Y represents the standardized data, and 101300.0 represents the magnitude of one standard atmosphere. After standardization, the pixel values ​​in the data represent overpressure. Where X represents the original data, Y represents the standardized data, and 101300.0 represents the magnitude of one standard atmosphere. After standardization, the pixel values ​​in the data represent overpressure. Treat each grid cell as a pixel and assign the grid cell value to the pixel cell. Each raw data set contains N consecutive time-stamped VTK format data, which are then converted into N frames of 3D image data. In step S2, the ratio of training data to test data is 2:1, and there are significant differences between the scenarios of training data and test data to verify the generalization ability of the model. Each training data set contains N frames of 3D image data. m training samples are sampled from the midpoint. Each training sample consists of 8 consecutive frames of 3D image data, with the first 4 frames used as model input and the last 4 frames used as model output. In step S3, a neural network is constructed based on the ConvGRU model commonly used in the video prediction field. The two-dimensional network layer in ConvGRU is transformed into a three-dimensional network. Four three-dimensional images are used as input, and the expected output is four subsequent three-dimensional images. In step S4, relative error is used as the standard for evaluating the quality of the prediction results. Since there are many background-irrelevant factors in the data, the threshold method is used to distinguish the key area from the background, and the relative error is used to evaluate the key area, which is more in line with actual needs. During testing, the test data is input from frames 1 to 4, and the subsequent 50 frames of data are predicted iteratively. Evaluation metrics are used to evaluate the prediction results to quantitatively judge the quality of the results. The model is evaluated using relative error, calculated using the following formula: Since there are many background-irrelevant factors in the data, a threshold method is used to distinguish the key area from the background, and the key area is evaluated using relative error. This is more in line with actual needs. In this invention, the threshold is set to 0, and the part greater than 0 is the positive pressure part, i.e. the key area. In step S5, calculation parameters and terrain mask are input, and three-dimensional image data of the pressure effect field distribution of frames 1 to 4 are calculated using empirical formulas combined with existing data information. Then, these 4 frames of data are input into the pressure effect field distribution prediction model to predict the 5th to 8th frames of data. Then, the 5th to 8th frames of data are used to predict the next 4 frames of data. The above process is repeated iteratively until the data for the next 50 frames is predicted.

Images