Geospace Radiation Belt Environment Modeling and Calculation Method Based on Deep Neural Network

Abstract

The present invention discloses a method for modeling and calculating the geospace radiation belt environment based on a deep neural network, belonging to the field of geospace radiation environment simulation modeling. The method includes the following steps: loading geospace radiation belt environment data, constructing flux intensity training data and flux distribution training data based on the geospace radiation belt environment data; building a regression prediction network and a binary classification network, optimizing the regression prediction network based on the flux intensity training data and optimizing the binary classification network based on the flux distribution training data to obtain a radiation belt ion flux intensity model and a radiation belt flux distribution model; a user inputs mission launch information, and constructs an input tensor based on the mission launch information; inputting the input tensor into the radiation belt ion flux intensity model and the radiation belt flux distribution model to obtain two prediction tensors; obtaining geospace radiation belt environment particle energy spectrum data based on the two prediction tensors.

Description

Technical Field

[0001] The present invention belongs to the technical field of geospace radiation environment simulation and modeling, and particularly relates to a method for modeling and calculating the geospace radiation belt environment based on a deep neural network. Background Art

[0002] The Earth's radiation belts are regions of high-energy particles trapped and maintained by the Earth's magnetic field, mainly divided into the inner radiation belt and the outer radiation belt. The former is mainly composed of high-energy protons, while the latter is mainly composed of high-energy electrons. Since these high-energy particles have an important impact on human activities and technical facilities, especially artificial satellites and spacecraft, the modeling research of the radiation belt environment is particularly crucial. The radiation belt environment modeling technology simulates the particle distribution, energy, and dynamic changes in the radiation belt through mathematical and physical models. These models usually rely on data from satellite observations (such as NASA's Van Allen Probes and ESA's Cluster missions) and ground-based observations (such as magnetometer networks and radio telescopes). The modeling methods include empirical models based on observational data (such as NASA's AE8 / AP8 model and AE9 / AP9 model), physical models based on physical theories (such as the Fokker-Planck equation, Salammbo model, and CRRESPRO / CRRESELE model), and numerical simulations using high-performance computing and numerical methods (such as the Particle-in-Cell method and the Monte Carlo method).

[0003] Although the prior art has made some progress, there are still some significant drawbacks. First, many empirical models are based on relatively old data and are difficult to reflect the latest changes in the current radiation belt environment, resulting in poor timeliness of the models. Second, although physical models provide a deeper understanding, they still have limitations in dealing with complex nonlinear processes and multi-scale couplings, leading to limited prediction accuracy. In addition, numerical simulation methods often require high-performance computing resources, with high computing costs and long calculation times, which limits their popularity in real-time prediction and applications. The spatio-temporal variability, nonlinear complexity, and multi-scale coupling of data increase the difficulty of modeling, making it difficult for existing models to accurately and quickly respond to sudden space weather events. Therefore, the present invention provides a method for modeling and calculating the geospace radiation belt environment based on a deep neural network. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a method for modeling and calculating the geospace radiation belt environment based on a deep neural network to solve the problems existing in the above prior art.

[0005] To achieve the above object, the present invention provides a method for modeling and calculating the geospace radiation belt environment based on a deep neural network, including the following steps:

[0006] Load the environmental data of the Earth's radiation belt, and construct flux intensity training data and flux distribution training data based on the environmental data of the Earth's radiation belt;

[0007] Build a regression prediction network and a binary classification network, optimize the regression prediction network based on the flux intensity training data, and optimize the binary classification network based on the flux distribution training data to obtain a radiation belt ion flux intensity model and a radiation belt flux distribution model;

[0008] Input the mission launch information, and construct an input tensor based on the mission launch information;

[0009] Input the input tensor into the radiation belt ion flux intensity model and the radiation belt flux distribution model to obtain two prediction tensors;

[0010] Obtain the particle energy spectrum data of the Earth's radiation belt environment based on the two prediction tensors.

[0011] Optionally, the process of constructing the flux intensity training data and the flux distribution training data based on the environmental data of the Earth's radiation belt includes:

[0012] Preprocess the environmental data of the Earth's radiation belt to obtain consistent data;

[0013] Perform coordinate transformation on the consistent data to obtain the environmental data of the Earth's radiation belt in the B-L magnetosphere coordinate system;

[0014] Combine the environmental data of the Earth's radiation belt in the B-L magnetosphere coordinate system into a two-dimensional matrix composed of magnetic field intensity, drift surface parameters, energy, and flux to obtain the flux intensity training data;

[0015] Combine the environmental data of the Earth's radiation belt in the B-L magnetosphere coordinate system into a two-dimensional matrix composed of magnetic field intensity, drift surface parameters, energy, and boolean values to obtain the flux distribution training data.

[0016] Optionally, the regression prediction network includes three fully connected hidden layers, ReLU, Leaky ReLU, and RReLU activation functions, and two Dropout layers;

[0017] The binary classification network includes three fully connected hidden layers, ReLU, Sigmoid activation functions, and one Dropout layer.

[0018] Optionally, the process of optimizing the regression prediction network and the binary classification network includes:

[0019] Use the first three columns of data in the flux intensity training data and the flux distribution training data as eigenvalue, and the first three columns of data include: magnetic field intensity, drift surface parameters, energy; use the fourth column of data as the label value;

[0020] Initialize the regression prediction network and the binary classification network and determine loss functions for the regression prediction network and the binary classification network; wherein, the loss function of the regression prediction network is the mean square error, and the loss function of the binary classification network is the cross entropy;

[0021] Input the first three columns of data of the flux intensity training data into the regression prediction network to obtain a first prediction value; calculate the mean square error between the first prediction value and the label value of the flux intensity training data, and cyclically train the regression prediction network based on the mean square error to obtain a radiation belt ion flux intensity model;

[0022] Input the first three columns of data of the flux distribution training data into the binary classification network to obtain a second prediction value; calculate the cross entropy between the second prediction value and the label value of the flux distribution training data, and cyclically train the binary classification network based on the cross entropy to obtain a radiation belt flux distribution model.

[0023] Optionally, the process of optimizing the regression prediction network and the binary classification network further includes adding regularization to the network training;

[0024] Wherein, the expression of the regularization is:

[0025]

[0026] In the formula, L is the original loss function, λ is the regularization intensity parameter, m is the number of samples, and w i is the model weight.

[0027] Optionally, the process of constructing an input tensor based on the mission launch information includes:

[0028] Calculate the flight trajectory data of the aircraft using the J2 perturbation model based on the mission launch information;

[0029] Perform coordinate transformation on the flight trajectory data to obtain the transformed flight trajectory data;

[0030] Perform data combination on the transformed flight trajectory data to obtain a two-dimensional tensor, and perform normalization processing on the two-dimensional tensor to obtain an input tensor.

[0031] Optionally, in the process of obtaining the earth's radiation belt environmental particle energy spectrum data based on the two prediction tensors, data post-processing is further included for the two prediction tensors; wherein, the process of data post-processing includes:

[0032] Convert the two prediction tensors into two-dimensional matrices respectively;

[0033] Convert the two-dimensional matrix of the radiation belt flux distribution model to obtain a flux distribution output;

[0034] Inverse normalization is performed on the two-dimensional matrix of the radiation belt ion flux intensity model to obtain the flux intensity output;

[0035] The corresponding terms of the flux distribution output and the flux intensity output are multiplied to obtain the original radiation belt particle flux matrix;

[0036] Based on the original radiation belt particle flux matrix, the average value of the flux is calculated for different energy intervals, and a two-dimensional matrix of energy and average flux is constructed.

[0037] Optionally, the process of converting the two-dimensional matrix of the radiation belt flux distribution model includes: setting it to 0 when the output probability of the radiation belt flux distribution model is less than 0.5, and setting it to 1 when it is greater than or equal to 0.5.

[0038] The present invention also provides a computer terminal device, including:

[0039] One or more processors;

[0040] A memory coupled to the processor for storing one or more programs;

[0041] When the one or more programs are executed by the one or more processors, the one or more processors implement a method for modeling and calculating the geospace radiation belt environment based on a deep neural network.

[0042] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, a method for modeling and calculating the geospace radiation belt environment based on a deep neural network is implemented.

[0043] Compared with the prior art, the present invention has the following advantages and technical effects:

[0044] The present application provides a new method for modeling and calculating the geospace radiation belt environment. This method utilizes the flexibility of the deep neural network and significantly improves the accuracy and effectiveness of model prediction by real-time updating the training data set. It solves the problem that traditional empirical models are difficult to update and maintain in a timely manner, ensuring that the model is always consistent with the dynamic changes of the current geospace radiation belt environment.

[0045] The present invention adopts a dual-model method of "flux intensity" and "flux distribution" to carry out energy spectrum simulation and modeling. By separately learning the data characteristics of the radiation belt flux intensity and flux distribution, it accurately predicts the radiation belt environment conditions (multiple charged particle flux values in a wide energy range) at each point of the flight trajectory.

[0046] This application realizes the end-to-end automated computing process from data processing to prediction output, greatly improving the efficiency and accuracy of simulation calculations, and providing solid basic data support for the stable operation of spacecraft in complex and changing radiation environments. Description of the Drawings

[0047] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0048] Figure 1 It is the flowchart of the modeling stage of the embodiment of the present invention;

[0049] Figure 2 It is the radiation belt flux intensity regression prediction network of the embodiment of the present invention;

[0050] Figure 3 It is the radiation belt flux distribution binary classification network of the embodiment of the present invention;

[0051] Figure 4 It is the flowchart of the earth radiation belt environment energy spectrum calculation of the embodiment of the present invention;

[0052] Figure 5 It is the output result of the radiation belt electron environment energy spectrum of the embodiment of the present invention;

[0053] Figure 6 It is the output result of the radiation belt proton environment energy spectrum of the embodiment of the present invention;

[0054] Figure 7 It is the output result of the radiation belt plasma electron environment energy spectrum of the embodiment of the present invention;

[0055] Figure 8 It is the output result of the radiation belt plasma proton environment energy spectrum of the embodiment of the present invention. Detailed Embodiments

[0056] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine with the embodiments to detail this application.

[0057] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0058] Embodiment 1

[0059] Radiation belt environment models have important applications in multiple fields, such as space mission planning, spacecraft design, and space weather forecasting, despite facing challenges such as the spatio-temporal variability, non-linear complexity, and multi-scale coupling of data. In the future, with the advancement of observation technologies and the improvement of computing capabilities, radiation belt environment modeling will develop towards higher accuracy, wider applications, and stronger prediction capabilities. Through more refined data integration, advanced numerical methods, and interdisciplinary cooperation, it will continuously break through bottlenecks and provide a solid guarantee for humanity's deeper exploration of space. Through continuous technological innovation and interdisciplinary cooperation, radiation belt environment modeling will continuously enhance its role in the safe operation of spacecraft and space science research, ultimately promoting humanity's comprehensive exploration and utilization of space.

[0060] In this embodiment, aiming at the problems of insufficient model timeliness, limited prediction accuracy, and high computational cost in the current technology, the present invention proposes a solution, that is, using a deep neural network to optimize and improve the modeling process of the Earth's radiation belt environment, so as to enhance the timeliness and prediction accuracy of the model while reducing the computational cost. By introducing a deep neural network model, the present invention can easily integrate the latest radiation belt environment data and achieve timely updates of the model, thereby improving the timeliness and accuracy of predictions. At the same time, the computational efficiency of the deep neural network is significantly improved, and it can complete complex computational tasks in a short time, occupying less computational resources and significantly reducing the computational cost. In this way, the present invention can not only provide more accurate and real-time radiation belt environment forecasts, but also be widely applied in fields such as space mission planning, spacecraft design, and space weather forecasting, solve the deficiencies of the existing technology, and promote the development of space science research and the safe operation of spacecraft.

[0061] This embodiment provides a method for modeling and calculating the Earth's space radiation belt environment based on a deep neural network, including the following steps:

[0062] The step process of the first stage is as shown in the appendix Figure 1 and includes: loading Earth's radiation belt environment data, constructing flux intensity training data and flux distribution training data based on the Earth's radiation belt environment data; building a regression prediction network and a binary classification network, and optimizing the regression prediction network based on the flux intensity training data and optimizing the binary classification network based on the flux distribution training data to obtain a radiation belt ion flux intensity model and a radiation belt flux distribution model.

[0063] (1) Loading data: Reading the Earth's radiation belt environment data generated by satellites or other modeling software, which contains time, geographical location coordinate information, particle energy, particle species, and ion flux information.

[0064] As a specific implementation of this embodiment, the radiation belt environment data can be the historical data of a detection satellite or the output data of other simulation models. Among them, the altitude range needs to be controlled within 100 - 65000 km, and the latitude range is controlled within 0 - 88°. The data composition should at least include: particle type, time, position coordinates relative to the Earth, energy, and flux. Different time resolutions need to be defined for data at different orbital altitudes. For the low Earth orbit (LEO) with an altitude within 2000 km, it is sampled once every 10 s; for the middle Earth orbit (MEO) with an altitude between 2000 - 35786 km, it is sampled once every 5 min; for the high orbit with an altitude above 35786 km, it is sampled once every 1 h. There is no limit on the amount of data, but it is necessary to cover the altitude and latitude ranges mentioned above as completely as possible. If the original data can cover a complete solar activity cycle, better prediction effects will be obtained.

[0065] (2) Data preprocessing: Preprocess the Earth's radiation belt environment data to obtain consistent data. To ensure the consistency of subsequent training data generation, a detailed type check needs to be performed on the loaded dataset. According to the results of these checks, unified data type conversion is performed on the data to ensure the standardization and compatibility of the dataset. In addition, the process of data type conversion will follow a unified standard to maintain the integrity and accuracy of the data and avoid introducing unnecessary biases or errors during the model training process. This standardized data preprocessing process is crucial for the performance and generalization ability of machine learning models.

[0066] As a specific implementation of this embodiment, since the loaded data comes from a wide range of sources, the data types will be more complex. For the quality of subsequent training data, it is necessary to perform unified conversion on the original data. Specifically, this method judges through the unit system of the read data and realizes the transformation of the following several common data types: 1) Conversion from simplified Julian date to decimal year; 2) Conversion from Cartesian coordinate system to WGS - 84 coordinate system; 3) Conversion from integral energy spectrum data to differential energy spectrum data. Through the above transformations, the data is unified into the form of {decimal year, longitude, latitude, altitude, energy, differential flux}. Note that the energy unit in the data is unified to MeV, and the differential flux unit is unified to MeV -1 ·cm -2 ·s -1 , and the following is the calculation formula for converting integral energy spectrum to differential energy spectrum:

[0067]

[0068] In the formula, E i (n) represents energy, F d (n) and F i(n) are the differential and integral fluxes of the n-th energy point. Note that one maximum energy point is lost when integrating the omnidirectional spectrum to obtain the differential spectrum.

[0069] (3) Coordinate transformation: Coordinate transformation is performed on the consistency data to obtain the Earth's radiation belt environmental data in the B-L magnetospheric coordinate system. The time and geographical location coordinates in the data are transformed into B-L coordinates (a magnetospheric coordinate system introduced to describe the intensity distribution of radiation belt particles). On the one hand, according to the formation mechanism of the radiation belt, the particle flux distribution is described by the magnetic field strength B and the parameter L that marks each drift surface. On the other hand, this operation reduces the strong dependence of the labeled data (particle flux data) on space and time, which is more conducive to the model convergence in the subsequent training process.

[0070] As a specific implementation manner of this embodiment, in order to more accurately describe the distribution of the radiation belt particle intensity, the B-L magnetospheric coordinate system is introduced to replace the traditional time-space coordinate system, where B represents the magnetic field strength and L is the parameter that marks each drift surface. The magnetic field strength B is obtained from the latest International Geomagnetic Reference Field (IGRF13) model, which is constructed based on the geomagnetic observation data worldwide. Since the Earth's magnetic field changes over time (including long-term and short-term changes), the IGRF model needs to be updated regularly to maintain its accuracy. Currently, the IGRF model is updated every 5 years to reflect the changes in the Earth's magnetic field. The calculation formula for the L value is as follows:

[0071]

[0072] where M is the dipole moment constant of the Earth's magnetic field, with a value of 7.9×10 30 nT·cm 3 , V represents the calculation parameter, and B0 is the magnetic field strength at the intersection of the magnetic field line at the calculation point and the equator line, both of which can be calculated and obtained through the IGRF13 model.

[0073] (4) Construct flux intensity training data: The data is combined into a two-dimensional matrix composed of B, L, energy, and particle differential flux. Meaningless values and outliers in the flux data are removed, and each column is normalized and saved as the flux intensity training data.

[0074] As a specific implementation manner of this embodiment, the data is combined into a two-dimensional matrix in the form of {B, L, energy, flux}. If the flux value is meaningless or incorrect, the entire row matrix is deleted to avoid contaminating the training data. After the above processing, each column of the matrix is normalized separately, and the numerical values are scaled between 0 and 1 to accelerate the convergence speed of the model during training. The normalization formula adopted is as follows:

[0075]

[0076] where Xtrain is the normalized output value, x is the input data, and X o represents the original data matrix.

[0077] (5) Construct the flux distribution training data: Combine the data into a two-dimensional matrix composed of B, L, energy, and particle differential flux. Set the normal data in the flux data to 1 and the abnormal data to 0, perform normalization processing on each column, and save it as the flux distribution training data.

[0078] As a specific implementation manner of this embodiment, combine the data into a two-dimensional matrix in the form of {B, L, energy, boolean value}. The boolean value 0 or 1 is determined by the flux value situation. If the flux is a meaningless or error value, set the flux value to 0, and the normal value to 1. After completing all row operations, perform normalization processing on the first three columns, and the method is the same as the previous step.

[0079] (6) Build a deep neural network model: According to the characteristics of the training data, two deep neural network models are built. For the flux intensity training data, a regression prediction network is built, and the structure is as Figure 2 shown, a deep neural network with three fully connected hidden layers, ReLU, Leaky ReLU, and RReLU activation functions, and two Dropout layers, with a total of 166,401 parameters. For the flux distribution training data, a binary classification network is built, and the structure is as Figure 3 shown, composed of two fully connected hidden layers, ReLU and Sigmoid activation functions, and one Dropout layer, with a total of 34,049 parameters.

[0080] As a specific implementation manner of this embodiment, according to the data distribution characteristics, a non-linear regression deep neural network and a binary classification deep neural network are built.

[0081] In the regression network, a linear layer (func1) with an input dimension of 3 and an output dimension of 512 was first created. Then, a Dropout layer with a retention rate of 0.5 was added to mitigate overfitting of the model. Next was a linear layer (func2) with an input dimension of 512 and an output dimension of 256, followed by another Dropout layer with a retention rate of 0.2. Finally, a linear layer (func3) with an input dimension of 256 and an output dimension of 128 was added, and a linear regression layer (regression) with an output dimension of 1 was connected for the regression task. During the forward propagation process, the input data was first passed through the ReLU activation function to func1, and then the previously defined Dropout layer was applied. Next, it was passed through the Leaky ReLU activation function to func2, and the Dropout layer was applied again. Finally, it was passed through the Randomized Leaky ReLU activation function to func3, and the prediction was ultimately made through the linear regression layer. The returned output is a one-dimensional tensor, and the first element is taken out through slicing as the final output result. The structure is as shown in the appendix Figure 2 as shown.

[0082] In the binary classification network, the structure is as shown in the appendix Figure 3 as shown, which contains three fully connected layers, namely func1, func2, and func3. During the initialization process, func1 with an input dimension of 3 and an output dimension of 256 was followed by a Dropout layer with a retention rate of 0.2. Next was func2 with an input dimension of 256 and an output dimension of 128. Finally, func3 with an input dimension of 128 and an output dimension of 1, and the output result was processed by the sigmoid function for the binary classification task. During the forward propagation process, the data was first passed through the ReLU activation function to func1, and then through the Dropout layer. Next, it was passed through the ReLU activation function to func2, and finally through the sigmoid function to func3 to obtain the classification result.

[0083] (7) Model training: The first three columns (B, L, and energy) of the training data were used as eigenvalue, and the fourth column, the particle differential flux, was used as the label value, and they were brought into the built deep neural network for training.

[0084] Further, the process of optimizing the regression prediction network and the binary classification network includes: using the first three columns of data in the flux intensity training data and the flux distribution training data as eigenvalue, where the first three columns of data include: magnetic field intensity, drift surface parameter, energy; using the fourth column of data as the label value; initializing the regression prediction network and the binary classification network and determining loss functions for the regression prediction network and the binary classification network; wherein, the loss function of the regression prediction network is mean square error, and the loss function of the binary classification network is cross entropy; inputting the first three columns of data of the flux intensity training data into the regression prediction network to obtain a first prediction value; calculating the mean square error between the first prediction value and the label value of the flux intensity training data, and cyclically training the regression prediction network based on the mean square error to obtain a radiation belt ion flux intensity model; inputting the first three columns of data of the flux distribution training data into the binary classification network to obtain a second prediction value; calculating the cross entropy between the second prediction value and the label value of the flux distribution training data, and cyclically training the binary classification network based on the cross entropy to obtain a radiation belt flux distribution model.

[0085] As a specific implementation manner of this embodiment, the present application selects the currently most popular PyTorch deep learning framework for model training. The training process is mainly divided into: 1) data preparation; 2) network loading; 3) selection of damage function and optimizer; 4) training loop; 5) verification. The specific description is as follows:

[0086] 1) In the data preparation process, first use the first three columns {B, L, energy} of the training data as eigenvalue, and the fourth column {flux} or {boolean value} as the label value. Subsequently, use the train_test_split function to divide the data into a training set and a test set, and the proportions of the two are 99% and 1% respectively. Finally, use the DataLoader function to randomly shuffle the training set and test set data respectively, and process them in batches into Tensor format. After multiple experiments, when the batch size batch_size = 512, the model training effect reaches the balance point of speed and effect.

[0087] 2) Load the deep neural network built in step six, and set the initial parameters of the network model. The present application uses the uniform distribution method for this.

[0088] 3) In the selection of the damage function, the two networks are different. For the regression network, the present application selects to use the mean square error loss function, and the formula is as follows:

[0089]

[0090] where y i is the true value of the verification set label, a i is the prediction value of the model, and n is the number of samples. For the binary classification network, the cross entropy loss function is used, and the formula is as follows:

[0091] L = -[y i * ln(a i ) + (1 - y i ) * ln(1 - a i )] (5)

[0092] In the formula, y i is the validation set label (0 or 1), and a i is the probability that the model predicts the positive class. In the selection of the optimizer, this application uses the Adam (Adaptive Moment Estimation) optimizer, which is an adaptive optimization algorithm that can adjust the learning rate according to historical gradient information.

[0093] In terms of parameter configuration, the initial learning rate is set to 0.001 to prevent the situation of non - convergence during network training. At the same time, to prevent overfitting in training, L2 regularization is added, and the formula is as follows:

[0094]

[0095] where L is the original loss function, λ is the regularization strength parameter, m is the number of samples, and w i is the model weight. During the gradient descent process, L2 regularization affects the update of the weight by adding the sum of the squares of the weights as a penalty term, so that when updating the weight, not only the gradient of the loss function is considered, but also the contribution of the regularization term to the gradient.

[0096] 4) In the training loop, first, forward propagation is performed. The data is input into the network to calculate the output of the network. Then, the defined loss function is used to calculate the loss value between the output of the network and the true label. Subsequently, according to the calculated loss, backpropagation is performed to calculate the gradient. Then, the optimizer is used to update the weights and biases of the network according to the calculated gradient. Finally, during the training process, metrics such as loss and accuracy are recorded for subsequent analysis. This application performs 100 - cycle loop training on the two networks respectively, and the learning rate is reduced by 50% every 20 cycles to obtain the optimal training result.

[0097] 5) The test and validation process is to evaluate the model using the validation set after each training epoch (epoch), which can help to understand the performance of the model on unseen data.

[0098] (8) Output the model: After successful training, two deep neural network models with corresponding weight values will be output, namely the radiation belt ion flux intensity model and the radiation belt flux distribution model.

[0099] As a specific implementation of this embodiment, the torch.save function is used to save the parameter weight values of the two types of model networks respectively, and they are stored in the file format of "xxx.pkl". Since different charged particles in the radiation belt environment need to be trained and saved separately, a total of 8 models are trained, namely: radiation belt proton flux intensity model, radiation belt proton flux distribution model, radiation belt electron flux intensity model, radiation belt electron flux distribution model, radiation belt plasma proton flux intensity model, radiation belt plasma proton flux distribution model, radiation belt plasma electron flux intensity model, and radiation belt plasma electron flux distribution model.

[0100] The calculation process in the second stage is as shown in the appendix Figure 4 and includes: the user inputs the mission launch information, and an input tensor is constructed based on the mission launch information; the input tensor is input into the radiation belt ion flux intensity model and the radiation belt flux distribution model to obtain two prediction tensors; the particle energy spectrum data of the Earth's radiation belt environment is obtained based on the two prediction tensors.

[0101] (1) User input: To calculate the radiation belt environment data, the user needs to input the launch information of the aircraft (launch date, apogee, perigee, inclination, right ascension of the ascending node, argument of perigee, mean anomaly, mission period, and time interval), the type of charged particles in the radiation belt environment to be calculated, and the energy range.

[0102] As a specific implementation of this embodiment, to drive the algorithm to perform calculations, the user needs to input detailed mission launch information, including the launch date of the aircraft, the expected apogee and perigee distances, orbital inclination, right ascension of the ascending node, argument of perigee, mean anomaly, as well as the specific mission period and time interval. In addition, we also need the user to specify the type of charged particles in the radiation belt environment to be calculated and the corresponding energy range, so as to more accurately evaluate the radiation environment that the aircraft may encounter during on-orbit operation. These input data are crucial for ensuring that the algorithm can accurately predict and plan the flight trajectory, mission execution efficiency, and safety of the aircraft. For example: The launch date is at 0:00 on January 1, 2004, the apogee is 23000 km, the perigee is 3000 km, the inclination is 5 degrees, the other orbital elements are 0 degrees, the flight duration is 7 days, and the time interval is 60 s. Calculate the electron and proton environments (including plasma) in the radiation belt, the proton energy range is 1.15 keV - 2 GeV, and the electron energy range is 1 keV - 10 MeV.

[0103] (2) Flight trajectory calculation: According to the aircraft launch information input by the user, the time, longitude, latitude, and altitude datasets of the aircraft during the mission period are calculated using the J2 perturbation model.

[0104] As a specific implementation of this embodiment, according to the user input in the previous step, the J2 perturbation model is used to calculate the flight trajectory of the aircraft. In the model, the J2 term represents the main influence of the elliptical oblateness of the Earth's equator on the satellite orbit. When calculating, it is first necessary to determine the initial orbital elements of the satellite, and then apply the perturbation effects caused by the J2 term to these orbital elements. This includes periodic corrections to parameters such as the right ascension of the ascending node, argument of perigee, and orbital inclination of the satellite orbit. By solving the perturbed orbital equation, the actual flight trajectory of the satellite under the influence of the non-uniform mass distribution of the Earth can be predicted. This method usually uses numerical integration techniques to process the perturbation equations, so as to accurately simulate the orbital changes of the satellite over a long period of time.

[0105] (3) Coordinate transformation: The same as the coordinate transformation in the first stage; transform the traditional time-space coordinate system (time, longitude, latitude, and altitude) into the B-L magnetosphere coordinate system.

[0106] (4) Construct the input tensor: Combine the data into a two-dimensional tensor composed of B, L, and energy, and perform normalization processing on each column respectively. The method is the same as step four in the modeling stage. Use the torch.tensor function to convert the matrix into the Tensor form required by the model.

[0107] (5) Data prediction: Load the flux intensity model and flux distribution model generated in the previous stage, set them to the evaluation mode, inject the input tensor into the two models in batches, and obtain two prediction tensors.

[0108] (6) Data post-processing: First, convert the tensors output by the two models into two-dimensional matrices. Secondly, convert the probability values output by the flux distribution model. Those less than 0.5 are set to 0, and vice versa. Then multiply the corresponding terms of the two matrices to obtain the radiation belt flux intensity distribution matrix. Finally, calculate the average value of the flux for each energy point.

[0109] Furthermore, the process of data post-processing includes: converting the two prediction tensors into two-dimensional matrices respectively; converting the two-dimensional matrix of the radiation belt flux distribution model to obtain the flux distribution output; performing inverse normalization processing on the two-dimensional matrix of the radiation belt ion flux intensity model to obtain the flux intensity output; multiplying the corresponding terms of the flux distribution output and the flux intensity output to obtain the original radiation belt particle flux matrix; based on the original radiation belt particle flux matrix, calculate the average value of the flux for different energy intervals, and construct a two-dimensional matrix of energy and average flux.

[0110] As a specific implementation of this embodiment, first convert the two tensors output in the previous step into matrix form, and then further process the output of the flux distribution model. Set the probability values less than 0.5 to 0 and those greater than 0.5 to 1. Then perform inverse normalization processing on the output of the flux intensity model and multiply it by the corresponding item of the output of the flux distribution model to obtain the original radiation belt particle flux matrix. Finally, calculate the average value of the flux for different energy intervals to construct a two-dimensional matrix of energy and average flux.

[0111] (7) Output energy spectrum: Sort the energies from smallest to largest and output the environmental particle energy spectrum data of the Earth's radiation belt.

[0112] As a specific implementation of this embodiment, write the two-dimensional matrix constructed in the previous step into a file, name it according to the calculation environment category input by the user in step one, and save it in the form of "xxx.dat". Attached Figures 5 - 8 is to plot the output results and compare them with the latest radiation belt environment model (AE9 / AP9 / SPM) of the US Naval Research Laboratory.

[0113] Among them, Figure 5 is the output result of the radiation belt electron environment energy spectrum; Figure 6 is the output result of the radiation belt proton environment energy spectrum; Figure 7 is the output result of the radiation belt plasma electron environment energy spectrum; Figure 8 is the output result of the radiation belt plasma proton environment energy spectrum.

[0114] This method makes full use of GPU acceleration technology, greatly improving the efficiency of model calculation. In previous long-term flight missions, the calculation of the radiation belt energy spectrum took a long time, bringing many inconveniences to the simulation calculation of the radiation environment of spacecraft. With the help of the powerful computing power of the GPU in this embodiment, this problem has been fundamentally solved, and the calculation speed has been significantly improved, providing strong support for the real-time decision-making and adjustment of spacecraft.

[0115] Embodiment 2

[0116] This embodiment also provides a computer terminal device, including:

[0117] One or more processors;

[0118] A memory, coupled to the processor, for storing one or more programs;

[0119] When one or more programs are executed by one or more processors, one or more processors implement a method for modeling and calculating the Earth's space radiation belt environment based on a deep neural network.

[0120] Embodiment 3

[0121] This embodiment also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, a method for modeling and calculating the geospace radiation belt environment based on a deep neural network is implemented.

[0122] The above is only a preferred specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for modeling and calculating the geospace radiation belt environment based on a deep neural network, characterized in that, Including the following steps: Loading the environmental data of the Earth's radiation belt, and constructing flux intensity training data and flux distribution training data based on the environmental data of the Earth's radiation belt; wherein, the process of constructing the flux intensity training data and the flux distribution training data includes: preprocessing the environmental data of the Earth's radiation belt to obtain consistent data; performing coordinate transformation on the consistent data to obtain the environmental data of the Earth's radiation belt in the B-L magnetosphere coordinate system; combining the environmental data of the Earth's radiation belt in the B-L magnetosphere coordinate system into a two-dimensional matrix composed of magnetic field intensity, drift surface parameters, energy, and flux to obtain the flux intensity training data; combining the environmental data of the Earth's radiation belt in the B-L magnetosphere coordinate system into a two-dimensional matrix composed of magnetic field intensity, drift surface parameters, energy, and boolean values to obtain the flux distribution training data; Building a regression prediction network and a binary classification network, and optimizing the regression prediction network based on the flux intensity training data and optimizing the binary classification network based on the flux distribution training data to obtain a radiation belt ion flux intensity model and a radiation belt flux distribution model; Inputting the mission launch information, and constructing an input tensor based on the mission launch information; Inputting the input tensor into the radiation belt ion flux intensity model and the radiation belt flux distribution model to obtain two prediction tensors; Obtaining the particle energy spectrum data of the Earth's radiation belt environment based on the two prediction tensors; the process of obtaining the particle energy spectrum data of the Earth's radiation belt environment based on the two prediction tensors further includes post-processing the two prediction tensors; wherein, the process of post-processing the data includes: respectively converting the two prediction tensors into two-dimensional matrices; converting the two-dimensional matrix of the radiation belt flux distribution model to obtain a flux distribution output; performing inverse normalization processing on the two-dimensional matrix of the radiation belt ion flux intensity model to obtain a flux intensity output; multiplying the corresponding items of the flux distribution output and the flux intensity output to obtain an original radiation belt particle flux matrix; based on the original radiation belt particle flux matrix, calculating the average value of the flux for different energy intervals, and constructing a two-dimensional matrix of energy and average flux.

2. The method for modeling and calculating the Earth's space radiation belt environment based on a deep neural network according to claim 1, wherein The regression prediction network includes three fully connected hidden layers, ReLU, Leaky ReLU, and RReLU activation functions, and two Dropout layers; The binary classification network includes three fully connected hidden layers, ReLU, Sigmoid activation functions, and one Dropout layer.

3. The method for modeling and calculating the geospace radiation belt environment based on a deep neural network according to claim 2, characterized in that, The process of optimizing the regression prediction network and the binary classification network includes: Taking the first three columns of data in the flux intensity training data and the flux distribution training data as eigenvalue, the first three columns of data include: magnetic field intensity, drift surface parameters, energy; taking the fourth column of data as the label value; Initializing the regression prediction network and the binary classification network and determining loss functions for the regression prediction network and the binary classification network; wherein, the loss function of the regression prediction network is mean square error, and the loss function of the binary classification network is cross entropy. Input the first three columns of the flux intensity training data into the regression prediction network to obtain a first prediction value; calculate the mean square error between the first prediction value and the label value of the flux intensity training data, and cyclically train the regression prediction network based on the mean square error to obtain a radiation belt ion flux intensity model; Input the first three columns of the flux distribution training data into the binary classification network to obtain a second prediction value; calculate the cross entropy between the second prediction value and the label value of the flux distribution training data, and cyclically train the binary classification network based on the cross entropy to obtain a radiation belt flux distribution model.

4. The method for geospace radiation belt environment modeling and calculation based on a deep neural network according to claim 3, wherein The process of optimizing the regression prediction network and the binary classification network further includes adding regularization to the network training; wherein, the expression of the regularization is: Where L is the original loss function, λ is the regularization strength parameter, m is the number of samples, and w i is the model weight.

5. The method for geospace radiation belt environment modeling and calculation based on a deep neural network according to claim 1, wherein The process of constructing an input tensor based on the mission launch information includes: Calculating the flight trajectory data of the aircraft using the J2 perturbation model based on the mission launch information; Performing a coordinate transformation on the flight trajectory data to obtain the transformed flight trajectory data; Performing data combination on the transformed flight trajectory data to obtain a two-dimensional tensor, and performing normalization processing on the two-dimensional tensor to obtain an input tensor.

6. The method for geospace radiation belt environment modeling and calculation based on a deep neural network according to claim 1, wherein The process of converting the two-dimensional matrix of the radiation belt flux distribution model includes: setting it to 0 when the output probability of the radiation belt flux distribution model is less than 0.5, and setting it to 1 when it is greater than or equal to 0.

5.

7. A computer terminal device, characterized in that, Including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for modeling and calculating the geospace radiation belt environment based on a deep neural network as described in any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for modeling and calculating the geospace radiation belt environment based on a deep neural network as described in any one of claims 1-6.

Images