A method for establishing an ionospheric electron density model based on a neural network

Abstract

The application discloses a method for establishing an ionospheric electron density model based on a neural network, which comprises the following steps: (1) performing quality control and normalization on acquired occultation observation data; (2) training Profile_net by using the occultation observation data; (3) training GAIN by using the occultation observation data and data provided by an ionospheric reference model; (4) outputting nmF2 and hmF2 maps by GAIN; and (5) inputting the nmF2 and hmF2 data of a single point into Profile_net, so that the nmF2 and hmF2 maps become three-dimensional observation data, and a global four-dimensional ionospheric electron density model is established. The application trains a neural network by deep learning, theoretical data provided by IRI and a large amount of accumulated occultation observation data, realizes a time evolution model in a horizontal dimension and a single-point profile dimension, combines the two to construct an electron density space-time distribution, and can realize a global F layer key parameter map, restore a single-point electron density profile from the F layer key parameters, and establish a global ionospheric electron density four-dimensional model.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of middle and high atmosphere science, and particularly relates to a method for establishing an ionospheric electron density model based on a neural network. BACKGROUND

[0002] Under the action of ionizing radiation, part of the neutral particles in the middle and high atmosphere is ionized to form the ionosphere. Among them, the electron density is an important environmental parameter in the middle and high atmosphere, which has an important influence on new technical systems such as short-wave communication, satellite communication, navigation and positioning. The spatial distribution, time variation and other characteristics have always been an important topic in the research of ionospheric physics and engineering application. With the development of radar detection technology and the maturation of ionospheric occultation detection technology, a large amount of F-layer electron density observation data is being accumulated, which brings new opportunities for F-layer electron density modeling research. Different from the traditional physical-based ionospheric electron density parameter modeling technology, the neural network can mine deep features of data by using the big data advantage, and construct an electron density model. Compared with the traditional modeling method, the neural network shows certain advantages in simulation performance and calculation efficiency. At present, a large number of related deep learning researches focus on establishing a global TEC map, rather than a three-dimensional electron density model.

[0003] There are many prior art schemes that apply deep learning to ionospheric modeling [1-4] . Liu et al. [5] explored a multi-layer tomographic imaging method for ionospheric modeling on a local GPS reference network to predict TEC. Tulunay et al. [6] used the Middle East Technical University Neural Network (METU-NN) to predict ionospheric TEC. García-Rigo et al. [7, 8] obtained better correlation between past and future TEC values by considering the values of Kp index, Dst index and Ap index, using various autoregressive orders and discrete cosine transform. Habarulema et al. [9] used artificial neural networks (ANN) to establish a TEC prediction model in the South African region by using multiple feature inputs. Sun et al.

[10] used a Long Short-Term Memory Network (LSTM) to predict the Beijing ionospheric vertical TEC. Wang et al.

[11] developed an adaptive autoregressive model to predict global TEC maps, and the difference between the obtained local TEC data and the standard data was about 3 TECU in low solar activity and possibly more than 6 TECU in medium solar activity. Pérez et al.

[12] Using the six input parameters of Kp index, solar flux, longitude and latitude, annual cumulative days, and time, a global TEC prediction model was constructed based on a multi-layer perceptron to predict global TEC in the next one to several days. Liu et al. [5] By adding the input of the LSTM model to predict TEC, the results show that it performs well in both quiet and stormy periods. Chen et al.

[13] TEC was predicted based on various improved LSTMs, and the results showed that the multi-level auxiliary forecast model has good generalization performance, good stability and low error during geomagnetic storms and quiet times. Yilmaz et al.

[14] The nonlinear modeling capabilities of Multilayer Perceptron (MLP) and Radial Basis Function Networks (RBFN) were studied to interpolate the local TEC map. Chen et al.

[15] The MIT-TEC map is completed by a regularized deep convolutional generative adversarial network, which overcomes the problem of missing TEC data in some areas and achieves better results.

[0004] References

[0005] [1]Williscroft LA, Poole AW V. Neural networks, foF2, sunspot number and magnetic activity [J]. Geophysical Research Letters, 2013, 23(24): 3659-3662. DOI: 10.1029 / 96GL03472.

[0006] [2]Altinay O, Tulunay E, Tulunay Y. Forecasting of ionospheric critical frequency using neural networks[J]. Geophysical Research Letters, 1997. DOI: info: doi / 10.1029 / 97GL01381.

[0007] [3] Wintoft P, Cander L R. Short-term prediction of foF2 using time-delay neural network [J]. Physics and Chemistry of the Earth, Part C: Solar, Terrestrial & Planetary Science, 1999, 24(4). DOI: 10.1016 / S1464-1917(99)00009-4.

[0008] [4] Nm. F, Ag. B, Ds. B, et al. Nonlinear prediction of the ionospheric parameter f(o)F(2) on hourly, daily, and monthly timescales [J]. Journal of Geophysical Research. Biogeosciences, 2000 (A6): 105.

[0009] [5] Liu Z, Gao Y. Ionospheric TEC predictions over a local area GPS reference network [J]. GPS Solutions, 2004, 8(1): 23-29. DOI: 10.1007 / s10291-004-0082-x.

[0010] [6] Tulunay E, Senalp E T, Radicella S M, et al. Forecasting total electron content maps by neural network technique [J]. Radio Science, 2016, 41(4): -. DOI: 10.1029 / 2005RS003285.

[0011] [7] A. García-Rigo, Monte E, M. Hernández-Pajares, et al. Prediction of Global Ionospheric TEC Maps: First results on a UPC forecast product [C] / / EGU General Assembly Conference Abstracts. EGU General Assembly Conference Abstracts, 2009.

[0012] [8] A. García-Rigo, E. Monte, M. Hernández-Pajares, et al. Global prediction of the vertical total electron content of the ionosphere based on GPS data [J]. Radio Science, 2011. DOI:10.1029 / 2010rs004643.

[0013] [9] Habarulema J B, Lee-Anne McKinnell, Opperman B D L. Regional GPS TEC modeling; Attempted spatial and temporal extrapolation of TEC using neural networks [J]. Journal of Geophysical Research Space Physics, 2011, 116. DOI:10.1029 / 2010JA016269.

[0014]

[10] Tomás Soler. GPS / GNSS current bibliography [J]. GPS Solutions, 2005, 9(3):243-245. DOI:10.1007 / s10291-004-0119-1.

[0015]

[11] Wang C, Xue K, Wang Z, et al. Global ionospheric maps forecasting based on an adaptive autoregressive modeling of grid point VTEC values [J]. Astrophysics and Space Science, 2020. DOI: 10.1007 / s10509-020-03760-2.

[0016]

[12] Perez R O. Using TensorFlow-based Neural Network to estimate GNSS single frequency ionospheric delay (IONONet) - ScienceDirect [J]. Advances in Space Research, 2019, 63(5): 1607-1618. DOI: 10.1016 / j.asr.2018.11.011.

[0017]

[13] Chen Z, Liao W, Li H, et al. Prediction of Global Ionosphere TEC based on Deep Learning [J]. 2021. DOI: 10.1002 / essoar.10507605.1.

[0018]

[14] A, Yilmaz. Regional TEC mapping using neural networks [J]. Radio Science, 2009. DOI: 10.1029 / 2008RS004049.

[0019]

[15] Chen Z, Jin M, Deng Y, et al. Improvement of a Deep Learning Algorithm for Total Electron Content Maps: Image Completion [J]. Journal of Geophysical Research: Space Physics, 2019, 124(1). DOI: 10.1029 / 2018JA026167. SUMMARY

[0020] Invention purposes: In view of the above problems, the application provides a method for establishing an ionospheric electron density model based on a neural network. The method splits a three-dimensional space model into a two-dimensional space model of a plane and a profile through a neural network, and then combines the two to establish a four-dimensional ionospheric electron density model that changes with time, longitude, latitude, and height.

[0021] Technical solutions: To achieve the purpose of the application, the technical solution adopted by the application is a method for establishing an ionospheric electron density model based on a neural network, comprising the following steps:

[0022] (1) performing quality control and normalization on the obtained occultation observation data;

[0023] (2) training Profile_net using the occultation observation data;

[0024] The Profile_Net is a multi-layer perceptron MLP model stacked by multiple neurons, the input of which is two parameters of the ionosphere: electron density nmF2 and electron height hmF2, and the output of which is an ionospheric electron density profile, and Profile_Net is trained using the electron density profile provided by the occultation observation data;

[0025] (3) training GAIN using the data provided by the occultation observation data and the ionospheric reference model;

[0026] (4) outputting nmF2 and hmF2 maps by GAIN;

[0027] Global map data of nmF2 and hmF2 after normalization of preset time resolution nmF2 and hmF2 provided by the occultation data to a global grid, noise z, and calibration matrix M are input into the trained GAIN network to obtain generated global nmF2 and hmF2 maps;

[0028] (5) reading the nmF2 and hmF2 maps of preset time resolution generated in (4), regarding each point as an independent object, inputting the obtained nmF2 and hmF2 two parameter data at each point into the trained Profile_Net to obtain the electron density profile data of the point, integrating the electron density profile data of all points to obtain the spatial three-dimensional electron density distribution of preset time resolution;

[0029] nmF2 and hmF2 maps output by GAIN are combined with Profile_net to establish a global four-dimensional ionospheric electron density model, and the ionospheric electron density change of the observation data is restored.

[0030] Further, a neural network GAIN for completing the incomplete global nmF2, hmF2 map provided by the occultation data is established, the GAIN is a neural network model established based on a generative adversarial interpolation network;

[0031] The input of the generator of the GAIN is the incomplete global map data of nmF2, hmF2 of a preset time resolution provided by the occultation data after being normalized to a global grid, noise z and a calibration matrix M, wherein the calibration matrix M indicates the grid point condition of the incomplete global map data with data and without data;

[0032] The output of the generator is a complete global nmF2, hmF2 map, and the trained generator realizes input of the incomplete global nmF2, hmF2 map data, noise z and the calibration matrix M, and output of the complete global nmF2, hmF2 map.

[0033] Further, in the training process, the nmF2, hmF2 reference data provided by the ionospheric reference model is normalized to a grid with the same accuracy, and is used to constrain the normalized result of the nmF2, hmF2 map generated by the generator, and the incomplete nmF2, hmF2 map input is used to constrain the generator input.

[0034] Further, the quality control process of the obtained occultation data includes elimination of abnormal peaks, height control, Chapman fitting and interpolation smoothing; and the data is normalized to a standard grid with an accuracy of 3 hours per hour, 12 degrees per longitude and 6 degrees per latitude.

[0035] Beneficial effects: Compared with the prior art, the technical scheme of the present application has the following beneficial technical effects:

[0036] The present application provides a complete four-dimensional ionospheric electron density distribution, and is closer to the actual observation compared with the IRI model based on theory, and provides a better auxiliary effect for ionospheric physical work and engineering application. Deep learning and space science are combined to establish a four-dimensional ionospheric electron density model. Previously, a large number of related deep learning researches focused on establishing a global TEC map, rather than a three-dimensional electron density model. In establishing the four-dimensional electron density model, a large amount of data observed by the RO is used as the big data basis for neural network learning, the temporal and spatial distribution of electron density is decomposed into a horizontal dimension and a single-point profile dimension time evolution model, and a global four-dimensional ionospheric electron density model is established by gradually combining. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The workflow diagram of the present application.

[0038] Figure 2(a) is a visualization of the nmF2 map provided by a certain occultation observation.

[0039] Figure 2(b) is a visualization of the calibration matrix M.

[0040] Figure 3 is a quality control workflow diagram.

[0041] Figure 4 is a Profile_Net internal structure diagram.

[0042] Figure 5 is an example of the electron density profile restored by Profile_Net.

[0043] Figure 6 is a 0-3 hour nmF2 map generated by GAIN.

[0044] Figure 7 is a 3-6 hour nmF2 map generated by GAIN.

[0045] Figure 8 is a 6-9 hour nmF2 map generated by GAIN.

[0046] Figure 9 is a 9-12 hour nmF2 map generated by GAIN.

[0047] Figure 10 is a 12-15 hour nmF2 map generated by GAIN.

[0048] Figure 11 is a 15-18 hour nmF2 map generated by GAIN.

[0049] Figure 12 is a 18-21 hour nmF2 map generated by GAIN.

[0050] Figure 13 is a 21-24 hour nmF2 map generated by GAIN. DETAILED DESCRIPTION

[0051] The technical solutions of the present application are further described below in conjunction with the accompanying drawings and examples.

[0052] The method for establishing an ionospheric electron density model based on a neural network splits a target, combines a horizontal dimension and a single-point profile dimension time evolution model to construct a four-dimensional electron density space-time distribution model, and gradually restores ionospheric electron density variation close to observation data. The obtained radio occultation (RO) observation data is subjected to quality control and normalization. The quality control process includes removing abnormal peaks, height control, Chapman fitting, and interpolation smoothing. Data normalization is to normalize the data that passes the quality control to a standard grid with an accuracy of 3 hours per hour, 12 degrees per longitude, and 6 degrees per latitude. The established neural network-based electron density profile model (Profile_Net) is a multilayer perceptron model (MLP) based on a neuron stack, which is trained with radio occultation observation electron density profile data to restore the electron density profile through two characteristic parameters of electron density nmF2 and electron height hmF2. A neural network is established based on a generative adversarial imputation network (GAIN) to complete the incomplete nmF2 and hmF2 maps provided by the radio occultation data. The radio occultation observation data and the data provided by the international ionospheric reference model (IRI model) are trained to complete the completion of the incomplete nmF2 and hmF2 maps. The established global four-dimensional ionospheric electron density model combines the nmF2 and hmF2 maps output by the GAIN with the Profile_net to establish an electron density model with an accuracy of 3 hours per hour, 12 degrees per longitude, and 6 degrees per latitude.

[0053] The embodiment provides a method for establishing an ionospheric electron density model based on a neural network. The neural network splits a three-dimensional space model into a two-dimensional space model of a plane and a profile, and then combines the two to establish a four-dimensional electron density model. The method flow of the present application is as follows Figure 1 , and specifically includes the following steps:

[0054] (1) Quality control and normalization of the obtained radio occultation observation data;

[0055] The quality control process of the obtained radio occultation data includes removing abnormal peaks, height control, Chapman fitting, and interpolation smoothing. Data normalization is to normalize the data after quality control to a standard grid with an accuracy of 3 hours per hour, 12 degrees per longitude, and 6 degrees per latitude. The quality control process of the radio occultation observation data is as follows Figure 3As shown in the figure, first, we need to remove abnormal peaks from the profile to ensure that it conforms to basic physical laws. Second, we need to control the profile's altitude range to be between 100 and 800 km. Then, to ensure the profile is valid, we perform a Chapman fit using nmF2 and hmF2. Profiles with a similarity of at least 95% are considered qualified.

[0056] (2) Use occultation observation data to train Profile_net;

[0057] The Profile_Net is a multi-layer perceptron (MLP) model composed of multiple neurons stacked together. Its input is two ionosphere parameters: electron density nmF2 and electron height hmF2. Its output is the ionosphere electron density profile. The electron density profile provided by the occultation observation data is used to train Profile_Net.

[0058] The internal structure of Profile_net that restores the single-point electron density profile is as follows Figure 4 In the embodiment of the present invention, there are five layers in total, and the structure is as follows: Embedding layer → first linear layer → second linear layer → third linear layer → fourth linear layer.

[0059] The parameters of each layer are set as follows: the number of neurons in the first to third linear layers is set to 64, 128, 256, and 512 respectively, and each linear layer performs linear operations and LeakyReLU activation processing; the Embedding layer performs encoding operations on one-hot vectors, and the training batch size is selected as 32 and the number of training epochs is selected as 300. For the optimizer, the Adam optimizer is selected, and the initial learning rate is set to 1×10 -3 , and decays by 0.8 every 30 epochs.

[0060] Figure 5 This is an example of an electron density profile restored by Profile_Net. The Chapman function and the theoretical model IRI are used as references. Here, pred profile represents the output of Profile_Net, raw profile represents the original occultation electron density profile, Chapman profile represents the Chapman function fitted profile, and iri profile represents the profile data provided by IRI.

[0061] For Profile_Net training, to facilitate subsequent network feature learning, the nmF2 and hmF2 values ​​of qualified profiles were binned into 71 intervals, forming one-hot encoded vectors. The qualified profile data was interpolated into a vector of length 71, corresponding to the electron density at every 10 km altitude between 100 and 800 km. To address cases where the observed data did not cover the electron density at 100 or 800 km, the electron density at both ends was confirmed by extending the tangent line at the nearest 30 km altitude, ensuring that its value was greater than 0. The one-hot vector and the original profile data were saved in separate files for subsequent training.

[0062] (3) Build a neural network GAIN to complete the incomplete global nmF2 and hmF2 maps provided by occultation data. GAIN is a neural network model based on a generative adversarial interpolation network. GAIN is trained using data provided by occultation observations and an ionospheric reference IRI model.

[0063] For GAIN training, qualified data are divided into standard grids with an accuracy of every 3 hours in UTC, every 12° in longitude, and every 6° in latitude and saved in a separate file for subsequent training. The input of the GAIN generator is the incomplete global map data after the three-hour time resolution nmF2 and hmF2 provided by the occultation data are standardized on the global grid, the noise z and the calibration matrix M, where the calibration matrix M indicates the grid points with and without data in the incomplete global map data. The calibration matrix M is visualized as follows Figures 2(a) to 2(b) , where Figure 2(a) shows the visualization of the nmF2 map provided by a certain occultation observation, and the corresponding calibration matrix M is visualized as shown in Figure 2(b), that is, the value of the grid point with data at the corresponding position in the matrix M is 1, and the calibration value of other grid points without data is 0.

[0064] The output of the generator is a complete global nmF2 and hmF2 map. The trained generator takes incomplete global nmF2 and hmF2 map data, noise z, and a calibration matrix M as input and outputs complete global nmF2 and hmF2 maps. During training, the nmF2 and hmF2 reference data provided by the ionospheric reference model are normalized to a grid of the same precision. This normalization is then used to constrain the normalized nmF2 and hmF2 maps generated by the generator. The incomplete nmF2 and hmF2 maps are then used to constrain the generator input.

[0065] The generator and discriminator structure and parameters of the neural network GAIN for reducing the two-dimensional F layer feature map are shown in Tables 1 and 2, respectively. In the embodiment of the present application, the training batch Batch_size is selected to be 16, and the training round Epoch is selected to be 800. For the optimizer, the Adam optimizer is selected for both the Generator and the Discriminator, and the initial learning rate is set to be 1x10 -4 Unlike the conventional training method, the batch alternation training method is adopted. First, the parameters of the Discriminator are fixed, loss1 and loss2 are taken as the loss of the Generator, and the Generator is trained for 200 rounds. Then, the parameters of the Generator are fixed, the Discriminator is trained for 100 rounds. Finally, loss3 is added as the loss of the Generator, and the Discriminator and the Generator are alternately trained for 500 rounds.

[0066] Table 1 Generator structure

[0067]

[0068]

[0069] Table 2 Discriminator structure

[0070]

[0071] (4) Output nmF2 and hmF2 maps through GAIN;

[0072] The global map data of nmF2 and hmF2 normalized to the global grid with a time resolution of three hours provided by the occultation data, noise z and calibration matrix M are input into the trained GAIN network, and the generated global nmF2 and hmF2 maps are obtained. Taking the nmF2 map as an example, the contour map of the nmF2 map output by GAIN with a time resolution of three hours within a single day on May 2, 2011 is shown in Figures 6 to 13

[0073] ​(5) read the nmF2, hmF2 map of three hours time resolution generated in (4), take each point as an independent object, input the two parameter data of nmF2, hmF2 obtained at each point into the trained Profile_Net, obtain the electron density profile data of the point, integrate the electron density profile data of all points, and obtain the spatial three-dimensional electron density distribution of three hours time resolution. Combine the nmF2, hmF2 map output by GAIN with Profile_net to establish a global four-dimensional ionospheric electron density model and restore the ionospheric electron density variation of the observation data.

Claims

1. A method for establishing an ionospheric electron density model based on a neural network, characterized in that, The method comprises the following steps: (1) quality control and normalization of the acquired occultation observation data; (2) training of Profile_net using the occultation observation data; The Profile_Net is a multi-layer perception MLP model stacked by multiple neurons, the input of which is two parameters of the ionosphere: electron density nmF2 and electron height hmF2, and the output of which is the ionospheric electron density profile, the Profile_Net being trained using the electron density profile provided by the occultation observation data; (3) training of GAIN using the data provided by the occultation observation data and the ionospheric reference model; (4) output of nmF2 and hmF2 maps by GAIN; The global map data of nmF2 and hmF2 with a preset time resolution provided by the occultation data, the noise z and the calibration matrix M after normalization to the global grid are input into the trained GAIN network, so that the generated global nmF2 and hmF2 maps are obtained; (5) reading of the nmF2 and hmF2 maps with a preset time resolution generated in (4), regarding each point as an independent object, inputting the two parameter data of nmF2 and hmF2 obtained at each point into the trained Profile_Net, so that the electron density profile data of the point are obtained, and the spatial three-dimensional electron density distribution with a preset time resolution is obtained by integrating the electron density profile data of all points; The nmF2 and hmF2 maps output by GAIN are combined with Profile_net, so that a global four-dimensional ionospheric electron density model is established, and the ionospheric electron density variation of the observation data is restored. 2.The method of claim 1, wherein, A neural network GAIN for completing the incomplete global nmF2 and hmF2 maps provided by the occultation data is established, and the GAIN is a neural network model established based on a generative adversarial interpolation network; The input of the generator of GAIN is the incomplete global map data of nmF2 and hmF2 with a preset time resolution provided by the occultation data after normalization to the global grid, the noise z and the calibration matrix M, wherein the calibration matrix M indicates the grid point conditions of the incomplete global map data with data and without data; The output of the generator is the complete global nmF2 and hmF2 maps, and the trained generator realizes input of the incomplete global nmF2 and hmF2 map data, the noise z and the calibration matrix M, and output of the complete global nmF2 and hmF2 maps. 3.The method of claim 1, wherein, In the training process, the nmF2 and hmF2 reference data provided by the ionospheric reference model are normalized to the grid with the same accuracy, and are used to constrain the normalized results of the nmF2 and hmF2 maps generated by the generator, and the incomplete nmF2 and hmF2 maps input by the generator are also used for constraint.

4. The method of claim 1-3, wherein, The process of quality control of the acquired occultation data comprises elimination of abnormal peak values, height control, Chapman fitting and interpolation smoothing; and the data normalization comprises normalization of the data after quality control to a standard grid with an accuracy of every 3 hours in universal time, every 12° in longitude and every 6° in latitude.

Images