Ionized layer TEC prediction method in near space

Abstract

The invention relates to an ionosphere TEC prediction method in a near space, and belongs to the technical field of ionospheres. The method comprises the following steps: acquiring ionosphere environment data at three consecutive moments; wherein the ionosphere environment data comprises ionosphere TEC data and earth magnetic field data; inputting the ionosphere environment data at the three moments into the trained ionosphere TEC prediction model to obtain ionosphere TEC data at the next moment; wherein the ionosphere TEC prediction model comprises an encoder, a condition generator and a decoder; the encoder is used for encoding the ionized layer TEC data at the current moment to obtain potential representation of the ionized layer TEC data; the condition generator is used for performing inversion coding on the ionosphere environment data at the three moments to obtain ionosphere electron density information; and the decoder is used for reconstructing the ionospheric TEC data at the next moment from the potential representation of the ionospheric TEC data at the current moment by using the ionospheric electron density information. The state of the ionosphere is predicted through inversion of the electron density information of the ionosphere.

Description

Technical Field

[0001] The present invention relates to the technical field of ionosphere, and in particular to an ionosphere TEC prediction method in near space. Background Art

[0002] The ionosphere is an important component of the Earth's near-space. The total electric contents (TEC) of the ionosphere is one of the important physical parameters that characterize the state of the ionosphere. Accurate modeling and prediction of TEC will help us further understand the ionosphere in scientific research and engineering activities, so as to better avoid ionospheric hazards and utilize the value of the ionosphere.

[0003] Ionospheric electron density is the basis for calculating ionospheric TEC. Real-time and accurate acquisition of ionospheric electron density data can provide the latest ionospheric state information, which helps to improve the accuracy of ionospheric TEC prediction. However, the ionosphere is a highly dynamic and complex environment, and its electron density changes rapidly with time, geographic location, solar activity, geomagnetic activity and other factors. This complexity makes real-time monitoring and prediction of electron density a challenge. Existing methods for observing ionospheric electron density usually rely on ground-based radio observatories, satellite-borne instruments or GNSS signals. These observation methods may be limited by geographic location, instrument accuracy, signal coverage and data update frequency. Summary of the invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide an ionospheric TEC prediction method in near space, so as to solve the problem that in the existing ionospheric prediction process, the change of ionospheric electron density cannot be accurately obtained in real time, and it is difficult to predict the ionospheric state by directly observing data.

[0005] The purpose of the present invention is mainly achieved through the following technical solutions:

[0006] The present invention provides an ionospheric TEC prediction method in near space, comprising the following steps:

[0007] Acquire ionospheric environment data at the current moment and two moments before the current moment; wherein the ionospheric environment data includes: ionospheric TEC data and earth magnetic field data;

[0008] The ionospheric environmental data at three moments are input into a trained ionospheric TEC prediction model to obtain the ionospheric TEC data at the next moment; wherein the ionospheric TEC prediction model includes an encoder, a condition generator and a decoder; the encoder is used to encode the ionospheric TEC data at the current moment to obtain a potential representation of the ionospheric TEC data; the condition generator is used to inversely encode the ionospheric environmental data at the current moment and at two moments before the current moment to obtain ionospheric electron density information; the decoder is used to reconstruct the ionospheric TEC data at the next moment from the potential representation of the ionospheric TEC data at the current moment using the ionospheric electron density information.

[0009] Further, encoding the ionospheric TEC data at the current moment to obtain a potential representation of the ionospheric TEC data includes:

[0010] The ionospheric TEC data at the current moment is extracted from the local features of the ionospheric TEC data and the spatial dimension of the features is reduced through the initial feature extraction module of the encoder to obtain an initial feature map; wherein the initial feature extraction module includes a convolution layer and a maximum pooling layer;

[0011] The initial feature map is sequentially passed through four encoding modules of the encoder with the same structure to map the data to a latent space, thereby obtaining a latent representation of the ionospheric TEC data; wherein the encoding module includes four convolution modules, each of which includes a convolution layer, a batch processing layer, and an activation layer.

[0012] Furthermore, in a coding module:

[0013] The input feature map of the current encoding module is sequentially extracted through two convolution modules, and the output feature map is fused with the input feature map of the current encoding module in the channel dimension to obtain a fused feature map;

[0014] The fused feature map is sequentially passed through two convolution modules for feature extraction again, and the obtained output feature map is fused with the fused feature map in the channel dimension to obtain the output feature map of the current encoding module.

[0015] Further, the inversion encoding of the ionospheric environment data at the current moment and two moments before the current moment to obtain the ionospheric electron density information includes:

[0016] The position embedding layer of the ionospheric environment data at the current moment and two moments before the current moment in the condition generator is added bit by bit to the position weight matrix to obtain ionospheric environment feature encoding data; wherein the position weight matrix is ​​calculated by the position vector in the ionospheric environment data through sine and cosine functions;

[0017] The ionospheric environment feature coded data is sequentially scale-segmented and fused and outputted through two multi-scale attention modules with the same structure of the condition generator to obtain the ionospheric electron density information.

[0018] Furthermore, in a multi-scale attention module:

[0019] The input data of the current multi-scale attention module is split into two sets of half-scale feature transformation data through the multi-scale separation layer of the multi-scale attention module;

[0020] The first group of the half-scale feature transformation data is self-attentionally calculated in the first head of the attention layer of the multi-scale attention module to obtain a first calculation result;

[0021] The first calculation result and the second group of the half-scale feature transformation data are bitwise added in the second head of the attention layer of the multi-scale attention module, and then self-attention calculation is performed to obtain a second calculation result;

[0022] The first calculation result and the second calculation result are spliced ​​in a splicing mapping layer of the attention layer of the multi-scale attention module to obtain output data of the multi-scale attention module.

[0023] Further, the reconstructing the ionospheric TEC data at the next moment from the potential representation of the ionospheric TEC data at the current moment by using the ionospheric electron density information comprises:

[0024] The potential representation of the ionospheric TEC data is spliced ​​with the ionospheric electron density information to obtain ionospheric enhancement characteristics;

[0025] The ionospheric enhancement feature is sequentially passed through four decoding modules of the same structure of the decoder of the ionospheric TEC prediction model to gradually reconstruct the detailed information of the feature and obtain a decoding module output feature map; wherein the decoding module includes two convolution modules and one deconvolution module; each convolution module includes a convolution layer, a batch processing layer and an activation layer; the deconvolution module includes a deconvolution layer and a batch processing layer;

[0026] The decoding module outputs a feature map, which is passed through the feature reconstruction module of the decoder to obtain the ionospheric TEC data at the next moment; wherein the feature reconstruction module includes a first full convolution layer, a convolution layer, and a second full convolution layer.

[0027] Furthermore, in a decoding module:

[0028] The input feature map of the current decoding module is sequentially passed through a convolution module, a deconvolution module, and a convolution module for upsampling, so that the input feature map of the current decoding module is restored to a high-resolution feature map;

[0029] The high-resolution feature map is fused with the input feature map of the current decoding module in the channel dimension to obtain the output feature map of the current decoding module.

[0030] Furthermore, the Earth's magnetic field data include: northward magnetic field component, planetary magnetic activity index, disturbance storm time and solar radiation index.

[0031] Furthermore, the ionospheric TEC prediction model is trained by the following method:

[0032] Constructing a training data set for the ionospheric TEC prediction model; wherein the training data set includes a number of ionospheric environmental data at three consecutive moments as sample data and the corresponding ionospheric TEC data at the next moment as labels;

[0033] Preliminarily training an ionospheric TEC prediction model based on the training data set, updating parameters of the ionospheric TEC prediction model, and obtaining a preliminarily trained ionospheric TEC prediction model;

[0034] Training a discriminator based on the training set and the preliminarily trained ionospheric TEC prediction model, updating discriminator parameters, and obtaining a preliminarily trained discriminator;

[0035] Based on the training set, the preliminarily trained ionospheric TEC prediction model and the preliminarily trained discriminator are trained simultaneously to obtain a trained ionospheric TEC prediction model.

[0036] Further, the preliminary training of the ionospheric TEC prediction model includes: loading the training data set into the ionospheric TEC prediction model, training the ionospheric TEC prediction model using a reconstruction loss function and a KL divergence loss function, and obtaining a preliminary trained ionospheric TEC prediction model;

[0037] The preliminary training discriminator includes: loading the training set into the preliminary trained ionospheric TEC prediction model to generate predicted ionospheric TEC data at the next moment as a false sample input to the discriminator; loading the corresponding ionospheric TEC data at the next moment in the training data set as a true sample input to the discriminator; training the discriminator using a binary cross entropy loss function to obtain a preliminary trained discriminator;

[0038] The method of simultaneously training the preliminarily trained ionospheric TEC prediction model and the preliminarily trained discriminator comprises: loading the training set into the preliminarily trained ionospheric TEC prediction model to generate predicted ionospheric TEC data at the next moment as a false sample input to the discriminator; loading the corresponding ionospheric TEC data at the next moment in the training data set as a true sample input to the discriminator; and using a reconstruction loss function, a KL divergence loss function, and a binary cross entropy loss function to simultaneously train the ionospheric TEC prediction model and the discriminator to obtain a trained ionospheric TEC prediction model.

[0039] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0040] 1. The technical solution of the present invention uses the encoder model to predict real-time ionospheric state data. The electron density information obtained by inversion and the prediction results of the encoder can reconstruct the entire ionospheric state and provide reliable ionospheric state prediction for satellite communications, navigation systems and other space activities.

[0041] 2. The technical solution of the present invention inversely encodes the electron density information through a condition generator, accurately extracts and represents the ionospheric electron density information in the input data, and can predict the state of the ionosphere based on the sampled real-time data, so that the model can monitor and predict the changes in the ionosphere in real time and discover potential ionospheric disturbances in advance.

[0042] 3. The ionospheric TEC prediction model proposed in the technical solution of the present invention can learn complex nonlinear relationships by using the conditional variational autoencoder (CVAE) and generative adversarial network (GAN) in the deep learning model, so as to better simulate the dynamic changes of the ionosphere.

[0043] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0045] Figure 1 It is a schematic flow chart of a method for predicting ionospheric TEC in near space in an embodiment of the present invention;

[0046] Figure 2Schematic diagram of the structure of the ionospheric TEC prediction model in an embodiment of the present invention;

[0047] Figure 3 Schematic diagram of the structure of the encoder of the ionospheric TEC prediction model in an embodiment of the present invention;

[0048] Figure 4 Schematic diagram of the structure of the condition generator of the ionospheric TEC prediction model in an embodiment of the present invention;

[0049] Figure 5 Schematic diagram of the structure of the decoder of the ionospheric TEC prediction model in an embodiment of the present invention;

[0050] Figure 6 Schematic diagram of the training process of the ionospheric TEC prediction model in an embodiment of the present invention;

[0051] Figure 7 Schematic diagram of the structure of the discriminator in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0053] A specific embodiment of the present invention discloses a method for predicting ionospheric TEC in near space, such as Figure 1 As shown, the steps S1 to S2 are included:

[0054] Step S1, obtaining ionospheric environment data at a current moment and two moments before the current moment; wherein the ionospheric environment data includes: ionospheric TEC data and earth magnetic field data.

[0055] Specifically, the data from 16 global navigation satellite system (GNSS) observation stations provided by the China Crustal Movement Observation Network (CMONOC) are used to obtain ionospheric environmental data at several times of the day. Among them, the Earth's magnetic field data include: northward magnetic field component, planetary magnetic activity index, disturbance storm time and solar radiation index; the ionospheric TEC data is the total electron content of the ionosphere, which refers to the sum of the electronic components of the ionosphere within a certain volume or mass unit.

[0056] Among them, the northward magnetic field component (Bz) is a component of the interplanetary magnetic field, and its influence on the ionosphere is particularly significant. When Bz is positive, that is, when the interplanetary magnetic field is northward, it can penetrate into the Earth's magnetosphere, causing the plasma in the ionosphere to sink, thus affecting the distribution and total amount of TEC.

[0057] The planetary magnetic activity index (Kp) is a parameter of the level of geomagnetic disturbances, which is related to the interaction between the solar wind and the Earth's magnetic field. Disturbances in the Earth's magnetic field can lead to significant changes in the ionosphere.

[0058] The disturbance storm time (Dst) is a parameter of the global geomagnetic disturbance, which is related to the violent disturbance of the ionosphere. During the magnetic storm, the ionospheric TEC will change significantly.

[0059] The solar radiation index (F10.7 index) is an indicator to measure the intensity of solar ultraviolet and X-ray radiation. It is related to the intensity of solar activity. Changes in solar radiation directly affect the ionization rate of the ionosphere, thereby affecting the ionospheric TEC data.

[0060] Step S2, inputting the ionospheric environmental data at three moments into the trained ionospheric TEC prediction model to obtain the ionospheric TEC data at the next moment; wherein the ionospheric TEC prediction model includes an encoder, a condition generator and a decoder; the encoder is used to encode the ionospheric TEC data at the current moment to obtain a potential representation of the ionospheric TEC data; the condition generator is used to inversely encode the ionospheric environmental data at the current moment and at two moments before the current moment to obtain ionospheric electron density information; the decoder is used to reconstruct the ionospheric TEC data at the next moment from the potential representation of the ionospheric TEC data at the current moment using the ionospheric electron density information.

[0061] Specifically, the ionospheric TEC prediction model of this embodiment is as follows: Figure 2 As shown, the obtained ionospheric environment data is used to invert the electron density information of the ionosphere, and the inverted electron density information is used as conditional input. A deep learning model based on conditional variational autoencoder (CVAE) is used to reconstruct the global state of the ionosphere. At the same time, the adversarial idea of ​​generative adversarial network (GAN) can be combined to generate higher quality and more realistic data.

[0062] Further, such as Figure 3 As shown, the encoder includes an initial feature extraction module and four encoding modules with the same structure connected in sequence, which are used to encode the ionospheric TEC data at the current moment to obtain a potential representation of the ionospheric TEC data.

[0063] Furthermore, the ionospheric TEC data at the current moment is subjected to an initial feature extraction module of the encoder to extract local features of the ionospheric TEC data and reduce the spatial dimension of the features to obtain an initial feature map; wherein the initial feature extraction module includes a convolution layer and a maximum pooling layer.

[0064] Specifically, the ionospheric TEC data at the current moment is convolved through a convolution layer with a convolution kernel of 5×5 and a step size of 2 in the initial feature extraction module, and then passes through a maximum pooling layer with a pooling window of 3×3 and a step size of 2 to reduce the resolution of the input feature map, reduce the amount of calculation, and improve the invariance of the features to obtain the initial feature map.

[0065] Furthermore, the initial feature map is sequentially passed through four encoding modules of the encoder with the same structure to map the data to a latent space, thereby obtaining a latent representation of the ionospheric TEC data; wherein the encoding module includes four convolution modules, each of which includes a convolution layer, a batch processing layer, and an activation layer.

[0066] Specifically, using four encoding modules in the encoder provides sufficient depth to extract deep features of the data. As the number of network layers increases, the network can extract more abstract and semantic features, thereby better capturing the information in the data. However, the more encoding modules, the better. Too many encoding modules will increase the complexity of the model, resulting in increased computation and longer training time.

[0067] Furthermore, in a coding module:

[0068] The input feature map of the current encoding module is sequentially extracted through two convolution modules, and the output feature map is fused with the input feature map of the current encoding module in the channel dimension to obtain a fused feature map;

[0069] The fused feature map is sequentially passed through two convolution modules for feature extraction again, and the obtained output feature map is fused with the fused feature map in the channel dimension to obtain the output feature map of the current encoding module.

[0070] Specifically, the input feature map of the current encoding module passes through a convolution layer with a convolution kernel of 3×3 and a step size of 2, a batch layer, and an activation layer in the first convolution module to extract the detailed features of the input feature map and perform downsampling, thereby reducing the spatial dimension of the feature map, reducing the amount of calculation, and increasing the receptive field of the model, so that the model can capture a wider range of features.

[0071] The output feature map of the first convolution module passes through a convolution layer with a convolution kernel of 3×3 and a stride of 1, a batch layer, and an activation layer in the second convolution module to further extract higher-level features while maintaining the spatial resolution of the features.

[0072] The output feature map of the second convolution module is fused with the input feature map of the current encoding module in the channel dimension to retain the original information directly transmitted from the input feature map and enhance the information content of the feature map by combining features at different levels.

[0073] The fused feature map passes through the third convolution module and a convolution layer with a convolution kernel of 3×3 and a step size of 1 in the fourth convolution module, a batch processing layer and an activation layer in turn to further extract the fused features.

[0074] The output feature map of the fourth convolution module is fused with the fused feature map in the channel dimension, and the features at different levels are combined again to enhance the information content of the feature map to obtain the output feature map of the current encoding module.

[0075] Further, such as Figure 4 As shown, the condition generator is an improved Transformer model, including a position embedding layer and two multi-scale attention modules with the same structure, which are used to inversely encode the ionospheric environment data at the current moment and two moments before the current moment to obtain ionospheric electron density information.

[0076] Furthermore, the position embedding layer of the ionospheric environment data at the current moment and two moments before the current moment in the condition generator is added bit by bit to the position weight matrix to obtain ionospheric environment feature encoding data; wherein the position weight matrix is ​​calculated by the position vector in the ionospheric environment data through sine and cosine functions.

[0077] Specifically, the position embedding layer is used to provide the position information of each element in the sequence. In the Transformer model, the self-attention mechanism is used to process the input sequence, and the self-attention mechanism itself does not consider the order of elements in the sequence. Position embedding ensures that the model can distinguish elements at different positions and capture patterns and dependencies in the sequence data.

[0078] It should be noted that the position weight matrix of the Transformer model The calculation process is as follows:

[0079]

[0080]

[0081] Among them, PE represents the position vector; pos represents the position of the posth time point in the time series, starting from 0; d model represents the dimension of the ionospheric environment data; i represents the i-th vector dimension, i∈[0,d model / 2].

[0082] Furthermore, the ionospheric environment feature encoded data is scale-segmented and then fused and outputted in turn through two multi-scale attention modules of the same structure of the condition generator to obtain the ionospheric electron density information.

[0083] Specifically, the multi-scale attention module includes a multi-scale separation layer and an attention layer, which divides the input feature map into feature maps of different scales, so that the model can capture multi-level information from details to the global level. By applying the attention mechanism at different scales, the model can pay more attention to the most important features for the current task, while suppressing less important information, which helps to improve the quality of feature representation.

[0084] Furthermore, in a multi-scale attention module:

[0085] The input data of the current multi-scale attention module is split into two sets of half-scale feature transformation data through the multi-scale separation layer of the multi-scale attention module.

[0086] Specifically, the multi-scale separation layer includes a combination of a convolutional layer, an activation layer, and a pooling layer, which divides the input feature map into transformed data of different scales and generates feature representations with different receptive fields, which can help the model extract features at different levels and realize nonlinear transformation of data.

[0087] The attention layer includes a two-head attention module and a concatenated mapping layer. Through the attention mechanism, the model can pay more attention to the most important features in the current task.

[0088] Further, the first group of the half-scale feature transformation data is subjected to self-attention calculation in a first head of the attention layer of the multi-scale attention module to obtain a first calculation result;

[0089] The first calculation result and the second group of the half-scale feature transformation data are bitwise added in the second head of the attention layer of the multi-scale attention module, and then self-attention calculation is performed to obtain a second calculation result;

[0090] The first calculation result and the second calculation result are spliced ​​in a splicing mapping layer of the attention layer of the multi-scale attention module to obtain output data of the multi-scale attention module.

[0091] Exemplarily, first, assume that the size of the input data is (3, 4096), where 3 indicates that there are data at three time points in the input data, and 4096 indicates the data dimension at each time point.

[0092] In the multi-scale segmentation layer, two calculations are performed on the data simultaneously to obtain data of sizes (3, 2048) respectively, and the calculation results are represented by a1 and a2.

[0093] The calculation results a1 and a2 are placed in the attention layer according to Figure 4 The calculation process shown is used for calculation. The result b1 is obtained by self-attention calculation of a1, and the size is (3, 2048); a1 and b1 are added according to the same element position, and b2 is obtained by self-attention calculation, and the size is (3, 2048); b1 and b2 are spliced ​​in the splicing mapping layer to obtain the final output data with a size of (3, 4096).

[0094] It should be noted that, through the cascading method, the output of the attention module of the previous head is fused with the input of the attention module of the next head, which realizes the interaction and information transmission between data and improves the ability of feature extraction.

[0095] Furthermore, the potential representation of the ionospheric TEC data is spliced ​​with the ionospheric electron density information to obtain ionospheric enhancement features.

[0096] Specifically, the potential representation of the ionospheric TEC data is spliced ​​with the ionospheric electron density information to fuse different types of data, so that the model can simultaneously consider the measured value of ionospheric TEC and the physical meaning of electron density, thereby improving the accuracy of the prediction.

[0097] Further, such as Figure 5 As shown, the decoder includes four decoding modules and feature reconstruction modules with the same structure connected in sequence, which are used to reconstruct the ionospheric TEC data at the next moment from the potential representation of the ionospheric TEC data at the current moment by using the ionospheric electron density information.

[0098] Furthermore, the ionospheric enhancement feature is sequentially passed through four decoding modules with the same structure of the decoder of the ionospheric TEC prediction model to gradually reconstruct the detailed information of the feature and obtain a decoding module output feature map; wherein the decoding module includes two convolution modules and one deconvolution module; each convolution module includes a convolution layer, a batch processing layer and an activation layer; the deconvolution module includes a deconvolution layer and a batch processing layer.

[0099] Specifically, using four decoding modules in the decoder can gradually refine the representation of the latent space into a feature representation that is closer to the original input data, thereby achieving high-quality prediction.

[0100] Furthermore, in a decoding module:

[0101] The input feature map of the current decoding module is sequentially passed through a convolution module, a deconvolution module, and a convolution module for upsampling, so that the input feature map of the current decoding module is restored to a high-resolution feature map;

[0102] The high-resolution feature map is fused with the input feature map of the current decoding module in the channel dimension to obtain the output feature map of the current decoding module.

[0103] Specifically, the input feature map of the current decoding module is extracted from the local features of the input feature map in the first convolution module through a convolution layer with a convolution kernel of 3×3 and a stride of 1, a batch layer, and an activation layer, and these features are converted into a new feature representation.

[0104] The output feature map of the first convolution module is restored to a high-resolution feature map through a deconvolution layer with a convolution kernel of 3×3 and a stride of 1 and a batch layer in the deconvolution module.

[0105] The output feature map of the deconvolution module is passed through a deconvolution layer with a convolution kernel of 3×3 and a stride of 1, a batch layer, and an activation layer in the second convolution module to further refine and enhance the features of the local area of ​​the feature map after the size increase.

[0106] The output feature map of the second convolution module is fused with the input feature map of the current decoding module in the channel dimension. This jump connection can help the network better restore the detailed information of the image and avoid the problem of semantic holes.

[0107] Furthermore, the decoding module outputs the feature map and passes it through the feature reconstruction module of the decoder to obtain the ionospheric TEC data at the next moment; wherein the feature reconstruction module includes a first full convolution layer, a convolution layer and a second full convolution layer.

[0108] Specifically, the feature map output by the decoding module passes through a full convolution layer with a convolution kernel of 3×3 and a step size of 2 of the feature reconstruction module to increase the size of the feature map and achieve upsampling; the feature map output by the full convolution layer passes through a convolution layer with a convolution kernel of 3×3 and a step size of 1 to extract local features, and the feature map output by the full convolution layer is mapped to a new feature space to extract higher-level feature representations; the output feature map of the convolution layer passes through a full convolution layer with a convolution kernel of 2×2 and a step size of 2 again to further fuse features, and the number of channels of the feature map is adjusted to reconstruct the ionospheric TEC data at the next moment.

[0109] Furthermore, the ionospheric TEC prediction model is as follows: Figure 6 The method shown is trained to obtain:

[0110] Specifically, in this embodiment, the ionospheric TEC prediction model is trained by a three-stage training method, which is respectively the training of the ionospheric TEC prediction model in the first stage, that is, the encoder, condition generator and decoder structure of the model are used to encode and decode the training data to reconstruct and generate ionospheric state data. The adversarial training of the discriminator in the second stage improves the authenticity of the generated data through the discriminator, that is, the discriminator learns how to accurately distinguish between real data and generated data by comparing the ionospheric TEC data generated by the ionospheric TEC prediction model with the actual observed data, and evaluates the authenticity of the generated data. The discriminator calculates the adversarial loss by comparing the real data and the generated data to train the ionospheric TEC prediction model to generate more authentic data. The third stage is the joint training of the ionospheric TEC prediction model and the discriminator. During the joint training process, the loss of the ionospheric TEC prediction model and the loss of the discriminator are optimized at the same time. The training of the ionospheric TEC prediction model ensures the rationality of the generated samples in the latent space and the structure of the data, and the training of the discriminator further improves the authenticity of the generated results.

[0111] Step S101, constructing a training data set for the ionospheric TEC prediction model; wherein the training data set includes ionospheric environment data of a number of consecutive three moments as sample data and corresponding ionospheric TEC data of the next moment as labels.

[0112] Specifically, data from 16 global navigation satellite system (GNSS) observation stations provided by the China Crustal Movement Observation Network (CMONOC) are used to obtain ionospheric environment data at three consecutive moments as sample input data, and ionospheric TEC data at the next moment of the three moments are used as labels for training data.

[0113] Step S102: preliminarily training an ionospheric TEC prediction model based on the training data set, updating parameters of the ionospheric TEC prediction model, and obtaining a preliminarily trained ionospheric TEC prediction model.

[0114] Furthermore, the first stage of training is to train the ionospheric TEC prediction model, including:

[0115] Loading the ionospheric environment data of three consecutive moments in the training set into the condition generator of the preliminary training ionospheric TEC prediction model to generate ionospheric electron density information;

[0116] Loading the ionospheric TEC data at the latest moment among the ionospheric environmental data at three consecutive moments in the training set into the encoder of the preliminary training ionospheric TEC prediction model, and generating the predicted ionospheric TEC data at the next moment through the decoder of the preliminary training ionospheric TEC prediction model together with the ionospheric electron density information;

[0117] Loading the corresponding ionospheric TEC data of the next moment in the training set as a label;

[0118] The ionospheric TEC prediction model is trained using a reconstruction loss function and a KL divergence loss function, and the parameters of the ionospheric TEC prediction model are updated by back propagation using a stochastic gradient descent method to obtain a preliminarily trained ionospheric TEC prediction model.

[0119] Specifically, the reconstruction loss function is a function that measures the difference between the data reconstructed by the model and the original data, including but not limited to the mean square error loss function (MSE), the cross-entropy loss function (Cross-Entropy Loss) and the mean absolute error loss function (MAE).

[0120] The KL divergence loss function is a function that measures the difference between two probability distributions. It is used to measure the difference between the probability distribution predicted by the model and the probability distribution of the real data, and is usually used in the deep learning model of the conditional variational autoencoder (CVAE).

[0121] Based on the above, the training loss function of the ionospheric TEC prediction model is as follows:

[0122] Loss TEC =Loss Reconstruction +Loss KL

[0123] Among them, Loss TEC Represents the training loss function of the ionospheric TEC prediction model; Loss Reconstruction Represents the reconstruction loss function; Loss KL represents the KL divergence loss function.

[0124] It should be noted that since the generator often performs poorly in the initial stage, the generated prediction structure is easily identified as a false sample by the discriminator, so the loss of the discriminator is very small and close to 0, and the discriminator cannot play its role. In order to solve this problem, in the first stage of training, the generator is first trained, and the ionospheric TEC prediction model is independently trained for multiple iterations to gradually improve its generation ability. In this way, when entering the second stage of training, the ionospheric TEC prediction model already has a certain generation ability.

[0125] Step S103: training a discriminator based on the training set and the preliminarily trained ionospheric TEC prediction model, updating discriminator parameters, and obtaining a preliminarily trained discriminator.

[0126] Furthermore, the second stage of training is to train the discriminator, including:

[0127] Loading the ionospheric environment data of three consecutive moments in the training set to the preliminarily trained ionospheric TEC prediction model to generate predicted ionospheric TEC data of the next moment as a false sample input discriminator;

[0128] Loading the corresponding ionospheric TEC data of the next moment in the training data set as a true sample input to the discriminator;

[0129] The parameters of the initially trained ionospheric TEC prediction model are fixed, and the discriminator is trained using the binary cross entropy loss function as the loss function. The discriminator parameters are updated by minimizing the binary cross entropy loss so that the discriminator can accurately divide the input data into true samples or false samples. After the training, the discriminator parameters are saved.

[0130] Specifically, Figure 7 As shown, the discriminator includes a fully convolutional neural network and a fully connected network, which is used to discriminate whether the ionospheric TEC data at the next moment generated by the ionospheric TEC prediction model is a true sample or a false sample.

[0131] The real ionospheric TEC data at the next moment and the ionospheric TEC data at the next moment generated by the ionospheric TEC prediction model are converted into feature vectors in the fully convolutional neural network layer of the discriminator through multiple convolution operations; the feature vector is obtained through the fully connected layer to determine whether the generated ionospheric TEC data at the next moment belongs to a true sample or a false sample.

[0132] It should be noted that in the second stage of training, the discriminator is trained separately, and the discriminator compares the ionospheric TEC data at the next moment generated by the ionospheric TEC prediction model with the real ionospheric TEC data at the next moment, and gradually learns how to accurately distinguish between real samples and generated samples. The training process of the discriminator helps to improve the discriminant ability of the discriminator, so that it can better evaluate the authenticity of the ionospheric TEC data at the next moment generated by the ionospheric TEC prediction model.

[0133] Step S104: training the preliminarily trained ionospheric TEC prediction model and the preliminarily trained discriminator simultaneously based on the training set to obtain a trained ionospheric TEC prediction model.

[0134] Furthermore, the third stage of training is to simultaneously train the preliminarily trained ionospheric TEC prediction model and the preliminarily trained discriminator, including:

[0135] Loading the ionospheric environment data of three consecutive moments in the training set to the preliminarily trained ionospheric TEC prediction model to generate predicted ionospheric TEC data of the next moment as a false sample input discriminator;

[0136] Loading the corresponding ionospheric TEC data of the next moment in the training data set as a true sample input discriminator;

[0137] The ionospheric TEC prediction model and the discriminator are trained simultaneously using a reconstruction loss function, a KL divergence loss function and a binary cross entropy loss function, and the parameters of the ionospheric TEC prediction model and the parameters of the discriminator are updated to obtain a trained ionospheric TEC prediction model.

[0138] It should be noted that in the third stage of training, through the adversarial training between the ionospheric TEC prediction model and the discriminator, the discriminator continuously provides feedback on the output of the ionospheric TEC prediction model to help the ionospheric TEC prediction model generate more realistic ionospheric TEC data. This game process prompts the ionospheric TEC prediction model to gradually improve its generation ability, making the generated ionospheric TEC data at the next moment closer to the real ionospheric TEC data, thereby further improving the performance of the ionospheric TEC prediction model.

[0139] In summary, the ionospheric TEC prediction method in near space according to an embodiment of the present invention has the following beneficial effects:

[0140] 1. The technical solution of the present invention uses the encoder model to predict real-time ionospheric state data. The electron density information obtained by inversion and the prediction results of the encoder can reconstruct the entire ionospheric state and provide reliable ionospheric state prediction for satellite communications, navigation systems and other space activities.

[0141] 2. The technical solution of the present invention inversely encodes the electron density information through a condition generator, accurately extracts and represents the ionospheric electron density information in the input data, and can predict the state of the ionosphere based on the sampled real-time data, so that the model can monitor and predict the changes in the ionosphere in real time and discover potential ionospheric disturbances in advance.

[0142] 3. The ionospheric TEC prediction model proposed in the technical solution of the present invention can learn complex nonlinear relationships by using the conditional variational autoencoder (CVAE) and generative adversarial network (GAN) in the deep learning model, so as to better simulate the dynamic changes of the ionosphere.

[0143] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for predicting ionospheric TEC in near space, characterized in that: The steps include: Acquire ionospheric environment data at the current moment and two moments before the current moment; wherein the ionospheric environment data includes: ionospheric TEC data and earth magnetic field data; The ionospheric environmental data at three moments are input into a trained ionospheric TEC prediction model to obtain the ionospheric TEC data at the next moment; wherein the ionospheric TEC prediction model includes an encoder, a condition generator and a decoder; the encoder is used to encode the ionospheric TEC data at the current moment to obtain a potential representation of the ionospheric TEC data; the condition generator is used to inversely encode the ionospheric environmental data at the current moment and two moments before the current moment to obtain ionospheric electron density information; the decoder is used to reconstruct the ionospheric TEC data at the next moment from the potential representation of the ionospheric TEC data at the current moment using the ionospheric electron density information.

2. The method according to claim 1, characterized in that: The step of encoding the ionospheric TEC data at the current moment to obtain a potential representation of the ionospheric TEC data comprises: The ionospheric TEC data at the current moment is extracted from the local features of the ionospheric TEC data and the spatial dimension of the features is reduced through the initial feature extraction module of the encoder to obtain an initial feature map; wherein the initial feature extraction module includes a convolution layer and a maximum pooling layer; The initial feature map is sequentially passed through four encoding modules of the encoder with the same structure to map the data to a latent space, thereby obtaining a latent representation of the ionospheric TEC data; wherein the encoding module includes four convolution modules, each of which includes a convolution layer, a batch processing layer, and an activation layer.

3. The method according to claim 2, characterized in that: In a coding module: The input feature map of the current encoding module is sequentially extracted through two convolution modules, and the output feature map is fused with the input feature map of the current encoding module in the channel dimension to obtain a fused feature map; The fused feature map is sequentially passed through two convolution modules for feature extraction again, and the obtained output feature map is fused with the fused feature map in the channel dimension to obtain the output feature map of the current encoding module.

4. The method according to claim 1, characterized in that: The inversion encoding of the ionospheric environment data at the current moment and two moments before the current moment to obtain the ionospheric electron density information includes: The position embedding layer of the ionospheric environment data at the current moment and two moments before the current moment in the condition generator is added bit by bit to the position weight matrix to obtain ionospheric environment feature encoding data; wherein the position weight matrix is ​​calculated by the position vector in the ionospheric environment data through sine and cosine functions; The ionospheric environment feature coded data is sequentially scale-segmented and fused and outputted through two multi-scale attention modules with the same structure of the condition generator to obtain the ionospheric electron density information.

5. The method according to claim 4, characterized in that: In a multi-scale attention module: The input data of the current multi-scale attention module is split into two sets of half-scale feature transformation data through the multi-scale separation layer of the multi-scale attention module; The first group of the half-scale feature transformation data is self-attention calculated in the first head of the attention layer of the multi-scale attention module to obtain a first calculation result; The first calculation result and the second group of the half-scale feature transformation data are bitwise added in the second head of the attention layer of the multi-scale attention module, and then self-attention calculation is performed to obtain a second calculation result; The first calculation result and the second calculation result are spliced ​​in a splicing mapping layer of the attention layer of the multi-scale attention module to obtain output data of the multi-scale attention module.

6. The method according to claim 1, characterized in that: The method of reconstructing the ionospheric TEC data at the next moment from the potential representation of the ionospheric TEC data at the current moment by using the ionospheric electron density information comprises: The latent representation of the ionospheric TEC data is spliced ​​with the ionospheric electron density information to obtain ionospheric enhancement features; The ionospheric enhancement feature is sequentially passed through four decoding modules of the same structure of the decoder of the ionospheric TEC prediction model to gradually reconstruct the detailed information of the feature and obtain a decoding module output feature map; wherein the decoding module includes two convolution modules and one deconvolution module; each convolution module includes a convolution layer, a batch processing layer and an activation layer; the deconvolution module includes a deconvolution layer and a batch processing layer; The decoding module outputs a feature map, which is passed through the feature reconstruction module of the decoder to obtain the ionospheric TEC data at the next moment; wherein the feature reconstruction module includes a first full convolution layer, a convolution layer, and a second full convolution layer.

7. The method according to claim 6, characterized in that: In a decode module: The input feature map of the current decoding module is sequentially passed through a convolution module, a deconvolution module, and a convolution module for upsampling, so that the input feature map of the current decoding module is restored to a high-resolution feature map; The high-resolution feature map is fused with the input feature map of the current decoding module in the channel dimension to obtain the output feature map of the current decoding module.

8. The method according to any one of claims 1 to 7, characterized in that: The earth magnetic field data include: northward magnetic field component, planetary magnetic activity index, disturbance storm time and solar radiation index.

9. The method according to claim 1, characterized in that: The ionospheric TEC prediction model is trained by the following method: Constructing a training data set for the ionospheric TEC prediction model; wherein the training data set includes a number of ionospheric environmental data at three consecutive moments as sample data and the corresponding ionospheric TEC data at the next moment as labels; Preliminarily training an ionospheric TEC prediction model based on the training data set, updating parameters of the ionospheric TEC prediction model, and obtaining a preliminarily trained ionospheric TEC prediction model; Training a discriminator based on the training set and the preliminarily trained ionospheric TEC prediction model, updating discriminator parameters, and obtaining a preliminarily trained discriminator; Based on the training set, the preliminarily trained ionospheric TEC prediction model and the preliminarily trained discriminator are trained simultaneously to obtain a trained ionospheric TEC prediction model.

10. The method according to claim 9, characterized in that: The preliminary training of the ionospheric TEC prediction model comprises: loading the training data set into the ionospheric TEC prediction model, training the ionospheric TEC prediction model using a reconstruction loss function and a KL divergence loss function, and obtaining a preliminary trained ionospheric TEC prediction model; The preliminary training discriminator includes: loading the training set into the preliminary trained ionospheric TEC prediction model to generate predicted ionospheric TEC data at the next moment as a false sample input to the discriminator; loading the corresponding ionospheric TEC data at the next moment in the training data set as a true sample input to the discriminator; training the discriminator using a binary cross entropy loss function to obtain a preliminary trained discriminator; The method of simultaneously training the preliminarily trained ionospheric TEC prediction model and the preliminarily trained discriminator comprises: loading the training set into the preliminarily trained ionospheric TEC prediction model to generate predicted ionospheric TEC data at the next moment as a false sample input to the discriminator; loading the corresponding ionospheric TEC data at the next moment in the training data set as a true sample input to the discriminator; and using a reconstruction loss function, a KL divergence loss function, and a binary cross entropy loss function to simultaneously train the ionospheric TEC prediction model and the discriminator to obtain a trained ionospheric TEC prediction model.

Images