A low earth orbit satellite channel prediction method

Abstract

The application discloses a low-orbit satellite channel prediction method, which comprises the following steps: S1, periodic invariant information is extracted by using decoupling characteristic learning; S2, historical channel state information is preprocessed, and the historical channel state information is fused with the periodic invariant information when the satellite switches; the preprocessed historical channel state information is converted, and after being processed in parallel in a Doppler domain, a frequency domain and a time delay domain, the related features in the three domains are fused; S3, after the historical channel state information fused with the features is blocked and position encoded, the historical channel state information is input into an encoder module, and final channel state information is obtained through data recombination; S4, according to a loss function, the error between the channel state information predicted in the step S3 and the real channel state information is calculated, and through back propagation, a trained channel prediction model is obtained; the model adopts a normalized root mean square error. The application can significantly improve the link reliability and resource utilization efficiency in a non-ground network.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication technology, and in particular relates to a method for predicting low-Earth orbit satellite channels. Background Technology

[0002] In recent years, the large-scale deployment of LEO satellite constellations has led to rapid development of satellite-based communication systems in terms of coverage, service continuity, and service types. Due to the high-speed motion of satellites, communication links exhibit significant characteristics such as large Doppler shift, long latency, and short channel coherence time. Communication systems typically rely on accurate channel state information to perform critical physical layer functions such as beamforming, power allocation, and resource scheduling. However, under the aforementioned conditions, the acquisition and updating of channel state information are severely constrained.

[0003] In existing technologies, traditional model-based channel prediction methods typically rely on linear extrapolation, parameter fitting, or simplified multipath structure descriptions, which can meet certain prediction requirements in low-speed or medium-speed environments. However, in LEO scenarios, due to the high relative speed between satellites and ground terminals and the rapid changes in geometric relationships, these methods require frequent updates to channel model parameters, making it difficult to maintain stable prediction performance in real-world systems. Furthermore, the structural assumptions of the models often fail to effectively capture the complex coupling characteristics of the satellite-to-ground link caused by variations in time, frequency, Doppler, and time delay, leading to a decrease in prediction accuracy.

[0004] With the development of deep learning technology, a number of data-driven methods have emerged that treat channel prediction as a sequence learning problem. While these methods can extract temporal structures from historical observations, they are typically limited to single features in the time or frequency domains, failing to simultaneously model multi-domain characteristics such as Doppler and time delay, and lacking adaptability to the rapidly evolving channel states in LEO environments. Furthermore, in satellite-to-ground communication, frequent satellite handovers are inherent, each handover causing abrupt changes in communication geometry, resulting in significant differences in the statistical distribution between historical and future channel state information. Existing methods generally treat channel state information before and after handover as homogeneous data, lacking effective mechanisms to bridge the distributional offset between different satellites, leading to significant degradation in prediction performance during handover periods.

[0005] In summary, existing technologies generally suffer from the following shortcomings:

[0006] 1. It is difficult to adapt to the rapid changes in the channel caused by the high-speed movement of LEO. Traditional models require frequent parameter updates, which has poor practical feasibility.

[0007] 2. The lack of joint modeling capability for multi-domain channel characteristics such as time, frequency, Doppler, and time delay leads to insufficient prediction accuracy.

[0008] 3. In satellite handover scenarios, the statistical distribution differences between different satellites cannot be handled, resulting in a significant decrease in channel prediction performance during handover.

[0009] The aforementioned defects directly affect the link reliability and resource utilization efficiency of satellite communication systems in high-speed mobile environments. Therefore, there is an urgent need for a channel prediction technology that can maintain prediction accuracy under complex multi-domain conditions and effectively handle distributed offsets during handover. Summary of the Invention

[0010] To address the aforementioned problems, this invention proposes a low-Earth orbit satellite channel prediction method, aiming to solve the technical deficiencies of existing technologies that are difficult to accurately predict channel status under high-speed motion and frequent switching conditions, and that cannot take into account Doppler / delay / frequency multi-domain characteristics and effectively cope with satellite switching distribution offset.

[0011] The above objectives are achieved through the following technical solutions:

[0012] A method for predicting low-Earth orbit satellite channels includes the following steps:

[0013] S1. Extracting periodically invariant information using decoupled representation learning ;

[0014] S2. Preprocess historical channel state information and fuse historical channel state information with periodically invariant information during satellite handover. The preprocessed historical channel state information is transformed and processed in parallel through Doppler domain, frequency domain, and time delay domain, and then the relevant features of the three domains are fused; the historical channel state information is obtained through channel estimation.

[0015] S3. After the historical channel state information after feature fusion is divided into blocks and encoded by location, it is input into the encoder module, and the final channel state information is obtained through data recombination.

[0016] S4. Based on the loss function, calculate the error between the channel state information predicted in step S3 and the actual channel state information, and obtain the trained channel prediction model through backpropagation; the loss function is the normalized root mean square error (NMSE), and its expression is:

[0017] ,

[0018] in, Represents the loss function. Expressing expectations, Represents the actual channel state information, This represents the predicted channel state information. This represents the square of the Frobenius norm.

[0019] Step S1 includes the following sub-steps:

[0020] S101. During the service period of a specific area covered by satellite, consider the total Each time slice corresponds to a snapshot of downlink channel state information. These snapshots were randomly divided into two mutually exclusive and equal-sized subsets, denoted as ; and This serves as the input domain for the decoupled representation learning module; then, the decoupled representation learning module uses data from... samples and from samples Combinations as input;

[0021] S102. Each input is converted into real-valued form by concatenating the real and imaginary parts along the vertical axis; subsequently, the channel state information matrices from different times are concatenated along the channel dimension;

[0022] S103. Input data is processed through an information extraction network to extract periodically invariant information. The information extraction network splits the channel state information to extract periodically invariant information. .

[0023] Step S2 includes the following sub-steps:

[0024] S201. When satellites switch, historical channel status information is transmitted through the data fusion module. With periodic invariant information The data fusion module integrates the data, enabling the output data to include historical channel state information and periodically invariant information. This can be achieved through weighted combination, as expressed by the formula:

[0025] ,

[0026] in, Data fusion module The output, The time length indicating historical channel state information. Indicates the number of subcarriers. Indicates the number of antennas. Represents the historical channel information matrix. This is an adjustable hyperparameter. To achieve weighted combination, period-invariant information... Broadcast over time and replicate. Second-rate.

[0027] To reduce computational complexity, antenna processing is parallelized, allowing for independent prediction of channel state information for each antenna. The channel state information is rearranged into a dimension of... Channel state information matrix It is worth noting that the data fusion module It only operates during satellite handover until the newly arrived satellite collects no less than Uplink channel state information for a given duration. Otherwise, when a satellite continues to provide service to user equipment, the historical channel state information will be directly rearranged. ;

[0028] S202. Will The channel state information matrix in the Doppler domain is obtained by transforming the data through Discrete Fourier Transform (DFT). ,

[0029] ,

[0030] in, The dimension is The DFT matrix. Then... Converted to real-valued tensors First, Normalization is performed: ,in and These represent the mean and standard deviation within the batch, respectively; subsequently, Reorganized into a real-valued channel state information matrix in the Doppler domain ;

[0031] Then, The input is fed into the feature extraction module to capture discriminative features of channel state information in the Doppler domain. This feature extraction module captures Doppler domain features of the input data, which can be implemented using a convolutional neural network (CNN). This module consists of three main parts: initial... Convolution, a set of convolutional paths for parallel operations, and the final... Convolution. Each convolutional part consists of a convolutional layer, followed by batch normalization and a ReLU activation function.

[0032] initial The convolution expands the input to four channels. Each residual block consists of a bottleneck structure, with kernel sizes of 3, 5, and 7 set in the parallel branches. Within each residual block, batch normalization is applied and residual connections are introduced.

[0033] Finally, the outputs of the three branches are aggregated and passed through a final... Compound convolution compresses the representation from four channels to a single channel, yielding the final feature representation. In mathematical terms, the overall operation of the feature extraction module can be summarized as follows:

[0034] ,

[0035] ,

[0036] ,

[0037] in, , and middle This represents a composite function, with subscripts 3, 5, and 7 indicating the kernel size, respectively. This indicates the number of repetitions in this module. Second-rate; Indicates the initial The output matrix of convolution, This represents the matrix resulting from the aggregation of the outputs of the three branches.

[0038] S203. [The following appears to be a separate, unrelated sentence:] Will The process is then transferred to the time delay domain; specifically, it is obtained through the inverse discrete Fourier transform (IDFT). Delay domain form ,

[0039] ,

[0040] in, The dimension is The DFT matrix, superscript Indicates conjugate transpose; subsequently, The feature extraction modules with the same structure but different network parameters are obtained. ;

[0041] S204. In parallel with S203, Convert to real-valued tensors Then it is passed to the frequency attention module. This is used to extract preliminary features in the frequency domain; the frequency attention module extracts high-frequency and low-frequency features of the data, which can be implemented through a variant of the CBAM neural network, and its expression is:

[0042] ,

[0043] in, This is the output of the feature attention module. This represents the Sigmoid function. Indicates splicing along the channel. This represents element-wise multiplication. During the multiplication process, the spatial attention value is broadcast along the channel dimension;

[0044] Subsequently, the output of the frequency attention module is added to its input. Finally, the resulting tensors are rearranged to combine the feature dimensions, yielding the final output in the frequency domain. ;

[0045] S205. By feature fusion module For the output of S202 S203 output and S204 output Integration is performed; the feature fusion module Will , , The fusion process outputs a single dataset containing all features, which can be achieved using a Convolutional Neural Network (CNN). This module consists of two parts: a complementary fusion part and a frequency enhancement part. The three feature representations are concatenated along the channel dimension to obtain the input tensor of the complementary fusion part. In mathematics, the overall process of the feature fusion module is represented as:

[0046] ,

[0047] in The output of the feature fusion module, This represents a composite neural network with 3 convolutional kernels, whose input is... The specific formula is as follows:

[0048] ,

[0049] in, Represents the RuLU activation function. Indicates batch normalization. Indicates that the convolution kernel is The convolution operation.

[0050] Step S3 includes the following sub-steps:

[0051] S301. Represent the fused features Divided into blocks of size [size] by the segmentation module. non-overlapping blocks, to obtain ,in Indicates the number of blocks;

[0052] S302. [The following appears to be a separate, unrelated sentence:] Will... Rearranged as Then, projected through a single fully connected layer as ,in This represents the feature dimension required by the encoder; subsequently, the positional encoding module is introduced, whose expression is:

[0053] ,

[0054] ,

[0055] in Indicates the index of the patch in the sequence. This represents the index within the feature dimension of each embedded patch. Indicates position code, Indicates that the first dimension is And the second dimension is The corresponding position code value, Indicates that the first dimension is And the second dimension is The corresponding position-coded value is then used to obtain the position-coded channel state information representation. ,

[0056] ;

[0057] S303. Input After processing by the encoder, the output is... Represented as:

[0058] ,

[0059] Subsequently, two fully connected layers are applied to... Predictive representation converted into channel state information ,in The predicted time length of the future channel state information is represented; then the predicted channel state information is represented... Converting from real-valued representation to complex-valued representation Finally, the prediction results of all antennas are summarized to obtain the final channel state information. .

[0060] The advantages of this invention compared to the prior art are: it can maintain stable channel prediction accuracy under conditions of high-speed movement and frequent satellite switching, realize joint modeling of multi-domain information such as Doppler, time delay and frequency, and effectively reduce the impact of distribution offset between different satellites, thereby significantly improving the link reliability and resource utilization efficiency in non-terrestrial networks. Attached Figure Description

[0061] Figure 1 This is a framework diagram of the method of the present invention.

[0062] Figure 2 This is an architecture diagram of the encoder module.

[0063] Figure 3 This is a comparison chart of the prediction error NMSE between the method of this invention and other comparative methods. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention are described in detail below with reference to the accompanying drawings. These embodiments are implemented based on the technical solutions of this invention, providing detailed implementation methods and specific operating procedures. It should be understood that the specific examples described herein are merely illustrative of this invention, but the scope of protection of this invention is not limited to the following embodiments.

[0065] A method for predicting low-Earth orbit satellite channels includes the following steps:

[0066] S1. Extracting periodically invariant information using decoupled representation learning ;

[0067] S2. Preprocess historical channel state information and fuse historical channel state information with periodically invariant information during satellite handover. The preprocessed historical channel state information is transformed and processed in parallel through Doppler domain, frequency domain, and time delay domain, and then the relevant features of the three domains are fused; the historical channel state information is obtained through channel estimation.

[0068] S3. After the historical channel state information after feature fusion is divided into blocks and encoded by location, it is input into the encoder module, and the final channel state information is obtained through data recombination.

[0069] S4. Based on the loss function, calculate the error between the channel state information predicted in step S3 and the actual channel state information, and obtain the trained channel prediction model through backpropagation; the loss function is the normalized root mean square error (NMSE), and its expression is:

[0070] ,

[0071] in, Represents the loss function. Expressing expectations, Represents the actual channel state information, This represents the predicted channel state information. This represents the square of the Frobenius norm.

[0072] Step S1 includes the following sub-steps:

[0073] S101. During the service period of a specific area covered by satellite, consider the total Each time slice corresponds to a snapshot of downlink channel state information. , Indicates time slice The corresponding information snapshot samples. These snapshots are randomly divided into two mutually exclusive and equal-sized subsets, denoted as . and This serves as the input domain for the decoupled representation learning module. For example... Figure 1 As shown in the decoupled representation learning module, the network uses data from... samples and from samples Combinations as input;

[0074] S102. Each input is converted to real-valued form by concatenating its real and imaginary parts along the vertical axis. Subsequently, the channel state information matrices from different times are concatenated along the channel dimension;

[0075] S103. Subsequently, the input data will be processed through a deep learning network to extract periodic invariant information. The information extraction network splits the channel state information to extract periodically invariant information. This information extraction network can be implemented using a convolutional neural network. The input data is first processed through repeated... The residual blocks are of the second type. Each residual block is a... The process begins with a convolutional layer, followed by batch normalization and ReLU activation, and then through a... Convolutional layers and batch normalization; further, residual connections are introduced within each residual block, and ReLU activation is performed; the output of the residual block. Represented as:

[0076] ,

[0077] in, To represent a composite function, Indicates the kernel size. For the number of repetitions, For the input of the residual block, the operator This indicates the composition of functions, i.e. Each residual function Defined as:

[0078] ,

[0079] in, Represents the RuLU activation function. Indicates batch normalization. Indicates that the convolution kernel is Convolution operations;

[0080] Subsequently, the features obtained from the residual network are integrated and refined through feature aggregation, and the dual-channel feature representation is then processed... The convolution is projected onto a single channel. Finally, the real and imaginary parts are merged using an R / I to complex conversion module to obtain the final period-invariant information. Similarly, the same deep learning network architecture is used to obtain... residual information in or In However, the difference lies in the fact that the input to the neural network becomes... or .

[0081] S104. Utilizing Neural Networks estimate and Mutual information between Neural Networks estimate Channel state pair Mutual information between and neural networks estimate and Mutual information between ,in Indicates the first Each sample. The neural network can be implemented in various ways, including but not limited to MINE and MINDE neural networks for mutual information estimation. Here, this invention uses a MINDE neural network as the mutual information estimator;

[0082] S105. Calculate the loss function for decoupled representation learning. Its form is:

[0083] ,

[0084] in, The total number of training samples, To balance the hyperparameters of the two terms, iterative optimization and decoupling of representation learning will extract periodic invariant information. .

[0085] Step S2 includes the following sub-steps:

[0086] S201. When satellite switching occurs, historical channel status information is transmitted via... Figure 1 Data fusion module With periodic invariant information The data fusion module integrates historical channel state information and periodically invariant information into the output data. This integration is achieved through weighted combination, as expressed by the formula:

[0087] ,

[0088] in, Data fusion module The output, The time length indicating historical channel state information. Indicates the number of subcarriers. Indicates the number of antennas. Represents the historical channel information matrix. This is an adjustable hyperparameter. To achieve weighted combination, period-invariant information... Broadcast over time and replicate. Second-rate.

[0089] To reduce computational complexity, antenna processing is parallelized, allowing for independent prediction of channel state information for each antenna. The channel state information is rearranged into a dimension of... Channel state information matrix It is worth noting that the data fusion module It only operates during satellite handover until the newly arrived satellite collects no less than Uplink channel state information for a given duration. Otherwise, when a satellite continues to provide service to user equipment, the historical channel state information will be directly rearranged. ;

[0090] S202. For example Figure 1 As shown, The channel state information matrix in the Doppler domain is obtained by transforming the data through Discrete Fourier Transform (DFT). ,

[0091] ,

[0092] in, The dimension is The DFT matrix. Then... Converted to real-valued tensors To facilitate network training, first... Normalization is performed: ,in and These represent the mean and standard deviation within the batch, respectively. Then, Reorganized into a real-valued channel state information matrix in the Doppler domain This is to facilitate neural network processing.

[0093] Then, The input data is fed into the feature extraction module, whose main function is to capture discriminative features of channel state information in the Doppler domain. This feature extraction module captures Doppler domain features of the input data, which can be implemented using a convolutional neural network (CNN). This module consists of three main parts: initial... Convolution, a set of convolutional paths for parallel operations, and the final... Convolution. Each convolutional part consists of a convolutional layer, followed by batch normalization and a ReLU activation function.

[0094] initial The convolution expands the input to four channels. Each residual block consists of a bottleneck structure, with kernel sizes of 3, 5, and 7 set in the parallel branches. Within each residual block, batch normalization is applied and residual connections are introduced.

[0095] Finally, the outputs of the three branches are aggregated and passed through a final... Compound convolution compresses the representation from four channels to a single channel, yielding the final feature representation. In mathematical terms, the overall operation of the feature extraction module can be summarized as follows:

[0096] ,

[0097] ,

[0098] ,

[0099] in, , and middle This represents a composite function, with subscripts 3, 5, and 7 indicating the kernel size, respectively. This indicates the number of repetitions in this module. Second-rate; Indicates the initial The output matrix of convolution, This represents the matrix resulting from the aggregation of the outputs of the three branches.

[0100] S203. For example... Figure 1 As shown, The process is then transferred to the time delay domain; specifically, it is obtained through the inverse discrete Fourier transform (IDFT). Delay domain form ,

[0101] ,

[0102] in The dimension is The DFT matrix, superscript This indicates the conjugate transpose. Subsequently, The feature extraction modules with the same structure but different network parameters are obtained. ;

[0103] S204. In parallel with S203, Converted to real-valued tensors Then pass in Figure 1 Frequency Attention Module This is used to extract preliminary features in the frequency domain. The frequency attention module extracts high-frequency and low-frequency features of the data, which can be implemented through a variant of the CBAM neural network, and its expression is:

[0104] ,

[0105] in, This is the output of the feature attention module. This represents the Sigmoid function. Indicates splicing along the channel. This represents element-wise multiplication. During the multiplication process, the spatial attention value is broadcast along the channel dimension;

[0106] Subsequently, the output of the frequency attention module is added to its input. Finally, the resulting tensors are rearranged to combine the feature dimensions, yielding the final output in the frequency domain. ;

[0107] S205. By Figure 1 Feature fusion module in For output , and The feature fusion module integrates the three input data points, outputting a single dataset containing all features. This dataset can be implemented using a Convolutional Neural Network (CNN). The module consists of two parts: a complementary fusion part and a frequency enhancement part. The three feature representations are concatenated along the channel dimension to obtain the input tensor of the complementary fusion part. In mathematics, the overall process of the feature fusion module can be represented as:

[0108] ,

[0109] in The output of the feature fusion module, This represents a composite neural network with 3 convolutional kernels, whose input is... The specific formula is as follows:

[0110] ,

[0111] in, Represents the RuLU activation function. Indicates batch normalization. Indicates that the convolution kernel is The convolution operation.

[0112] Step S3 includes the following sub-steps:

[0113] S301. Represent the fused features Divided into blocks of size [size] by the segmentation module. non-overlapping blocks, to obtain ,in Indicates the number of blocks;

[0114] S302. [The following appears to be a separate, unrelated sentence:] Will... Rearranged as Then, projected through a single fully connected layer as ,in This represents the feature dimensions required by the encoder. Subsequently, the positional encoding module is introduced, with the following expression:

[0115] ,

[0116] ,

[0117] in Indicates the index of the patch in the sequence. This represents the index within the feature dimension of each embedded patch. Indicates position code, Indicates that the first dimension is And the second dimension is The corresponding position code value, Indicates that the first dimension is And the second dimension is The corresponding position-coded value is then obtained. Finally, the channel state information representation with position coding is obtained. ,

[0118] ;

[0119] S303. Input After processing by the encoder, the output is... Represented as:

[0120] ,

[0121] The network structure of the encoder is as follows: Figure 2As shown, the encoder consists of four parts: multi-head attention, addition and normalization, feedforward neural network, and addition and normalization, forming a whole. Second-rate.

[0122] Subsequently, two fully connected layers are applied to... Predictive representation converted into channel state information ,in This represents the time length of the predicted future channel state information. Then, the predicted channel state information is represented... Converting from real-valued representation to complex-valued representation Finally, the prediction results from all antennas are aggregated to obtain the final channel state information. .

[0123] Figure 3 The relationship between the proposed method and existing methods is presented under different channel estimation error interference conditions on historical channel state information. The comparison methods are: iTransformer (first method), PatchTST (second method), Crossformer (third method), and FEDformer (fourth method). Under different SNR noise interference conditions on historical channel state information, the normalized mean square error (NMSE) of the proposed method is smaller than that of the existing methods. Specifically, when the SNR is 15 dB, the NMSE of the proposed method is less than 0.1, while the NMSE of the other four comparison methods is greater than 0.1.

Claims

1. A method for predicting low-Earth orbit satellite channels, characterized in that, The method includes the following steps: S1. Extracting periodically invariant information using decoupled representation learning ; S2. Preprocess historical channel state information and fuse historical channel state information with periodically invariant information during satellite handover. ; The preprocessed historical channel state information is transformed and processed in parallel in the Doppler domain, frequency domain, and time delay domain, and then the relevant features of the three domains are fused. The historical channel state information is obtained through channel estimation; S3. After the historical channel state information after feature fusion is divided into blocks and encoded by location, it is input into the encoder module, and the final predicted channel state information is obtained through data recombination. S4. Based on the loss function, calculate the error between the channel state information predicted in step S3 and the actual channel state information, and obtain the trained channel prediction model through backpropagation; the loss function is the normalized root mean square error (NMSE), and its expression is: ， in, Represents the loss function. Expressing expectations, Represents the actual channel state information, This represents the predicted channel state information. Denotes the square of the Frobenius norm; The specific implementation steps of step S2 include: S201. When satellites switch, historical channel status information is transmitted through the data fusion module. With periodic invariant information The data fusion module integrates the data, enabling the output data to include historical channel state information and periodically invariant information. During the processing of uplink channel state information collected by the satellite, antenna processing is parallelized, with the channel state information of each antenna predicted independently. The channel state information is then rearranged into a dimensional array. Channel state information matrix ,in The time length indicating historical channel state information. Indicates the number of subcarriers; S202. Will The channel state information matrix in the Doppler domain is obtained by converting the data to the Doppler domain using the Discrete Fourier Transform (DFT). : ， in, The dimension is The DFT matrix; subsequently, Converted to real-valued tensors First, Normalization is performed: ,in and These represent the mean and standard deviation within the batch, respectively; subsequently, Reorganized into a real-valued channel state information matrix in the Doppler domain ; Then, The input is fed into the feature extraction module to capture discriminative features of channel state information in the Doppler domain, and the final output is the extracted feature representation. The feature extraction module captures the Doppler domain features of the input data. S203. [The following appears to be a separate, unrelated sentence:] Will The process is then transferred to the time delay domain; specifically, it is obtained through the inverse discrete Fourier transform (IDFT). Delay domain form : ， in, The dimension is The DFT matrix, superscript Indicates conjugate transpose; subsequently, The feature extraction modules with the same structure but different network parameters are obtained. ; S204. In parallel with S203, Convert to real-valued tensors Then it is passed to the frequency attention module. This is used to extract preliminary features in the frequency domain, and then compares the extracted preliminary features with... Add them together to get the final output. The frequency attention module extracts high-frequency and low-frequency features from the data. S205. By feature fusion module For the output of S202 S203 output and S204 output Integration is performed; the feature fusion module Will , , The data is then fused to output a single dataset containing all the features.

2. The low-Earth orbit satellite channel prediction method according to claim 1, characterized in that, The specific implementation steps in step S1 include: S101. During the service period of a specific area covered by satellite, consider the total Each time slice corresponds to a snapshot of downlink channel state information. , Indicates time slice The corresponding information snapshot samples; these snapshots are randomly divided into two mutually exclusive and equal-sized subsets, denoted as . and This serves as the input domain for the decoupled representation learning module; then, the decoupled representation learning module uses data from... samples and from samples Combinations as input; S102. Each input is converted into real-valued form by concatenating the real and imaginary parts along the vertical axis; subsequently, the channel state information matrices from different times are concatenated along the channel dimension; S103. Input data is processed through an information extraction network to extract periodically invariant information. The information extraction network splits the channel state information to extract periodically invariant information. .

3. The low-Earth orbit satellite channel prediction method according to claim 2, characterized in that, The information extraction network described in step S103 can be implemented using a convolutional neural network (CNN). The specific implementation process is as follows: Input data is first processed through repeated... The residual blocks are of the order of sub-blocks; each residual block is a single... The process begins with a convolutional layer, followed by batch normalization and ReLU activation, and then through a... Convolutional layers and batch normalization; further, residual connections are introduced within each residual block, and ReLU activation is performed; the output of the residual block. Represented as ， in, To represent a composite function, Indicates the kernel size. For the number of repetitions, For the input of the residual block, the operator This indicates the composition of functions, i.e. Each residual function Defined as ， in, Represents the RuLU activation function. Indicates batch normalization. Indicates that the convolution kernel is Convolution operations; Subsequently, the features obtained from the residual network are integrated and refined through feature aggregation, and the dual-channel feature representation is then processed... The convolution is projected onto a single channel; finally, the real and imaginary parts are merged using an R / I to complex conversion module to obtain the final period-invariant information. .

4. The low-Earth orbit satellite channel prediction method according to claim 1, characterized in that, The data fusion module described in step S201 It only operates during satellite handover until the newly arrived satellite collects no less than Uplink channel state information for a duration of time; Otherwise, when a satellite continues to provide service to user equipment, the historical channel state information will be directly rearranged. .

5. The low-Earth orbit satellite channel prediction method according to claim 1, characterized in that, The data fusion module described in step S201 This can be achieved through weighted combination, and its formula is expressed as follows: ， in, Data fusion module The output, Indicates the number of antennas. Represents the historical channel information matrix. These are adjustable hyperparameters; to achieve weighted combination, Broadcast over time and replicate. Second-rate.

6. The low-Earth orbit satellite channel prediction method according to claim 1, characterized in that, The feature extraction module described in step S202 can be implemented using a convolutional neural network (CNN). This module consists of three main parts: an initial... Convolution, a set of convolutional paths for parallel operations, and the final... Convolution; each convolutional part consists of a convolutional layer, followed by batch normalization and a ReLU activation function; initial The convolution expands the input to four channels; each residual block consists of a bottleneck structure, with kernel sizes of 3, 5, and 7 set in the parallel branches; within each residual block, batch normalization is applied and residual connections are introduced. Finally, the outputs of the three branches are aggregated and passed through a final... Compound convolution compresses the representation from four channels to a single channel, yielding the final feature representation. In mathematics, the overall operation of the feature extraction module can be summarized as follows: ， ， ， in, , and middle This represents a composite function, with subscripts 3, 5, and 7 indicating the kernel size, respectively. This indicates the number of repetitions in this module. Second-rate; Indicates the initial The output matrix of convolution, This represents the matrix resulting from the aggregation of the outputs of the three branches.

7. The low-Earth orbit satellite channel prediction method according to claim 1, characterized in that, The frequency attention module described in step S204 can be implemented using a variant of the CBAM neural network, and its expression is: ， in, This is the output of the feature attention module. This represents the Sigmoid function. Indicates splicing along the channel. This indicates element-wise multiplication; during the multiplication process, the spatial attention value is broadcast along the channel dimension. Subsequently, the output of the frequency attention module is added to its input; finally, the resulting tensors are rearranged to combine the feature dimensions, yielding the final output in the frequency domain. .

8. The low-Earth orbit satellite channel prediction method according to claim 1, characterized in that, The feature fusion module described in step S205 can be implemented using a convolutional neural network (CNN). This module consists of two parts. Composition: complementary fusion section and frequency enhancement section; The three feature representations are concatenated along the channel dimension to obtain the input tensor of the complementary fusion part. In mathematics, the overall process of the feature fusion module is represented as: ， in The output of the feature fusion module, This represents a composite neural network with 3 convolutional kernels, whose input is... The specific formula is as follows: ， in, Represents the RuLU activation function. Indicates batch normalization. Indicates that the convolution kernel is The convolution operation.

9. A method for predicting low-Earth orbit satellite channels according to claim 1, characterized in that, The specific implementation steps in step S3 include: S301. Represent the fused features Divided into blocks of size [size] by the segmentation module. non-overlapping blocks, to obtain ,in Indicates the number of blocks; S302. [The following appears to be a separate, unrelated sentence:] Will... Rearranged as Then projected through a single fully connected layer to ,in This represents the feature dimensions required by the encoder; subsequently, the positional encoding module is introduced, whose expression is: ， ， in Indicates the index of the patch in the sequence. This represents the index within the feature dimension of each embedded patch. Indicates position code, Indicates that the first dimension is And the second dimension is The corresponding position code value, Indicates that the first dimension is And the second dimension is The corresponding position-coded value is then used to obtain the position-coded channel state information representation. : ； S303. Input After processing by the encoder, the output is... Represented as ， Subsequently, two fully connected layers are applied to... Predictive representation converted into channel state information ,in The predicted time length of the future channel state information is represented; then the predicted channel state information is represented... Converting from real-valued representation to complex-valued representation Finally, the prediction results of all antennas are summarized to obtain the final channel state information. , Indicates the number of antennas.

Images