Channel prediction method and application thereof in reconfigurable intelligent surface-assisted satellite communication

Abstract

The invention relates to the technical field of wireless communication, and discloses an intelligent channel prediction method based on a deep multi-layer perceptron architecture in order to solve the problem of high-dimensional near-field cascade channel prediction in a reconfigurable intelligent surface assisted low-orbit satellite communication system. Designing a frequency domain MLP architecture to learn global time sequence characteristics of a cascade channel; secondly, considering high-dimensional spatial domain characteristics of a near-field cascade channel, decomposing the cascade channel into a plurality of feature dimensions, and proposing a multi-branch MLP architecture to realize global spatial domain correlation modeling; and finally, by taking the frequency domain MLP and the multi-branch MLP as basic network feature extraction modules, designing a channel prediction network model based on an encoder-decoder architecture. According to the method disclosed by the invention, the accurate prediction of the downlink CSI can be realized by utilizing the historical channel state information estimated by the uplink communication, and the proposed channel prediction model has better channel prediction precision and model convergence speed.

Description

Technical Field

[0001] The invention relates to the technical field of wireless communications, and in particular to a channel prediction method for a reconfigurable intelligent surface-assisted satellite communication system. Background Art

[0002] The integrated air-ground-space network integrates ground, air and space resources to achieve seamless coverage and high-speed interconnection on a global scale, thereby improving the coverage, transmission rate and service quality of the 6G communication network and providing better support for various 6G application scenarios. The non-ground network based on low-orbit satellite communication (LEO) is one of the important components of the 6G integrated air-ground-space network. It can directly provide services to handheld terminal users, thereby significantly reducing transmission delays. Due to the serious signal propagation loss caused by long-distance communication links, LEO communications currently mainly support communication scenarios with good line-of-sight (LoS) links in remote areas and coastal waters. However, for complex urban environments, the application of satellite communications in dense urban areas is limited because LoS transmission is hindered by large-scale buildings. Although researchers have proposed corresponding methods to alleviate the LoS blocking problem in satellite communications, such as introducing high-altitude platforms or drones as aerial relays to assist satellite signal coverage, these methods often introduce high deployment costs and equipment energy consumption.

[0003] In order to solve the problem of effective coverage of LEO satellite signals, the Reconfigurable Intelligent Surface (RIS) technology is expected to improve the quality of satellite services at a lower cost. RIS is a new type of artificial electromagnetic metasurface that has the characteristics of flexible control of the frequency, amplitude, phase, polarization and propagation direction of electromagnetic waves. By configuring the reflection coefficient of RIS in real time, effective passive beamforming can be achieved. RIS has shown great potential in various ground mobile communication scenarios. It is worth noting that the optimal beamforming design of RIS often requires the communication system to obtain accurate channel state information (CSI). However, due to the passive characteristics of RIS and the high mobility of LEO satellites, obtaining CSI for RIS-assisted LEO satellite communication systems has become a very challenging problem. First, the passive characteristics of RIS make RIS have no independent signal processing and channel estimation capabilities, so the cascade channel can only be estimated at the transceiver. Since passive RIS needs to integrate a large number of reflection units to improve the passive beamforming gain, the number of cascade channel coefficients that need to be estimated will increase proportionally with the increase in the number of RIS units, resulting in a huge pilot overhead for channel estimation. Secondly, due to the natural high dynamic characteristics of LEO satellites, huge Doppler frequency shifts will be introduced, making the channel exhibit fast time-varying characteristics, especially for highly mobile millimeter wave communication systems with shorter channel coherence time. At the same time, the huge propagation delay caused by long-distance satellite-to-ground link transmission will cause obvious obsolescence problems for both the CSI estimated by the satellite in the time division duplex mode and the CSI fed back by users in the frequency division duplex (FDD) mode, resulting in a mismatch between the estimated channel quality and the current channel quality. Therefore, an effective channel prediction algorithm for LEO satellite communication systems assisted by RIS is urgently needed. In particular, with the increase of RIS unit dimensions and array aperture, electromagnetic wave propagation modeling will shift from traditional far-field plane wavefront propagation to new near-field spherical wavefront propagation. In far-field communication, the response of large-scale antenna arrays depends only on the angle of the signal, while the response of antenna arrays in near-field communication is related to both distance and angle, which makes the typical far-field channel estimation and tracking algorithm based on angle domain sparsity no longer applicable to near-field communication. Summary of the invention

[0004] The present invention is to solve the problems of large channel prediction deviation, slow convergence speed and low convergence accuracy in the prior art, and proposes a channel prediction method for a reconfigurable intelligent surface-assisted satellite communication system. The content includes the following steps:

[0005] Step 1: Build a RIS-assisted low-orbit satellite communication system simulation platform and model the reconfigurable intelligent surface (RIS)-assisted low-orbit satellite communication system channel.

[0006] Step 2: Receive the uplink pilot signal of the RIS-assisted low-orbit satellite communication system and collect the corresponding channel information to build the training data set required for the intelligent channel prediction framework.

[0007] Step 3: Based on the RIS-assisted low-orbit satellite communication system channel modeling method proposed in step 1, and using the training data set constructed in step 2, an intelligent channel prediction framework is developed to achieve accurate prediction of the downlink CSI of the RIS-assisted LEO satellite frequency division duplex (FDD) communication system.

[0008] Furthermore, the specific contents of step 2 are as follows:

[0009] Step 2.1: Assume that the LEO satellite is equipped with M antennas and the RIS is equipped with N reflection units. Let G represent the RIS to LEO satellite channel, h k represents the channel from the kth user to RIS, represents the reflection vector of RIS, where β n ∈{0,1} and θ n ∈[0,2π](n=1,2,…,N) respectively represent the reflection amplitude and phase of the nth unit of RIS. The received signal of LEO satellite in the qth (1≤q≤Q) time slot can be expressed as

[0010]

[0011] Where diag(·) represents vector diagonalization or diagonal matrix vectorization, s q,k Represents the pilot signal sent by the kth user, satisfying represents additive Gaussian white noise. Let H k =Gdiag(h k ) represents the cascade channel of the k-th user signal reflected from the RIS to the base station. For multi-user channel estimation, the typical orthogonal pilot transmission strategy is widely used. For example, when different users transmit pilot signals in orthogonal time slots, the multi-user channel estimation can be separated into K independent executions of single-user channel estimation in sequence. Considering that this application mainly focuses on the cascade channel prediction problem in high-speed mobile LEO satellite communication scenarios, the classic least squares channel estimation framework based on discrete Fourier transform (DFT) is used to obtain the uplink cascade channel. When the k-th user completes Q (Q ≥ N) pilot transmission time slots, the pilot signal received by the LEO satellite is Then the estimated cascade channel can be expressed as

[0012] H k =Y k Θ H (ΘΘ H ) -1

[0013] in is the DFT matrix The first N rows of , that is, the (p1, p2)th element of the RIS reflection matrix Θ satisfies

[0014] Step 2.2: For the RIS-assisted LEO satellite communication system, due to the high mobility of LEO satellites and the transmission delay of the satellite-to-ground link, the uplink and downlink channel reciprocity widely used in the ground mobile communication system is no longer valid. This application proposes an efficient channel prediction network, which uses the estimated historical P u Frame uplink concatenation channel information To predict the future d Frame downlink cascade channel information Considering that existing neural networks mainly support real number operations, the input tensor of the channel prediction network is and data labels Designed to and The real number form of . and Respectively represent the real part and imaginary part of the complex data, then the network input tensor is designed to be Data labels are designed to Since channel estimation is a regression task, the commonly used loss functions are the minimum square error loss function based on the L2 norm and the minimum absolute error loss function based on the L1 norm. The L1 norm has better convergence performance than the L2 norm. Therefore, the neural network loss function of this operation is defined as

[0015]

[0016] Among them, f w (·) represents a neural network model with network weight w, B represents the number of batch samples in one neural network parameter iteration, and |·| represents the 1 norm. represents the output of the neural network, then the final channel prediction result can be expressed as (m=1,2,…,M; n=1,2,…,N; p=1,2,…,P d ).

[0017] Furthermore, the specific contents of step 3 are as follows:

[0018] Step 3.1: Improve channel prediction accuracy. Since the traditional time-domain MLP architecture only learns the time-domain features of the cascaded channel, it is easy to cause point-wise mapping and information bottlenecks. Therefore, the input tensor is first We perform discrete Fourier transform along the time domain dimension to generate frequency domain features, and then develop a frequency domain MLP architecture to achieve global correlation modeling of cascaded channels. For the time domain input tensor H(t), its frequency domain tensor It can be expressed as:

[0019]

[0020] According to the Pasval theorem in Fourier transform, the energy conservation formula is

[0021]

[0022] If most of the energy of a time series is concentrated in a few frequency components, then only these components can be used to accurately represent the time series. For RIS-assisted satellite-to-ground communication systems, the scatterers in the communication link are limited, and only the LoS channel and a limited number of paths in the channel matrix have energy. Therefore, the frequency domain MLP architecture designed in this application can well learn the effective characteristics of the cascaded channel and improve the channel prediction performance. In the time domain, it is a discrete time channel matrix, so DFT transform is used along the time domain to transform the input tensor for

[0023]

[0024] According to the conjugate symmetry property of DFT, we have Therefore, only use Half of the information All frequency domain information can be represented. Furthermore, due to Contains real and imaginary information, so this application designs a frequency domain MLP architecture specifically for learning complex tensor features, and the output features of its first layer can be expressed as

[0025]

[0026] in, and They represent the activation function, complex weights and biases of the first layer of the MLP. As mentioned earlier, since it is difficult to directly construct complex weights and biases in a neural network, according to the principle of complex multiplication, for two complex variables and Its complex multiplication satisfies

[0027]

[0028] Based on this complex number multiplication principle, can be rewritten as

[0029]

[0030] Furthermore, the frequency domain features are restored to the time domain using the Inverse Discrete Fourier Transform (IDFT), and then the output result F of the frequency domain MLP is obtained by concatenating the real and imaginary data of the time domain data into a real feature tensor, that is,

[0031] Step 3.2, speed up the convergence of the model; Traditional deep learning-based channel estimation methods often use convolutional neural networks (CNN) to model the spatial characteristics of the channel. However, due to its limited perception, convolutional networks also have difficulty in achieving global spatial feature modeling. In order to model the long-distance dependency information of cascade channels in the spatial domain, this application develops a multi-branch MLP architecture to achieve spatial feature extraction, and extracts features from the cascade channels from different dimensions in each branch, including satellite antenna dimensions, RIS unit dimensions, original channel dimensions, and 2D Fourier transform dimensions. For the output F of the frequency domain MLP, first use a convolutional layer to increase the dimension to obtain the input tensor of the multi-branch MLP. By introducing a dimensionality transformation factor S = C / C s ,(1≤C s ≤C), in branch 1 of the multi-branch MLP, F i is transformed into a tensor Then, a weight is The linear layer is used to extract features, and finally restored to the original dimension to obtain the output features of branch one. In branch 2 of the multi-branch MLP, F i is transformed into a tensor Then, a weight is The linear layer performs feature extraction and finally restores it to the original dimension to obtain the output features of branch 2 In branch 3 of the multi-branch MLP, the input tensor F i No operation is performed to preserve the original features. In branch 4 of the multi-branch MLP, the spatial domain of the cascaded channel is first transformed using a two-dimensional discrete Fourier transform (2D DFT), i.e.

[0032]

[0033] Furthermore, a learnable filter Ψ is introduced to optimize the spectral representation of F4, namely

[0034]

[0035] Finally, the two-dimensional inverse discrete Fourier transform (2D IDFT) is used to restore the spatial domain representation of the concatenated channels. By concatenating the outputs of the four branches, we can get a feature tensor Then use a weight of The linear layer performs feature dimensionality reduction and performs feature fusion with the input tensor to obtain the final output tensor of the multi-branch MLP module. Right now

[0036]

[0037] Step 3.3, reduce model complexity; based on the MLP feature extraction module designed above, the proposed cascade channel prediction overall network architecture adopts a D-layer encoder-decoder architecture to explore the potential sparsity of the input data and suppress the noise contained in the input data. First, the frequency domain MLP architecture is used to extract the input tensor The time domain features are then dimensionally upgraded to serve as the first layer input features of the encoder For the D-layer encoder, the feature The M×N spatial dimension is gradually reduced to M / 2 by a sampling factor of 2. D ×N / 2 D , and The channel dimension of is gradually increased from C by a sampling factor of 2, and finally a feature tensor is obtained. Each layer of the encoder contains a multi-branch MLP, a frequency domain MLP, and a convolution layer with a convolution kernel sliding step of 2 to implement the downsampling operation. The decoder has a symmetrical structure with the encoder. By using D-1 decoders, the encoded features are Restore to The decoder of each layer contains a multi-branch MLP, a frequency domain MLP, and an upsampling layer based on bilinear interpolation. Finally, a convolution kernel with number P d The 1×1 convolutional layer is used to achieve time domain alignment, thereby obtaining the final output of the channel prediction network

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. Improved channel prediction accuracy: The present invention proposes an intelligent channel prediction model based on a deep MLP architecture, which achieves accurate prediction of downlink CSI by utilizing the historical CSI estimated by uplink communication, thereby greatly improving the accuracy of channel prediction.

[0040] 2. Accelerated model convergence accuracy: This paper proposes a new algorithm to model the time domain correlation of RIS cascade channels and designs a frequency domain MLP architecture to learn the global timing characteristics of cascade channels. Furthermore, considering the high-dimensional spatial characteristics of near-field cascade channels, the cascade channels are decomposed into multiple feature dimensions, and a multi-branch MLP architecture is proposed to achieve global spatial correlation modeling, which greatly accelerates the model convergence speed.

[0041] 3. Reduced model complexity: The present invention uses frequency domain MLP and multi-branch MLP as basic network feature extraction modules, designs a channel prediction network model based on an encoder-decoder architecture, and greatly reduces the model convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the algorithm flow chart;

[0043] Figure 2 It is a RIS-assisted LEO satellite communication system;

[0044] Figure 3 It is a multi-branch MLP architecture;

[0045] Figure 4 It is the channel prediction network architecture;

[0046] Figure 5 It is the channel prediction NMSE performance under different signal-to-noise ratios;

[0047] Figure 6 It is the NMSE performance of channel prediction under different prediction frame numbers;

[0048] Figure 7 It is a comparison of the convergence performance of different channel prediction models. DETAILED DESCRIPTION

[0049] In order for professionals in the field to have a deeper understanding of the technical content of the present application, detailed explanation and description are provided through the drawings of the embodiments of the present application. Please note that the embodiments shown only represent part of the present application and do not cover all possible implementation methods. Based on these embodiments, all other embodiments obtained by ordinary technicians in the field without the need for creative work should be included in the protection scope of the present application.

[0050] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods.

[0051] Combination Figure 1 , the present invention comprises the following steps:

[0052] Step 1: Build the RIS-assisted satellite communication system architecture, design and configure the RIS system, including setting hardware parameters such as satellite antenna, RIS elements, and user terminals. Set the carrier frequency for uplink and downlink communications, and determine the physical parameters of the satellite and RIS, such as altitude and speed. Build the channel model from RIS to satellite and user terminal, assuming that the RIS to satellite link is far-field line-of-sight transmission and the user to RIS is near-field non-line-of-sight transmission.

[0053] Step 2: Obtain the uplink pilot signal of the RIS-assisted satellite communication system and construct a training data set. Extract the channel state information of multiple time frames from the uplink communication channel matrix as the input of the model. Assume that the number of historical uplink frames is P u , then extract P from the uplink channel data d Frame channel state information forms a historical uplink CSI data set. Channel state information of multiple future time frames is extracted from the downlink communication channel matrix and used as the output label of the model. Assume that the number of future downlink frames is P d , then extract P from the downlink channel data d The frame channel state information is used to form the future downlink CSI data set. The complex channel data in the historical uplink CSI and the future downlink CSI are split into real and imaginary parts. The split real and imaginary parts are used as independent input features and output labels respectively, so that the model can process real data.

[0054] Step 3 Intelligent channel prediction framework

[0055] Step 3.1: Frequency domain feature modeling. Perform discrete Fourier transform on the input time domain channel data to generate frequency domain features. Use Pasval's theorem to verify energy conservation and ensure that frequency domain features accurately reflect time domain channel information. Design a frequency domain MLP architecture, including multiple neural network layers, each containing complex weights and biases. Learn the global time domain characteristics of the cascaded channel through the frequency domain MLP architecture. Use the inverse discrete Fourier transform to restore the frequency domain features to the time domain. Concatenate the real and imaginary parts of the time domain data to generate a real feature tensor and output the frequency domain MLP results.

[0056] Step 3.2: Spatial feature modeling, design a multi-branch MLP architecture, each branch extracts features from the cascade channel from different dimensions, including satellite antenna dimension, RIS unit dimension, original channel dimension, and two-dimensional Fourier transform dimension. In each branch, dimension transformation, linear feature extraction, and restoration of the original dimension are performed. Perform a two-dimensional Fourier transform on the cascade channel data of one of the branches and perform spectrum analysis. Introduce a learnable filter to optimize the spectrum and improve the accuracy of feature extraction. Concatenate the features output by each branch to form a comprehensive feature tensor. Reduce the dimension of the comprehensive feature tensor through a linear layer to generate the final output features of the multi-branch MLP.

[0057] Step 3.3, develop the encoder-decoder architecture, design the encoder, input the uplink CSI, extract the time domain features through the frequency domain MLP, and perform dimensionality upscaling. The encoder contains multiple layers, each layer consists of a multi-branch MLP, a frequency domain MLP, and a convolution layer. The spatial dimension is reduced layer by layer and the number of channels is increased through downsampling operations. Design the decoder, whose structure is symmetrical with the encoder. Use bilinear interpolation for upsampling and restore to the original dimension layer by layer. Achieve time domain alignment through a 1×1 convolution layer to ensure the time domain consistency of the output features and the downlink CSI. Output the final downlink CSI prediction result.

[0058] The following is an experimental demonstration of the method of the present invention with specific examples, and the specific contents are as follows:

[0059] 1. Experimental environment:

[0060] In order to verify the positioning accuracy of the algorithm, MATLAB R2020a and Python 3.9 were used for simulation. The parameter settings of the RIS-assisted satellite communication system are shown in Table 1. In the channel prediction network structure proposed in this invention, the dimension transformation factor S = 4, the number of encoder layers D = 3, and the number of initial channels C = 64. The adaptive moment estimation optimization algorithm is used for network training. The learning rate descent rule obeys the cosine function, and the initial learning rate is 1×10 -3 , the final learning rate is 1×10 -5 , the batch size is set to 32, and the number of training cycles is 50. This paper uses LSTM, CNN-LSTM, Transformer, and Temporal Convolutional Network (TCN) as benchmark models to evaluate the performance of the proposed channel prediction network.

[0061] Table 1 RIS-assisted satellite communication system parameters

[0062]

[0063]

[0064] 2. Evaluation Criteria

[0065] Normalized mean square error (NMSE) is used as an evaluation index of channel prediction performance and can be expressed as:

[0066]

[0067] Among them, ||·|| F represents the Frobenius norm.

[0068] 3. Experimental results

[0069] 1) Discuss the NMSE performance of different channel prediction algorithms as the uplink signal-to-noise ratio (SNR) changes. Figure 5 Due to the high-dimensional characteristics of the near-field cascade channel, the traditional LSTM method is difficult to apply directly. Although some improved methods, such as CNN-LSTM and TCN that introduce convolutional networks to learn the spatial characteristics of the cascade channel and the Transformer model based on spatial attention, can appropriately improve the cascade channel prediction performance, their NMSE accuracy is still limited. Compared with the existing benchmark scheme, the channel prediction network proposed in this invention has better channel prediction accuracy in the entire SNR range.

[0070] 2) Discuss the different channel prediction models in the test phase for different prediction frame numbers P d The NMSE performance under the condition where the number of prediction frames in the network training phase is fixed to P d =5. Figure 6 As the number of predicted downlink channels increases, the temporal correlation with the historical uplink channels will be lower, so the accuracy of NMSE will decrease. For different numbers of channels, the channel prediction network proposed in the present invention has better NMSE performance.

[0071] 3) Discuss the convergence performance of the loss function of different channel prediction models in the network training phase, such as Figure 7 , it can be observed that the proposed channel prediction network has better convergence speed and convergence value. Table 2 summarizes the parameter amount and floating-point operation number of different network models. Compared with traditional time series prediction models, such as LSTM, CNN-LSTM and TCN, the channel prediction network proposed in this paper has fewer parameters. Compared with the emerging Transformer model, the channel prediction network proposed in this paper has a lower number of floating-point operations. Therefore, compared with the existing algorithms, the algorithm proposed in this invention has a better balance between spatial complexity and computational complexity while achieving better channel prediction accuracy.

[0072] Table 2 Complexity analysis of different channel prediction models

[0073]

[0074] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any technician familiar with the profession, without departing from the scope of the technical solution of the present invention, makes any simple modifications, equivalent changes and modifications to the above embodiments based on the technical essence of the present invention, which still fall within the scope of the technical solution of the present invention.

Claims

1. A channel prediction method for a reconfigurable intelligent surface-assisted satellite communication system, characterized in that: The following steps are involved: Step 1, modeling the reconfigurable intelligent surface-assisted low-orbit satellite communication system channel; Build a RIS-assisted low-orbit satellite communication system simulation platform; Step 2: Receive the uplink pilot signal of the RIS-assisted low-orbit satellite communication system and collect the corresponding channel information to build the training data set required for the intelligent channel prediction framework; Step 3: Based on the RIS-assisted low-orbit satellite communication system channel modeling method proposed in step 1 and using the training data set constructed in step 2, an intelligent channel prediction framework is developed to achieve accurate prediction of the downlink CSI of the RIS-assisted LEO satellite frequency division duplex communication system.

2. The method for channel prediction of a reconfigurable intelligent surface-assisted satellite communication system according to claim 1, characterized in that: Step 1 includes the following, Step 1.1, modeling the reconfigurable intelligent surface-assisted low-orbit satellite communication system channel; assuming that the height of the satellite's orbit is H S , the height of RIS is H R , a 3D Cartesian coordinate system is established with the projection of the center unit of RIS on the ground as the origin of coordinates. The coordinates of the center unit of RIS are (0,0,H R ), the initial coordinates of the LEO satellite are set to (0,0,H S ); K users with height H U Randomly distributed in the positive semicircle of RIS, the coordinates of the kth user are assumed to be (X k ,Y k ,H U ); Assume that the LEO satellite is configured with M = M1 × M2 uniform planar array antennas and the RIS is configured with N = N1 × N2 uniform planar array elements; Referring to the channel modeling method of millimeter wave Massive MIMO satellite system, a ray-based block fading geometric channel modeling framework is adopted. Assuming that the RIS to LEO satellite link is LoS transmission, for the uplink carrier frequency f u The channel from RIS to the satellite at the tth time block is It can be expressed as: in,() T represents the vector transpose, j is the imaginary unit, and satisfies j 2 =-1, and They represent the arrival azimuth (elevation) of the LEO satellite and the departure azimuth (elevation) of the RIS in the tth time block respectively; is the radiation pattern of the RIS unit; represents the path loss from RIS to LEO satellite; and They represent the transmit array response of the RIS unit to the LOS path and the receive array response of the LEO satellite antenna respectively; T s represents the signal sampling period, represents the signal propagation delay, Indicates the Doppler shift of the line-of-sight signal; Assume that there is S in the RIS to ground link UR scatterers, the channel h from near-field user k to RIS k Contains virtual LoS path components constructed by RIS and the scattered path component It can be expressed as: in, represents the distance from user k to the RIS center; for simplicity, an indicator variable u∈{LoS,l} is introduced, and the variable They represent the LoS path and the scattering path in the t-th time block user k to RIS link 0≤l≤S UR Arrival azimuth and elevation angles, RIS unit radiation pattern, path fading, Doppler shift and delay; It obeys the complex Gaussian distribution and represents the complex path gain of the lth scattering path; is the normalization factor in the multipath channel model; is the near-field array response of the RIS units in the link from user k to RIS.

3. The method for channel prediction of a reconfigurable intelligent surface-assisted satellite communication system according to claim 1, characterized in that: The steps of building a signal model of a reconfigurable intelligent surface-assisted low-orbit satellite communication system in step 1 include the following: Step 1.2, let represents the reflection vector of RIS, where β n ∈{0,1} and θ n ∈[0,2π](n=1,2,,N) respectively represent the reflection amplitude and phase of the nth unit of RIS; the received signal of LEO satellite in the qth (1≤q≤Q) time slot can be expressed as: Among them, diag() means vector diagonalization or diagonal matrix vectorization, s q,k Represents the pilot signal sent by the kth user, satisfying represents additive Gaussian white noise; let H k =Gdiag(h k ) represents the cascade channel of the k-th user signal reflected from the RIS to the base station. For multi-user channel estimation, the typical orthogonal pilot transmission strategy is widely used. For example, different users transmit pilot signals in orthogonal time slots, and the multi-user channel estimation can be separated into K independent executions of single-user channel estimation in sequence. The classic least squares channel estimation framework based on discrete Fourier transform is used to obtain the uplink cascade channel. When the k-th user completes Q (Q ≥ N) pilot transmission time slots, the pilot signal received by the LEO satellite is Then the estimated cascade channel can be expressed as: H k =Y k I H (TH) H ) -1 ; in is the DFT matrix The first N rows of , that is, the (p1, p2)th element of the RIS reflection matrix Θ satisfies 4. The method for channel prediction of a reconfigurable intelligent surface-assisted satellite communication system according to claim 1, characterized in that: The training data set required for constructing the intelligent channel prediction framework in step 2 includes the following steps: Step 2.1, propose an efficient channel prediction network by utilizing the estimated historical P u Frame uplink concatenation channel information To predict the future d Frame downlink cascade channel information Considering that existing neural networks mainly support real number operations, the input tensor of the channel prediction network is and data labels Designed to and The real form of and Respectively represent the operation of obtaining the real and imaginary parts of complex data, then the network input tensor is designed to be Data labels are designed to The neural network loss function is defined as: Among them, f w () represents a neural network model with a network weight of w, B represents the number of batch samples in one iteration of neural network parameters, and represents the 1 norm; represents the output of the neural network, then the final channel prediction result can be expressed as 5. The method for channel prediction of a reconfigurable intelligent surface-assisted satellite communication system according to claim 1, characterized in that: The intelligent channel prediction framework in step 3 is specifically as follows: Step 3.1, improve channel prediction accuracy; first, the input tensor A discrete Fourier transform is performed along the time domain dimension to generate frequency domain features, and then a frequency domain MLP architecture is developed to achieve global correlation modeling of cascaded channels; For the time domain input tensor H(t), its frequency domain tensor It is expressed as: According to the Pasval theorem in Fourier transform, the energy conservation formula is If most of the energy of a time series is concentrated on a few frequency components, then only using these components can accurately represent the time series; for RIS-assisted satellite-to-ground communication systems, the scatterers in the communication link are limited, and only the LoS channel and a limited number of paths in the channel matrix have energy; therefore, the frequency domain MLP architecture designed in this paper can well learn the effective characteristics of the cascade channel and improve the channel prediction performance; due to the actual digital communication system, In the time domain, it is a discrete time channel matrix, so DFT transform is used along the time domain to transform the input tensor for According to the conjugate symmetry property of DFT, we have Therefore, only use Half of the information All frequency domain information can be represented; Contains real and imaginary part information, a frequency domain MLP architecture that specializes in learning complex tensor features. The output features of its lth layer can be expressed as: Among them, σ l , and denote the activation function, complex weight and bias of the MLP layer l respectively; for two complex variables and Its complex multiplication satisfies Based on this complex number multiplication principle, can be rewritten as The frequency domain features are restored to the time domain using the inverse discrete Fourier transform, and then the output result F of the frequency domain MLP is obtained by concatenating the real and imaginary data of the time domain data into a real feature tensor, that is, 6. The method for channel prediction of a reconfigurable intelligent surface-assisted satellite communication system according to claim 1, characterized in that: The intelligent channel prediction framework in step 3 is specifically as follows: Step 3.2: Accelerate the convergence of the model. A multi-branch MLP architecture is developed to realize spatial feature extraction. In each branch, the cascade channel is feature extracted from different dimensions, including the satellite antenna dimension, the RIS unit dimension, the original channel dimension, and the 2D Fourier transform dimension. For the output F of the frequency domain MLP, a convolutional layer is first used to increase the dimension to obtain the input tensor of the multi-branch MLP. By introducing a dimensionality transformation factor S = C / C s ,(1≤C s ≤C), in branch 1, F i is transformed into a tensor Then, a weight is The linear layer is used to extract features, and finally restored to the original dimension to obtain the output features of branch one. In branch 2, F i is transformed into a tensor Then, a weight is The linear layer performs feature extraction and finally restores it to the original dimension to obtain the output features of branch 2 In branch three, input tensor F i No operation is performed to retain the original features; in branch 4, the spatial domain of the cascaded channel is first transformed using a two-dimensional discrete Fourier transform (2D DFT), i.e. A learnable filter Ψ is introduced to optimize the spectral representation of F4, namely The 2D inverse discrete Fourier transform is used to restore the spatial domain representation of the cascaded channels By concatenating the outputs of the four branches, we can get a feature tensor Then use a weight of The linear layer performs feature dimensionality reduction and performs feature fusion with the input tensor to obtain the final output tensor of the multi-branch MLP module. Right now 7. The method for channel prediction of a reconfigurable intelligent surface-assisted satellite communication system according to claim 1, characterized in that: The intelligent channel prediction framework in step 3 is specifically as follows: Step 3.3, reduce model complexity; the overall network architecture of the cascade channel prediction proposed in this paper adopts a D-layer encoder-decoder architecture to explore the potential sparsity of the input data and suppress the noise contained in the input data; first, the frequency domain MLP architecture is used to extract the input tensor The time domain features are then dimensionally upgraded to serve as the first layer input features of the encoder For the D-layer encoder, feature F1 e The M×N spatial dimension is gradually reduced to M / 2 by a sampling factor of 2. D ×N / 2 D , And F1 e The channel dimension of is gradually increased from C by a sampling factor of 2, and finally a feature tensor is obtained. The encoder of each layer contains a multi-branch MLP, a frequency domain MLP, and a convolution layer with a convolution kernel sliding step of 2 to implement the downsampling operation; the decoder has a symmetrical structure with the encoder, and the encoded features are converted into Restore to The decoder of each layer consists of a multi-branch MLP, a frequency domain MLP, and an upsampling layer based on bilinear interpolation; finally, a convolution kernel with P d The 1×1 convolutional layer is used to achieve time domain alignment, thereby obtaining the final output of the channel prediction network

Images