Satellite channel estimation method, system, equipment and medium

Abstract

The invention relates to the technical field of satellite communication, and discloses a satellite channel estimation method, system and device and a medium, and the method comprises the steps that a terminal device receives a first signal from a satellite; acquiring a corresponding first channel response based on the first signal; inputting the first channel response into a trained noise prediction model to obtain a first noise prediction value of noise contained in the first channel response; and performing denoising processing on the first channel response based on the first noise prediction value to obtain a denoised channel response estimation value. According to the invention, the precision and efficiency of satellite channel estimation are improved.

Description

Technical Field

[0001] This application relates to the field of satellite communication technology, and in particular to a satellite channel estimation method, system, device and medium. Background Technology

[0002] Existing low Earth Orbit (LEO) satellite channel estimation techniques typically fail to achieve ideal estimation accuracy in low signal-to-noise ratio environments, resulting in large channel estimation errors that severely impact system communication quality. They are also ill-suited to the high-speed movement and rapid changes characteristic of LEO satellite channels. Furthermore, existing technical solutions often rely on fixed models and assumptions, lacking flexibility and robustness. Summary of the Invention

[0003] In view of this, this application provides a satellite channel estimation method, system, device and medium.

[0004] This application discloses a satellite channel estimation method, which includes: The terminal device receives the first signal from the satellite; Based on the first signal, obtain the corresponding first channel response; The first channel response is input into the trained noise prediction model to obtain the first noise prediction value of the noise contained in the first channel response. The first channel response is denoised based on the first noise prediction value to obtain the denoised channel response estimate.

[0005] Furthermore, the training process of the noise prediction model includes: Obtain the training sample set; With the goal of minimizing the loss function, the noise prediction network is trained using the training sample set, and its network parameters are optimized to obtain a trained noise prediction model.

[0006] Furthermore, the process of obtaining the training sample set includes: Generate multiple second-channel responses from the satellite to the terminal device; Each of the plurality of second channel responses is subjected to noise step by step to obtain the corresponding plurality of third channel responses; The set of the multiple third channel responses is used as the training sample set of the noise prediction network; wherein each third channel response corresponds to a training sample in the training sample set.

[0007] Furthermore, the process of obtaining the loss function includes: Based on the predicted noise value, the actual noise, the physical consistency term, and the prior region regularization constraint, the loss function of the noise prediction network is constructed; where: The noise prediction value is the result output by the noise prediction network. The actual noise is the noise contained in each of the training samples; The physical consistency term is used to constrain the gradient smoothness of the output of the noise prediction network. The prior region regularization constraint is used to guide the noise prediction network to fuse context information during the process of recovering the second channel response from each of the training samples.

[0008] Furthermore, the process of obtaining the physical consistency item includes: Based on the noise prediction value and each training sample, a first channel response estimate is obtained; The second channel response estimate is obtained based on multiple first channel response estimates; The physical consistency term is obtained based on the preset first factor and the second channel response estimate.

[0009] Furthermore, the process of obtaining the prior region regularization constraints includes: Based on the noise prediction value and each training sample, a first channel response estimate is obtained; The second channel response estimate is obtained based on multiple first channel response estimates; The prior region regularization constraint is obtained based on the mask region subset and the second channel response estimate; the mask region subset is used to specify the location in the second channel response where a masking operation needs to be performed to simulate a physical scenario where some channel information is missing.

[0010] Furthermore, the noise prediction network includes an encoder, a connection structure, and a decoder connected in sequence; The encoder is used to perform layer-by-layer feature abstraction on each training sample in the input training sample set to obtain a high-level feature representation; The connection structure includes a first fully connected mapping layer, a nonlinear activation function, and a second fully connected mapping layer connected in sequence, which are used to perform dimensionality reduction, nonlinear transformation, and projection on the high-level feature representation. The decoder, whose structure is symmetrical to and functionally inverse of the encoder, is used to combine the delay and power distribution of the channel response to progressively reconstruct the features output by the connection structure, thereby obtaining an accurate estimate of the noise contained in each training sample.

[0011] Furthermore, the encoder includes a first embedding module, a first Swing Transformer module, a data integration layer, a first multilayer perceptron, and a second Swing Transformer module connected in sequence. The first embedding module is used to divide each training sample in the training sample set into data blocks and perform linear embedding to obtain an initial feature representation; The first Swin Transformer module is used to extract the correlation features of the satellite channel on the local time delay scale based on the initial feature representation through a window and shift window attention mechanism; the correlation features are used to indicate the primary feature representation of the satellite channel within the local time delay-time window; The data integration layer is used to downsample the correlation features to obtain a downsampled tensor; The first multilayer perceptron is used to perform position-by-position nonlinear transformation on the downsampled tensor; The second Swing Transformer module is used to fuse features between different local windows through window and shift window attention mechanisms to capture the power distribution features with long time delay span caused by multipath effects in the satellite channel and output a high-level feature representation.

[0012] Furthermore, the decoder includes a third Swing Transformer module, a second multilayer perceptron, a fourth Swing Transformer module, and a second embedding module connected in sequence. The third Swin Transformer module is used to perform preliminary reconstruction of the features output by the connection structure and restore the contextual dependencies of the features through window and shift window attention mechanisms. The second multilayer perceptron is used to perform nonlinear inverse transformation and dimensionality upscaling on the reconstructed features; The fourth Swing Transformer module is used to guide noise estimation to conform to the channel physical model through its window and shift window attention mechanism. The second embedding module reverses and projects the high-dimensional feature sequence output by the fourth Swing Transformer module to obtain a noise prediction value with the same dimension as the training sample.

[0013] This application also discloses a satellite channel estimation system, which includes: The signal receiving module is used by the terminal device to receive the first signal from the satellite; The data acquisition module is used to acquire the corresponding first channel response based on the first signal; The noise prediction module is used to input the first channel response into the trained noise prediction model to obtain the first noise prediction value of the noise contained in the first channel response. The channel response estimation module is used to perform denoising processing on the first channel response based on the first noise prediction value to obtain the denoised channel response estimate value.

[0014] This application also discloses an electronic device including a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, implements the method described above.

[0015] This application also discloses a computer-readable storage medium comprising a computer program or instructions that, when executed on a computer, cause the computer to perform the methods described above.

[0016] Due to the adoption of the above technical solution, this application has the following advantages: This application can better adapt to the rapid changes and dynamic environment of LEO satellite channels, and improve the accuracy and efficiency of satellite channel estimation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart illustrating a satellite channel estimation method according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the SwinCE-DM model according to an embodiment of this application; Figure 3 This is a block diagram of a satellite channel estimation system according to an embodiment of this application; Figure 4 This is a block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0019] The present application will be further described in conjunction with the accompanying drawings and embodiments. The described embodiments are only some, not all, of the embodiments of the present application. All other embodiments obtained by those skilled in the art should fall within the protection scope of the embodiments of the present application.

[0020] See Figure 1 This application provides an embodiment of a satellite channel estimation method, which includes: Step 101: The terminal device receives the first signal from the satellite; For example, the first signal is a downlink navigation / communication signal broadcast continuously or periodically by the satellite, containing satellite ephemeris, time information, and system parameters. Due to long transmission distances, large path losses, and the potential for multipath effects, atmospheric attenuation, or ground electromagnetic interference, the first signal is typically weak and has a low signal-to-noise ratio when it reaches the terminal device.

[0021] Step 102: Based on the first signal, obtain the corresponding first channel response; For example, assume the first signal is a downlink synchronization signal block (SSB) in a 5G NTN (non-terrestrial network). The terminal device first searches for and captures the SSB in the time-frequency domain, extracting the primary synchronization signal (PSS) and secondary synchronization signal (SSS) as known reference sequences. Subsequently, the received PSS / SSS sequence is cross-correlated with the local sequence in the time domain to obtain a time-domain impulse response. A noise threshold is set on this response (e.g., paths below -20dB of the peak value are zeroed), the three strongest multipath components are identified, and their relative delays and complex gains are recorded. This set of parameterized information serves as the first channel response in this embodiment. This response can be used for preliminary time synchronization, frequency synchronization, and can be used to assess the delay spread and multipath strength of the current link.

[0022] Step 103: Input the first channel response into the trained noise prediction model to obtain the first noise prediction value of the noise contained in the first channel response; In one embodiment of this application, the training process of the noise prediction model includes: Obtain the training sample set; With the goal of minimizing the loss function, the noise prediction network is trained using the training sample set, and its network parameters are optimized to obtain a well-trained noise prediction model.

[0023] In one embodiment of this application, the process of obtaining a training sample set includes: Generate multiple second-channel responses from the satellite to the terminal device; Noise is added to each of the multiple second channel responses step by step to obtain the corresponding multiple third channel responses; The set of multiple third-channel responses is used as the training sample set for the noise prediction network; where each third-channel response corresponds to a training sample in the training sample set.

[0024] In one embodiment of this application, the process of obtaining the loss function includes: Based on the predicted noise value, the actual noise, the physical consistency term, and the prior region regularization constraint, a loss function for the noise prediction network is constructed; where: The noise prediction value is the result output by the noise prediction network. The actual noise is the noise contained in each training sample; The physical consistency term is used to constrain the gradient smoothness of the output of the noisy prediction network. Prior region regularization constraints are used to guide the noise prediction network to fuse contextual information during the process of recovering the second channel response from each training sample.

[0025] In one embodiment of this application, the process of obtaining the physical consistency item includes: Based on the noise prediction value and each training sample, the first channel response estimate is obtained; The second channel response estimate is obtained based on multiple first channel response estimates; The physical consistency term is obtained based on the preset first factor and second channel response estimates.

[0026] In one embodiment of this application, the process of obtaining prior region regularization constraints includes: Based on the noise prediction value and each training sample, the first channel response estimate is obtained; The second channel response estimate is obtained based on multiple first channel response estimates; Based on the mask region subset and the second channel response estimate, the prior region regularization constraint is obtained; the mask region subset is used to specify the location in the second channel response where the masking operation needs to be performed, in order to simulate the physical scenario where some channel information is missing.

[0027] For example, multiple second-channel responses from a satellite to a terminal device can be generated in the following manner: The LEO satellite channel is modeled as follows: Regarding large-scale fading, according to 3GPP 38811-f40, the path loss calculation for non-terrestrial networks depends on the transmit gain of the satellite antenna and the receive gain of the terminal antenna. (dBi) can be represented as:

[0028] in, This represents the maximum transmit gain of the satellite antenna. Denotes the first-order Bessel function of the first kind. This indicates the angle between the ground node and the beam line of sight. For antenna aperture, For wave number.

[0029] The terminal antenna gain (dBi) can be expressed as:

[0030] in, For the antenna element's receiving gain, This represents the number of receiving antenna elements.

[0031] Regarding the calculation of path loss and shadow fading, the signal path between the satellite and the terminal undergoes several stages of propagation and attenuation. Path loss consists of the following components:

[0032] in, Total path loss, in dB. This represents the basic path loss, expressed in dB. The attenuation caused by the atmosphere is expressed in dB. The attenuation caused by ionospheric or tropospheric scintillation is expressed in dB. This represents the penetration loss of the building, measured in dB.

[0033] Basic path loss includes signal free space propagation loss (FSPL), ground feature loss (CL), and shadow fading (SF).

[0034]

[0035] Given the distance d between the user and the satellite and the signal frequency, the free space loss can be expressed as:

[0036] The distance d between the ground user and the satellite can be determined by the satellite altitude, the Earth's radius, and the user's elevation angle α to the satellite:

[0037] Based on the above equations and parameter settings, an overall channel modeling framework for the LEO scenario is formed, which generates time-delay-Doppler-frequency domain consistent channel responses (second channel responses) for multiple scenarios (such as sea surface, dense city, rural area, etc.) in the simulation platform.

[0038] For example, the third channel response is obtained based on the second channel response in the following manner: During forward diffusion, let the original channel response (second channel response) be... By every time Add Gaussian noise to construct a Markov chain:

[0039] in, Original channel response To the final signal The channel response diffusion process (of the third channel response); Noise scheduling factor; express arrive The probability density function obtained; The identity matrix is ​​used to represent the variance of the noise. The equation of a Markov chain is equivalent to... Conditional Gaussian distribution, used to describe signals Given the signal from the previous time moment The distribution below.

[0040] Forward diffusion process in the form of a Markov chain This study characterizes the joint distribution of the channel response throughout the progressive noise addition process from a probabilistic modeling perspective. Due to the transition at each step... All variables are conditionally Gaussian distributed. This Markov chain has an analyzable marginalization property, meaning that intermediate variables can be marginalized. Eliminating the noise allows us to obtain the noisy channel response at any given time relative to the original channel response. Closed edge distribution:

[0041] in, This represents the cumulative signal retention coefficient from the initial time to step t. Based on the above closed-edge distribution form, the random sampling process can be explicitly represented using the reparameterization trick as follows:

[0042] in, It is the signal during the diffusion process, representing the original signal during forward diffusion. and noise According to a certain weighting factor The combined signal. It is a parameter that controls the amount of noise added, and it usually changes with time step.

[0043] Based on the forward diffusion process, the noisy channel tensor was obtained. Noisy channel tensor It will be directly applied to the backsampling process.

[0044] During the reverse sampling process, the response from the noisy channel Starting with using neural networks to predict noise terms Step by step estimate Finally, the estimated channel response (first channel response estimate) is obtained. :

[0045] in, For fixed or learnable inverse variance; The output of the noise prediction network is denoted as , where represents the noise component in the original channel response predicted by the noise prediction network at time t. denoted as the noise standard deviation during the reverse sampling process, and z as the noise term of the standard normal distribution, used to weight the noise components during the reverse process.

[0046] The channel response after noise removal can be obtained from the above formula. Ultimately multiple This constitutes the final channel estimation tensor (the second channel response estimate). .

[0047] To ensure the output of the noise prediction network To improve the accuracy of the noise prediction network, the loss function and training objective are given below: The KL divergence minimization objective can be rewritten as a noise prediction loss function in the expected form:

[0048] in, KL divergence, representing the difference between the actual noise distribution and the noise distribution predicted by the noise prediction network, This represents the actual noise distribution during the forward diffusion process. The noise distribution predicted by the network; The noise is predicted by the noise prediction network, which is based on the current signal response and time step.

[0049] The ultimate training objective is:

[0050] in, This represents the expectation of the noise predicted for all channel responses and corresponding time steps, i.e., calculating the loss over all data and time steps.

[0051] The original channel response is forward-diffused to add noise, and the noisy channel tensor is backsampled. During the backsampling process, noise is predicted using a noise prediction network. Noise reduction is performed to ultimately recover the accurate channel response.

[0052] In one embodiment of this application, the noise prediction network includes an encoder, a connection structure, and a decoder connected in sequence. An encoder is used to perform layer-by-layer feature abstraction on each training sample in the input training sample set to obtain a high-level feature representation; The connection structure includes a first fully connected mapping layer, a nonlinear activation function, and a second fully connected mapping layer connected in sequence, which are used to perform dimensionality reduction, nonlinear transformation, and projection on the high-level feature representation; The decoder, whose structure is symmetrical to and functionally inverse of the encoder, is used to combine the delay and power distribution of the channel response to progressively reconstruct the features output by the connected structure, thereby obtaining an accurate estimate of the noise contained in each training sample.

[0053] In one embodiment of this application, the encoder includes a first embedding module, a first SwinTransformer module, a data integration layer, a first multilayer perceptron, and a second Swin Transformer module connected in sequence. The first embedding module is used to divide each training sample in the training sample set into data blocks and perform linear embedding to obtain an initial feature representation. The first Swing Transformer module is used to extract the correlation features of the satellite channel at the local time delay scale through window and shift window attention mechanisms on the initial feature representation; the correlation features are used to indicate the primary feature representation of the satellite channel within the local time delay-time window. The data integration layer is used to downsample the relevant features to obtain the downsampled tensor; The first multilayer perceptron is used to perform position-by-position nonlinear transformation on the downsampled tensor; The second Swing Transformer module is used to fuse features between different local windows through window and shift window attention mechanisms to capture the power distribution features with long time delay spans caused by multipath effects in the satellite channel and output high-level feature representations.

[0054] In one embodiment of this application, the decoder includes a third Swing Transformer module, a second multilayer perceptron, a fourth Swing Transformer module, and a second embedding module connected in sequence. The third Swin Transformer module is used to perform preliminary reconstruction of the features output by the connection structure through window and shift window attention mechanisms, and to restore the contextual dependencies of the features. The second multilayer perceptron is used to perform nonlinear inverse transformation and dimensionality increase on the reconstructed features; The fourth Swing Transformer module is used to guide noise estimation to conform to the channel physical model through its window and shift window attention mechanism; The second embedding module reverses and projects the high-dimensional feature sequence output by the fourth Swing Transformer module to obtain noise prediction values ​​with the same dimension as the training samples.

[0055] For example, to efficiently capture local delay structures and cross-time global correlation features in dynamic channel responses, a stacked Swing Transformer is introduced as the backbone feature extractor in the coding path. See [link to relevant documentation] Figure 2 The SwinCE-DM (Swin Transformer for Channel Estimation with Diffusion Model) model maintains a spatially symmetric structure, supports multi-scale modeling and cross-delay interaction, and significantly improves noise inference capabilities under high dynamic conditions. During forward diffusion, the original channel response... Gaussian noise is gradually added to generate a noisy tensor. This is used to simulate perturbation mechanisms in real-world channels. The noisy tensor is deeply modeled using an encoder. Specifically, the noisy tensor sequentially passes through a first embedding module, a first Swin Transformer module, a data integration layer, a first multilayer perceptron, a second Swin Transformer module, a first fully connected mapping layer, and a nonlinear activation function to extract contextual representations. The first and second Swin Transformer modules are mainly used to compute self-attention within local windows and interact with information through shifted windows. During backsampling, the noise prediction network at each time step... Input current channel response With current time Predict its corresponding noise The estimated channel response is generated using an inverse diffusion method within the framework of the Denoising Diffusion Probabilistic Model (DDPM). The Swin Transformer module maintains a spatially symmetrical structure and effectively captures channel delay and power distribution through the Swing Block decoding process, achieving high-fidelity reconstruction with low errors. The entire decoding process can be understood as a Patch Division process, which functions to divide the high-dimensional features obtained from the encoder. According to the original channel response The dimensions are inversely rearranged into low-dimensional feature blocks to restore the original channel dimensions in order to output the estimated channel response. High-dimensional features The feature space comprehensively characterizes the overall structural information of the channel response in the time and delay dimensions, including the relative distribution of multipath components, delay aggregation characteristics and their correlation with time.

[0056] Specifically, the encoder is designed as follows: The encoder perturbs the input. Starting with the first embedding module, the token representation (initial feature representation) is mapped to linear embedding through partitioning:

[0057] in, Indicates the size of the patch (data block). Let R represent the embedding dimension, and R represent the real number field. This represents the number of tokens obtained after dividing the data using patches over a time dimension. This indicates the number of tokens after partitioning in terms of time delay dimension. Indicates the length of the time window. For the number of discrete delay points, For linear embedding mapping, For rearrangement operations, perturb the input. This represents the channel data after noise has been added through the diffusion process. This is the initial feature representation after data block partitioning and linear embedding, reflecting the local features of each data block; the local features comprehensively characterize the energy distribution, time delay clustering, and time-varying local structure of the multipath components within the mask region.

[0058] Each Swing Transformer Block (any one of the first, second, third, and fourth Swing Transformer modules) contains two layers of local attention mechanisms and a feedforward network structure. Let the input be... The attention mapping process is as follows:

[0059] in, and This represents the intermediate output within each Swing Transformer Block. W-MSA and SW-MSA are the window and shift window attention mechanisms, respectively. Representation layer normalization, It consists of a two-layer feedforward network, which ultimately produces the output. ; This represents the primary characteristics (correlation characteristics) of the channel within the local delay-time window.

[0060] To achieve cross-scale abstraction, a data integration layer (Patch Merging) is used between the first and second Swin Transformer modules to perform spatial downsampling and channel dimension expansion:

[0061] in, express The downsampled tensor after the Patch Merging layer has a higher level of abstraction; and They represent Spatial resolution in the time and latency dimensions This indicates the number of corresponding feature channels.

[0062] Feature extraction is then performed using the second Swing Transformer module. The processing flow of the second Swing Transformer module is consistent with that of the first Swing Transformer module, with the input being... Finally, the context representation (high-level feature representation) is extracted. To accommodate the high-dimensional characteristics of delay structure and power distribution. Used to comprehensively characterize the global delay structure and power distribution characteristics of the channel.

[0063] High-level contextual features of encoder output Dimension alignment and feature compression are performed through a first fully connected mapping layer (FC), and a nonlinear activation function S is introduced to enhance representational power. Finally, a second fully connected mapping layer is used to obtain the latent variable representation for the backdiffusion process, providing conditional input for the subsequent noise inversion module. .

[0064] The main goal of the decoder is to model and predict the noise introduced during the diffusion process. Its overall structure is symmetrical to that of the encoder in terms of hierarchical organization and feature processing flow.

[0065] in, This represents the noise prediction result generated by the decoder; This represents the decoder section, which is responsible for predicting noise from the input features. SwinBlock represents a feature transformation module based on a hierarchical window self-attention mechanism, used to model the correlation between channel features within a local window and to achieve cross-regional information interaction through window shifting.

[0066] To achieve high-precision recovery of the original channel response, this application proposes the SwinCE-DM model, which replaces the traditional U-Net noise prediction network with a Swin Transformer within the diffusion model framework. Compared to the original diffusion model based on DDPM, SwinCE-DM uses DDIM as its basic framework. While ensuring efficient backsampling, it introduces a hierarchical window attention mechanism corresponding to each Swin Transformer module: within the same Swin Transformer module, W-MSA models the local delay characteristics of subcarriers multiplied by the delay grid within a fixed window; different Swin Transformer modules expand the receptive field step by step through the shift window of SW-MSA combined with downsampling / upsampling, and combine time-step coding to capture long-range time dependencies across time slots and diffusion steps, thereby more robustly completing the noise residual prediction for channel estimation.

[0067] The SwinCE-DM architecture enhances the network's ability to express and generalize complex LEO satellite channel characteristics, resulting in superior estimation performance and robustness when facing typical scenarios such as multipath fading and frequency selectivity. This optimization not only improves the accuracy of channel reconstruction but also significantly reduces the dependence on hardware resources, meeting the engineering requirements of LEO satellite communication systems for low latency and high real-time performance.

[0068] The channel estimation model proposed in this application can more effectively capture the local delay characteristics and long-range correlations in LEO satellite channels, thereby achieving more accurate and robust channel reconstruction under high dynamics and low pilot configuration. Combined with DDPM's efficient backsampling mechanism, SwinCE-DM significantly improves estimation accuracy and generalization ability while maintaining low computational complexity.

[0069] The embodiments of this application design a clear input-output tensor specification and a unified loss function system, which integrates the requirements of physical consistency and diffusion modeling.

[0070] Each time step of the diffusion process The input tensor is defined as:

[0071] in, The 2 in the equation represents the concatenation of the real and imaginary parts of the complex channel. This indicates the number of samples processed in parallel during each iteration. The time window length (the number of time sampling points for observation) corresponds to the "time dimension" resolution. This represents the number of discrete time delay taps.

[0072] After residual propagation through the network encoder and decoder, the noise prediction network outputs the channel response estimate (second channel response estimate) for the corresponding time step: During the training phase, the prediction noise network uses the noise prediction error as the main supervision signal and introduces a physical consistency term. With prior domain regularization constraints Construct the following loss function:

[0073] in, This represents the expectation over all training samples, noise terms, and time steps. To deduce an approximate noise-free tensor from the noise prediction results; This represents a regularization term, designed to constrain the gradient of a noisy prediction network. Represents the local gradient operator; This indicates that additional constraints on the noise prediction network are used to guide it to consider more contextual information when recovering the signal. These are the weighting coefficients for the masked region reconstruction loss term, used to balance the noise prediction loss and the mask constraint term. The relative contribution to the overall loss function; For the cross-entropy reconstruction term of the mask region, Indicates the original channel response The above describes a subset of masked regions obtained according to a predefined masking rule, used to simulate scenarios where some channel information is missing or unobservable. For the predefined masking rule, we use a continuous delay block masking rule, selecting several consecutive tap intervals as masked regions along the delay dimension. During the training of the noise prediction network, when the total number of delay taps can be 64, the length of the continuous mask interval can be set to 8–16 taps. By minimizing the cross-entropy reconstruction term, the noise prediction network can more accurately recover the masked data.

[0074] The aforementioned loss function is applied to channel estimation tasks in low-Earth orbit satellite communication scenarios. The final channel response is derived from the noise prediction results. Apply the following constraints:

[0075] in, The gradient constraint threshold is used to limit the magnitude of gradient changes in the channel estimation results in the time and delay dimensions, so as to avoid drastic fluctuations that do not conform to physical characteristics.

[0076] A reconstruction error constraint is introduced within the mask region, which satisfies:

[0077] in, The upper bound threshold for the reconstruction error of the masked region is used to limit the prediction error of the noise prediction network in this region to not exceed a preset range, thereby guiding the noise prediction network to make full use of context information to complete channel recovery.

[0078] In summary, during training, minimizing the aforementioned loss function... For model parameters The optimization process minimizes the overall loss function itself, and the direct output of the noise prediction network is the predicted noise result at the corresponding time step. Ultimately, the noise prediction network can more fully utilize contextual information to complete channel recovery while meeting the physical characteristics of low-Earth orbit satellite communication scenarios, thereby improving the generalization ability and reconstruction accuracy of the overall channel estimation.

[0079] This application's embodiments, through a gradual addition of noise and denoising process, can effectively recover the channel response in complex channel environments, especially performing exceptionally well in low signal-to-noise ratio (SNR) environments. Compared to existing technologies, this application's embodiments, by introducing a channel estimation model, provide stronger robustness and accuracy in channel estimation, significantly improving performance in dynamic and complex channel environments.

[0080] Step 104: Denoise the first channel response based on the first noise prediction value to obtain the denoised channel response estimate.

[0081] For example, the process of obtaining the final channel response estimate from the noise prediction value output by the third channel response and noise prediction network, as given in the foregoing embodiments, can be referred to to obtain the denoised channel response estimate.

[0082] In traditional DDIM, the sampling step size in the inverse denoising process is fixed. As a variation, an adaptive sampling step size strategy can be designed to dynamically adjust the sampling step size based on signal complexity and noise levels during different iterations. This variation helps improve sampling efficiency and reduce computational complexity in complex channel environments. Furthermore, this variation better adapts to the channel characteristics of different environments, improving performance in dynamic and rapidly changing LEO satellite channels while reducing computational overhead.

[0083] In the technical solution of this application, an architecture combining DDIM and the Swing Transformer module is adopted to achieve channel estimation. Other types of generative models, such as Variational Autoencoders (VAEs) or Generative Adversarial Networks (GANs), can be considered for integration with the Swing Transformer module to provide different channel modeling approaches, exhibiting different advantages in channel estimation tasks, especially in scenarios with complex channel distribution or scarce samples. By combining the powerful feature extraction capabilities of the Swing Transformer module with the modeling advantages of generative models, the channel recovery accuracy in low SNR and high-noise environments can be further improved.

[0084] In addition to using the Swin Transformer module, other neural network architectures, such as Long Short-Term Memory (LSTM) networks or Transformer variants like ViT, can be integrated to jointly learn temporal channel variations and spatial channel characteristics. LSTM effectively models dynamic changes in time series, while the Transformer structure excels at capturing long-range dependencies and global features. Through collaborative design with the multi-level windowing mechanism of the Swin Transformer, this joint architecture enhances the ability to characterize the spatiotemporal variations of LEO satellite channels, thereby improving the accuracy and robustness of estimations, making it particularly suitable for channel scenarios in high-speed or dynamic environments.

[0085] This application is the first to apply a diffusion model to LEO satellite channel estimation. Traditional channel estimation methods struggle to cope with the rapid changes and high dynamism of LEO satellite channels, while the diffusion model, through a gradual noise addition and denoising process, can effectively recover the true channel response. By customizing the modeling of LEO satellite channel characteristics, the diffusion model can accurately capture the complex characteristics of the channel, improve channel estimation accuracy, and possess stronger robustness, adapting to rapidly changing channel environments. This is an innovative solution that existing technologies cannot provide.

[0086] The improved architecture SwinCE-DM, based on the Swin Transformer module, utilizes its hierarchical window attention mechanism to capture the local and global structural features of channel data, thereby enhancing the network's ability to perceive channel delay characteristics, frequency selectivity, and spatial correlation. This improved architecture not only enhances channel estimation accuracy but also possesses stronger generalization capabilities and structural flexibility, making it particularly suitable for resource-constrained terminals and low signal-to-noise ratio environments. Thanks to the efficient attention computation and strong representation capabilities of the Swin Transformer module, stable recovery of channel response under dynamic channels can be achieved while maintaining model computational efficiency, effectively addressing the insufficient accuracy and robustness of existing diffusion estimation models in LEO communication scenarios.

[0087] See Figure 3 This application also provides a satellite channel estimation system, which includes: The signal receiving module is used by the terminal device to receive the first signal from the satellite; The data acquisition module is used to acquire the corresponding first channel response based on the first signal; The noise prediction module is used to input the first channel response into the trained noise prediction model to obtain the first noise prediction value of the noise contained in the first channel response. The channel response estimation module is used to perform denoising processing on the first channel response based on the first noise prediction value to obtain the denoised channel response estimate value.

[0088] See Figure 4 This application also provides an electronic device including a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the methods described in the above embodiments. As an example, the electronic device may include multiple processors. A processor may refer to one or more devices, circuits, and / or computing units for processing data (e.g., computer programs). The processor can invoke the computer program stored in the memory to implement the methods described in the above embodiments. Figure 4 Taking an electronic device consisting of one processor and one memory as an example, the processor and memory are used to indicate a type of device or equipment, and the quantity of each type of device or equipment can be determined according to business needs.

[0089] This application also provides a computer-readable storage medium, which includes a computer program or instructions that, when executed on a computer, cause the computer to perform the methods described in the above embodiments.

[0090] It should be noted that, in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0091] Those skilled in the art should clearly understand that, for the sake of convenience and brevity, the specific working processes of the satellite channel estimation system, electronic equipment, and computer-readable storage medium described in the above embodiments can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0092] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware, or by a program instructing the relevant hardware to implement them. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0093] The above are merely optional embodiments of this application, used only to illustrate the technical solution of this application and not to limit it. Any modifications, equivalent substitutions, improvements, etc., to the specific implementation of this application without departing from the spirit and scope of this application should be covered within the protection scope of this application.

Claims

1. A satellite channel estimation method, characterized in that, include: The terminal device receives the first signal from the satellite; Based on the first signal, obtain the corresponding first channel response; The first channel response is input into the trained noise prediction model to obtain the first noise prediction value of the noise contained in the first channel response. The first channel response is denoised based on the first noise prediction value to obtain the denoised channel response estimate.

2. The method according to claim 1, characterized in that, The training process of the noise prediction model includes: Obtain the training sample set; With the goal of minimizing the loss function, the noise prediction network is trained using the training sample set, and its network parameters are optimized to obtain a trained noise prediction model.

3. The method according to claim 2, characterized in that, The process of obtaining the training sample set includes: Generate multiple second-channel responses from the satellite to the terminal device; Each of the plurality of second channel responses is subjected to noise step by step to obtain the corresponding plurality of third channel responses; The set of the multiple third channel responses is used as the training sample set of the noise prediction network; wherein each third channel response corresponds to a training sample in the training sample set.

4. The method according to claim 2, characterized in that, The process of obtaining the loss function includes: Based on the predicted noise value, the actual noise, the physical consistency term, and the prior region regularization constraint, the loss function of the noise prediction network is constructed; where: The noise prediction value is the result output by the noise prediction network. The actual noise is the noise contained in each of the training samples; The physical consistency term is used to constrain the gradient smoothness of the output of the noise prediction network. The prior region regularization constraint is used to guide the noise prediction network to fuse context information during the process of recovering the second channel response from each of the training samples.

5. The method according to claim 4, characterized in that, The process of obtaining the physical consistency item includes: Based on the noise prediction value and each training sample, a first channel response estimate is obtained; The second channel response estimate is obtained based on multiple first channel response estimates; The physical consistency term is obtained based on the preset first factor and the second channel response estimate.

6. The method according to claim 4 or 5, characterized in that, The process of obtaining the prior region regularization constraints includes: Based on the noise prediction value and each training sample, a first channel response estimate is obtained; The second channel response estimate is obtained based on multiple first channel response estimates; The prior region regularization constraint is obtained based on the mask region subset and the second channel response estimate; the mask region subset is used to specify the location in the second channel response where a masking operation needs to be performed to simulate a physical scenario where some channel information is missing.

7. The method according to claim 2, characterized in that, The noise prediction network includes an encoder, a connection structure, and a decoder connected in sequence. The encoder is used to perform layer-by-layer feature abstraction on each training sample in the input training sample set to obtain a high-level feature representation; The connection structure includes a first fully connected mapping layer, a nonlinear activation function, and a second fully connected mapping layer connected in sequence, which are used to perform dimensionality reduction, nonlinear transformation, and projection on the high-level feature representation. The decoder, whose structure is symmetrical to and functionally inverse of the encoder, is used to combine the delay and power distribution of the channel response to progressively reconstruct the features output by the connection structure, thereby obtaining an accurate estimate of the noise contained in each training sample.

8. The method according to claim 7, characterized in that, The encoder includes a first embedding module, a first Swing Transformer module, a data integration layer, a first multilayer perceptron, and a second Swing Transformer module connected in sequence. The first embedding module is used to divide each training sample in the training sample set into data blocks and perform linear embedding to obtain an initial feature representation; The first Swin Transformer module is used to extract the correlation features of the satellite channel on the local time delay scale based on the initial feature representation through a window and shift window attention mechanism; the correlation features are used to indicate the primary feature representation of the satellite channel within the local time delay-time window; The data integration layer is used to downsample the correlation features to obtain a downsampled tensor; The first multilayer perceptron is used to perform position-by-position nonlinear transformation on the downsampled tensor; The second Swing Transformer module is used to fuse features between different local windows through window and shift window attention mechanisms to capture the power distribution features with long time delay span caused by multipath effects in the satellite channel and output a high-level feature representation.

9. The method according to claim 7, characterized in that, The decoder includes a third Swing Transformer module, a second multilayer perceptron, a fourth Swing Transformer module, and a second embedding module connected in sequence. The third Swin Transformer module is used to perform preliminary reconstruction of the features output by the connection structure and restore the contextual dependencies of the features through window and shift window attention mechanisms. The second multilayer perceptron is used to perform nonlinear inverse transformation and dimensionality upscaling on the reconstructed features; The fourth Swing Transformer module is used to guide noise estimation to conform to the channel physical model through its window and shift window attention mechanism. The second embedding module reverses and projects the high-dimensional feature sequence output by the fourth Swing Transformer module to obtain a noise prediction value with the same dimension as the training sample.

10. A satellite channel estimation system, characterized in that, include: The signal receiving module is used by the terminal device to receive the first signal from the satellite; The data acquisition module is used to acquire the corresponding first channel response based on the first signal; The noise prediction module is used to input the first channel response into the trained noise prediction model to obtain the first noise prediction value of the noise contained in the first channel response. The channel response estimation module is used to perform denoising processing on the first channel response based on the first noise prediction value to obtain the denoised channel response estimate value.

11. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, implements the method of any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program or instructions that, when executed on a computer, cause the computer to perform the method of any one of claims 1-9.

Images