Method and device for predicting RSMA beamforming in non-terrestrial network based on deep learning

Abstract

The application relates to the field of beamforming control, and discloses a non-ground network RSMA beamforming prediction method and device based on deep learning, which comprises the following steps: acquiring historical CSI information of multiple channels; determining precoder construction parameters and common rate allocation parameters for beamforming of the multiple channels according to the historical CSI information of the multiple channels through a beamforming prediction model; wherein the beamforming prediction model comprises a Transformer unit, a CNN unit and a generation unit, the Transformer unit is used for extracting CSI time-dependent features in historical CSI data of the multiple channels, the CNN unit is used for extracting spatial correlation features between the historical CSI data of the multiple channels, and the generation unit is used for generating the precoder construction parameters and the common rate allocation parameters for the beamforming of the multiple channels. The application extracts space-time feature beamforming prediction from historical CSI through a transformer and a convolutional neural network, thereby improving the application effect of the RSMA technology.

Description

Technical Field

[0001] This application relates to the field of beamforming control technology, and more specifically, to a method and apparatus for RSMA beamforming prediction in non-terrestrial networks based on deep learning. Background Technology

[0002] In communication between low Earth orbit (LEO) satellites and unmanned aerial vehicles (UAVs), rate-splitting multiple access (RSMA) technology enables the feasibility of non-terrestrial networks (NTNs) due to its interference management and reliable data transmission. However, the precoder design in RSMA technology still relies on accurate channel state information (CSI) feedback and complex optimization, which poses a significant challenge to the practical deployment of RSMA technology.

[0003] In the communication process between low-Earth orbit (LEO) satellites and drones, beamforming is necessary to ensure communication quality. In traditional predictive beamforming designs, when the LEO satellite performs beamforming prediction for each drone, in the previous round of communication between the LEO satellite and the drone (corresponding to time t-1), each drone first reports its own Channel State Information (CSIT) to the LEO satellite. After collecting the CSIs of all drones reported in the previous round of communication (corresponding to time t-1), the LEO satellite continues to predict the Channel State Information at the transmitter (CSIT) of the drone in the next round of communication (corresponding to time t+1). This allows the LEO satellite to utilize the CSIT prediction results from the prediction phase to coordinate and reduce channel interference between multiple drones using RSMA technology. However, in the process of achieving beamforming and reducing channel interference through RSMA using this beamforming prediction method, data needs to be transmitted between different modules, resulting in long data processing times and ultimately poor application effectiveness of RSMA technology in reducing channel interference. Summary of the Invention

[0004] The purpose of this application is to provide a method and apparatus for RSMA beamforming prediction in non-terrestrial networks based on deep learning. This method solves the technical problems of traditional beamforming prediction, such as the need for data to be transmitted between different modules and the long data processing time. It achieves the technical effect of improving the application effect of RSMA technology by extracting spatiotemporal features from historical CSI through transformers and convolutional neural networks for beamforming prediction.

[0005] This application provides a deep learning-based RSMA beamforming prediction method for non-terrestrial networks. The method includes: acquiring historical CSI information of multiple channels; and determining precoder construction parameters and common rate allocation parameters for beamforming multiple channels based on the historical CSI information of multiple channels using a beamforming prediction model. The beamforming prediction model includes a Transformer unit, a CNN unit, and a generation unit. The Transformer unit is used to extract CSI time-dependent features from the historical CSI data of multiple channels, the CNN unit is used to extract spatial correlation features between the historical CSI data of multiple channels, and the generation unit is used to generate precoder construction parameters and common rate allocation parameters for beamforming multiple channels.

[0006] In one possible implementation, the method further includes: representing the real and imaginary parts of the historical CSI information of multiple channels using real numbers, and inputting the real and imaginary parts of the historical CSI information represented by real numbers into the Transformer unit of the beamforming prediction model; the generation unit outputs the complex-valued matrix corresponding to the precoder construction parameters used for beamforming multiple channels.

[0007] In another possible implementation, the Transformer unit includes a multi-head attention layer, a first residual connection and a normalization layer, a feedforward network layer and a second residual connection and a normalization layer arranged in sequence; the CNN unit includes a convolutional layer, a pooling layer and a flattening layer arranged in sequence; and the generation unit includes a fully connected layer and a generation layer.

[0008] In another possible implementation, the common rate allocation parameters output by the beamforming prediction model are normalized and the total power limit of the low-Earth orbit satellite is satisfied.

[0009] In another possible implementation, when the common rate allocation parameters output by the beamforming prediction model are normalized and the total power limit of low-Earth orbit satellites is satisfied, it can be expressed by the following formula:

[0010]

[0011] in, This represents the channel between a low Earth orbit (LEO) satellite and the k-th UAV, where k represents the UAV number or channel number, γ k This represents the signal-to-noise ratio (SINR) of the common stream decoded by the drone or channel k. This indicates the total number of drones. Let o represent the transposed channel vector between the low-orbit satellite and the k-th UAV. c Indicates a common linear precoder. Let σ represent the m-th private linear precoder.k Let represent the variance of the distribution of additive white Gaussian noise at the k-th drone. Indicates instantaneous reachability. O represents the complex-valued matrix output by the beamforming prediction model, and O represents the precoding matrix. Indicates output The calculated achievable common rate, This represents the achievable common rate set by the beamforming prediction model, and min indicates finding the minimum value. Indicates the parameter size is Located in the set of complex numbers, C k This represents the common rate portion of the k-th drone. P represents the normalized absolute value, tr represents the trace of the solution matrix, and P represents the trace of the solution matrix. t This indicates the launch power of low-Earth orbit satellites.

[0012] In another possible implementation, the method for training the beamforming prediction model on-orbit for a low-Earth orbit (LEO) satellite includes: acquiring historical CSI information for multiple channels; training the beamforming prediction model on-orbit based on the historical CSI information for multiple channels; wherein the loss function of the beamforming prediction model is expressed by the following formula:

[0013]

[0014] Where L represents the loss function value, u represents the expected total downlink rate obtained by the k-th drone. k The weights representing the obtained rates, and Φ representing the number of samples in the historical CSI information. C represents the instantaneous rate obtained under the corresponding sample. k This represents the common rate portion of the k-th drone, where k represents the drone number or channel number. Indicates the total number of drones;

[0015] The process of updating the beamforming prediction model can be represented by the following formula:

[0016] w n+1 =w n -η n g n

[0017] η n =η0*λ n

[0018] Where, η n Let η0 represent the learning rate in the nth training round, and λ0 represent the initial learning rate. n w represents the decay rate of the learning rate.n w represents the weight parameters of the global model in the nth round of training. n+1 g represents the weight parameters of the global model in the (n+1)th training round. n This represents a local estimate of the gradient of the loss function in satellite s.

[0019] This application also provides an RSMA beamforming prediction apparatus for non-terrestrial networks based on deep learning, including a unit for performing the method described in any of the preceding claims.

[0020] This application also provides an RSMA beamforming prediction device for non-terrestrial networks based on deep learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method described in any of the preceding claims.

[0021] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any of the preceding claims.

[0022] This application also provides a low-orbit satellite that uses the RSMA beamforming prediction method in a non-terrestrial network based on deep learning as described in any of the preceding embodiments when performing beamforming prediction.

[0023] This application also provides a method for training a low-Earth orbit satellite in orbit using the beamforming prediction model as described above.

[0024] The beneficial effects of the embodiments of this application compared with the prior art are:

[0025] This application provides a deep learning-based RSMA beamforming prediction method for non-terrestrial networks. The method includes: acquiring historical CSI information for multiple channels; and determining precoder construction parameters and common rate allocation parameters for beamforming multiple channels based on the historical CSI information using a beamforming prediction model. The beamforming prediction model includes a Transformer unit, a CNN unit, and a generation unit. The Transformer unit extracts CSI time-dependent features from the historical CSI data of multiple channels, the CNN unit extracts spatial correlation features between the historical CSI data of multiple channels, and the generation unit generates the precoder construction parameters and common rate allocation parameters for beamforming multiple channels. The method in this application can accurately predict beamforming, reduce data transmission time, and improve the application effect of reducing channel interference through RSMA technology. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart illustrating a deep learning-based RSMA beamforming prediction method in a non-terrestrial network, provided as an embodiment of this application;

[0028] Figure 2 A schematic diagram illustrating the working scenario of a deep learning-based RSMA beamforming prediction method in a non-terrestrial network, provided in an embodiment of this application;

[0029] Figure 3 This is a schematic diagram of the structure of a beamforming prediction model provided in an embodiment of this application;

[0030] Figure 4 A performance comparison chart of an RSMA beamforming prediction method based on deep learning in a non-terrestrial network provided in this application embodiment;

[0031] Figure 5 A flowchart illustrating an on-orbit training method for a beamforming prediction model provided in an embodiment of this application;

[0032] Figure 6 A schematic diagram of the logic structure of an RSMA beamforming prediction device in a non-terrestrial network based on deep learning, provided for an embodiment of this application;

[0033] Figure 7 This is a schematic diagram of the physical structure of an RSMA beamforming prediction device in a non-terrestrial network based on deep learning, provided as an embodiment of this application. Detailed Implementation

[0034] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0035] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0036] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0037] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0038] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0039] In existing technologies, when beamforming is achieved through beamforming prediction methods and channel interference is reduced using RSMA technology, data needs to be transmitted between different modules, resulting in long data processing times and poor application performance of reducing channel interference using RSMA technology.

[0040] Based on the above reasons, this application provides a deep learning-based RSMA beamforming prediction method for non-terrestrial networks. This method includes: acquiring historical CSI information for multiple channels; and determining precoder construction parameters and common rate allocation parameters for beamforming multiple channels based on the historical CSI information using a beamforming prediction model. The beamforming prediction model includes a Transformer unit, a CNN unit, and a generation unit. The Transformer unit extracts CSI time-dependent features from the historical CSI data of multiple channels, the CNN unit extracts spatial correlation features between the historical CSI data of multiple channels, and the generation unit generates the precoder construction parameters and common rate allocation parameters for beamforming multiple channels. The method in this application can accurately predict beamforming, reduce data transmission time, and improve the application effect of reducing channel interference through RSMA technology.

[0041] In some scenarios, the RSMA beamforming prediction method based on deep learning in non-terrestrial networks according to embodiments of this application can be applied to RSMA-supported communication processes in non-terrestrial networks composed of low-Earth orbit satellites and UAVs. It can accurately predict beamforming, reduce data transmission time, and improve the application effect of reducing channel interference through RSMA technology.

[0042] In other scenarios, the deep learning-based RSMA beamforming prediction method in non-terrestrial networks according to embodiments of this application can also be applied to the communication process between low-Earth orbit satellites and ground mobile terminals that support RSMA. It can accurately predict beamforming during the communication process between low-Earth orbit satellites and ground mobile terminals, reduce data transmission time, and improve the application effect of reducing channel interference through RSMA technology.

[0043] The following specific examples illustrate a deep learning-based RSMA beamforming prediction method for non-terrestrial networks provided in this application.

[0044] Figure 1 A flowchart illustrating a deep learning-based RSMA beamforming prediction method in a non-terrestrial network, as provided in this application embodiment, is shown below. Figure 1 As shown, this method includes S110 to S120, and S110 to S120 will be described in detail below.

[0045] Figure 2 This application provides a schematic diagram illustrating the working scenario of a deep learning-based RSMA beamforming prediction method in a non-terrestrial network, as shown in the embodiments of this application. Figure 2 As shown, the method in this embodiment can be applied to communication between low-Earth orbit satellites and multiple drones. When communicating, low-Earth orbit satellites and multiple drones can use Rate Split Multiple Access (RSMA) technology to manage interference and ensure reliable data transmission, thus ensuring the feasibility of non-terrestrial networks (NTNs).

[0046] S110: Obtain historical CSI information for multiple channels.

[0047] Figure 3 This is a schematic diagram of the structure of a beamforming prediction model provided in an embodiment of this application, as shown below. Figure 3 As shown, during operation, the method in this embodiment of the application uses deep learning technology to predict beamforming for multiple channels. Therefore, it can first obtain the historical CSI information of multiple channels. One historical CSI information contains the channel state information (CSI) of one channel.

[0048] like Figure 2 As shown, when low-orbit satellites and multiple UAVs communicate, there can be multiple channels. Each channel can obtain the channel state information (CSI) corresponding to that channel, and then obtain the historical CSI information of that channel. In turn, the beamforming can be predicted using the historical CSI information.

[0049] S120. Using a beamforming prediction model, based on historical CSI information from multiple channels, determine the precoder construction parameters and common rate allocation parameters for beamforming multiple channels. The beamforming prediction model includes a Transformer unit, a CNN unit, and a generation unit. The Transformer unit extracts CSI time-dependent features from the historical CSI data of multiple channels, the CNN unit extracts spatial correlation features between the historical CSI data of multiple channels, and the generation unit generates the precoder construction parameters and common rate allocation parameters for beamforming multiple channels.

[0050] like Figure 3 As shown, the beamforming prediction model can directly determine the precoder construction parameters and common rate allocation parameters for beamforming multiple channels based on the historical CSI information of multiple channels. The precoder construction parameters obtained from beamforming prediction can be directly used to adjust the transmission direction and intensity of the signal in RSMA technology to improve the signal transmission efficiency and reception quality. At the same time, the common rate allocation parameters are used to split the transmission rate of each user into multiple parts to adapt to different transmission conditions and requirements.

[0051] like Figure 3 As shown, the beamforming prediction model can include a Transformer unit, a CNN unit, and a generator unit. During operation, after inputting historical CSI information of multiple channels, the Transformer unit (Transformer module), the CNN unit (convolution module), and the generator unit (generator module) sequentially complete the feature extraction of historical CSI data of multiple channels and generate pre-encoder construction parameters and common rate allocation parameters for beamforming.

[0052] During operation, the Transformer unit extracts CSI time-dependent features from historical CSI data across multiple channels, capturing complex time-series dependencies within these historical CSI data. Through an attention mechanism, the Transformer dynamically extracts features from different parts of the historical CSI data during prediction.

[0053] During operation, the CNN unit is used to extract spatial correlation features between historical CSI data of multiple channels. The convolutional neural network (CNN) can extract complex nonlinear features from the matrix input. Since beamforming prediction also needs to handle interference management between different user channels, the spatial dependence of beamforming prediction between different channels can be handled by the CNN unit.

[0054] During operation, the generation unit is used to generate precoder construction parameters and common rate allocation parameters for beamforming multiple channels.

[0055] It should be noted that the beamforming prediction model in this application embodiment can be abbreviated as the TranCN model. TranCN integrates the Transformer and CNN architectures, providing a new method to improve the effectiveness of predicting beamforming strategies.

[0056] Figure 4 A performance comparison chart of a deep learning-based RSMA beamforming prediction method in a non-terrestrial network, provided in an embodiment of this application, is shown below. Figure 4 As shown, Figure 4 The horizontal axis represents the signal-to-noise ratio (SNR), and the vertical axis represents the weighted ergodic sum rate (WESR) performance under different LEO satellite SNRs. The TranCN model in this embodiment outperforms the performance of only the Transformer model, only the CNN model, and only the SCA model.

[0057] The beneficial effects of the above implementation method are that it can directly generate precoder construction parameters and common rate allocation parameters for beamforming of multiple channels, accurately predict beamforming, reduce data transmission time, and improve the application effect of reducing channel interference through RSMA technology.

[0058] The beneficial effect of the above implementation method is that the integration of Transformer unit and CNN unit can effectively capture the spatial and temporal dependency features in CSI data, thereby improving the overall performance of beamforming prediction of RSMA in NTNs.

[0059] In some implementations, the above method also includes S111 and S121, which will be explained in detail below.

[0060] S111. Represent the real and imaginary parts of the historical CSI information of multiple channels using real numbers, and input the real and imaginary parts of the historical CSI information of multiple channels represented by real numbers into the Transformer unit of the beamforming prediction model.

[0061] In order to accurately reflect the historical CSI information of multiple channels during operation, the real and imaginary parts of the historical CSI information of multiple channels can be represented by real numbers, and the real and imaginary parts of the historical CSI information represented by real numbers can be input into the Transformer unit of the beamforming prediction model.

[0062] For example, such as Figure 3 As shown, the historical CSI information of multiple channels can be multiple historical CSI information from time t-1 to time t-T0, and the multiple historical CSI information from time t-1 to time t-T0 is used as the basic input of the beamforming prediction model (TranCN).

[0063] For example, multiple historical CSI information can be represented using a matrix H, where H represents historical channel coefficients, and H can be expressed as... The historical channel coefficient of the first drone is: superscript Indicates the number of rounds in the communication cycle. The subscript 1 indicates the UAV's serial number; the historical channel coefficients of other UAVs are represented similarly. Here, H represents a matrix corresponding to multiple historical CSI information entries, and T0 represents time. This represents the historical channel coefficient from the previous communication cycle. Indicates the previous communication cycle. Let N represent the drone with serial number 1, t represent time, k represent the drone with serial number k, and N represent the drone with serial number k. t Indicates the number of antennas. K represents the total number of drones, and K represents the maximum number of drones. This indicates a size of T0×KN t The set of complex numbers, It means "... is defined as...".

[0064] In order to ensure the availability of gradient information during training, multiple historical CSI information can be split into real and imaginary parts when performing beamforming prediction, and the real and imaginary parts of multiple historical CSI information can be represented by real numbers.

[0065] When inputting the model, the real and imaginary parts of the historical CSI information of multiple channels, represented by real numbers, can be represented as a time series, denoted as . in, The matrix represents the historical CSI information, where the real and imaginary parts of the channel coefficients are represented by real numbers, and R represents the set of real numbers. This indicates a size of T0×2KN t The set of real numbers.

[0066] This conversion method can encapsulate the real and imaginary parts of historical CSI information from multiple channels, while retaining the basic channel angle information of the historical CSI information from multiple input channels.

[0067] S121. The generation unit outputs a complex-valued matrix corresponding to the precoder construction parameters used for beamforming multiple channels.

[0068] When the generation unit is working, it can output the complex-valued matrix corresponding to the precoder construction parameters used for beamforming multiple channels. When constructing the precoder, the complex-valued matrix corresponding to the precoder construction parameters can be split to obtain the real and imaginary parts of the precoder construction parameters.

[0069] For example, the complex-valued matrix corresponding to the precoder construction parameters can be represented as a complex-valued matrix. in, This represents the precoder construction parameters represented using complex numbers. Indicates a size of N t The set of complex numbers ×(K+1).

[0070] The beneficial effect of the above implementation method is that by splitting multiple historical CSI information into real and imaginary parts, the availability of gradient information is guaranteed during the training process.

[0071] The beneficial effect of the above implementation method is that the generation unit can output the complex-valued matrix corresponding to the precoder construction parameters used for beamforming multiple channels, which facilitates the direct construction of the precoder.

[0072] In some implementations, the Transformer unit includes a multi-head attention layer, a first residual connection and normalization layer, a feedforward network layer, and a second residual connection and normalization layer arranged sequentially.

[0073] like Figure 3 As shown, a Transformer unit may include a multi-head attention layer, a first residual connection and normalization layer, a feedforward network layer and a second residual connection and normalization layer arranged in sequence.

[0074] Structurally, the multi-head attention layer contains D self-attention layers. The original input (multiple historical CSI information) undergoes a linear transformation to produce three matrices: in, Let T represent the set of real numbers of size T0×ι1. These transformations can be expressed by formula (1):

[0075]

[0076] In formula (3), A matrix representing the historical CSI information of multiple channels. The real and imaginary parts of the channel coefficients are represented by real numbers, where i represents the number of heads in the multi-head attention layer, Z represents the parameter matrix, and ι1 represents the dimension of the feature. This indicates a size of 2KN. t The set of real numbers ×ι1, M represents the set of multi-head attention layers, and Q, V and K represent the Q matrix, V matrix and K matrix in the Transformer unit, respectively.

[0077] During operation, the output of the self-attention layer can be represented by formula (2):

[0078]

[0079] In Equation (2), Attention(Q,V,K) represents the output of the self-attention layer, and softmax represents the normalization function.

[0080] During operation, the outputs of all self-attention layers are connected in the multi-head attention layer using a concat layer. The output of the multi-head attention layer can be represented by formula (3):

[0081] X1 = Concat(ω1,…,ω) D )Z o (3)

[0082] In formula (3), X1 represents the output of the multi-head attention layer. D represents the number of attention layers, and Concat represents concat layers.

[0083] Furthermore, to enhance the model's attention to the difference between the input (X) and the output (F(X)), embodiments of this application introduce residual connections and normalization layers. The residual connections and normalization layers contain a normalization layer, which can be represented by equation (4):

[0084] Add&Norm(X,F(X))=Norm(F(X))+X (4)

[0085] In formula (4), X represents the input of the same layer, F(X) represents the output of the multi-head attention layer or feedforward layer, and Add&Norm(X,F(X)) represents normalizing X and F(X) and performing residual connection.

[0086] In this embodiment, a feedforward layer is constructed using two fully connected (FC) layers, which can be represented by formula (5):

[0087] Feed Forward(X2)=max(A1X1+b1,0)A2+b2 (5)

[0088] In formula (5), A1X1+b1 represents a linear transformation of X1, where A1 and b1 are the linear transformation parameters, A2 and b2 are the linear transformation parameters, and max(A1X1+b1,0) represents taking the maximum value. If A1X1+b1>0, then A1X1+b1 is output; otherwise, 0 is output. X2 represents the output of the Transformer unit.

[0089] In some implementations, a CNN unit includes a convolutional layer, a pooling layer, and a flattening layer arranged in sequence.

[0090] like Figure 3 As shown, the CNN unit includes a convolutional layer, a pooling layer and a flattening layer arranged in sequence. The CNN can then be used to extract complex nonlinear features from the matrix input. The beamforming prediction model in this application can collect the spatial correlation between channels with different interferences by merging CNN modules.

[0091] like Figure 3 As shown, the initial layer of a CNN unit employs convolution operations, using a 2×2 filter to convolve with the input data, thereby enabling the identification of basic features across different interference channels. Subsequently, to further simplify the generated features and reduce their spatial dimensionality, average pooling is introduced in the pooling layer, which helps extract key information while preserving the basic features required for subsequent processing. The final layer of the CNN unit performs a flattening operation, reshaping the output into a one-dimensional vector. This reshaping facilitates seamless integration with subsequent layers, ensuring the continuity of information flow in the beamforming prediction model. The output information of CNN units can capture nonlinear features from interfering channels.

[0092] In some implementations, the generation unit includes a fully connected layer and a generation layer.

[0093] like Figure 3As shown, the generation unit comprises a fully connected layer and a generation layer, which work together to parameterize the beamforming strategy. The fully connected layer is the first part, enabling the model to learn details from the extracted features. This is achieved through fully connected operations and LeakyReLU. The fully connected layer captures nonlinear relationships in the feature space. LeakyReLU activation introduces a controllable amount of nonlinearity, enhancing the model's ability to recognize and absorb complex patterns.

[0094] Following the fully connected layer, the generation layer is seamlessly integrated to reshape the output, thereby obtaining the precoder construction parameters. and common rate allocation parameters Fully connected layers and generative layers can refine the learned features into precise and actionable parameters.

[0095] The beneficial effect of the above implementation is that using multiple attention heads in the Transformer unit enhances its functionality. Each attention head in the Transformer unit can learn different relationships in the data, enabling the model to capture a wide range of dependencies, including long-range dependencies. The Transformer unit plays a crucial role in extracting high-dimensional features and temporal correlations embedded in historical CSI information, thus improving the predictive power of beamforming predictions.

[0096] The beneficial effect of the above implementation method is that, in the Transformer unit, residual connections and normalization layers are introduced, which enhances the attention to the difference between the input (X) and output (F(X)) of each layer and improves the effect of beamforming prediction using multiple historical CSI information.

[0097] In some implementations, the common rate allocation parameters output by the beamforming prediction model are normalized and the total power limit of low-Earth orbit satellites is satisfied.

[0098] During operation, the common rate allocation parameters output by the beamforming prediction model can be normalized first to facilitate the calculation of the sum of the common rate allocation parameters output by the beamforming prediction model. Furthermore, the common rate allocation parameters output by the beamforming prediction model satisfy the total power limit of low-Earth orbit satellites, thereby enabling common rate allocation in RSMA under the total power limit of low-Earth orbit satellites.

[0099] The beneficial effect of the above implementation method is that the common rate allocation parameters output by the beamforming prediction model meet the total power limit of low-Earth orbit satellites, thus enabling effective common rate allocation in RSMA under the total power limit of low-Earth orbit satellites.

[0100] In some implementations, when the common rate allocation parameters output by the beamforming prediction model are normalized and the total power limit of low-Earth orbit satellites is met, they are expressed by formulas (5) to (9):

[0101]

[0102] tr(O H O)=P t (12)

[0103] In formulas (6) to (12), This represents the channel between a low Earth orbit (LEO) satellite and the k-th UAV or channel k, where k represents the UAV number or channel number. k This represents the signal-to-noise ratio (SINR) of the common stream decoded by the drone or channel k. K represents the total number of drones, and K represents the maximum number of drones. Let o represent the transposed channel vector between the low-orbit satellite and the k-th UAV. c Indicates a common linear precoder. Let σ represent the m-th private linear precoder. k Let represent the variance of the distribution of additive white Gaussian noise at the k-th drone. Indicates instantaneous reachability. O represents the complex-valued matrix output by the beamforming prediction model, and O represents the precoding matrix. Indicates output The calculated achievable common rate, min represents finding the minimum value, C k This represents the common rate portion of the k-th drone. P represents the normalized absolute value, tr represents the trace of the solution matrix, and P represents the trace of the solution matrix. t This indicates the launch power of low-Earth orbit satellites.

[0104] The beneficial effect of the above implementation method is that it facilitates the calculation of the common rate allocation parameters output by the beamforming prediction model in order to meet the total power limit of low-Earth orbit satellites.

[0105] This application also provides an on-orbit training method for beamforming prediction models, used for on-orbit training of beamforming prediction models for low-Earth orbit satellites. Figure 5 This is a flowchart illustrating an on-orbit training method for a beamforming prediction model provided in an embodiment of this application, as shown below. Figure 5 As shown, this method includes S210 to S220, and S210 to S220 will be described in detail below.

[0106] S210. Obtain historical CSI information for multiple channels.

[0107] The S220 low-Earth orbit satellite trains the beamforming prediction model in orbit based on historical CSI information from multiple channels. The low-Earth orbit satellite possesses on-orbit computing capabilities, and the loss function of the beamforming prediction model is expressed by formulas (13) and (14):

[0108]

[0109] In formulas (13) and (14), L represents the value of the loss function. u represents the expected total downlink rate obtained by the k-th drone. k The weights representing the obtained rates, and Φ representing the number of samples in the historical CSI information. This represents the instantaneous rate obtained under the corresponding sample.

[0110] Using the above formulas (13) and (14), the original optimization problem P0 is transformed into a trainable loss function, enabling the model to learn and adjust its parameters according to the desired objective.

[0111] In some implementations, the process of updating the beamforming prediction model can be represented by equations (15) and (16):

[0112] w n+1 =w n -η n g n (15)

[0113] η n =η0*λ n (16)

[0114] In formulas (15) and (16), η n Let η0 represent the learning rate in the nth training round, and λ0 represent the initial learning rate. n w represents the decay rate of the learning rate. n w represents the weight parameters of the global model in the nth round of training. n+1 g represents the weight parameters of the global model in the (n+1)th training round. n This represents a local estimate of the gradient of the loss function in satellite s.

[0115] The beneficial effect of the above implementation method is that by continuing the training method until the model converges, the training effect of the model is guaranteed.

[0116] This application also provides an RSMA beamforming prediction apparatus for non-terrestrial networks based on deep learning, including a unit for performing the method described in any of the preceding claims.

[0117] Figure 6A schematic diagram of the logic structure of an RSMA beamforming prediction device in a deep learning-based non-terrestrial network, as provided in an embodiment of this application, is shown below. Figure 6 As shown, the apparatus 1 in this embodiment includes a processing unit 11, a storage unit 12, and a transceiver unit 13. The processing unit 11 is used to process data, the storage unit 12 is used to store data, and the transceiver unit 13 is used to send and receive data. The processing unit 11, the storage unit 12, and the transceiver unit 13 cooperate with each other to implement the above-described method. The beneficial effects brought about by the embodiments of this application have been described in the above-described method and will not be repeated here.

[0118] This application also provides an RSMA beamforming prediction device for non-terrestrial networks based on deep learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method described in any of the preceding claims.

[0119] Figure 7 A schematic diagram of the physical structure of an RSMA beamforming prediction device in a deep learning-based non-terrestrial network, as provided in an embodiment of this application, is shown below. Figure 7 As shown, the device 2 of this embodiment includes: at least one processor 20 ( Figure 7 Only one processor 20, a memory 21, and a computer program 22 stored in the memory 21 and executable on the at least one processor 20 are shown. When the processor 20 executes the computer program 22, it implements the steps in any of the above method embodiments. The beneficial effects of the embodiments of this application have been described in the above methods and will not be repeated here.

[0120] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0121] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0122] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any of the preceding claims.

[0123] This application also provides a low-orbit satellite that uses the RSMA beamforming prediction method in a non-terrestrial network based on deep learning as described in any of the preceding embodiments when performing beamforming prediction.

[0124] This application also provides a method for training a low-Earth orbit satellite in orbit using the beamforming prediction model as described above.

[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0126] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0127] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0128] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0129] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0130] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A deep learning-based RSMA beamforming prediction method in non-terrestrial networks, characterized in that, The method includes: Acquire historical CSI information for multiple channels; Based on historical CSI information from multiple channels, a beamforming prediction model is used to determine the precoder construction parameters and common rate allocation parameters for beamforming multiple channels. The beamforming prediction model includes a Transformer unit, a CNN unit, and a generator unit. The Transformer unit is used to extract the CSI time-dependent features from the historical CSI data of multiple channels, the CNN unit is used to extract the spatial correlation features between the historical CSI data of multiple channels, and the generator unit is used to generate the precoder construction parameters and common rate allocation parameters for beamforming multiple channels. The method further includes: The real and imaginary parts of the historical CSI information of multiple channels are represented by real numbers, and the real and imaginary parts of the historical CSI information represented by real numbers are input into the Transformer unit of the beamforming prediction model. The generation unit outputs a complex matrix corresponding to the precoder construction parameters used for beamforming multiple channels.

2. The method as described in claim 1, characterized in that, The Transformer unit includes a multi-head attention layer, a first residual connection and normalization layer, a feedforward network layer, and a second residual connection and normalization layer arranged sequentially. A CNN unit consists of a convolutional layer, a pooling layer, and a flattening layer arranged sequentially. The generation unit includes a fully connected layer and a generation layer.

3. The method as described in claim 2, characterized in that, The common rate allocation parameters output by the beamforming prediction model are normalized and meet the total power limit of low-Earth orbit satellites.

4. The method as described in claim 3, characterized in that, When the common rate allocation parameters output by the beamforming prediction model are normalized and meet the total power limit for low-Earth orbit satellites, they are expressed by the following formula: ash H O)=P t in, This represents the channel between a low Earth orbit (LEO) satellite and the k-th UAV, where k represents the UAV number or channel number, γ k This represents the signal-to-noise ratio (SINR) of the common stream decoded by the drone or channel k. This indicates the total number of drones. Let o represent the transposed channel vector between the low-orbit satellite and the k-th UAV. c Indicates a common linear precoder. Let σ represent the m-th private linear precoder. k Let represent the variance of the distribution of additive white Gaussian noise at the k-th drone. Indicates instantaneous reachability. O represents the complex-valued matrix output by the beamforming prediction model, and O represents the precoding matrix. Indicates output The calculated achievable common rate, This represents the achievable common rate set by the beamforming prediction model, and min indicates finding the minimum value. Indicates the parameter size is Located in the set of complex numbers, C k This represents the common rate portion of the k-th drone. P represents the normalized absolute value, tr represents the trace of the solution matrix, and P represents the trace of the solution matrix. t This indicates the launch power of low-Earth orbit satellites.

5. An on-orbit training method for a beamforming prediction model, characterized in that, The method for training beamforming prediction models for low-Earth orbit satellites in orbit includes: Acquire historical CSI information for multiple channels; The low-Earth orbit (LEO) satellite trains the beamforming prediction model in orbit based on historical CSI information from multiple channels. The LEO satellite possesses on-orbit computing capabilities, and the loss function of the beamforming prediction model is expressed by the following formula: Where L represents the loss function value, u represents the expected total downlink rate obtained by the k-th drone. k The weights representing the obtained rates, and Φ representing the number of samples in the historical CSI information. C represents the instantaneous rate obtained under the corresponding sample. k This represents the common rate portion of the k-th drone, where k represents the drone number or channel number. Indicates the total number of drones; The process of updating the beamforming prediction model can be represented by the following formula: w n+1 =w n -or n g n or n =η0*λ n Where, η n Let η0 represent the learning rate in the nth training round, and λ0 represent the initial learning rate. n w represents the decay rate of the learning rate. n w represents the weight parameters of the global model in the nth round of training. n+1 g represents the weight parameters of the global model in the (n+1)th training round. n This represents a local estimate of the gradient of the loss function in satellite s.

6. A deep learning-based RSMA beamforming prediction device in a non-terrestrial network, characterized in that, Includes a unit for performing the method according to any one of claims 1 to 4.

7. A deep learning-based RSMA beamforming prediction device for non-terrestrial networks, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 4.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 4.

9. A low-orbit satellite, characterized in that, When performing beamforming prediction, the RSMA beamforming prediction method based on deep learning in non-terrestrial networks as described in any one of claims 1 to 4 is used; or when training the beamforming prediction model, the on-orbit training method of the beamforming prediction model described in claim 5 is used.

Images