Transform-based precoding method and device
By using a Transformer-based precoding method and historical information for prediction, the problem of high pilot and feedback overhead in large-scale MIMO systems is solved, achieving high-precision precoding and improving system throughput and spectral efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-03-31
AI Technical Summary
In large-scale MIMO systems, existing technologies suffer from high pilot and feedback overhead, CSI delay mismatch, and high precoding computational complexity, which affect the system's spectral efficiency and throughput performance.
We employ a Transformer-based precoding method. By obtaining the precoding matrix of historical resource blocks, we use a pre-trained precoding prediction model for prediction. We combine a multi-head self-attention encoder and a causal mask to capture global dependencies, reduce computational complexity, and achieve high-precision precoding.
Without increasing pilot and feedback overhead, it improves system throughput and spectral efficiency, reduces computational complexity and latency, and is suitable for large-scale antenna and broadband scenarios.
Smart Images
Figure CN121770567A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of communication technology, and in particular to a precoding method and apparatus based on Transformer. Background Technology
[0002] In modern communication systems, MIMO (Multiple-Input Multiple-Output) technology significantly improves spectral efficiency and data transmission rates by using multiple antennas at both the transmitter and receiver. In physical layer communication, the design of the precoding matrix typically relies on accurate downlink channel state information (CSI). When a system employs a large-scale antenna array and operates in a broadband environment, acquiring and maintaining high-precision CSI requires substantial pilot and feedback resources. Limited by the channel's coherence time and bandwidth, these overheads significantly encroach on time-frequency resources available for service data, thus impacting the system's overall spectral efficiency and throughput performance.
[0003] Existing CSI acquisition and precoding designs mainly include two approaches: First, under the TDD (Time Division Duplex) system, downlink CSI is inferred based on uplink pilots and channel reciprocity. However, in actual systems, radio frequency link calibration and compensation are still required, and there is a "channel aging" problem in high-speed or rapidly changing scenarios, resulting in a time mismatch between the estimation results and the actual downlink channel. Second, under the FDD (Frequency Division Duplex) system, downlink pilot estimation is based on terminal feedback of fine codebook index or quantization vector. However, the feedback overhead increases approximately linearly with the number of antennas and subcarriers, which can easily lead to signaling burden and latency accumulation in large-scale systems.
[0004] In terms of precoding implementation, traditional methods often employ codebook-based finite feedback schemes or instantaneous CSI-based linear precoding. Codebook methods require a trade-off between quantization resolution and feedback overhead, making it difficult to simultaneously achieve accuracy and latency in high-dimensional scenarios. Linear precoding relies on stable and timely CSI, is significantly affected by estimation errors and latency, and is accompanied by high matrix computation complexity. Summary of the Invention
[0005] This invention provides a Transformer-based precoding method and apparatus to address the shortcomings of existing technologies, such as high pilot and feedback overhead, channel aging, and high precoding computational complexity. This invention achieves high-precision precoding prediction without relying on complete channel state information and by reducing pilot overhead, thereby reducing computational complexity and latency, and ultimately improving system throughput and spectral efficiency.
[0006] This invention provides a Transformer-based precoding method, comprising: obtaining a precoding matrix of historical resource blocks; inputting the precoding matrix of the historical resource blocks into a pre-trained precoding prediction model to obtain a precoding matrix of future resource blocks output by the precoding prediction model; wherein the precoding prediction model is based on a Transformer network and trained using precoding training samples.
[0007] According to a Transformer-based precoding method provided by the present invention, the step of inputting the precoding matrix of the historical resource block into a pre-trained precoding prediction model to obtain the precoding matrix of the future resource block output by the precoding prediction model includes: inputting the precoding matrix of the historical resource block into a multi-head self-attention encoder layer to capture global dependencies and calculate attention weights; inputting the output of the multi-head self-attention encoder layer into a feedforward layer for feature transformation after residual summation and layer normalization; adding the output of the feedforward layer with the residual and performing layer normalization to obtain the precoding matrix prediction result of the future resource block; and determining the precoding matrix of the future resource block based on the prediction result.
[0008] According to a Transformer-based precoding method provided by the present invention, before inputting the precoding matrix of the historical resource block into a multi-head self-attention encoder layer, the method further includes: splitting the precoding matrix of the historical resource block into real and imaginary parts and quantizing them to obtain a vectorized sequence; performing input embedding on the vectorized sequence to obtain an embedding vector; performing position encoding on the vectorized sequence and superimposing it with the embedding vector to obtain an input sequence, so as to execute the step of inputting the multi-head self-attention encoder layer based on the input sequence.
[0009] According to a Transformer-based precoding method provided by the present invention, the step of inputting the precoding matrix of the historical resource block into a multi-head self-attention encoder layer to capture global dependencies and calculate attention weights includes: generating query, key, and value vectors respectively through linear transformation based on the precoding matrix of the historical resource block; calculating attention weights using scaled dot product attention and weighted summing of the value vectors; using a causal mask with a lower triangular structure in the attention weight calculation stage; and concatenating and linearly mapping the parallel outputs of multi-head attention.
[0010] According to a Transformer-based precoding method provided by the present invention, determining the precoding matrix of the future resource block based on the prediction result includes: performing complex reconstruction and power normalization on the prediction result to obtain the precoding matrix of the future resource block.
[0011] According to the Transformer-based precoding method provided by the present invention, the loss function of the precoding prediction model is: , in, For the number of subcarriers, It is the identity matrix. For the first Each time step The number of historical resource blocks, For noise power density, For the first One resource block, The number of transmitting antennas, For the channel matrix, This is the precoding matrix.
[0012] The present invention also provides a Transformer-based precoding apparatus, comprising: an acquisition module for acquiring a precoding matrix of historical resource blocks; and a prediction module for inputting the precoding matrix of the historical resource blocks into a pre-trained precoding prediction model to obtain a precoding matrix of future resource blocks output by the precoding prediction model; wherein the precoding prediction model is based on a Transformer network and trained using precoding training samples.
[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the Transformer-based precoding method as described above.
[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the Transformer-based precoding method as described above.
[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the Transformer-based precoding method as described above.
[0016] The present invention provides a Transformer-based precoding method and apparatus that can achieve high-precision precoding prediction without relying on complete channel state information and reducing pilot overhead, thereby reducing computational complexity and latency, and thus improving system throughput and spectral efficiency. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0018] Figure 1 This is a flowchart illustrating the Transformer-based precoding method provided by the present invention.
[0019] Figure 2 This is a schematic diagram illustrating the principle of the Transformer-based precoding method provided by the present invention.
[0020] Figure 3 This is a schematic diagram illustrating the principle of the precoding prediction model provided by the present invention.
[0021] Figure 4 This is a diagram showing the performance comparison of precoding methods.
[0022] Figure 5 This is a schematic diagram of the structure of the Transformer-based precoding device provided by the present invention.
[0023] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0025] In recent years, with the rapid development of mobile communication technology, communication systems are facing higher requirements for spectral efficiency and energy efficiency. Massive MIMO technology, with its superior spatial multiplexing and diversity gain capabilities, has significantly improved system performance and has become one of the core technologies of the next-generation mobile communication standard. Precoding, as a key signal processing technology in wireless communication systems, effectively improves the reliability and spectral efficiency of signal transmission by performing complex data transformations at the transmitting end. It plays a crucial role in optimizing signal transmission quality, reducing interference, and improving overall system performance.
[0026] In physical layer communication, the design of precoding matrices typically relies on accurate downlink channel state information (CSI). When a system employs a large-scale antenna array and operates in a broadband environment, acquiring and maintaining high-precision CSI requires significant pilot and feedback resources. Limited by the channel's coherence time and bandwidth, these overheads significantly encroach on time-frequency resources available for service data, thereby affecting the system's overall spectral efficiency and throughput performance.
[0027] Existing CSI acquisition and precoding designs mainly include two approaches: First, under the TDD system, downlink CSI is inferred based on uplink pilots and channel reciprocity. However, in actual systems, radio frequency link calibration and compensation are still required, and there is a "channel aging" problem in high-speed or rapidly changing scenarios, resulting in a time mismatch between the estimation results and the actual downlink channel. Second, under the FDD system, downlink pilot estimation is based on terminal feedback of fine codebook index or quantization vector. However, the feedback overhead increases approximately linearly with the number of antennas and subcarriers, which can easily lead to signaling burden and latency accumulation in large-scale systems.
[0028] In terms of precoding implementation, traditional methods often employ codebook-based finite feedback schemes (such as DFT / Grassmannian codebooks) or linear precoding based on instantaneous CSI (such as ZF and MMSE). Codebook methods require a trade-off between quantization resolution and feedback overhead, making it difficult to simultaneously achieve accuracy and latency in high-dimensional scenarios. Linear precoding relies on stable and timely CSI, is significantly affected by estimation errors and latency, and is accompanied by high matrix computation complexity (such as large-scale matrix inversion or singular value decomposition).
[0029] Existing technologies still face challenges in scenarios involving massive MIMO, broadband, and rapid time-varying conditions, including high pilot and feedback overhead, time delay mismatch between CSI and usage time, high precoding computational complexity, and insufficient modeling of long-range correlations and causality. Therefore, it is necessary to propose a precoding prediction method and apparatus that utilizes only historically available information and satisfies temporal causality without adding or significantly reducing CSI detection and feedback, thereby reducing implementation complexity and supporting real-time online deployment.
[0030] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the Transformer-based precoding method provided by the present invention.
[0031] This invention provides a Transformer-based precoding method, comprising: 101: Obtain the precoding matrix of historical resource blocks.
[0032] In this embodiment, the user terminal generates a precoding matrix sequence of historical resource blocks and reports or accesses it. Specifically, the base station side or network side collects the most recent data within the target cell. The precoding matrix used in each resource block / slot / subframe is denoted as ,in It is a complex matrix whose dimensions are consistent with the number of transmit antennas and spatial layers of the system.
[0033] As a preferred embodiment, before inputting the precoding matrix of the historical resource block into the multi-head self-attention encoder layer, the method further includes: splitting the precoding matrix of the historical resource block into real and imaginary parts and quantizing them to obtain a vectorized sequence; embedding the vectorized sequence into an input to obtain an embedding vector; performing position encoding on the vectorized sequence and superimposing it with the embedding vector to obtain an input sequence, so as to perform the step of inputting the multi-head self-attention encoder layer based on the input sequence.
[0034] To convert complex matrices into a format suitable for neural network processing, in this embodiment, the precoding matrix of each historical resource block is split into real and imaginary parts, resulting in a real part matrix and an imaginary part matrix. The real and imaginary parts matrices are then expanded into one-dimensional vectors column-wise (or row-wise) to form a vectorized sequence. .
[0035] For vectorized sequences Perform a linear mapping to obtain The embedding representation (embedding vector) is mapped using a block-linear layer to process data from different historical moments. The embedding vector can be shared or configured independently across different historical moments to adapt to different channel dynamics. To characterize time and / or frequency location, a location vector with the same dimension as the embedding vector is generated and added element-wise to the embedding vector to form an input sequence incorporating location information. The positional encoding expression is: in, This represents the position index, used to indicate the position of an element in the input sequence. The channel pair index represents the position encoding and is used to determine the corresponding sine / cosine scaling factor.
[0036] Furthermore, positional encoding not only includes fixed encoding based on sine / cosine functions, but also introduces a learnable positional encoding vector. This vector is dynamically adjusted during the training process to better adapt to the time-varying characteristics of the channel. Channel quality indication information can also be incorporated into the input embedding process, embedding it into the input vector to enhance the model's ability to perceive channel quality.
[0037] Vectorized sequences can also be normalized before input embedding. The normalization parameters are dynamically adjusted based on historical channel states to optimize the input stability of the model.
[0038] Please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating the principle of the Transformer-based precoding method provided by the present invention. In the diagram, t represents a timestamp, and A and B are both resource blocks in wireless communication. A represents those resource blocks that will be acquired and precoded by the system, while B represents those resource blocks that will be predicted by the model.
[0039] Please refer to Figure 3 , Figure 3 This is a schematic diagram illustrating the principle of the precoding prediction model provided by the present invention.
[0040] 102: Input the precoding matrix of historical resource blocks into the pre-trained precoding prediction model to obtain the precoding matrix of future resource blocks output by the precoding prediction model; wherein, the precoding prediction model is based on the Transformer network and is trained using precoding training samples.
[0041] In a preferred embodiment, the precoding matrix of historical resource blocks is input into a pre-trained precoding prediction model to obtain the precoding matrix of future resource blocks output by the precoding prediction model. This includes: inputting the precoding matrix of historical resource blocks into a multi-head self-attention encoder layer to capture global dependencies and calculate attention weights; inputting the output of the multi-head self-attention encoder layer into a feedforward layer for feature transformation after residual summation and layer normalization; summing the output of the feedforward layer with the residuals and normalizing the output to obtain the prediction result of the precoding matrix of future resource blocks; and determining the precoding matrix of future resource blocks based on the prediction result.
[0042] As a preferred embodiment, the precoding matrix of historical resource blocks is input into the multi-head self-attention encoder layer to capture global dependencies and calculate attention weights, including: generating query, key, and value vectors respectively through linear transformation based on the precoding matrix of historical resource blocks; calculating attention weights using scaled dot product attention and weighted summing of the value vectors; using a causal mask with a lower triangular structure in the attention weight calculation stage; and concatenating and linearly mapping the parallel outputs of multi-head attention.
[0043] In this embodiment, the input sequence The input multi-head self-attention encoder layer applies a strict lower triangular causal mask during the attention weight calculation stage to suppress the relevance after the target time step, achieving the temporal causal consistency required for online prediction. The mask and attention matrix are applied bitwise to restrict the information flow. Specifically, in the multi-head self-attention encoder layer, query, key, and value vectors are generated through linear transformations; attention weights are calculated using scaled dot product attention and the value vectors are weighted and summed; the parallel outputs of the multi-head attention are concatenated and linearly mapped.
[0044] The output of the multi-head self-attention encoder layer is first compared with the input sequence. X The residuals are summed to preserve input information and prevent data loss. The results after residual concatenation are further processed by layer normalization to ensure the stability of the training process.
[0045] The result after residual summation and layer normalization is input into the feedforward layer, which consists of two fully connected layers containing a non-linear activation function (such as ReLU). The function of the feedforward layer is to perform a non-linear transformation on the input features, further extracting and transforming the features.
[0046] The output of the feedforward layer is added to the input residual and normalized again to obtain the precoding matrix prediction result for the future resource block. The output dimension is kept consistent with the input dimension to facilitate the reconstruction of the precoding matrix.
[0047] In addition, the precoding prediction model is trained using precoding training samples, supports precoding matrix prediction in multi-user scenarios, and can generate personalized precoding matrices for multiple user devices at the same time to improve the overall performance of multi-user systems.
[0048] Causal masking uses an offset mask, where for the J-th position in the sequence, its self-attention can access the preceding positions of the input sequence. Positions (i.e., indexes) );when When the sequence length is exceeded, the process degenerates into accessing all historical inputs. This way, the first output sees the first and second inputs, the second output sees the first three inputs, and so on. This design allows the model to better perceive the wireless sounding cycle.
[0049] As a preferred embodiment, determining the precoding matrix of future resource blocks based on the prediction results includes: performing complex reconstruction and power normalization on the prediction results to obtain the precoding matrix of future resource blocks.
[0050] In this embodiment, the precoding matrix prediction result of future resource blocks is mapped to a matrix of length [length missing]. real-valued vectors (The first half is the real part, and the second half is the imaginary part), rearranged column-wise to obtain the real part matrix. With the imaginary part matrix This forms a complex precoding matrix: , Then, power normalization is performed to ensure it meets the transmitter constraints.
[0051] During online inference, the historical sequence is updated in a sliding window manner. When a new resource block is sent, the actual usage at that moment is pre-encoded (or a confirmed reference value) and stored in the database for prediction at subsequent moments. To reduce latency, key-value caching can be used to reuse intermediate results from previous attention.
[0052] In addition, a dynamic resource allocation mechanism can be used during the inference phase to dynamically adjust the allocation strategy of resource blocks based on the current channel quality and system load in order to optimize system performance.
[0053] For unknown resource blocks, the base station performs downlink transmission using a predicted precoding matrix.
[0054] Data simulation was performed based on the 3GPP CDL series models. Continuous historical precoding sequences and corresponding radio channels were collected. The precoding matrix corresponding to the historical channel was solved by the Singular Value Decomposition (SVD) precoding algorithm. At the same time, the channel to be predicted was collected and its precoding matrix was solved as a tag.
[0055] With the goal of improving system throughput, a loss function that is monotonically correlated with transmission rate is constructed.
[0056] As a preferred embodiment, the loss function of the precoding prediction model is: , in, For the number of subcarriers, It is the identity matrix. For the first Each time step The number of historical resource blocks, For noise power density, For the first One resource block, The number of transmitting antennas, For the channel matrix, This is the precoding matrix.
[0057] Furthermore, the training objective function of the precoding prediction model not only includes maximizing system throughput, but can also combine a multi-objective optimization strategy that minimizes bit error rate and optimizes power efficiency.
[0058] Regarding the model training strategy, a supervisory signal is calculated based on the channel information of the reference resource block, and a cosine annealing learning rate scheduling method is adopted. This allows the learning rate to maintain a large exploratory capability in the early stages of training and gradually converge smoothly in the later stages, thereby improving the model's stability and final performance. Furthermore, gradient pruning and layer normalization mechanisms are introduced to suppress gradient explosion, improve the training dynamics of deep networks, and update network parameters to achieve end-to-end learning. The combination of these two mechanisms ensures that the model maintains the stability of parameter updates while rapidly reducing loss, avoiding unstable oscillations.
[0059] To facilitate model deployment in real-world systems, a structured pruning approach, using column-based pruning, was further employed to reduce network complexity. Structured pruning directly removes entire columns of parameters, making it easier for hardware to accelerate and reducing storage burden. This approach can significantly reduce model size with minimal loss of performance, improving inference speed on edge devices or in online environments.
[0060] Overall, this training strategy achieves a balance between stability, convergence efficiency, and deployment friendliness, providing a solid foundation for its subsequent real-time application in communication systems.
[0061] Please refer to Figure 4 , Figure 4 This diagram illustrates the performance comparison of precoding methods. The horizontal axis represents the number of training epochs, and the vertical axis represents the gain in sum rate brought by the predicted precoding matrix compared to zero-order preservation (baseline).
[0062] This invention achieves high-precision prediction of the target resource block precoding matrix based on historically available information without increasing or significantly reducing pilot and feedback overhead; it alleviates the latency mismatch problem caused by channel aging by jointly modeling multi-head self-attention and causal masking; while ensuring performance, online inference can be combined with engineering techniques such as key-value caching to reduce computational complexity and latency, making it easy to deploy in large-scale antenna and broadband scenarios.
[0063] The Transformer-based precoding apparatus provided by the present invention will be described below. The Transformer-based precoding apparatus described below can be referred to in correspondence with the Transformer-based precoding method described above.
[0064] Please refer to Figure 5 , Figure 5 A schematic diagram of the structure of the Transformer-based precoding device provided by the present invention.
[0065] The present invention also provides a Transformer-based precoding device, comprising: an acquisition module 501 for acquiring the precoding matrix of historical resource blocks; and a prediction module 502 for inputting the precoding matrix of historical resource blocks into a pre-trained precoding prediction model to obtain the precoding matrix of future resource blocks output by the precoding prediction model; wherein the precoding prediction model is based on a Transformer network and is trained using precoding training samples.
[0066] In a preferred embodiment, the precoding matrix of historical resource blocks is input into a pre-trained precoding prediction model to obtain the precoding matrix of future resource blocks output by the precoding prediction model. This includes: inputting the precoding matrix of historical resource blocks into a multi-head self-attention encoder layer to capture global dependencies and calculate attention weights; inputting the output of the multi-head self-attention encoder layer into a feedforward layer for feature transformation after residual summation and layer normalization; summing the output of the feedforward layer with the residuals and normalizing the output to obtain the prediction result of the precoding matrix of future resource blocks; and determining the precoding matrix of future resource blocks based on the prediction result.
[0067] As a preferred embodiment, before inputting the precoding matrix of the historical resource block into the multi-head self-attention encoder layer, the method further includes: splitting the precoding matrix of the historical resource block into real and imaginary parts and quantizing it to obtain a vectorized sequence; embedding the vectorized sequence into an input to obtain an embedding vector; performing position encoding on the vectorized sequence and superimposing it with the embedding vector to obtain an input sequence, so as to perform the step of inputting the multi-head self-attention encoder layer based on the input sequence.
[0068] As a preferred embodiment, the precoding matrix of historical resource blocks is input into the multi-head self-attention encoder layer to capture global dependencies and calculate attention weights, including: generating query, key, and value vectors respectively through linear transformation based on the precoding matrix of historical resource blocks; calculating attention weights using scaled dot product attention and weighted summing of the value vectors; using a causal mask with a lower triangular structure in the attention weight calculation stage; and concatenating and linearly mapping the parallel outputs of multi-head attention.
[0069] As a preferred embodiment, determining the precoding matrix of future resource blocks based on the prediction results includes: performing complex reconstruction and power normalization on the prediction results to obtain the precoding matrix of future resource blocks.
[0070] As a preferred embodiment, the loss function of the precoding prediction model is: , in, For the number of subcarriers, It is the identity matrix. For the first Each time step The number of historical resource blocks, For noise power density, For the first One resource block, The number of transmitting antennas, For the channel matrix, This is the precoding matrix.
[0071] Figure 6 An example is a schematic diagram of the structure of an electronic device, such as... Figure 6 As shown, the electronic device may include a processor 601, a communications interface 602, a memory 603, and a communication bus 604. The processor 601, communications interface 602, and memory 603 communicate with each other via the communication bus 604. The processor 601 can call logical instructions in the memory 603 to execute a Transformer-based precoding method. This method includes: obtaining the precoding matrix of historical resource blocks; inputting the precoding matrix of historical resource blocks into a pre-trained precoding prediction model to obtain the precoding matrix of future resource blocks output by the precoding prediction model; wherein the precoding prediction model is based on a Transformer network and trained using precoding training samples.
[0072] Furthermore, the logical instructions in the aforementioned memory 603 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0073] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the Transformer-based precoding method provided by the above methods. The method includes: obtaining a precoding matrix of historical resource blocks; inputting the precoding matrix of historical resource blocks into a pre-trained precoding prediction model to obtain a precoding matrix of future resource blocks output by the precoding prediction model; wherein the precoding prediction model is based on a Transformer network and trained using precoding training samples.
[0074] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the Transformer-based precoding method provided by the above methods. The method includes: obtaining a precoding matrix of historical resource blocks; inputting the precoding matrix of historical resource blocks into a pre-trained precoding prediction model to obtain a precoding matrix of future resource blocks output by the precoding prediction model; wherein the precoding prediction model is based on a Transformer network and trained using precoding training samples.
[0075] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0076] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.
Claims
1. A Transformer-based precoding method, characterized in that, include: Obtain the precoding matrix of historical resource blocks; The precoding matrix of the historical resource block is input into the pre-trained precoding prediction model to obtain the precoding matrix of the future resource block output by the precoding prediction model. The pre-encoded prediction model is based on the Transformer network and is trained using pre-encoded training samples.
2. The Transformer-based precoding method according to claim 1, characterized in that, The step of inputting the precoding matrix of the historical resource block into a pre-trained precoding prediction model to obtain the precoding matrix of the future resource block output by the precoding prediction model includes: The precoding matrix of the historical resource block is input into the multi-head self-attention encoder layer to capture global dependencies and calculate attention weights; The output of the multi-head self-attention encoder layer is added to the residual and normalized by the layer before being input into the feedforward layer for feature transformation. The output of the feedforward layer is added to the residual and then normalized by the layer to obtain the precoding matrix prediction result of the future resource block. Based on the prediction results, the precoding matrix of the future resource block is determined.
3. The Transformer-based precoding method according to claim 2, characterized in that, Before inputting the precoding matrix of the historical resource block into the multi-head self-attention encoder layer, the method further includes: The precoding matrix of the historical resource block is split into real and imaginary parts and quantized to obtain a vectorized sequence; The vectorized sequence is embedded to obtain an embedding vector; The vectorized sequence is positionally encoded and superimposed with the embedding vector to obtain an input sequence, which is then used to execute the steps of the input multi-head self-attention encoder layer.
4. The Transformer-based precoding method according to claim 2, characterized in that, The step of inputting the precoding matrix of the historical resource block into the multi-head self-attention encoder layer to capture global dependencies and calculate attention weights includes: Based on the precoding matrix of the historical resource blocks, query, key, and value vectors are generated respectively through linear transformation; The attention weights are calculated using scaled dot product attention, and the value vectors are then summed using weighted summation. A causal mask with a lower triangular structure is used in the attention weight calculation stage; The parallel outputs of multi-head attention are concatenated and linearly mapped.
5. The Transformer-based precoding method according to claim 2, characterized in that, Determining the precoding matrix of the future resource block based on the prediction result includes: The prediction results are reconstructed using complex numbers and normalized by power to obtain the precoding matrix of the future resource block.
6. The Transformer-based precoding method according to any one of claims 1 to 5, characterized in that, The loss function of the precoding prediction model is: ， in, For the number of subcarriers, It is the identity matrix. For the first Each time step The number of historical resource blocks, For noise power density, For the first One resource block, The number of transmitting antennas, For the channel matrix, This is the precoding matrix.
7. A Transformer-based precoding apparatus, characterized in that, include: The acquisition module is used to acquire the precoding matrix of historical resource blocks; The prediction module is used to input the precoding matrix of the historical resource block into a pre-trained precoding prediction model to obtain the precoding matrix of the future resource block output by the precoding prediction model. The pre-encoded prediction model is based on the Transformer network and is trained using pre-encoded training samples.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the Transformer-based precoding method as described in any one of claims 1 to 6.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the Transformer-based precoding method as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the Transformer-based precoding method as described in any one of claims 1 to 6.