A neural network-based precoding matrix indication selection method
By preprocessing the downlink channel matrix of the 5G-A system and learning through neural networks, PMI parameters are directly output, solving the problem of high complexity in PMI selection in existing technologies and achieving low-complexity PMI selection and efficient CSI feedback.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2026-05-28
- Publication Date
- 2026-06-23
AI Technical Summary
The existing precoding matrix indication (PMI) selection method in 5G-A systems requires traversing and searching a large number of candidate codewords, which results in high computational complexity and makes it difficult to balance real-time performance with low complexity requirements.
A neural network-based precoding matrix indicator selection method is adopted. By downsampling, performing piecewise discrete Fourier transform and complex-to-real-domain preprocessing on the downlink channel matrix, and using neural network for end-to-end learning, the codeword number is directly output and converted into PMI parameters, reducing the computational complexity of the user end.
While maintaining the predefined codebook mapping mechanism on the base station side, the complexity of PMI selection on the user side is significantly reduced, and the real-time performance of CSI feedback and downlink precoding transmission efficiency are improved.
Smart Images
Figure CN122268436A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a precoding matrix indicator selection method based on neural networks, belonging to the field of communication technology. Background Technology
[0002] In 5G-A mobile communication systems, Channel State Information (CSI) feedback is a core component ensuring efficient downlink transmission, and Precoding Matrix Indicator (PMI) selection is a crucial step in the CSI feedback process. Current 5G-A systems employ a PMI feedback mechanism based on a predefined codebook. The user terminal must traverse and search all candidate precoding matrices in the codebook, calculate the matching degree based on criteria such as maximum mutual information, and finally select the optimal PMI and feed it back to the base station. The base station then uses the same codebook to recover the corresponding precoding matrix for downlink beamforming transmission. While this method conforms to existing standard procedures, the user terminal needs to traverse and search a large number of candidate PMIs, resulting in high computational complexity. As the size of base station antennas increases, the size of the candidate codebook expands significantly, making it difficult for traditional search-based PMI selection methods to simultaneously meet the requirements of real-time performance and low complexity.
[0003] The aforementioned issues severely restrict the efficiency and real-time performance of CSI feedback in 5G-A systems, becoming a key bottleneck for air interface optimization. Therefore, achieving fast PMI selection with low complexity while ensuring PMI selection accuracy is a pressing technical problem that needs to be solved. Neural networks possess strong feature learning and pattern recognition capabilities, enabling them to directly predict the optimal PMI from the original channel matrix, thus avoiding a traversal of all candidate codewords. However, achieving low-complexity, high-efficiency PMI selection while maintaining compatibility with existing codebook feedback mechanisms still requires further research and engineering implementation. Summary of the Invention
[0004] Objective: To address the problems of high computational complexity and real-time limitations in existing PMI selection methods based on predefined codebooks, which typically require traversal searching of candidate codewords at the user end, this invention proposes a neural network-based precoding matrix indication PMI selection method. This method, while maintaining the base station-side predefined codebook mapping mechanism, utilizes a neural network to learn and infer the preprocessed downlink channel characteristics, directly determining the codeword number corresponding to each subband and converting it to obtain PMI parameters. This reduces the complexity of PMI selection at the user end, improves PMI feedback efficiency, and achieves compatibility with existing codebook feedback mechanisms.
[0005] Technical solution: The present invention provides a precoding matrix indicator selection method based on a neural network, comprising the following steps:
[0006] (1) The user terminal obtains the downlink channel matrix H, and performs a combination of downsampling, piecewise discrete Fourier transform, and complex-to-real-domain feature preprocessing on the channel matrix H to obtain a real-valued feature tensor that is adapted to the neural network input.
[0007] (2) Input the real number feature tensor into the end-to-end neural network model trained offline, and the model directly outputs the evaluation probability of each sub-band codeword, and determines the optimal codeword number according to the probability distribution;
[0008] (3) Based on the codeword number, the PMI parameter is converted according to the predefined codebook rules, the PMI parameter is fed back to the base station, and the base station obtains the precoding matrix according to the predefined codebook consistent with the user terminal, which is used for downlink beamforming transmission.
[0009] In one embodiment, the downsampling includes: dividing the original channel state information within the broadband into multiple sub-bands, and selecting some frequency points within each sub-band at preset intervals to reduce data redundancy caused by the correlation between adjacent frequency points and reduce the complexity of subsequent processing.
[0010] The piecewise discrete Fourier transform includes: dividing the downsampled channel feature matrix into multiple segments, and performing a spatial-domain to angular-domain transformation on each segment, the expression of which is:
[0011]
[0012] in, This represents the segmented spatial domain channel matrix. Represents the discrete Fourier transform matrix The conjugate transpose of . Represents the angle-domain channel matrix.
[0013] The process of converting complex domain channel features into real number input features includes: extracting the real and imaginary parts of each segment and concatenating them along the column dimension.
[0014] In one embodiment, the neural network model includes a feature extraction module and a codeword mapping module. The feature extraction module is used to extract discriminative features for PMI selection from preprocessed channel features, and the codeword mapping module is used to output codeword numbers corresponding to each sub-band based on the discriminative features. The neural network model performs codeword prediction for each sub-band and constructs broadband PMI feedback results based on the codeword prediction results for each sub-band.
[0015] In one embodiment, the feature extraction module processes each subband separately, extracts local features through a neural network module containing convolutional layers and residual connections, and flattens them into a one-dimensional feature vector; the codeword mapping module maps the one-dimensional feature vector to an output space of equivalent codebook size through a fully connected layer, outputs the probability distribution of each codeword through a Softmax activation function, and selects the codeword number corresponding to the highest probability.
[0016] In one embodiment, PMI parameter transformation includes numbering the codeword of the k-th subband. Convert to PMI parameter triples ( The conversion formula is:
[0017]
[0018]
[0019]
[0020]
[0021] in, This indicates the total number of codewords in the current codebook structure; These represent PMI parameters. The number of possible values for the parameter is determined by the antenna configuration, the predefined codebook type, and the transmission rank. This is a round-down operation. mod (·) represents the modulus extraction operation.
[0022] In one embodiment, the neural network model is deployed using an offline training and online inference approach; the labels of the training samples are the optimal codeword numbers obtained through traversal search under the same predefined codebook constraints. The neural network model training process employs a cross-entropy loss function, expressed as:
[0023]
[0024] Where M is the total number of training samples, K is the number of subbands, and C is the number of codeword categories; The true label for the m-th sample, the k-th subband, and the c-th codeword; Let be the predicted probability of the m-th sample, the k-th subband, and the c-th codeword.
[0025] In one embodiment, the present invention replaces the traversal search-based PMI selection process with the neural network model at the user end, while the precoding matrix mapping mechanism based on a predefined codebook remains unchanged at the base station. The base station recovers the corresponding precoding matrix from the same predefined codebook as the user end according to the PMI parameters for downlink beamforming transmission.
[0026] Compared with existing technologies, the advantages of this invention are as follows: Addressing the problem of high computational complexity at the user end due to the need for traversing and searching candidate codewords in existing codebook-driven PMI selection methods, this invention utilizes a neural network to perform end-to-end learning and inference on preprocessed channel features, directly outputting codeword numbers and PMI parameters corresponding to predefined codebook rules, thereby significantly reducing the complexity of PMI selection at the user end. Simultaneously, this invention preprocesses the original channel matrix through downsampling and piecewise discrete Fourier transform, reducing input redundancy and improving feature extraction efficiency by leveraging the sparsity of the angle domain; by performing codeword prediction on sub-bands separately, it better adapts to the frequency selectivity of broadband channels; and this invention only replaces the PMI selection process at the user end, while maintaining the precoding matrix mapping mechanism based on the predefined codebook at the base station side. Therefore, it can be quickly coupled and deployed with existing 5G-A systems, improving CSI feedback real-time performance and downlink precoding transmission efficiency while ensuring PMI selection performance.
[0027] In summary, this invention achieves low-complexity PMI selection based on neural networks within the existing codebook feedback framework, providing a compatible and deployable implementation scheme for efficient CSI feedback and precoded transmission in 5G-A systems. Attached Figure Description
[0028] Figure 1 This outlines the overall deployment and operation process of the present invention.
[0029] Figure 2 This is a neural network structure diagram of the feature extraction and codeword mapping module in one embodiment of the present invention;
[0030] Figure 3 This is a flowchart illustrating the offline training and online deployment process of the present invention. Detailed Implementation
[0031] The embodiments of the present invention will now be described with reference to the accompanying drawings.
[0032] This invention designs a precoding matrix indication (PMI) selection method based on a neural network. While maintaining the predefined codebook mapping mechanism on the base station side, it replaces the user-side traversal search-based PMI selection process with a neural network, thereby reducing the computational complexity on the user side, improving the real-time performance of PMI feedback, and enhancing downlink precoding transmission efficiency. This method is applicable to codebook-driven CSI feedback scenarios. The following section uses a broadband multi-antenna downlink transmission scenario as an example to illustrate the specific implementation of this invention.
[0033] Consider a wideband multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) downlink system. The total number of bandwidth resource blocks is 144, divided into 12 sub-bands, each containing 12 resource blocks. The transmitter is equipped with 8 antennas, and the receiver with 2 antennas. The precoding matrix indicator uses a Type I codebook containing 128 candidate codewords. The PMI feedback result is used for precoding matrix recovery at the base station. In traditional schemes, the user terminal needs to iterate and evaluate all candidate PMIs one by one based on a predefined codebook, resulting in high computational complexity. Using the method described in this invention, the user terminal directly inputs the channel matrix into a pre-trained neural network model, which outputs the corresponding codeword number and further converts it into PMI parameters to be fed back to the base station.
[0034] The method is implemented according to the following steps:
[0035] (1) such as Figure 1 As shown, the user terminal first obtains the spatial domain channel matrix H of the downlink through the channel estimation module. In one embodiment, the original dimension of the channel matrix can be represented as the number of frequency points × the number of receive antennas × the number of transmit antennas. Since the CSI of adjacent frequency points within the broadband is highly correlated, to reduce input redundancy and subsequent computational complexity, downsampling processing is first performed on the original channel matrix. Specifically, the broadband channel is divided into 12 sub-bands, each containing 12 frequency points. Downsampling is performed on each sub-band at fixed intervals, for example, retaining 1 frequency point every 4 frequency points, thereby compressing the 12 frequency points of each sub-band into 3 frequency points.
[0036] After downsampling, the CSI data of each sub-band is reshaped into a two-dimensional feature matrix. In this embodiment, the downsampled CSI data of a single sub-band is reshaped into a 3×16 two-dimensional matrix to facilitate subsequent segmentation processing. Then, a piecewise discrete Fourier transform is performed on the two-dimensional feature matrix to convert the spatial domain channel features into angular domain channel features. Specifically, the 3×16 two-dimensional matrix is divided into four 3×4 segments by columns, and a discrete Fourier transform is performed on each segment, with the expression:
[0037]
[0038] in, This represents the segmented spatial domain channel matrix. This represents the conjugate transpose of the discrete Fourier transform matrix. This represents the channel matrix in the angle domain. Through the above transformation, the sparsity of the wireless channel in the angle domain can be utilized to improve the ability of subsequent neural networks to extract effective channel features.
[0039] Since the angle domain channel features obtained by piecewise discrete Fourier transform are in complex form, while neural network models typically use real tensors as input, the complex domain channel features are further converted into real input features. Specifically, the real and imaginary parts of each segment are extracted and concatenated along the column dimension to obtain a 3×8 real matrix; then, the real matrices corresponding to the four segments are concatenated to form a 3×32 real feature matrix for a single sub-band; finally, the feature matrices of the 12 sub-bands are stacked to form a real input tensor with dimensions of sample number × 12 × 3 × 32, which serves as the input to the subsequent neural network model.
[0040] (2) such as Figure 1 As shown, the neural network model includes a feature extraction module and a codeword mapping module. The feature extraction module is used to extract discriminative features for PMI selection from the preprocessed real-number input tensor, and the codeword mapping module is used to output the codeword evaluation results corresponding to each subband based on the discriminative features.
[0041] like Figure 2 As shown, in one embodiment, the feature extraction module first processes subbands 1 to K separately. It sequentially extracts a 3×32 feature matrix for each subband from the input tensor of dimensions (sample number × 12 × 3 × 32) and transforms its dimensions to fit the input format of the convolutional layer. Specifically, the 3×32 two-dimensional feature matrix of a single subband is converted into a 3×32×1 three-dimensional tensor. Subsequently, a convolutional neural network is used to extract local features from this three-dimensional tensor. In this embodiment, the feature extraction module sequentially passes through convolutional layers Conv1 (3×3 kernel) and Conv2 (3×3 kernel), with a residual connection between the two convolutional layers: the output of Conv1 is adjusted for the number of channels through a 1×1 convolution, added to the output of Conv2, and then output after passing through a ReLU activation function. This residual connection preserves shallow feature information, thereby enhancing the model's ability to express complex channel features. After convolution processing, the output feature map is flattened into a one-dimensional feature vector, which serves as the input to the codeword mapping module.
[0042] In the codeword mapping module, the flattened one-dimensional feature vector is first compressed and mapped to a low-dimensional latent space using a fully connected layer FC1 to retain the core discriminative features related to PMI selection. Subsequently, a fully connected layer FC2 further maps the latent space features to an output space corresponding to the codebook size. Finally, the probability distribution of each codeword is output through a Softmax activation function, and the user selects the codeword number corresponding to the highest probability as the prediction result for that sub-band. In this embodiment, the output space dimension is consistent with the total number of candidate codewords, i.e., 128 dimensions. The 128-dimensional output result represents the evaluation result of each candidate codeword corresponding to the current sub-band. After performing the above operations on each of the 12 sub-bands, the codeword evaluation result corresponding to each sub-band can be obtained, and the codeword number of each sub-band can be determined accordingly, thus constituting the broadband PMI feedback result.
[0043] (3) such as Figure 1 As shown, after obtaining the codeword numbers of each sub-band, the codeword numbers are converted into PMI parameters according to predefined codebook rules. In this invention, the codeword number of the k-th sub-band is used... For example, convert it into a PMI parameter triplet ( ), which satisfies:
[0044]
[0045]
[0046]
[0047]
[0048] in, This indicates the total number of codewords in the current codebook structure; These represent PMI parameters. The number of possible values for the parameter is determined by the antenna configuration, the predefined codebook type, and the transmission rank. This is a round-down operation. mod (·) represents the modulus extraction operation. For the configuration of 8 transmit antennas, 2 receive antennas, Type I codebook, and rank 2 described in this embodiment, the parameter triplet is used to uniquely determine the corresponding PMI information and to allow the base station to recover the precoding matrix according to the same codebook rules.
[0049] (4) such as Figure 3As shown, the neural network model employs offline training and online deployment. During the training phase, the neural network model is first built on a neural network training platform, for example, using TensorFlow as the backend and Keras for network construction. Training data can be generated by a channel simulation tool; the input is the channel matrix H under the corresponding scenario, and the output label is the optimal codeword number obtained through traversal search under the same predefined codebook constraints. The traversal search can evaluate candidate precoding matrices one by one according to evaluation criteria such as maximum mutual information to ensure the optimality of the training labels.
[0050] In one embodiment, the training process uses a cross-entropy loss function to constrain the difference between the model output and the training labels, the expression of which is:
[0051]
[0052] Where M is the total number of training samples, K is the number of subbands, and C is the number of codeword categories; The true label for the m-th sample, the k-th subband, and the c-th codeword; Let be the predicted probability of the m-th sample, the k-th subband, and the c-th codeword. During training, the Adam optimization algorithm can be used to update the neural network parameters, and the learning rate, batch size, and number of training epochs can be set according to the training convergence.
[0053] (5) After completing offline training, the trained neural network model is deployed to the user terminal hardware platform for real-time inference during actual operation. After receiving the downlink reference signal, the user terminal first completes channel estimation and obtains the current downlink channel matrix H; then, it performs downsampling, piecewise discrete Fourier transform and real-domain conversion according to step (1) to form neural network input features; then, it determines the codeword number corresponding to each sub-band and converts it to obtain PMI parameters through steps (2) and (3); finally, the user terminal feeds back the PM parameters to the base station through the uplink control channel. The base station does not need to modify the existing codebook mapping mechanism, but still uses the same predefined codebook as the user terminal, recovers the corresponding precoding matrix according to the received PMI parameters, and performs downlink beamforming transmission accordingly. It can be seen that the present invention only changes the user terminal PMI selection method, without changing the precoding matrix recovery process based on the codebook on the base station side, and thus can be quickly coupled and deployed with the existing system.
[0054] The embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.
Claims
1. A method for selecting precoding matrix indicators based on neural networks, characterized in that, Includes the following steps: (1) The user terminal obtains the downlink channel matrix H, and performs a combination of downsampling, piecewise discrete Fourier transform, and complex-to-real-domain feature preprocessing on the channel matrix H to obtain a real-valued feature tensor that is adapted to the neural network input. (2) Input the real number feature tensor into the end-to-end neural network model trained offline, and the neural network model directly outputs the evaluation probability of each sub-band codeword. The user terminal determines the optimal codeword number based on the probability distribution. (3) The precoding matrix indication parameter is obtained by converting the codeword number according to the predefined codebook rules, the precoding matrix indication parameter is fed back to the base station, and the base station obtains the precoding matrix by mapping according to the predefined codebook consistent with the user terminal.
2. The precoding matrix indication selection method based on neural networks according to claim 1, characterized in that, The downsampling includes: dividing the original channel state information within the broadband into multiple sub-bands, and selecting a portion of frequency points within each sub-band at preset intervals; the piecewise discrete Fourier transform includes: dividing the downsampled channel feature matrix into multiple segments, and performing a spatial-domain to angular-domain transformation on each segment, expressed as: ,in, This represents the segmented spatial domain channel matrix. It is the discrete Fourier transform matrix. for The conjugate transpose of . The channel matrix in the angle domain is represented; the transformation from the complex domain to the real domain includes: extracting the real and imaginary parts of each segment and concatenating them along the column dimension.
3. The precoding matrix indication selection method based on neural networks according to claim 1, characterized in that, The neural network model includes a feature extraction module and a codeword mapping module; the feature extraction module is used to extract discriminant features for precoding matrix indication selection from the preprocessed channel features, and the codeword mapping module is used to output the codeword number corresponding to each subband according to the discriminant features.
4. The precoding matrix indication selection method based on neural networks according to claim 3, characterized in that, The feature extraction module processes each subband separately, extracting local features through a neural network module containing convolutional layers and residual connections, and flattening them into a one-dimensional feature vector. The codeword mapping module maps the one-dimensional feature vector to an output space of similar size to the codebook through a fully connected layer, outputs the probability distribution of each codeword through a Softmax activation function, and selects the codeword number corresponding to the highest probability.
5. The method for selecting a precoding matrix indicator based on a neural network according to claim 1, characterized in that, The precoding matrix indicator parameter transformation includes numbering the codewords of the k-th subband. Convert to precoded matrix indicator parameters The conversion formula is: ； ； ； ； in, This indicates the total number of codewords in the current codebook structure; These represent the precoding matrix indicator parameters. The number of possible values is determined by the antenna configuration, the predefined codebook type, and the transmission rank; This is a round-down operation. mod (·) represents the modulus extraction operation.
6. The precoding matrix indication selection method based on neural networks according to claim 1, characterized in that, The neural network model is deployed using an offline training and online inference approach; the labels of the training samples are the optimal codeword numbers obtained through traversal search under the same predefined codebook constraints.
7. The precoding matrix indication selection method based on neural networks according to claim 6, characterized in that, The neural network model uses a cross-entropy loss function during training, the expression of which is: ； Where M is the total number of training samples; K is the number of subbands; and C is the number of codeword categories. The true label for the m-th sample, the k-th subband, and the c-th codeword; Let be the predicted probability of the m-th sample, the k-th subband, and the c-th codeword.
8. The precoding matrix indication selection method based on neural networks according to claim 1, characterized in that, The neural network model performs codeword prediction on each subband and constructs a wideband precoding matrix to indicate the feedback result based on the codeword prediction results of each subband.
9. The method for selecting a precoding matrix indicator based on a neural network according to claim 1, characterized in that, The neural network model is used at the user end to replace the precoding matrix indication selection process based on traversal search, while the precoding matrix mapping mechanism based on a predefined codebook remains unchanged at the base station side.
10. The neural network-based precoding matrix indication selection method according to any one of claims 1 to 9, characterized in that, The base station recovers the corresponding precoding matrix from the same predefined codebook as the user terminal according to the precoding matrix indication parameters, for use in downlink beamforming transmission.