CSI (Channel State Information) quantification method and device, electronic equipment and computer readable storage medium
By dividing the CSI compressed channel vector into multiple sub-compressed channel vectors and quantizing them with independent codebooks, and by optimizing the codebook with a sparsity strategy, the problem of large CSI quantization error is solved, and the CSI quantization accuracy and system performance are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-29
- Publication Date
- 2026-05-01
AI Technical Summary
The existing CSI quantization method has a large quantization error, which affects the performance of FDD large-scale MIMO systems.
The compressed channel vector of CSI is divided into multiple sub-compressed channel vectors, and each sub-compressed channel vector is quantized using an independent codebook. The quantized compressed channel vector is obtained by concatenating the sub-compressed channel vectors. A sparsity strategy is introduced to optimize the codebook and reduce quantization error.
By reducing the quantization error of the sub-compressed channel vector, the accuracy of CSI quantization and the overall system performance are improved, the quantization error is reduced, and the accuracy of CSI feedback is increased.
Smart Images
Figure CN121966632A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing, and more particularly to a CSI quantization method, apparatus, electronic device, and computer-readable storage medium. Background Technology
[0002] In Frequency Division Duplexing (FDD) massive Multi-Input Multi-Output (MIMO) systems, deep learning-based methods can be employed to effectively feedback Channel State Information (CSI). Research has been conducted on CSI compression and feedback using autoencoder models in relevant protocols. In this model, the raw CSI is compressed into a quantized bitstream at the user terminal (UE) using an encoder, and then transmitted over the air interface to network equipment (e.g., base station, BS). A decoder then decompresses the bitstream to recover the reconstructed CSI. The key to this process lies in the selection of the quantization strategy, which directly affects the overall system performance.
[0003] In related technologies, the CSI quantization method is often used as vector quantization. Vector quantization largely depends on the codebook. In the current vector quantization process, the entire CSI is mapped to a codeword in the codebook (which can also be understood as an embedded vector / quantization feature vector, which is a quantization feature representation of CSI). This can easily lead to large quantization errors. Summary of the Invention
[0004] This application provides a CSI quantization method, apparatus, electronic device, computer-readable storage medium, and computer program product to solve the problem of large existing quantization errors.
[0005] To solve the above-mentioned technical problems, this application is implemented as follows:
[0006] In a first aspect, embodiments of this application provide a Channel State Information (CSI) quantization method, applied to a terminal, the method comprising:
[0007] The feedback CSI is compressed to obtain the first compressed channel vector;
[0008] The first compressed channel vector is divided into M first sub-compressed channel vectors, where M is an integer greater than 1;
[0009] Obtain M codebooks, each of which corresponds one-to-one with one of the M first sub-compressed channel vectors;
[0010] For each first sub-compressed channel vector, the first sub-compressed channel vector is quantized based on the codebook of the first sub-compressed channel vector to obtain the quantized vector of the first sub-compressed channel vector;
[0011] The quantized vectors of the M first sub-compressed channel vectors are concatenated to obtain the quantized compressed channel vector of the CSI to be fed back.
[0012] Secondly, embodiments of this application provide a CSI quantization method applied to network devices, the method comprising:
[0013] The receiving terminal sends a quantized compressed channel vector of the CSI to be fed back, wherein the quantized compressed channel vector is a vector obtained by concatenating the quantized vectors of the M first sub-compressed channel vectors of the CSI to be fed back, where M is an integer greater than 1.
[0014] Thirdly, embodiments of this application provide a CSI quantization device applied to a terminal, the device comprising:
[0015] The compression module is used to compress the feedback CSI to obtain the first compressed channel vector;
[0016] The segmentation module is used to segment the first compressed channel vector into M first sub-compressed channel vectors, where M is an integer greater than 1;
[0017] The codebook acquisition module is used to acquire M codebooks, wherein each of the M codebooks corresponds one-to-one with one of the M first sub-compressed channel vectors;
[0018] The quantization module is used to quantize the first sub-compressed channel vector based on the codebook of the first sub-compressed channel vector for each first sub-compressed channel vector, so as to obtain the quantized vector of the first sub-compressed channel vector;
[0019] The splicing module is used to splice the quantized vectors of the M first sub-compressed channel vectors to obtain the quantized compressed channel vector of the CSI to be fed back.
[0020] Fourthly, embodiments of this application provide a CSI quantization device applied to a network device, the device comprising:
[0021] The receiving module is used to receive the quantized compressed channel vector of the CSI to be fed back sent by the terminal. The quantized compressed channel vector is the vector obtained by concatenating the quantized vectors of the M first sub-compressed channel vectors of the CSI to be fed back, where M is an integer greater than 1.
[0022] Fifthly, embodiments of this application provide an electronic device, including a transceiver and a processor.
[0023] The processor is used for:
[0024] The feedback CSI is compressed to obtain the first compressed channel vector;
[0025] The first compressed channel vector is divided into M first sub-compressed channel vectors, where M is an integer greater than 1;
[0026] Obtain M codebooks, each of which corresponds one-to-one with one of the M first sub-compressed channel vectors;
[0027] For each first sub-compressed channel vector, the first sub-compressed channel vector is quantized based on the codebook of the first sub-compressed channel vector to obtain the quantized vector of the first sub-compressed channel vector;
[0028] The quantized vectors of the M first sub-compressed channel vectors are concatenated to obtain the quantized compressed channel vector of the CSI to be fed back.
[0029] Sixthly, embodiments of this application provide an electronic device, including a transceiver and a processor.
[0030] The processor is used for:
[0031] The receiving terminal sends a quantized compressed channel vector of the CSI to be fed back, wherein the quantized compressed channel vector is a vector obtained by concatenating the quantized vectors of the M first sub-compressed channel vectors of the CSI to be fed back, where M is an integer greater than 1.
[0032] In a seventh aspect, embodiments of this application provide an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the CSI quantization method described in the first or second aspect above.
[0033] Eighthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the CSI quantization method described in the first or second aspect above.
[0034] Ninthly, embodiments of this application provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the method described above.
[0035] In this embodiment, the compressed channel vector of CSI can be divided into M first sub-compressed channel vectors with reduced dimension. Then, for each of the M first sub-compressed channel vectors, a corresponding codebook is used to perform independent quantization. That is, quantization can be performed independently in the codebook corresponding to each first sub-compressed channel vector, which can better preserve the local structure of the data, reduce quantization error, and thus improve the quantization effect. Attached Figure Description
[0036] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application 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.
[0037] Figure 1 This is one of the flowcharts of a CSI quantization method provided in the embodiments of this application;
[0038] Figure 2 This is a second flowchart of a CSI quantization method provided in an embodiment of this application;
[0039] Figure 3 This is a CSI feedback principle diagram provided in an embodiment of this application;
[0040] Figure 4 This is a schematic diagram of a parameter learning principle provided in an embodiment of this application;
[0041] Figure 5 This application provides a quantitative effect comparison chart in its embodiments;
[0042] Figure 6 This is one of the structural schematic diagrams of a CSI quantization device provided in the embodiments of this application;
[0043] Figure 7 This is a second schematic diagram of the structure of a CSI quantization device provided in the embodiments of this application;
[0044] Figure 8 This is one of the structural schematic diagrams of an electronic device provided in the embodiments of this application;
[0045] Figure 9 This is a second schematic diagram of the structure of an electronic device provided in the embodiments of this application. Detailed Implementation
[0046] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0047] See Figure 1 , Figure 1 This is a flowchart of a CSI quantization method provided in an embodiment of this application, which can be applied to a terminal. For example... Figure 1 As shown, the CSI quantization method provided in this embodiment includes the following steps:
[0048] Step 101: Compress the CSI to be fed back to obtain the first compressed channel vector.
[0049] This application provides various methods for compressing the feedback CSI in its embodiments, without specific limitations. For example, compression can be performed through, but is not limited to, a trained encoder.
[0050] Step 102: Divide the first compressed channel vector into M first sub-compressed channel vectors, where M is an integer greater than 1.
[0051] Step 103: Obtain M codebooks, each of which corresponds one-to-one with one of the M first sub-compressed channel vectors;
[0052] It should be understood that the first compressed channel vector can be divided into M sub-compressed channel vectors in subspaces. Each of the M first sub-compressed channel vectors corresponds one-to-one with one of the M subspaces (M independent subspaces). Each of the M first sub-compressed channel vectors includes the sub-compressed channel vectors of the M subspaces, and one sub-compressed channel vector in one subspace corresponds to one first sub-compressed channel vector. That is, in this embodiment, the first compressed channel vector can be split into M low-dimensional first sub-compressed channel vectors, reducing the dimensionality of each individual first sub-compressed channel vector compared to the first compressed channel vector.
[0053] In addition, M codebooks are obtained, one codebook for each subspace, with a one-to-one correspondence between the M subspaces and the M codebooks, that is, a one-to-one correspondence between the M codebooks and the M first sub-compression channel vectors. In other words, in this embodiment, for each subspace, a corresponding codebook is obtained, and the codebook corresponding to each subspace can also be understood as a sub-codebook, that is, M sub-codebooks can be obtained.
[0054] Step 104: For each first sub-compressed channel vector, quantize the first sub-compressed channel vector based on the codebook of the first sub-compressed channel vector to obtain the quantized vector of the first sub-compressed channel vector.
[0055] Step 105: Concatenate the quantized vectors of the M first sub-compressed channel vectors to obtain the quantized compressed channel vector to be fed back as CSI.
[0056] For each first sub-compressed channel vector, quantization is performed using the codebook corresponding to that first sub-compressed channel vector from among M codebooks to obtain the quantized vector of that first sub-compressed channel vector, thus completing the quantization of that first sub-compressed channel vector. Each of the M first sub-compressed channel vectors is quantized in a similar manner, resulting in M quantized vectors of the first sub-compressed channel vectors. In this way, for each first sub-compressed channel vector with reduced dimension, quantization is performed independently using the corresponding sub-codebook, meaning that each first sub-compressed channel vector with reduced subspace dimension is quantized independently. Then, the quantized vectors of the M first sub-compressed channel vectors are concatenated to obtain the quantized compressed channel vector (i.e., the quantized compressed channel vector) to be fed back as CSI.
[0057] For example, if the compressed channel vector is D (an integer greater than 1) dimensional, the D-dimensional space can be divided into M subspaces. For instance, each subspace can have a dimension of D / M, and the codebook corresponding to each subspace can be obtained, i.e., M codebooks can be obtained. The compressed channel vector is then divided into the same dimension as the subspace to obtain M first sub-compressed channel vectors. For each first sub-compressed channel vector, the corresponding codebook is used for independent quantization, and then the quantized vectors of the M first sub-compressed channel vectors are concatenated.
[0058] In this embodiment, the compressed channel vector of CSI can be divided into M first sub-compressed channel vectors with reduced dimensionality. Then, for each of the M first sub-compressed channel vectors, a corresponding codebook is used to perform independent quantization. That is, quantization can be performed independently in the codebook corresponding to each first sub-compressed channel vector, which can better preserve the local structure of the data, reduce quantization error, and thus improve the quantization effect. At the same time, during the quantization process, M quantized vectors are obtained. Combining the M quantized vectors yields the final quantized compressed channel vector of the CSI to be fed back. In this way, by combining M different quantized vectors, the error of the quantization vector in a single subspace can be compensated to a certain extent, thereby improving the overall quantization accuracy and quantization effect.
[0059] In some embodiments, the first sub-compressed channel vector is quantized based on the codebook of the first sub-compressed channel vector to obtain the quantized vector of the first sub-compressed channel vector, including:
[0060] Calculate the distance between the first sub-compressed channel vector and each codeword in the codebook of the first sub-compressed channel vector;
[0061] The target codeword is determined as the quantization vector of the first sub-compressed channel vector. The target codeword is the codeword in the codebook of the first sub-compressed channel vector that is closest to the first sub-compressed channel vector.
[0062] That is, the first sub-compressed channel vector is mapped to the codeword in the corresponding codebook that is closest to it, thus completing the quantization of the first sub-compressed channel vector. A similar quantization process is performed for each first sub-compressed channel vector, mapping each first sub-compressed channel vector to the codeword in its corresponding codebook, completing the independent quantization of M first sub-compressed channel vectors, reducing information detail loss and improving quantization performance.
[0063] In some embodiments, the M codebooks are determined through the following training method:
[0064] Obtain CSI samples;
[0065] The CSI samples are compressed to obtain the second compressed channel vector;
[0066] The second compressed channel vector is segmented to obtain M second sub-compressed channel vectors;
[0067] Initialize the codebook to obtain M initial codebooks;
[0068] For each second sub-compressed channel vector of the second compressed channel vector, the second sub-compressed channel vector is quantized based on the initial codebook of the second sub-compressed channel vector to obtain the quantized vector of the second sub-compressed channel vector;
[0069] The target loss is determined based on the M second sub-compressed channel vectors and the quantization vectors of the M second sub-compressed channel vectors;
[0070] Based on the target loss, the M initial codebooks are adjusted to determine the M codebooks.
[0071] It should be understood that the M initial codebooks are the initial codebooks of the M subspaces, and there is a one-to-one correspondence between the M initial codebooks and the M subspaces. The M codebooks are the result of training and adjusting the M initial codebooks using the training method of this embodiment. The M second sub-compressed channel vectors of the second compressed channel vector can be understood as the second sub-compressed channel vectors of the M subspaces, corresponding one-to-one with the M subspaces. That is, the M second sub-compressed channel vectors correspond one-to-one with the M initial codebooks.
[0072] Additionally, it should be noted that the total number of CSI samples can be N, where N is a positive integer, meaning there can be one or more CSI samples (N greater than 1). When there are multiple CSI samples (N greater than 1), each CSI sample is compressed to obtain a corresponding second compressed channel vector, resulting in N second compressed channel vectors, each corresponding to one of the N CSI samples. Each second compressed channel vector can be further segmented to obtain M corresponding second sub-compressed channel vectors, meaning each second compressed channel vector corresponds to M second sub-compressed channel vectors. For each CSI sample, each of the M second sub-compressed channel vectors corresponding to that CSI sample is quantized using its corresponding initial codebook to obtain its quantized vector, thus obtaining the quantized vectors of the M second sub-compressed channel vectors corresponding to that CSI sample. Finally, based on the M second sub-compressed channel vectors corresponding to each of the N CSI samples and their quantized vectors, the target loss is determined.
[0073] For example, quantizing the second sub-compressed channel vector based on the initial codebook of the second sub-compressed channel vector may include: calculating the distance between the second sub-compressed channel vector and each codeword in the initial codebook of the second sub-compressed channel vector; determining a reference codeword as the quantization vector of the second sub-compressed channel vector, wherein the reference codeword is the codeword in the initial codebook of the second sub-compressed channel vector that is closest to the second sub-compressed channel vector.
[0074] In this embodiment, each of the M second sub-compressed channel vectors can be independently quantized using a corresponding initial codebook. Through the M initial codebooks, the quantization of the M second sub-compressed channel vectors is completed, resulting in corresponding quantized vectors. Using the M second sub-compressed channel vectors and their quantized vectors, a target loss is determined. This target loss is then used to adjust the M initial codebooks, i.e., to optimize the codebooks, thus determining the trained codebooks, i.e., the optimized M codebooks. This completes codebook learning, enabling the learning of the CSI sample data distribution. Subsequently, the M codebooks learned from the codebooks are used to quantize the CSI samples to be fed back, improving the codebook quantization effect.
[0075] In some embodiments, the target loss is determined based on the M second sub-compressed channel vectors and the quantization vectors of the M second sub-compressed channel vectors, including:
[0076] Calculate the total quantization loss based on the M second sub-compressed channel vectors and the quantization vectors of the M second sub-compressed channel vectors;
[0077] The CSI sample is reconstructed based on the quantization vectors of the M second sub-compressed channel vectors to obtain the reconstructed CSI corresponding to the CSI sample;
[0078] Calculate the reconstruction loss based on the CSI samples and the reconstructed CSI corresponding to the CSI samples;
[0079] The target loss is determined based on the total loss and the reconstruction loss.
[0080] In calculating the training loss, on the one hand, quantization loss is considered. For each CSI sample, the quantization loss of that CSI sample can be calculated using the M second sub-compressed channel vectors and their corresponding quantization vectors, thus obtaining the total quantization loss of the CSI sample. On the other hand, after obtaining the quantization vectors of the M second sub-compressed channel vectors, CSI reconstruction can be performed based on the quantization vectors of the M second sub-compressed channel vectors obtained by quantization using the M codebooks, resulting in the reconstructed CSI corresponding to the CSI sample. Since losses may occur during the reconstruction process, i.e., there may be differences between the reconstructed CSI and the original CSI sample, and since the second sub-compressed channel vectors are obtained by quantization using the initial codebook, the quality of the initial codebook directly affects the quantization quality. Moreover, the more accurate the quantization vectors of the second sub-compressed channel vectors are, the smaller the difference between the reconstructed CSI and the original CSI sample. Therefore, during the codebook learning process, the loss between the reconstructed CSI and the original CSI sample is also considered, i.e., the CSI reconstruction loss is taken into account. Based on total loss and reconstruction loss, a target loss is determined for codebook learning to improve the codebook learning effect, thereby improving the quality of the M codebooks learned. CSI quantization is then performed using the M codebooks to improve the CSI quantization effect.
[0081] In some embodiments, the total quantitative loss is calculated as follows:
[0082]
[0083] in,
[0084] Indicates total loss. Let z represent the quantization loss of the i-th CSI sample, sg[·] represent the gradient stopping operation, α represent the first preset weight (α > 0 and < 1), β represent the second preset weight (β > 0 and < 1), ||·|| represents the L2 norm, ||·||1 represents the L1 norm, and z m,i Let z′ represent the m-th second sub-compressed channel vector corresponding to the i-th CSI in the CSI sample. m,i C represents the quantization vector of the m-th second sub-compressed channel vector corresponding to the i-th CSI in the CSI sample. m This represents the m-th initial codebook among M initial codebooks, where N is the total number of CSI samples.
[0085] The codebook includes multiple codewords, some of which may be redundant or unused during training. Therefore, this embodiment can reduce redundant codewords in the codebook through a sparsity strategy, and also enhance the separation between codewords in the codebook, making different codewords more distinguishable. This helps to more accurately capture subtle differences in information, thereby improving the quantization effect. In implementation, by controlling the quantization error (corresponding to...) Based on this, the L1 norm of the codebook is introduced to achieve sparsification of the codebook. The second preset weight β can also be called the sparsification weight, which can control the degree of sparsification of the codebook.
[0086] In some embodiments, the reconstruction loss is calculated as follows:
[0087]
[0088] Where N is the total number of CSI samples, H represents the reconstruction loss. i ′ represents the i-th CSI in the CSI sample, H″ i This represents the reconstructed CSI corresponding to the i-th CSI in the CSI sample.
[0089] In some embodiments, CSI samples are compressed by an encoder, the CSI to be fed back is compressed by a trained encoder, and the CSI samples are reconstructed by a decoder. The trained decoder is used to reconstruct the CSI to be fed back.
[0090] The encoder and decoder are trained in the following way:
[0091] Based on the target loss, the parameters of the encoder and decoder are adjusted to determine the trained encoder and decoder.
[0092] During training, not only is quantization using a codebook, but CSI compression via an encoder and CSI reconstruction via a decoder are also required. In this embodiment, not only the codebook is optimized, but also the parameters of the encoder and decoder are optimized. The parameters of the encoder and decoder are optimized using target loss, so that the trained encoder can compress the feedback CSI to improve the compression effect, and the trained decoder can reconstruct the feedback CSI to improve the reconstruction performance. In this way, a more accurate reconstructed CSI corresponding to the feedback CS is obtained.
[0093] In some embodiments, the method further includes:
[0094] Send the quantized compressed channel vector to the network device.
[0095] In other words, CSI quantization feedback is implemented, which reduces CSI feedback overhead and improves transmission performance. Furthermore, during the process of receiving the quantized compressed channel vector of the CSI from the terminal, M quantized vectors are obtained by independently quantizing M first sub-compressed channel vectors using M codebooks, resulting in M quantized vectors. This improves the quantization effect and yields more accurate quantized vectors. These M quantized vectors are then concatenated to obtain the quantized compressed channel vector, which is then fed back to the network equipment, thus improving CSI feedback performance.
[0096] See Figure 2 , Figure 2 This is a flowchart of a CSI quantization method provided in an embodiment of this application, applied to network devices (e.g., exemplary, including but not limited to base stations, etc.). Figure 2 As shown, the CSI quantization method provided in this embodiment includes the following steps:
[0097] Step 201: Receive the quantized compressed channel vector of the CSI to be fed back sent by the receiving terminal. The quantized compressed channel vector is the vector obtained by concatenating the quantized vectors of the M first sub-compressed channel vectors of the CSI to be fed back, where M is an integer greater than 1.
[0098] In some embodiments, the method further includes:
[0099] The quantized compressed channel vector is used to reconstruct the CSI to be fed back, thus obtaining the reconstructed CSI corresponding to the CS to be fed back.
[0100] It should be noted that the method described above for network devices is different from the method described above for terminals, and their technical features are similar, so they will not be described again here.
[0101] The process of the above method will be specifically described below with some specific embodiments.
[0102] Introduction to related technologies:
[0103] In FDD massive MIMO systems, deep learning-based methods can be used to effectively feed back Channel State Information (CSI, or Channel State Data). Research has been conducted on CSI compression and feedback using autoencoder models in relevant protocols. In this model, the raw CSI is compressed into a quantized bitstream at the UE using an encoder model and transmitted to the BS over the air interface. Then, a decoder model decompresses it into the recovered CSI (i.e., reconstructed CSI). The key to this process lies in the choice of quantization strategy, which directly affects the overall system performance. Commonly used quantization methods in related technologies include:
[0104] 1. CSI feedback based on scalar quantization:
[0105] To ensure compatibility of latent vectors with communication systems, they need to be converted into finite-length bit sequences. This leads to the proposal of CSI feedback based on scalar quantization. Scalar quantization can be divided into uniform quantization and non-uniform quantization. Uniform quantization can be viewed as a rounding operation, converting continuous values into discrete values using rounding after determining the quantization interval. Non-uniform quantization, on the other hand, uses a suitable companding function based on the signal distribution characteristics, employing different quantization intervals in different intervals to improve quantization performance.
[0106] 2. CSI feedback based on vector quantization:
[0107] When the input signal is non-uniformly distributed, uniform quantization can lead to significant quantization errors in certain intervals. The improved performance of non-uniform quantization relies on the companding function being able to fit the distribution of the input signal. However, in deep learning-based CSI feedback problems, the output distribution at the encoder is random, and directly using non-uniform quantization can actually degrade performance. Vector quantization is implemented by introducing a Vector Quantization Variational Autoencoder (VQ-VAE) framework. This framework enables end-to-end learning of the codeword distribution at the encoder output, improving the feedback performance of CSI.
[0108] CSI feedback based on vector quantization is highly dependent on the size and quality of the codebook. A codebook that is too small may lead to large quantization errors, while a codebook that is too large will cause quantization features to jump back and forth between different codewords, reducing the performance of the model.
[0109] This application proposes a product quantization CSI feedback method based on a codebook set. This method proposes a product quantization approach based on a codebook set, solving the problem of quantization features jumping back and forth between different codewords in vector quantization. Simultaneously, a sparsity strategy is introduced during the codebook set learning process, retaining relevant features (attributes related to the current learning task are called "relevant features") while removing a large number of irrelevant features (attributes unrelated to the current learning task are called "irrelevant features"), thereby improving the codebook learning efficiency.
[0110] This application proposes a product quantization CSI feedback method based on codebook sets, such as... Figure 3 As shown, it mainly includes:
[0111] A product quantization method based on codebook set is proposed to provide feedback on CSI;
[0112] A parameter learning method for codebook sets is proposed, which learns the CSI data distribution during end-to-end training by introducing a sparsity strategy.
[0113] Figure 3 A network element can be a network device, such as a base station.
[0114] For product quantization based on codebook set, the compressed channel vector is quantized as follows:
[0115] This embodiment differs from vector quantization. After compressing the CSI to obtain a compressed channel vector, the compressed channel vector is decomposed into M sub-vectors (i.e., M sub-compressed channel vectors), which can be understood as sub-vectors of M subspaces. Each sub-vector is then quantized using a corresponding codebook. Subsequently, M quantization results (quantization vectors) are obtained, and the quantization results of multiple subspaces are combined. In this way, even if there is some error in the quantization of each subspace, the combination of different subspaces can compensate for the deficiencies of a single subspace to a certain extent, thereby improving the overall quantization accuracy. This design, by splitting the compressed channel vector into multiple subspaces, each corresponding to a codebook, limits the problem of quantization features jumping back and forth between different codewords while achieving a smaller quantization error.
[0116] For the m-th subspace after splitting, configure a subcodebook C of size J×K. m This can also be called the m-th sub-codebook (if it's before training, it's called the m-th initial codebook; if it's after training and obtaining M codebooks, it's called the m-th codebook). Where C... m It contains K embedding vectors c of length J. m,k (i.e., code words). c m,k The parameters are generated randomly, and can be represented as C. m The k-th embedding vector, i.e., the k-th codeword, is continuously optimized during subsequent end-to-end training. The overall codebook C can be represented as the Cartesian product of all sub-codebooks:
[0117] C = C1 × C2 × … × C M ;
[0118] The codebook set C contains all possible codeword combinations; therefore, in the method of this application embodiment, the set of M codebooks is referred to as the "codebook set". Although the overall codebook size grows exponentially, the storage requirement is significantly reduced because this application embodiment does not require explicitly storing the entire Cartesian product, but only the codebooks of each subspace. The implementation process of quantizing latent feature vectors using the codebook set includes the following steps:
[0119] First: Compressed channel vector segmentation:
[0120] Using the data processing center E, the compressed channel vector z of CSI is obtained (E can be compressed using a common encoder in neural networks to obtain z; specific details are not elaborated here). Then, z is divided into M sub-compressed channel vectors:
[0121] z = [z1, z2, ..., zM ];
[0122] Wherein, each sub-compressed channel vector z m Belongs to a low-dimensional subspace This represents the m-th component of z, which can also be understood as the sub-compressed channel vector corresponding to the m-th subspace, i.e., the m-th sub-compressed channel vector.
[0123] Secondly: Obtain the feature index of the corresponding sub-compressed channel vector (i.e., the index of the codeword closest to the sub-compressed channel vector):
[0124] For each subvector z m Calculate its relationship with all embedding vectors c in the corresponding codebook. m,k Calculate the distance between them and obtain the shortest distance embedding vector c. m,k index q m (It can also be represented as k′) m ):
[0125] q m =argmin k ||z m -c m,k ||2;
[0126] Furthermore: the quantization vector of the sub-compressed channel vector is obtained based on the feature index:
[0127] Based on the feature index, from the subcodebook C m Extract the corresponding embedding vector to represent the sub-compressed channel vector:
[0128] z′ m =C m,qm
[0129] z′ m Represents the sub-compressed channel vector z m The corresponding quantization vector, i.e. the m-th quantization vector, the quantization vector of the m-th subspace, can also be called the m-th sub-quantization feature;
[0130] The quantized compressed channel vector can then be represented as the concatenation of all sub-quantized features:
[0131] z′=[z′1,z′2,…,z′ M ].
[0132] For parameter learning methods of codebook sets:
[0133] To enable the codebook to adapt to CSI feedback tasks in different scenarios, the method in this embodiment obtains the global loss. That is, the target loss is used, and existing deep learning parameter optimization methods are employed to jointly optimize the parameters of the codebook set, data processing center E (encoder), and data processing center D (decoder) in this embodiment. The main parameter learning process is as follows:
[0134] First: Obtain the global loss.
[0135] The quantization operation in the codebook-based CSI feedback method lacks gradients; therefore, during training, these gradients cannot be propagated to the data processing center E via backpropagation. This solution bypasses the codebook by directly estimating the gradients. Specifically, the gradients from data processing center D are directly copied to data processing center E. This reduces the quantization loss of the codebook. Defined as:
[0136]
[0137] Here, sg[·] is the gradient stopping operation, which is used to prevent the gradient from backpropagating to the quantization step. The first term ||sg[z e,m ]-z′ e,m || 2 The parameters used to update the codebook set are such that the embedding vector c m,k It moves in the direction of the sub-compressed channel vector to better represent the current data. The second term α||z e,m -sg[z′ e,m ]|| 2 This serves as a constraint loss, preventing the data processing center E from outputting arbitrarily and instead guiding it towards the codebook set. α is a weighting parameter used to balance the effects of these two losses. The value of α is a constant greater than 0 and less than 1, and its specific value needs to be adjusted according to the actual situation.
[0138] The codebook contains multiple embedding vectors, some of which may be redundant or unused during training. Therefore, this embodiment reduces redundant embedding vectors in the codebook through a sparsity strategy. Simultaneously, the sparsity strategy enhances the separation between embedding vectors in the codebook, making different embedding vectors more discriminative. This improved separation helps the model more accurately capture subtle differences in the data, thereby improving quantization performance. In implementation, this embodiment introduces the L1 norm to achieve sparsity of the codebook parameters, and the quantization loss... Defined as:
[0139]
[0140] Among them, ||C m ||1 is for retrieving the subcodebook C m The L1 norm, β, is the sparsity weight used to control the subcodebook C.m The sparsity of β is determined by the fact that β is a constant that is greater than 0 and less than 1, and the specific value needs to be adjusted according to the actual situation.
[0141] It should be noted that the above formula is the quantization loss for a specific CSI. The quantization loss for the i-th CSI is different. It can be represented as follows:
[0142]
[0143] In addition, to reduce the gap between the original CSI and the reconstructed CSI, this embodiment uses mean squared error (MSE) as the reconstruction loss.
[0144]
[0145] Where N is the total number of CSI samples used in the learning and training process, and H′ i and H″ i These are the values of the original data and the reconstructed data at the i-th point, respectively, i.e., H′. i H″ is the i-th CSI in the CSI sample. i Let be the reconstructed CSI corresponding to the i-th CSI in the CSI samples, also known as the i-th reconstructed CSI. Then, the global loss of the CSI feedback method based on the codebook set is... It can be defined as follows:
[0146]
[0147] Secondly: Parameter learning:
[0148] The process of learning the method parameters in this embodiment is as follows: Figure 4 As shown, a common optimizer used in deep learning methods (details omitted) is employed, and the optimization is based on the number of iterations, learning rate, and global loss. And so on, optimizing the codebook set (which can be understood as optimizing M initial codebooks to obtain M codebooks), the parameters of data processing center E and data processing center D.
[0149] The maximum number of iterations and learning rate can be preset, and the codebook set parameters can be initialized (i.e., M initial codebooks can be initialized).
[0150] Data processing center E compresses the CSI samples to generate compressed channel vectors for the CSI samples;
[0151] Quantization process: The compressed channel vector is segmented to obtain M sub-compressed channel vectors; the feature index corresponding to each sub-compressed channel vector in the codebook set is obtained; the sub-compressed channel vectors are quantized according to the feature index, i.e., compressed channel vector quantization; the quantization loss is obtained.
[0152] The quantized compressed channel vector is fed back to the network element, and the data processing center D in the network element reconstructs (i.e. reassembles) the CSI based on the quantized compressed channel vector.
[0153] Obtain the reconstruction loss;
[0154] Based on the quantization loss and reconstruction loss, obtain the global loss, i.e., the target loss;
[0155] Determine if the maximum number of iterations has been reached;
[0156] If the target is reached, training ends; otherwise, parameters are updated, such as updating the initial codebook, the parameters of data processing centers E and D (i.e., the encoder and decoder parameters), and the process returns to data processing center E for compression. The compression, quantization, reconstruction, and loss acquisition processes are repeated for the next iteration until the maximum number of iterations is reached, at which point the training process ends. In the case of training interpretation, the codebook, encoder, and decoder are learned, resulting in M codebooks, a trained encoder, and a trained decoder. Subsequently, the terminal can use the trained encoder to compress channel state data (e.g., CSI data to be fed back), and use the M codebooks to quantize the sub-compressed channel vectors segmented from the CSI data to be fed back. The network device can then reconstruct the quantized compressed channel vectors of the CSI data fed back by the terminal using the trained decoder, thus reconstructing the channel state data (CSI).
[0157] Effect verification:
[0158] In this embodiment, both the training and test samples can be generated using the COST 2100 channel model. This embodiment can be applied to different scenarios. Taking an indoor microcell scenario as an example, the operating frequency band is 5.3 GHz, and the base station (BS) is located at the center of a square area with dimensions of 20m × 20m. An example is illustrated by setting the number of BS antennas to 32, configuring the UE with a single antenna, and having 1024 subcarriers. Sampling each movement of the terminal yields an original CSI matrix with dimensions of 1024 × 32.
[0159] As an example, the total generated data is as follows: 100,000 training samples, 30,000 validation samples, and 20,000 test samples. The number of iterations, learning rate, and batch size can be set to 1000, 0.001, and 200, respectively, and the network parameters are optimized using the Adam optimizer. Subsequently, the same processing method as the baseline model CRNet (channel reconstruction network) using an autoencoder model is used to obtain an input tensor of dimension 2×32×32, which is then fed into the network for training.
[0160] To verify the effectiveness of the method proposed in this application, this application, based on CRNet, sets the compression ratio to 4 and introduces vector quantization operations for different embedding vector sizes, and trains it as a reference model. Subsequently, CRNet is compared with the method proposed in this application. For evaluation, Normalized Mean Square Error (NMSE) is used to assess the performance of different methods in CSI reconstruction.
[0161] Quantification effect comparison Figure 5 As shown, the horizontal axis K represents the number of codewords, and the vertical axis is NMSE in decibels (dB). When using Vector Quantization (VQ) to quantize the CRNet encoder output (compressed channel vector for CSI compression), the performance significantly degrades. This indicates that poor codebook design leads to a significant decrease in quantization performance. Due to codebook design issues, the VQ method may fail to effectively capture channel characteristics when compressing and reconstructing Channel State Information (CSI), resulting in larger errors, especially in complex channel environments. In contrast, the performance of CSI feedback is improved when using the method proposed in this application. This result demonstrates that the proposed method is better designed to match channel characteristics, improving quantization accuracy and reducing reconstruction errors.
[0162] This application proposes a novel CSI feedback method, which employs a product quantization method based on the codebook set to improve the accuracy and efficiency of CSI feedback. Furthermore, this application proposes a parameter optimization method that improves the efficiency of end-to-end training of the codebook set by introducing a sparsity strategy.
[0163] Compared to vector quantization methods in related technologies, the method in this application addresses the problem of quantized features jumping back and forth between different codewords by splitting the latent feature vector into multiple subspaces and designing a corresponding codebook for each subspace. Compared to the parameter learning methods mentioned in vector quantization, the method in this application further reduces the sparsity of codewords in the codebook during end-to-end training, thereby alleviating data redundancy in the codebook.
[0164] This application provides a novel approach to CSI feedback based on artificial intelligence (AI), which improves the feedback accuracy of existing feedback methods and can also save network overhead and increase the spectrum efficiency of 5G networks.
[0165] like Figure 6 As shown, Figure 6 This is a schematic diagram of a CSI quantization device provided in an embodiment of this application, which can be applied to a terminal, such as... Figure 6 As shown, the CSI quantization device 600 includes:
[0166] Compression module 601 is used to compress the CSI to be fed back to obtain a first compressed channel vector;
[0167] The segmentation module 602 is used to segment the first compressed channel vector into M first sub-compressed channel vectors, where M is an integer greater than 1;
[0168] The codebook acquisition module 603 is used to acquire M codebooks, and the M codebooks correspond one-to-one with the M first sub-compressed channel vectors;
[0169] The quantization module 604 is used to quantize the first sub-compressed channel vector based on the codebook of the first sub-compressed channel vector for each first sub-compressed channel vector, so as to obtain the quantized vector of the first sub-compressed channel vector;
[0170] The splicing module 605 is used to splice the quantized vectors of the M first sub-compressed channel vectors to obtain the quantized compressed channel vector to be fed back as CSI.
[0171] In some embodiments, the quantization module includes:
[0172] The distance calculation unit is used to calculate the distance between the first sub-compressed channel vector and each codeword in the codebook of the first sub-compressed channel vector;
[0173] The determining unit is used to determine the target codeword as the quantization vector of the first sub-compressed channel vector, wherein the target codeword is the codeword in the codebook of the first sub-compressed channel vector that is closest to the first sub-compressed channel vector.
[0174] In some embodiments, the M codebooks are determined through the following training method:
[0175] Obtain CSI samples;
[0176] The CSI samples are compressed to obtain the second compressed channel vector;
[0177] The second compressed channel vector is segmented to obtain M second sub-compressed channel vectors;
[0178] Initialize the codebook to obtain M initial codebooks;
[0179] For each second sub-compressed channel vector of the second compressed channel vector, the second sub-compressed channel vector is quantized based on the initial codebook of the second sub-compressed channel vector to obtain the quantized vector of the second sub-compressed channel vector;
[0180] The target loss is determined based on the M second sub-compressed channel vectors and the quantization vectors of the M second sub-compressed channel vectors;
[0181] Based on the target loss, the M initial codebooks are adjusted to determine the M codebooks.
[0182] In some embodiments, the target loss is determined based on the M second sub-compressed channel vectors and the quantization vectors of the M second sub-compressed channel vectors, including:
[0183] Calculate the total quantization loss based on the M second sub-compressed channel vectors and the quantization vectors of the M second sub-compressed channel vectors;
[0184] The CSI sample is reconstructed based on the quantization vectors of the M second sub-compressed channel vectors to obtain the reconstructed CSI corresponding to the CSI sample;
[0185] Calculate the reconstruction loss based on the CSI samples and the reconstructed CSI corresponding to the CSI samples;
[0186] The target loss is determined based on the total loss and the reconstruction loss.
[0187] In some embodiments, the total quantitative loss is calculated as follows:
[0188]
[0189] in,
[0190] Indicates total loss. Let z represent the quantization loss of the i-th CSI sample, sg[·] represent the gradient stopping operation, α represent the first preset weight (α > 0 and < 1), β represent the second preset weight (β > 0 and < 1), ||·|| represents the L2 norm, ||·||1 represents the L1 norm, and z m,iLet z′ represent the m-th second sub-compressed channel vector corresponding to the i-th CSI in the CSI sample. m,i C represents the quantization vector of the m-th second sub-compressed channel vector corresponding to the i-th CSI in the CSI sample. m This represents the m-th initial codebook among M initial codebooks, where N is the total number of CSI samples.
[0191] In some embodiments, CSI samples are compressed by an encoder, the CSI to be fed back is compressed by a trained encoder, and the CSI samples are reconstructed by a decoder. The trained decoder is used to reconstruct the CSI to be fed back.
[0192] The encoder and decoder are trained in the following way:
[0193] Based on the target loss, the parameters of the encoder and decoder are adjusted to determine the trained encoder and decoder.
[0194] In some embodiments, the apparatus further includes:
[0195] The transmitting module is used to send the quantized compressed channel vector to the network device.
[0196] The CSI quantization device 600 provided in this embodiment can realize each process of the various embodiments of the CSI quantization method applied to terminal devices described above. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, it will not be described again here.
[0197] See Figure 7 , Figure 7 This is a schematic diagram of a CSI quantization device provided in an embodiment of this application, which can be applied to network devices, such as... Figure 7 As shown, the CSI quantization device 700 includes:
[0198] The receiving module 701 is used to receive the quantized compressed channel vector of the CSI to be fed back sent by the terminal. The quantized compressed channel vector is the vector obtained by concatenating the quantized vectors of the M first sub-compressed channel vectors of the CSI to be fed back, where M is an integer greater than 1.
[0199] In some embodiments, the apparatus further includes:
[0200] The reconstruction module is used to reconstruct the CSI to be fed back based on the quantized compressed channel vector, so as to obtain the reconstructed CSI corresponding to the CS to be fed back.
[0201] The CSI quantization device 700 provided in this embodiment can implement the various processes of the above-described CSI quantization method applied to network devices. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, it will not be described again here.
[0202] This application also provides an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the various processes of the above-described CSI quantization method embodiment applied to the terminal and achieves the same technical effect. To avoid repetition, it will not be described again here.
[0203] For details, see Figure 8 This application also provides an electronic device, including a bus 801, a transceiver 802, an antenna 803, a bus interface 804, a processor 805, and a memory 806.
[0204] The processor 805 is used for:
[0205] The feedback CSI is compressed to obtain the first compressed channel vector;
[0206] The first compressed channel vector is divided into M first sub-compressed channel vectors, where M is an integer greater than 1;
[0207] Obtain M codebooks, each of which corresponds one-to-one with one of the M first sub-compressed channel vectors;
[0208] For each first sub-compressed channel vector, the first sub-compressed channel vector is quantized based on the codebook of the first sub-compressed channel vector to obtain the quantized vector of the first sub-compressed channel vector;
[0209] The quantized vectors of the M first sub-compressed channel vectors are concatenated to obtain the quantized compressed channel vector of the CSI to be fed back.
[0210] exist Figure 8 In this document, a bus architecture (represented by bus 801) is used. Bus 801 can include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 805 and memory represented by memory 806. Bus 801 can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 804 provides an interface between bus 801 and transceiver 802. Transceiver 802 can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 805 is transmitted over a wireless medium via antenna 803, which further receives data and transmits data to processor 805.
[0211] The processor 805 manages the bus 801 and handles general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. The memory 806 can be used to store data used by the processor 805 during operation.
[0212] Optionally, the processor 805 can be a CPU, ASIC, FPGA, or CPLD.
[0213] The processor 805 of the electronic device provided in this embodiment can implement each process of the above-described embodiments of the CSI quantization method applied to the terminal. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, it will not be described again here.
[0214] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the various processes of the CSI quantization method embodiment applied to a terminal described above, and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.
[0215] This application also provides an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the various processes of the above-described CSI quantization method embodiment applied to network devices and achieves the same technical effect. To avoid repetition, it will not be described again here.
[0216] For details, see Figure 9 As shown in the figure, this application embodiment also provides an electronic device, including a bus 901, a transceiver 902, an antenna 903, a bus interface 904, a processor 905, and a memory 906.
[0217] The processor 905 is used for:
[0218] The receiving terminal sends the quantized compressed channel vector of the CSI to be fed back. The quantized compressed channel vector is the vector obtained by concatenating the quantized vectors of the M first sub-compressed channel vectors of the CSI to be fed back, where M is an integer greater than 1.
[0219] exist Figure 9In this document, a bus architecture (represented by bus 901) is used. Bus 901 can include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 905 and memory represented by memory 906. Bus 901 can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 904 provides an interface between bus 901 and transceiver 902. Transceiver 902 can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 905 is transmitted over a wireless medium via antenna 903, which further receives data and transmits it to processor 905.
[0220] Processor 905 manages bus 901 and general processing, and also provides various functions, including timing, peripheral interface, voltage regulation, power management, and other control functions. Memory 906 can be used to store data used by processor 905 during operation.
[0221] Optionally, the processor 905 can be a CPU, ASIC, FPGA, or CPLD.
[0222] The processor 905 of the electronic device provided in this embodiment can implement each process of the various embodiments of the CSI quantization method applied to network devices described above. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, it will not be described again here.
[0223] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the various processes of the CSI quantization method embodiments described above for network devices, achieving the same technical effects. To avoid repetition, these details will not be repeated here. The computer-readable storage medium may be, for example, ROM, RAM, magnetic disk, or optical disk.
[0224] This application provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the method described in the embodiment. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, they will not be described again here.
[0225] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0226] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, 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 is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or first network device, etc.) to execute the methods of the various embodiments of this application.
[0227] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.
Claims
1. A method for quantizing Channel State Information (CSI), characterized in that, Applied to a terminal, the method includes: The feedback CSI is compressed to obtain the first compressed channel vector; The first compressed channel vector is divided into M first sub-compressed channel vectors, where M is an integer greater than 1; Obtain M codebooks, each of which corresponds one-to-one with one of the M first sub-compressed channel vectors; For each first sub-compressed channel vector, the first sub-compressed channel vector is quantized based on the codebook of the first sub-compressed channel vector to obtain the quantized vector of the first sub-compressed channel vector; The quantized vectors of the M first sub-compressed channel vectors are concatenated to obtain the quantized compressed channel vector of the CSI to be fed back.
2. The method according to claim 1, characterized in that, The step of quantizing the first sub-compressed channel vector based on the codebook of the first sub-compressed channel vector to obtain the quantized vector of the first sub-compressed channel vector includes: Calculate the distance between the first sub-compressed channel vector and each codeword in the codebook of the first sub-compressed channel vector; The target codeword is determined as the quantization vector of the first sub-compressed channel vector, wherein the target codeword is the codeword in the codebook of the first sub-compressed channel vector that is closest to the first sub-compressed channel vector.
3. The method according to claim 1, characterized in that, The M codebooks were determined through the following training method: Obtain CSI samples; The CSI samples are compressed to obtain a second compressed channel vector; The second compressed channel vector is segmented to obtain M second sub-compressed channel vectors of the second compressed channel vector; Initialize the codebook to obtain M initial codebooks; For each second sub-compressed channel vector of the second compressed channel vector, the second sub-compressed channel vector is quantized based on the initial codebook of the second sub-compressed channel vector to obtain the quantized vector of the second sub-compressed channel vector; The target loss is determined based on the M second sub-compressed channel vectors and the quantization vectors of the M second sub-compressed channel vectors; Based on the target loss, the M initial codebooks are adjusted to determine the M codebooks.
4. The method according to claim 3, characterized in that, The determination of the target loss based on the M second sub-compressed channel vectors and the quantization vector of the M second sub-compressed channel vectors includes: Calculate the total quantization loss based on the M second sub-compressed channel vectors and the quantization vectors of the M second sub-compressed channel vectors; The CSI sample is reconstructed based on the quantization vectors of the M second sub-compressed channel vectors to obtain the reconstructed CSI corresponding to the CSI sample; Calculate the reconstruction loss based on the CSI sample and the reconstructed CSI corresponding to the CSI sample; The target loss is determined based on the totalized loss and the reconstruction loss.
5. The method according to claim 4, characterized in that, The total loss is calculated as follows: in, The This represents the totalized loss. Let sg[·] represent the quantization loss of the i-th CSI in the CSI sample, sg[·] represent the gradient stopping operation, α represent the first preset weight (α > 0 and < 1), β represent the second preset weight (β > 0 and < 1), ||·|| represent the L2 norm, ||·||1 represent the L1 norm, and z m,i z represents the m-th second sub-compressed channel vector corresponding to the i-th CSI in the CSI sample. ′ m,i C represents the quantization vector of the m-th second sub-compressed channel vector corresponding to the i-th CSI in the CSI sample. m This represents the m-th initial codebook among the M initial codebooks, and N is the total number of CSI samples.
6. The method according to claim 3, characterized in that, The CSI sample is compressed by an encoder, the CSI to be fed back is compressed by the trained encoder, and the CSI sample is reconstructed by a decoder. The trained decoder is used to reconstruct the CSI to be fed back. The encoder and the decoder are trained in the following manner: Based on the target loss, the parameters of the encoder and the decoder are adjusted to determine the trained encoder and the trained decoder.
7. The method according to any one of claims 1-6, characterized in that, The method further includes: The quantized compressed channel vector is sent to the network device.
8. A CSI quantization method, characterized in that, Applied to network devices, the method includes: The receiving terminal sends a quantized compressed channel vector of the CSI to be fed back, wherein the quantized compressed channel vector is a vector obtained by concatenating the quantized vectors of the M first sub-compressed channel vectors of the CSI to be fed back, where M is an integer greater than 1.
9. The method according to claim 8, characterized in that, The method further includes: The CSI to be fed back is reconstructed based on the quantized compressed channel vector to obtain the reconstructed CSI corresponding to the CS to be fed back.
10. A CSI quantization device, characterized in that, Applied to a terminal, the device includes: The compression module is used to compress the feedback CSI to obtain the first compressed channel vector; The segmentation module is used to segment the first compressed channel vector into M first sub-compressed channel vectors, where M is an integer greater than 1; The codebook acquisition module is used to acquire M codebooks, wherein each of the M codebooks corresponds one-to-one with one of the M first sub-compressed channel vectors; The quantization module is used to quantize the first sub-compressed channel vector based on the codebook of the first sub-compressed channel vector for each first sub-compressed channel vector, so as to obtain the quantized vector of the first sub-compressed channel vector; The splicing module is used to splice the quantized vectors of the M first sub-compressed channel vectors to obtain the quantized compressed channel vector of the CSI to be fed back.
11. A CSI quantization device, characterized in that, Applied to network devices, the device includes: The receiving module is used to receive the quantized compressed channel vector of the CSI to be fed back sent by the terminal. The quantized compressed channel vector is the vector obtained by concatenating the quantized vectors of the M first sub-compressed channel vectors of the CSI to be fed back, where M is an integer greater than 1.
12. An electronic device, characterized in that, Including transceivers and processors, The processor is used for: The feedback CSI is compressed to obtain the first compressed channel vector; The first compressed channel vector is divided into M first sub-compressed channel vectors, where M is an integer greater than 1; Obtain M codebooks, each of which corresponds one-to-one with one of the M first sub-compressed channel vectors; For each first sub-compressed channel vector, the first sub-compressed channel vector is quantized based on the codebook of the first sub-compressed channel vector to obtain the quantized vector of the first sub-compressed channel vector; The quantized vectors of the M first sub-compressed channel vectors are concatenated to obtain the quantized compressed channel vector of the CSI to be fed back.
13. An electronic device, characterized in that, Including transceivers and processors, The processor is used for: The receiving terminal sends a quantized compressed channel vector of the CSI to be fed back, wherein the quantized compressed channel vector is a vector obtained by concatenating the quantized vectors of the M first sub-compressed channel vectors of the CSI to be fed back, where M is an integer greater than 1.
14. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method as claimed in any one of claims 1 to 7, or implements the steps of the method as claimed in any one of claims 8 to 9.
15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 7, or implements the steps of the method as described in any one of claims 8 to 9.
16. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1-9.