WO2025213896A1 - 信道状态信息压缩方法、重构方法和电子设备 - Google Patents

信道状态信息压缩方法、重构方法和电子设备

Info

Publication number
WO2025213896A1
WO2025213896A1 PCT/CN2025/070654 CN2025070654W WO2025213896A1 WO 2025213896 A1 WO2025213896 A1 WO 2025213896A1 CN 2025070654 W CN2025070654 W CN 2025070654W WO 2025213896 A1 WO2025213896 A1 WO 2025213896A1
Authority
WO
WIPO (PCT)
Prior art keywords
channel
common
feature
csi
features
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/CN2025/070654
Other languages
English (en)
French (fr)
Inventor
毕媛媛
郭化盐
刘坚能
张文凯
程敏
陈家璇
陈志堂
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huawei Technologies Co Ltd
Hong Kong University of Science and Technology
Original Assignee
Huawei Technologies Co Ltd
Hong Kong University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huawei Technologies Co Ltd, Hong Kong University of Science and Technology filed Critical Huawei Technologies Co Ltd
Publication of WO2025213896A1 publication Critical patent/WO2025213896A1/zh
Pending legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00—Arrangements for detecting or preventing errors in the information received
    • H04L1/02—Arrangements for detecting or preventing errors in the information received by diversity reception
    • H04L1/06—Arrangements for detecting or preventing errors in the information received by diversity reception using space diversity
    • H04L1/0618—Space-time coding
    • H04L1/0675—Space-time coding characterised by the signaling
    • H04L1/0693—Partial feedback, e.g. partial channel state information [CSI]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/0413—MIMO systems
    • H04B7/0452—Multi-user MIMO systems
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station
    • H04B7/0613—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission
    • H04B7/0615—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal
    • H04B7/0619—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal using feedback from receiving side
    • H04B7/0621—Feedback content
    • H04B7/0626—Channel coefficients, e.g. channel state information [CSI]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00—Arrangements for detecting or preventing errors in the information received
    • H04L1/02—Arrangements for detecting or preventing errors in the information received by diversity reception
    • H04L1/06—Arrangements for detecting or preventing errors in the information received by diversity reception using space diversity

Definitions

  • the present invention relates to the field of artificial intelligence or communication technology, and in particular to a channel state information compression method, a reconstruction method and an electronic device.
  • a base station serves a user equipment (UE)
  • UE user equipment
  • CSI channel state information
  • the UE compresses (or encodes) the CSI to obtain a bit stream, which is then transmitted to the BS.
  • the BS then decodes the bit stream and obtains the reconstructed CSI.
  • one type of scheme leverages expert experience.
  • the CSI matrices of multiple UEs may be jointly correlated due to sharing a common local scattering set. Therefore, this scheme designs a distributed CS compression and reconstruction framework that exploits the joint correlation between the CSI matrices of multiple UEs to further compress the CSI matrices.
  • such schemes rely heavily on expert experience and typically make strong assumptions. While they perform well when these assumptions are met, their performance degrades dramatically if these assumptions are violated, leaving room for improvement in their generalizability or versatility.
  • the embodiments of the present application provide a channel state information compression method, a reconstruction method, and an electronic device, which do not rely on expert experience and improve generalization capability and versatility.
  • an embodiment of the present application provides a channel state information compression method that can be applied to a first user equipment UE, first determining the first channel characteristics corresponding to the first UE; the first channel characteristics are obtained based on the first channel state information CSI to be encoded; and determining a mask matrix; based on the first channel characteristics and the mask matrix, obtaining a bit stream corresponding to the first UE.
  • the first UE is one of the multiple UEs.
  • the first CSI to be encoded is the original CSI matrix.
  • the base station obtains reconstructed channel state information (reconstructed CSI) according to the bit stream.
  • This method does not rely on expert experience or numerous assumptions, resulting in improved generalization and versatility. It is applicable to multiple units (MUs), multi-station, and multi-frequency scenarios, significantly improving generalization and versatility.
  • MUs multiple units
  • multi-station multi-station
  • multi-frequency scenarios significantly improving generalization and versatility.
  • a mask matrix is used to indicate which features are redundant, reducing the amount of communication resources occupied by these redundant features.
  • obtaining a bit stream corresponding to the first UE based on the first channel characteristic and the mask matrix may be performed by performing a mask operation on the first channel characteristic based on the mask matrix to obtain the bit stream corresponding to the first UE.
  • Masking operations can explicitly remove redundant features at the encoding end. For example, using a mask matrix to set redundant features to 0 makes the features obtained more sparse and have smaller entropy, thus achieving smaller reconstruction loss at the same compression ratio.
  • determining the mask matrix may be determining a first mask matrix corresponding to the first UE based on a first CSI corresponding to the first UE; or determining the same mask matrix corresponding to multiple UEs, where the multiple UEs include the first UE.
  • Multiple UEs can learn the same mask matrix, or multiple UEs can learn their own mask matrices respectively.
  • the embodiment of the present application proposes a new coding network solution for CSI compression and reconstruction scenarios.
  • a mask generator is added to the encoder.
  • the mask generator is used to learn redundant features.
  • the mask generator is used to indicate redundant features, thereby removing redundant features through masking operations.
  • the first channel feature corresponding to the first UE is determined by an encoder, which can be to output a mean feature and a scale feature through the encoder; based on the mean feature and the scale feature and the first parameter matrix, the first channel feature corresponding to the first UE is determined; wherein the first parameter matrix obeys a sparse distribution.
  • the encoder includes at least one encoding module, a first processing layer, a second processing layer, and a first normalization layer connected to the second processing layer; wherein the output of at least one encoding module is input to the first processing layer and the second processing layer in parallel; the first processing layer or the second processing layer is implemented based on a convolutional layer or an attention mechanism; through the encoder, the mean feature and the scale feature are output, which can be through the first processing layer to output the mean feature; through the second processing layer and the first normalization layer, the scale feature is output.
  • the mask generator shares at least one encoding module with the encoder; the mask generator also includes a third processing layer and a second normalization layer; the output of at least one encoding module is also input in parallel to the third processing layer; or, the mask generator includes at least one dedicated encoding module and a third processing layer and a second normalization layer; the mask matrix is determined by the mask generator, which can be output through the third processing layer and the second normalization layer.
  • the bit stream corresponding to the first UE is obtained, which can be based on the point product of the first channel characteristics and the mask matrix to obtain the hidden layer characteristics; based on the hidden layer characteristics, the bit stream corresponding to the first UE is obtained.
  • the method may also determine a first loss based on the first channel characteristics; determine a second loss based on the mask matrix; and optimize the trainable parameters in the encoder and the mask generator based on the loss function; wherein the loss function is the first loss and the second loss.
  • determining the first loss may be to determine a first prior distribution; the first prior distribution is a sparse distribution; based on the KL divergence loss function, calculating the distance between the first distribution corresponding to the first channel feature and the first prior distribution to obtain the first loss; based on the mask matrix, determining the second loss may be to determine a second prior distribution; the second prior distribution is a sparse distribution; based on the KL divergence loss function, calculating the distance between the second distribution corresponding to the mask matrix and the second prior distribution to obtain the second loss.
  • the first loss is the loss corresponding to the encoder
  • the second loss is the loss corresponding to the mask generator.
  • the prior distributions of the encoder and mask are designed separately to control their sparsity and prevent trivial solutions.
  • an embodiment of the present application also proposes a channel state information reconstruction method, which can be applied to a base station BS, first determining at least one second channel characteristic corresponding to at least one UE; the second channel characteristic is obtained based on the bit stream uploaded by the UE; based on the at least one second channel characteristic, a common characteristic is obtained; based on the common characteristic and the second channel characteristic corresponding to at least one UE, a second CSI corresponding to at least one UE is obtained.
  • the solution proposed in this embodiment decomposes the channel into two components for reconstruction: common features and personalized channel features. This eliminates the need to consider personalized features when reconstructing the common channel, while eliminating the interference of common features when reconstructing the personalized channel. Compared to directly learning a fused channel, this approach reduces the difficulty of network learning and is more conducive to obtaining a high-performance model, thereby achieving high-performance CSI information compression and reconstruction.
  • common channel state information can be obtained based on the common characteristics; after obtaining the second CSI corresponding to at least one UE, third CSI corresponding to at least one UE can be obtained based on the common channel state information and the second CSI corresponding to at least one UE.
  • the embodiments of the present application also propose extracting the common channel (common CSI), decomposing the channel into two parts: a personalized channel and a common channel.
  • the channels are reconstructed using personalized features and common features respectively, and finally fused to obtain the final reconstructed channel.
  • the common channel represents common information in the environment and belongs to low-frequency background information. In this way, superimposing the personalized channel on the common channel is equivalent to completing the details and high-frequency information, which can take into account both types of information at the same time, thereby achieving a smaller reconstruction loss.
  • the BS is deployed with a decoding network, which includes a decoder and a common feature extractor; at least one UE includes a first UE; based on at least one second channel feature, a common feature is obtained, which can be inputting at least one second channel feature into the common feature extractor, and obtaining the common feature through the common feature extractor; based on the common feature and the second channel feature corresponding to at least one UE, a second CSI corresponding to at least one UE is obtained, which can be inputting the common feature and the channel feature corresponding to the first UE into a decoder corresponding to the first UE, and obtaining the second CSI corresponding to the first UE through the decoder.
  • a decoding network which includes a decoder and a common feature extractor
  • at least one UE includes a first UE
  • a common feature is obtained, which can be inputting at least one second channel feature into the common feature extractor, and obtaining the common feature through the common feature extractor
  • the second CSI corresponding to at least one UE is obtained.
  • the second channel characteristics are input into a decoder to obtain decoded second channel characteristics, and then the decoded second channel characteristics and the common characteristics are input into the decoder again to obtain the second CSI.
  • the BS further includes a common channel generator; obtaining the common channel state information based on the common characteristics may be inputting the common characteristics into the common channel generator to obtain the common channel state information.
  • the embodiment of the present application proposes a new decoding network, which adds a common feature extractor and a common channel generator on the basis of the basic solution to extract common features and common channels respectively.
  • an embodiment of the present application further provides an electronic device, comprising: a processor, wherein the processor is configured to execute a computer program or instruction in a memory to implement any of the methods described above.
  • an embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored program, wherein when the program is executed by a processor, the method as described in any one of the above items is implemented.
  • an embodiment of the present application also provides a chip system, comprising: a communication interface for inputting and/or outputting data; a processor for executing a computer executable program so that a device equipped with the chip system executes any of the methods described above.
  • FIG1 is a schematic diagram of CSI compression and reconstruction between the UE and the BS in a channel estimation scenario
  • Figure 2 is an example diagram of the MU scenario
  • FIG3 is a diagram illustrating an example of a system architecture of a channel state information compression method provided in an embodiment of the present application
  • FIG4 is an exemplary diagram of the architecture design of the coding network in the channel state information compression method provided in an embodiment of the present application
  • FIG5 is another exemplary diagram of the architecture design of the coding network in the channel state information compression method provided in an embodiment of the present application.
  • FIG6 is an example diagram of an encoding module in a channel state information compression method provided in an embodiment of the present application.
  • FIG7 is a schematic diagram of a core network architecture in one embodiment of a channel state information compression method provided in an embodiment of the present application
  • FIG8 is a flow chart of an embodiment of a channel state information compression method provided in an embodiment of the present application.
  • FIG9 is an exemplary diagram of a specific architecture design of a coding network in one embodiment of a channel state information compression method provided in an embodiment of the present application;
  • FIG10 is a diagram illustrating an exemplary network architecture design of a decoder in one embodiment of a channel state information compression method provided in an embodiment of the present application
  • Figure 11a is a performance data table for a 2-UE scenario
  • Figure 11b is a performance data table for three UE scenarios
  • Figure 12 shows the NMSE comparison of two UEs, three UEs and a single UE in the MU scenario.
  • a base station Before a base station (BS) can serve a UE, it must perform precoding and other processing. This requires obtaining CSI between the BS and the UE.
  • CSI reflects environmental information between the BS and the UE, such as scattering and refraction caused by obstacles between them. Typically, the UE feeds CSI back to the BS. Since CSI is a complex matrix with dimensions equal to the number of antennas and the number of frequency bands, providing feedback based on the original dimensions of the CSI matrix results in significant communication overhead.
  • the original CSI matrix H needs to be compressed and quantized through a series of operations on the UE side to obtain a 01-bit stream h with controllable length.
  • the BS After the BS receives the 01-bit stream h, it reconstructs the CSI matrix Among them, the original CSI matrix, the compressed quantized bit stream, and the reconstructed CSI matrix are represented by H, h, CSI Compression indicates CSI compression.
  • the fifth generation mobile communication technology 5G era
  • multiple antennas can be installed on both the base station side and the UE side.
  • the "number of antennas" in the CSI is the product of the number of antennas of the BS and the UE. However, this will cause interference between adjacent antenna signals.
  • Massive MIMO technology is a technology used to reduce interference and has been widely used in 5G.
  • the same base station can serve multiple UEs at the same time, and the same UE can also receive services from multiple different base stations, thereby increasing the channel capacity when the spectrum resources are fixed.
  • a base station usually serves multiple different UEs, and a UE often communicates with multiple base stations.
  • multiple frequency bands Multi-Band
  • the frequency band here refers to the "center frequency point”. Channels on multiple frequency points can be collected around a center frequency point. These frequency point combinations are called "frequency bands" in CSI.
  • one base station serves multiple UEs at the same time, which can be referred to as multi-user (MU); multiple base stations serve one UE at the same time, which can be referred to as multi-station; and serving on multiple frequency bands can be referred to as multi-band.
  • MU multi-user
  • multi-station multiple base stations serve one UE at the same time, which can be referred to as multi-station
  • serving on multiple frequency bands can be referred to as multi-band.
  • the simultaneously transmitted channels have a certain degree of correlation.
  • the multiple channels in the multi-user, multi-station or multi-frequency application scenarios are collectively referred to as "correlated channels" or "correlated channel pairs".
  • the environment between the UE and the BS is complex and changeable, resulting in a variety of correlations between related channels. Therefore, high requirements are placed on the versatility and generalization of the compression and reconstruction scheme.
  • Each UE needs to share the same encoder. For example, the parameters of the encoding network need to be exactly the same, but it is hoped that the information extracted by each UE will not be repeated as much as possible.
  • NMSE Normalized Mean Squared Error
  • CSI compression and reconstruction were performed through expert design schemes.
  • some schemes combine distributed source coding technology and deep learning to jointly compress and reconstruct related channel pairs.
  • neural network-based solutions have significant advantages over expert white box solutions.
  • the uninterpretability of neural networks makes it difficult to integrate white box solutions with basic neural network solutions. This makes it difficult to expand neural network-based solutions from basic scenarios to multi-user, multi-station, or multi-frequency application scenarios.
  • a multi-user downlink CSI feedback method based on deep learning first sends a pilot signal, and the UE performs channel estimation to obtain the downlink CSI; then, the UE compresses and quantizes the CSI and feeds it back to the BS; the BS receives the downlink CSI feedback from all UEs and performs joint CSI reconstruction; the BS uses the reconstructed CSI to implement precoding and communicate with all UEs at the same time.
  • the UE-side CSI compression and quantization of this solution is identical to the basic compression and reconstruction solution for a single UE-to-single-base station scenario. This means that during the encoding phase, the UE-side compression does not account for the correlation between related channels in multi-user, multi-station, or multi-frequency scenarios.
  • the only difference from the basic compression and reconstruction solution for a single UE-to-single-base station is that the BS uses CSI information sent by other UEs when restoring the channel.
  • this multi-user downlink CSI feedback method based on deep learning does not consider the existence of other UEs when encoding on the UE side, and does not extract more comprehensive feature information during encoding. It completely relies on the decoding end to guide the encoding end to remove the information of other UEs during compression.
  • the decoding end decodes the bit stream transmitted by the UE.
  • the feature information extracted on the UE side is limited, it is difficult for the decoding end to remove the feature information of other UEs, which increases the training difficulty of the decoding end.
  • the neural network will automatically learn how to integrate the information of each UE.
  • the neural network is difficult to learn and train, and the training results are difficult to meet expectations, resulting in the trained model being difficult to apply in practice and not having the ability to solve practical problems.
  • some related neural network-based compression and reconstruction schemes have the network learn end-to-end which features are fed back by which UE, and then fuse the CSI matrices of all UEs at the decoding end. Because these schemes implicitly remove redundant features, there is no explicit distinction between redundant and non-redundant features, or the distinction is rather vague. This makes it difficult for the neural network to learn which features are redundant and which are non-redundant. If the neural network could be taught explicitly which features are redundant and can be reused, and which features require separate feedback, it would be more conducive to neural network learning and convergence, improving the performance of the trained model.
  • the embodiments of the present application propose a channel state information compression method and/or reconstruction method.
  • it can be a joint compression and reconstruction scheme of related channels (pairs) based on the AI autoencoder, which improves the ability of the autoencoder to compress features and reconstruct the CSI matrix, the generalization ability and versatility of the trained model in practical applications, and the performance of the model is improved.
  • one manifestation of improved model performance can be a lower NMSE at the same compression ratio. That is, at the same compression ratio (the compression ratio between the original CSI and the compressed bitstream is the same), the NMSE between the reconstructed CSI and the original CSI decreases. Alternatively, while the NMSE remains unchanged, the bitstream is transmitted using fewer communication resources, meaning the bitstream data volume is reduced and a higher compression ratio is achieved.
  • the method provided in the embodiments of the present application can be applied to any and various scenarios, including MU, multi-station, and multi-frequency scenarios.
  • it can be applied to the MU scenario shown in Figure 2.
  • the scenario shown in Figure 2 is j UEs per base station, where 1 ⁇ i ⁇ j, and i and j are both integers.
  • the original matrix corresponding to the i-th UEi is H i
  • the compressed bit stream is h i .
  • the following description will mostly use the MU scenario as an example.
  • the neural network scheme for single-UE channel compression feedback is defined as a basic network scheme in the embodiment of the present application.
  • the method proposed in the embodiment of the present application can be applied to multi-UE scenarios, and proposes how to utilize the common characteristics between multiple UEs to further improve the performance of compression reconstruction.
  • This multi-UE joint compression reconstruction scheme can be adapted to different basic network schemes. For example, it can be compatible with or combined with various basic network schemes.
  • the scheme proposed in the embodiment of the present application can, on the premise of improving the compression reconstruction performance in multi-UE scenarios, give full play to the superior performance of the basic network scheme itself.
  • the method provided in the embodiments of the present application explicitly divides CSI into common features and personalized features.
  • Common features can be understood as the common features of the channel characteristics corresponding to multiple UEs
  • personalized features can be understood as the unique features of the channel characteristics corresponding to each of the multiple UEs. This method reduces reconstruction error and can achieve a smaller NMSE when compressed to the same bit stream length.
  • the encoder can use a mask to set redundant features to zero, thereby reducing the number of features that need to be transmitted.
  • adjacent channel features in the spatial and/or frequency domains are collected to compensate for missing features, ultimately achieving a smaller NMSE with the same number of feedback bits.
  • the channel state information compression method provided in the embodiment of the present application can be executed based on the system architecture shown in Figure 3.
  • a Mask generator is added on the UE side to extract common features between related channels or related channel pairs, and to divide the channel features into personalized features and common features for separate learning.
  • the neural network deployed on the UE side can be referred to as a coding network, which includes at least an encoder and a Mask generator.
  • UE1 (the first UE) is used as an example.
  • UE1 is any of multiple UEs.
  • UE1 encodes the original CSI matrix through an encoder to obtain first channel features.
  • the first channel features can be understood as the personalized channel features extracted by the encoder.
  • the mask matrix is output.
  • one UE corresponds to a dedicated Mask generator, and the channel state information (i.e., original CSI) of multiple UEs is input into multiple Mask generators respectively to obtain multiple Mask matrices U.
  • the original CSI corresponding to UE1 is input into the Mask generator dedicated to UE1, and the mask matrix U1 (first mask matrix) corresponding to UE1 is output.
  • the Mask generator needs to take the original CSI corresponding to the corresponding UE as input.
  • multiple UEs can jointly learn the same mask matrix, and directly use the mask matrix U as a learnable variable. Multiple UEs only learn the same U. In this case, the Mask generator does not need to input the original CSI information corresponding to a single UE.
  • a bit stream corresponding to the first UE is obtained based on the first channel characteristics and the mask matrix.
  • a mask matrix is used to indicate mask positions.
  • the mask matrix can be used to perform masking processing on the encoded first channel features.
  • the purpose of the masking processing is to further remove redundant features in the channel features and reduce the amount of data that needs to be transmitted.
  • the masking processing performed on the encoded first channel features using the mask matrix can be performed by calculating the dot product of the mask matrix and the first channel feature matrix to achieve the purpose of setting the values of the elements at corresponding positions in the first channel feature matrix to 0.
  • obtaining a bit stream corresponding to the first UE based on the first channel characteristics and the mask matrix can be performed by, after performing masking, continuing to perform quantization, entropy coding, and other operations on the masked first channel characteristic matrix to obtain a bit stream to be transmitted.
  • the masking operation can remove redundant features, and the removal of redundant features can further reduce the amount of data to be transmitted, thereby obtaining a bit stream with a higher compression ratio.
  • the compression ratio here refers to the compression ratio between the original CSI and the bit stream.
  • the encoder may include at least one encoding module, a first processing layer, and a second processing layer, and a first normalization layer connected to the second processing layer.
  • the encoder includes encoding modules 100, 101, and 102, as well as processing layers 201 and 202, and a normalization layer 301 connected to processing layer 202.
  • Encoding modules 101 to 103 are connected in series, and the output of encoding module 102 is input in parallel to processing layers 201 and 202.
  • the output of processing layer 201 is used as the mean feature; a normalization layer is connected after processing layer 202, and scale features are obtained after normalization.
  • the mask generator can share at least one encoding module with the encoder.
  • the encoder and the mask generator share an encoding module. It should be noted that the number of encoding modules shown in FIG4 is three for example only. In other embodiments, the number of encoding modules can be appropriately increased or decreased.
  • the mask generator also includes a third processing layer and a second normalization layer connected to the third processing layer.
  • the output of at least one encoding module is also input in parallel to the third processing layer.
  • the mask generator includes processing layer 203 and normalization layer 302.
  • the output of encoding module 102 is input to processing layer 203. After passing through normalization layer 302, a mask matrix is obtained.
  • the Mask generator includes at least one dedicated encoding module, for example, encoding module 103 , encoding module 104 , and encoding module 105 , a processing layer 203 , and a normalization layer 302 .
  • the processing layer can be implemented based on at least one convolutional layer or an attention-based layer.
  • processing layers 201 to 203 are all convolutional layers.
  • Normalization layer 301 or normalization layer 302 can be implemented based on a sigmoid function.
  • an encoding module may adopt the architecture design shown in Figure 6.
  • an encoding module may include a convolutional layer 1001, a normalization layer 1002, and a PReLU layer 1003.
  • the normalization layer 1002 may specifically be a batch normalization layer.
  • the activation function PReLU Parametric Rectified Linear Unit
  • Leaky ReLU Leaky ReLU
  • ELU Extended Linear Unit
  • Maxout Maxout
  • the first channel feature corresponding to UE1 may be determined in the following manner:
  • the encoder outputs mean features and scale features.
  • the output of processing layer 201 is the mean feature
  • the output of normalization layer 301 is the scale feature.
  • the first parameter matrix is a matrix that follows a sparse distribution.
  • the first parameter matrix can follow a multivariate Laplace distribution.
  • a matrix L is obtained by sampling from the multivariate standard Laplace distribution. In other words, each element in matrix L follows a standard Laplace distribution.
  • M represents the first channel characteristic
  • M represents the mean feature (specifically, it can be the mean feature matrix)
  • the scale feature specifically, it can be a scale feature matrix
  • L is the first parameter matrix.
  • the first parameter matrix is not limited to obeying the Laplace distribution.
  • it may also obey a sparse distribution such as a normal distribution or a ⁇ distribution.
  • the masked hidden features are obtained based on the first channel features and the mask matrix. For example, the hidden features are obtained by calculating the dot product of the first channel features and the mask matrix. Subsequently, the hidden features are quantized and entropy coded to obtain a bitstream.
  • the function of the mask matrix is to set the elements at the corresponding positions in the channel feature matrix to 0, and masking can be achieved through the dot multiplication operation.
  • the loss function includes a partial loss corresponding to the channel features and a partial loss corresponding to the mask matrix.
  • the two losses are defined as the first loss and the second loss, respectively.
  • the first loss is obtained based on the output first channel features
  • the second loss is obtained based on the mask matrix.
  • the first losses may be determined in the following manners:
  • the first prior distribution can be a sparse distribution such as Laplace distribution, normal distribution, beta distribution, etc.
  • the first prior distribution is Laplace distribution.
  • the distance between the first distribution corresponding to the first channel feature and the first prior distribution is calculated to obtain the first loss.
  • the second loss may be determined in the following manner:
  • the second prior distribution can be a sparse distribution such as Laplace distribution, normal distribution, or beta distribution; for example, the second prior distribution is a Laplace distribution.
  • the distance between the second distribution corresponding to the mask matrix and the second prior distribution is calculated to obtain the second loss.
  • the loss function corresponding to the encoding network can be obtained.
  • a decoding network which includes a decoder, a common feature extractor, and a common channel generator.
  • the input to the common feature extractor can be the channel features corresponding to a UE. In this way, a common feature is obtained based on at least one second channel feature corresponding to at least one UE.
  • the following method 1 or method 2 may be used to obtain the second CSI corresponding to the at least one UE based on the common feature and the second channel feature corresponding to the at least one UE:
  • Method 1 The common feature and at least one second channel feature (undecoded) corresponding to at least one UE are input into a decoder, and decoded by the decoder to obtain decoded CSI (defined as second CSI to avoid confusion).
  • Method 2 Input the second channel characteristics corresponding to the first UE into the decoder, output the decoded second channel characteristics, and then input the common characteristics and the decoded second channel characteristics into the decoder corresponding to the first UE again, and obtain the second CSI corresponding to the first UE through the decoder.
  • the decoding network further includes a common channel generator, into which the common features are input, and which outputs common CSI. Then, based on the common channel state information and the second CSI corresponding to each UE, reconstructed CSI corresponding to each UE is obtained. For example, the common CSI is added to the second CSI to obtain reconstructed CSI (third CSI).
  • the second CSI is obtained by decoding the second channel characteristics and common characteristics.
  • the second channel characteristics are obtained by dequantizing the bitstream uploaded by each UE and can be understood as personalized channel characteristics.
  • Common characteristics are shared characteristics of multiple channels extracted based on the channel characteristics of multiple UEs. During decoding, the characteristics are explicitly divided into personalized characteristics and common characteristics for extraction and decoding, making the model easier to learn during the training phase and improving the decoding performance of the trained model.
  • common channel state information is further extracted based on common features, that is, common CSI is extracted, and then the common CSI is added to the second CSI.
  • the resulting reconstructed CSI contains both personalized channel state information and common channel state information, with a higher degree of restoration and a smaller NMSE compared to the original CSI.
  • the overall process can be implemented based on multiple modules or network architectures such as a data acquisition module, a preprocessing module, an encoding network, and a decoding network.
  • the main processing process can include data acquisition, preprocessing, encoding, and decoding.
  • this embodiment uses a data set generated by the wireless channel model COST2100, and obtains a sparse CSI matrix through preprocessing steps such as discrete Fourier transform (DFT) and matrix truncation.
  • the sparse CSI matrix is then fed into the core network architecture proposed in the embodiment of this application to obtain a reconstructed sparse CSI matrix.
  • post-processing steps restore the original CSI matrix.
  • the core network architecture is shown in Figure 7.
  • the encoding and decoding processes of the CSI matrix in this embodiment can be performed based on the core network architecture shown in Figure 7.
  • the channel model uses the COST2100 public data model, with a center frequency of 2.6 GHz.
  • NLOS non-line-of-sight
  • 20 UEs are randomly distributed within a 5m x 5m range.
  • the number of subbands, N c is set to 1024
  • the number of base station transmit antennas, N t is set to 32
  • the number of UE receive antennas is set to 1.
  • the generated original CSI matrix is denoted as H, where H is a complex matrix with dimensions of 1024 x 32.
  • the subsequent process may include the following steps:
  • the process of converting the original matrix into a sparse matrix can be performed by the preprocessing module.
  • the specific conversion method may be as follows:
  • H is a complex matrix with a dimension of 1024 ⁇ 32.
  • the sparse CSI matrix is input into the coding network, and the coding network encodes the sparse CSI matrix X into a bit stream.
  • the operations performed based on the coding network may specifically include steps 802 to 804.
  • the encoding network includes an encoder and a mask generator.
  • k sparse CSI matrices corresponding to K UEs are obtained, namely X 1 , X 2 , ...X K .
  • X 1 , X 2 , ...X K are input into K encoders, and the output is K compressed features, namely M 1 , M 2 , ...M K .
  • the sparse CSI matrix X of each UE is used as the input of the Mask generator, for example, X 1 , X 2 ..X K are respectively input into K Mask generators, and output U 1 , U 2 ..U K .
  • U can be directly used as a learnable variable, and multiple UEs can learn the same U together.
  • the Mask generator does not need to input the sparse CSI matrix X. It can be understood that multiple UEs share the parameters in the same mask generator, and the trainable parameters in the Mask generators trained in multiple UEs are consistent.
  • the encoder can be implemented based on a convolutional neural network (CNN), and X is input to an encoder implemented based on a fully convolutional network.
  • CNN convolutional neural network
  • the encoder can include three convolution modules, each of which includes a CNN layer, a batch normalization layer, and a PReLU nonlinear layer.
  • the encoder includes three encoding modules (i.e., convolution modules), namely encoding module 100a, encoding module 101a, and encoding module 102a, which are connected in series.
  • the encoding module 100a includes a convolution layer 1001, a batch normalization layer 10021, and a PReLU layer 1003;
  • the encoding module 101a includes a convolution layer 1011, a batch normalization layer 1012, and a PReLU layer 1013;
  • the encoding module 102a includes a convolution layer 1021, a batch normalization layer 1022, and a PReLU layer 1023.
  • the convolution kernel sizes of the convolution layers 1001, 1011, and 1021 are 9 ⁇ 9, 5 ⁇ 5, and 5 ⁇ 5, respectively.
  • zero padding can be used to keep the dimensions unchanged after the convolution operation.
  • the features output by the three encoding modules are fed into three convolutional layers in parallel, for example, into convolutional layer 2011, convolutional layer 2021, and convolutional layer 2031.
  • the convolution kernel size of the three convolutional layers is 5 ⁇ 5.
  • the output obtained by the convolution layer 2011 is the mean matrix
  • the mean matrix is a data form of the mean feature.
  • a Sigmoid layer 3011 is connected after the convolution layer 2021, and the output of the convolution layer 2021 is passed through the Sigmoid function to obtain the scale matrix
  • the scale matrix is a data form of scale features. Both the scale matrix and the mean matrix have dimensions of 32 ⁇ 32 ⁇ 2.
  • a sigmoid layer 3011 is connected after the convolution layer 2031. The output of the convolution layer 2031 is passed through a sigmoid function to obtain the mask matrix U, which has dimensions of 32 ⁇ 32 ⁇ 1.
  • S802 and S803. can be executed in parallel or in sequence.
  • S803 can be executed before or after S802.
  • S804 Obtain a masked hidden layer feature matrix according to the hidden layer feature matrix M and the mask matrix.
  • the matrix ⁇ M can be sampled from a uniform distribution of [-0.5, 0.5], and let
  • M′ is uniformly quantized in b dimensions to obtain the quantized For example, divide the interval b [0,1] into equal parts, and replace each element in M′ with the center point of the interval it falls into.
  • the trainable parameters in the base model are optimized: by optimizing the parameters of the neural network Make the elements in the hidden feature matrix M obey the Laplace distribution with mean 0 and scale ⁇ .
  • the Laplace distribution is selected as the prior distribution to control the sparsity.
  • KL divergence loss function Kullback-Leibler Divergence, KLD
  • ⁇ is a hyperparameter that can strictly control the entropy of M and thus the length of the bitstream.
  • j represents the element number in the hidden feature matrix M.
  • Lap represents the Laplace distribution.
  • the loss of each UE is calculated separately and then added together to obtain the total loss function value corresponding to multiple UEs. Based on the total loss, the trainable parameters in the encoder are optimized.
  • the second goal is to obtain trainable parameters in the Mask generator in multi-UE scenarios.
  • l is the sequence number of the element in U
  • ⁇ U is a hyperparameter that can control the sparsity of U.
  • Loss encoder Loss encoder
  • U + Loss encoder M
  • the trainable parameters in the encoding network are adjusted according to the value of the loss function until the model converges.
  • the bit stream is input into the entropy coding process and then transmitted to the BS where it is decoded to obtain the reconstructed CSI.
  • the following is an exemplary description of the decoding process on the BS side.
  • the input of common feature extraction is K UEs corresponding to K Right now
  • the features output by the common feature extractor are used as common features s.
  • the dimensions of common features s are 32 ⁇ 32 ⁇ 4, where 4 is a hyperparameter that can be set.
  • each UE Together with the common feature s, it is input into the decoder to obtain the respective channels
  • the personalized channel state information obtained after decoding is the second CSI.
  • the dimensions of the common channel H common are 32 ⁇ 32 ⁇ 2.
  • the common channel is the common channel state information, ie, the common CSI.
  • the decoder, common feature extractor, and/or common channel generator can be implemented based on a CNN or attention mechanism, or a combination of CNN and attention.
  • An example of a CNN-based implementation is listed below.
  • the common feature extractor may include a convolution layer 401, two residual modules, and a convolution layer 404.
  • the two residual modules are residual module 402 and residual module 403.
  • Residual module 402 may include convolution layers 4021 and 4022
  • residual module 403 may include convolution layers 4031 and 4032.
  • the convolution kernel size of each convolution layer in the residual module may be 5 ⁇ 5.
  • the input is sent to convolutional layer 401, which may have a convolution kernel size of 5 ⁇ 5.
  • the output is then copied twice, one of which is input to residual module 402.
  • the output of residual module 402 is then added to the input of residual module 402 and input to residual module 403.
  • the output of residual module 403 is then added to the other output of convolutional layer 401.
  • the output of residual module 403 is then passed through a 5 ⁇ 5 convolutional layer 404 to obtain the output, i.e., the common feature s.
  • the common channel generator can adopt the same or different network architecture as the common feature extractor.
  • the internal network architecture of the decoder and the common feature extractor can be the same or different.
  • the decoder can be appropriately adjusted based on the common extractor.
  • the decoder can use the same architecture as the common feature extractor to obtain a 32 ⁇ 32 ⁇ 16 output, where 16 is a hyperparameter.
  • the decoder output is then concatenated with the common s in the channel dimension to obtain a 32 ⁇ 32 ⁇ (16+4) output. This is then input into the decoder to obtain the personalized channel H individual , with dimensions of 32 ⁇ 32 ⁇ 2.
  • the reconstruction loss is then calculated based on the reconstructed sparse CSI matrix.
  • the mean squared error (MSE) or KLD loss can be selected as the loss function on the decoding side.
  • a post-processing module may also be deployed on the BS side to convert the reconstructed sparse CSI matrix into the original CSI matrix.
  • This matrix Input into the two-dimensional inverse DFT to obtain the reconstructed CSI matrix in the space-frequency domain.
  • Figure 11a shows the performance of a two-UE scenario at different compression rates
  • Figure 11b shows the performance of a three-UE scenario at different compression rates.
  • bitrate is the bit rate
  • compression rate is the compression rate or compression ratio
  • correlation is the accuracy.
  • the NMSE obtained by implementing the solution proposed in the embodiment of this application with two and three UEs in a MU scenario is compared with that obtained with a single UE.
  • a smaller NMSE indicates better performance.
  • MU 2UE indicates a MU scenario with two UEs
  • MU 3UE indicates a MU scenario with three UEs.
  • a Massive MIMO downlink system model can be first established, and the channel data can be collected through the model. Then the data is input into the neural network. The neural network of each UE, including the parameters, is the same, and there is no interaction between UEs. Finally, each UE sends the compressed and quantized bit stream to the base station, and the base station uses the CSI feedback information of multiple UEs to reconstruct the channels of multiple UEs. For example, a Massive MIMO link simulation system is first established, and then the CSI matrix is obtained on the UE side. The bit stream is calculated according to the CSI matrix through the encoding module; then, on the BS side, the bit streams of multiple UEs are collected through the decoding module, and the CSI matrix is jointly reconstructed.
  • the embodiment of the present application proposes a joint compression and reconstruction scheme based on an AI autoencoder for related channel pairs, which can remove redundant information, including multi-UE and multi-base station channels adjacent in space, and multi-frequency channels adjacent in the frequency domain.
  • the embodiment of the present application has made improvements in at least the following aspects and achieved corresponding technical effects:
  • the encoder network in the related art generally has the same solution for a single UE.
  • the embodiment of the present application proposes to directly and explicitly remove the redundancy in the hidden layer features: multiple UEs are considered at the encoder, and unnecessary features are set to 0 through the Mask matrix; a new encoder structure is proposed, which not only outputs features, but also outputs a mask matrix to control which features are redundant information and can be discarded.
  • the features of the related channels are jointly processed to extract the common features (s) and common channels (H common ).
  • the common features are also input into the decoder to reconstruct the personalized channels:
  • the channel is decomposed into two parts: a personalized channel and a common channel.
  • the channels are reconstructed using personalized features and common features respectively, and finally the final reconstructed channel is obtained by fusing them.
  • the solution proposed in the embodiment of the present application decomposes the channel into two parts for reconstruction. In this way, when reconstructing the common channel, there is no need to consider personalized features, and when reconstructing the personalized channel, the interference of common features is eliminated. Compared with directly learning a fused channel, the difficulty of network learning is relatively reduced.
  • a new decoder network is designed to simultaneously input the UE's own features and common features to calculate a personalized channel; that is, the channel is decomposed into two parts: a common channel and a personalized channel.
  • the common channel represents common information in the environment and belongs to low-frequency background information.
  • the embodiment of the present application continues the neural network solution and improves the generalization and versatility: the single-UE basic solution is well extended to the multi-UE solution, can be adapted to any neural network basic solution, and does not need to meet a series of assumptions, and has strong adaptability to the actual application environment.
  • An embodiment of the present application further provides an electronic device, comprising: a processor, wherein the processor is configured to execute a computer program or instruction in a memory to implement the method described in any of the above embodiments.
  • a processor may include one or more processing units, such as a neural-network processing unit (NPU), an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a digital signal processor (DSP), a baseband processor, and the like.
  • the different processing units may be independent devices or integrated into one or more processors.
  • the controller may generate operation control signals based on instruction opcodes and timing signals to control instruction fetching and execution.
  • the memory can be used to store computer executable program code, which includes instructions.
  • the internal memory may include a program storage area and a data storage area.
  • the program storage area may store an operating system, at least one application required for a function, etc.
  • the data storage area may store data (such as input data, output data) created during the use of the electronic device.
  • the internal memory may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, a universal flash storage (UFS), etc.
  • the processor executes various functional applications and data processing of the electronic device by running instructions stored in the internal memory and/or instructions stored in a memory provided in the processor.
  • the structure illustrated in the embodiment of the present invention is merely an example and does not limit the electronic device.
  • the electronic device in the embodiment of the present application may include more or fewer components than shown, or combine or separate certain components, or arrange the components differently.
  • the illustrated components may be implemented in hardware, software, or a combination of software and hardware.
  • An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored program, wherein when the program is executed by a processor, the method described in any of the above embodiments is implemented.
  • An embodiment of the present application further provides a computer program product, which includes a program.
  • the program When the program is executed by an electronic device, the electronic device implements the method described in any of the above embodiments.
  • An embodiment of the present application also provides a chip system, including: a communication interface for inputting and/or outputting data; and a processor for executing a computer executable program so that a device equipped with the chip system executes a method as described in any of the above embodiments.
  • the above-mentioned computer-readable storage medium may adopt any combination of one or more computer-readable media.
  • the computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium.
  • the computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any combination thereof.
  • Computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM) or flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.
  • a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus or device.
  • a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.
  • first and second are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as “first” or “second” may explicitly or implicitly include at least one of such features.
  • plural means at least two, for example, two, three, etc., unless otherwise specifically defined.
  • Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application belong.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请实施例涉及无线通信或人工智能技术领域，尤其涉及一种信道状态信息压缩方法、重构方法和电子设备，可以提高信道估计场景下CSI压缩重构的诸多性能。该方法可以先确定第一UE对应的第一信道特征；第一信道特征，基于待编码的第一信道状态信息CSI获得；确定掩码矩阵；基于第一信道特征和掩码矩阵，获得第一UE对应的比特流。

Description

信道状态信息压缩方法、重构方法和电子设备
本申请要求于2024年04月08日提交中国国家知识产权局、申请号为202410411951.1、申请名称为“信道状态信息压缩方法、重构方法和电子设备”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及人工智能或通信技术领域，尤其涉及一种信道状态信息压缩方法、重构方法和电子设备。
背景技术
基站(BaseStation，BS)为用户设备(User Equipment，UE)服务之前，需要获得BS和UE之间的信道状态信息(Channel State Information，CSI)，一般是UE将CSI执行压缩(或者说编码)等处理之后得到比特流，传输至BS，BS对比特流执行解码等操作，获得重构后的CSI。
相关信道进行联合压缩重构的场景下，一类方案利用专家经验进行联合压缩重构，例如，多用户大规模多输入多输出(Massive Multiple Input Multiple Output，massive MIMO)系统中，由于共享共同的局部散射集合，多UE的CSI矩阵可能具有联合的相关性，因此，该方案设计了一种分布式CS压缩重构框架，利用多UE的CSI矩阵之间的联合相关性进一步压缩CSI矩阵。但是，此类方案主要依赖于专家经验，通常具有较强的假设条件，当满足这些假设条件时，可以表现得很好；但是，一旦假设条件不满足，此类方案的性能就会急剧下降，方案的泛化性或者说通用性有待提升。
发明内容
本申请实施例提供一种信道状态信息压缩方法、重构方法和电子设备，无需依赖专家经验，提升泛化能力和通用性。
第一方面，本申请实施例提供了一种信道状态信息压缩方法，可以应用于第一用户设备UE，先确定第一UE对应的第一信道特征；第一信道特征，基于待编码的第一信道状态信息CSI获得；以及，确定掩码矩阵；基于第一信道特征和掩码矩阵，获得第一UE对应的比特流。
其中，第一UE为多个UE中的一个。待编码的第一CSI，即原始CSI矩阵。BS侧根据比特流获得重构信道状态信息(重构CSI)。
该方法无需依赖专家经验，不需要满足诸多假设条件，泛化性和通用性更好，能够适用于MU或多站或多频的场景，泛化性和通用性得到了提升。在编码时，通过掩码矩阵，指示哪些特征为冗余特征，减少冗余特征对于通信资源的占用。
在一些实施例中，基于第一信道特征和掩码矩阵，获得第一UE对应的比特流，可以是基于掩码矩阵，对第一信道特征执行掩码操作，获得第一UE对应的比特流。
掩码操作，可以在编码端显式进行冗余特征的去除，例如用Mask矩阵将冗余特征置0，这样获得的特征稀疏性更明显，具有更小的熵，因而能在同样的压缩比下，获得更小的重构损失。
在一些实施例中，确定掩码矩阵，可以是基于第一UE对应的第一CSI，确定第一UE对应的第一掩码矩阵；或者，确定多个UE对应的同一掩码矩阵，多个UE包括第一UE。
多个UE可以学习同一个掩码矩阵，或者多个UE分别学习各自的掩码矩阵。
在一些实施例中，第一UE部署有编码网络，编码网络包括掩码生成器和编码器。确定第一UE对应的第一信道特征，可以是通过编码器，确定第一UE对应的第一信道特征。确定掩码矩阵，可以是通过掩码生成器，确定掩码矩阵。
掩码生成器即Mask生成器。本申请实施例提出了一种CSI压缩重构场景下的新的编码网络方案，在编码器基础上增加了掩码生成器，在训练阶段通过掩码生成器学习冗余特征，在应用阶段，通过掩码生成器指示冗余特征，从而可以通过掩码操作去除冗余特征。
在一些实施例中，通过编码器，确定第一UE对应的第一信道特征，可以是通过编码器，输出均值特征和尺度特征；基于均值特征和尺度特征以及第一参数矩阵，确定第一UE对应的第一信道特征；其中，第一参数矩阵服从稀疏分布。
在一些实施例中，编码器包括至少一个编码模块，以及第一处理层和第二处理层，以及与第二处理层连接的第一归一化层；其中，至少一个编码模块的输出并行输入至第一处理层和第二处理层；第一处理层或第二处理层，基于卷积层或注意力机制实现；通过编码器，输出均值特征和尺度特征，可以是通过第一处理层，输出均值特征；通过第二处理层和第一归一化层，输出尺度特征。
在一些实施例中，掩码生成器与编码器共享至少一个编码模块；掩码生成器还包括第三处理层和第二归一化层；至少一个编码模块的输出还并行输入至第三处理层；或者，掩码生成器包括专用的至少一个编码模块和第三处理层以及第二归一化层；通过掩码生成器，确定掩码矩阵，可以是通过第三处理层和第二归一化层，输出掩码矩阵。
在一些实施例中，基于第一信道特征和掩码矩阵，获得第一UE对应的比特流，可以是基于第一信道特征和掩码矩阵的点乘，得到隐层特征；基于隐层特征，得到第一UE对应的比特流。
在一些实施例中，确定第一UE对应的第一信道特征，以及确定掩码矩阵之前，方法还可以是基于第一信道特征，确定第一损失；基于掩码矩阵，确定第二损失；基于损失函数，优化编码器和掩码生成器中的可训练参数；其中，损失函数第一损失和第二损失。
在一些实施例中，基于第一信道特征，确定第一损失，可以是确定第一先验分布；第一先验分布为稀疏分布；基于KL散度损失函数，计算第一信道特征对应的第一分布与第一先验分布之间的距离，获得第一损失；基于掩码矩阵，确定第二损失，可以是确定第二先验分布；第二先验分布为稀疏分布；基于KL散度损失函数，计算掩码矩阵对应的第二分布与第二先验分布之间的距离，获得第二损失。
其中，第一损失为编码器对应的损失，第二损失为Mask生成器对应的损失。分别设计编码器和Mask的先验分布，对其稀疏程度进行控制，防止平凡解。
第二方面，本申请实施例还提出一种信道状态信息重构方法，可以应用于基站BS，先确定至少一个UE对应的至少一个第二信道特征；第二信道特征基于UE上传的比特流获得；基于至少一个第二信道特征，获得公共特征；基于公共特征和至少一个UE分别对应的第二信道特征，获得至少一个UE分别对应的第二CSI。
其中，本申请实施例提出的方案将信道分解成了公共特征和个性化信道特征两个部分进行重构。如此，在重构公共信道时，无需考虑个性化特征，而重构个性化信道时，免除了公共特征的干扰。相比于直接学习一个融合后的信道，网络学习的难度相对降低，更有利于获得高性能的模型，从而实现高性能的CSI信息压缩和重构。
在一些实施例中，获得公共特征之后，还可以基于公共特征，获得公共信道状态信息；获得至少一个UE分别对应的第二CSI之后，还可以基于公共信道状态信息和至少一个UE分别对应的第二CSI，获得至少一个UE分别对应的第三CSI。
进一步地，本申请实施例还提出提取公共信道(公共CSI)，将信道分解为个性化信道和公共信道两个部分，通过个性化特征和公共特征分别重构信道，最后融合得到最终的重构信道。公共信道代表了环境中的共同信息，属于低频背景信息。如此，在公共信道上叠加个性化信道，相当于补全细节、高频信息，可以同时兼顾两种信息，从而获得更小的重构损失。
在一些实施例中，BS部署有解码网络，解码网络包括解码器和公共特征提取器；至少一个UE包括第一UE；基于至少一个第二信道特征，获得公共特征，可以是将至少一个第二信道特征输入至公共特征提取器，通过公共特征提取器，获得公共特征；基于公共特征和至少一个UE分别对应的第二信道特征，获得至少一个UE分别对应的第二CSI，可以是将公共特征和第一UE对应的信道特征，输入至第一UE对应的解码器，通过解码器，获得第一UE对应的第二CSI。
在其他实施例中，基于公共特征和至少一个UE分别对应的第二信道特征，获得至少一个UE分别对应的第二CSI，还可以是将第二信道特征输入至解码器，得到解码后的第二信道特征，然后将解码后的第二信道特征与公共特征一起再次输入至解码器，得到第二CSI。
在一些实施例中，BS还包括公共信道生成器；基于公共特征，获得公共信道状态信息，可以是将公共特征输入至公共信道生成器，获得公共信道状态信息。
本申请实施例提出了一种新的解码网络，解码网络在基础方案基础上增加了公共特征提取器和公共信道生成器，分别提取公共特征和公共信道。
第三方面，本申请实施例还提供一种电子设备，所述电子设备包括：处理器，所述处理器用于执行存储器中的计算机程序或指令以实现如上述任一项所述的方法。
第四方面，本申请实施例还提供一种计算机可读存储介质，所述计算机可读存储介质包括存储的程序，其中，所述程序被处理器执行时实现如上述任一项所述的方法。
第五方面，本申请实施例还提供一种芯片系统，包括：通信接口，用于输入和/或输出数据；处理器，用于执行计算机可执行程序，使得安装有所述芯片系统的设备执行如上述任一项所述的方法。
附图说明
图1为信道估计场景下UE侧与BS之间进行CSI压缩重构的示意图；
图2为MU场景示例图；
图3为本申请实施例提供的信道状态信息压缩方法的系统架构示例图；
图4为本申请实施例提供的信道状态信息压缩方法中编码网络的架构设计一个示例图；
图5为本申请实施例提供的信道状态信息压缩方法中编码网络的架构设计另一个示例图；
图6为本申请实施例提供的信道状态信息压缩方法中编码模块的一个示例图；
图7为本申请实施例提供的信道状态信息压缩方法的一个实施例中核心网络架构示意图；
图8为本申请实施例提供的信道状态信息压缩方法的一个实施例的流程示意图；
图9为本申请实施例提供的信道状态信息压缩方法的一个实施例中编码网络的具体架构设计一个示例图；
图10本申请实施例提供的信道状态信息压缩方法一个实施例中解码器的网络架构设计一个示例图；
图11a为2个UE场景下的性能数据表；
图11b为3个UE场景下的性能数据表；
图12为MU场景下2个UE以及3个UE与单UE分别对应的NMSE对比图。
具体实施方式
为了更好的理解本说明书的技术方案，下面结合附图对本申请实施例进行详细描述。
应当明确，所描述的实施例仅仅是本说明书一部分实施例，而不是全部的实施例。基于本说明书中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其它实施例，都属于本说明书保护的范围。
在本申请实施例中使用的术语是仅仅出于描述特定实施例的目的，而非旨在限制本说明书。在本申请实施例和所附权利要求书中所使用的单数形式的“一种”、“所述”和“该”也旨在包括多数形式，除非上下文清楚地表示其他含义。
基站(BS)为UE服务之前，需要对UE进行预编码等处理，因而需要获得基站和UE之间的CSI，CSI可以反映基站和UE之间的环境信息，例如二者之间障碍物造成的散射、折射等等。一般而言，是由UE将CSI反馈给BS，由于CSI是一个维度为天线数*频带数的复数矩阵，若按照CSI矩阵的原始维度进行反馈，通信开销很大。
如图1所示，为了解决反馈开销过大的问题，需要在UE端，将原始CSI矩阵H通过一系列操作进行压缩、量化，得到长度可控的01比特流h。当BS端接受到01比特流h后，再重构出CSI矩阵其中，原始CSI矩阵、压缩量化后的比特流、重构CSI矩阵分别用H、h、表示。CSI Compression表示CSI压缩。
在第五代移动通信技术(5th Generation Mobile Communication Technology，5G)时代，基站侧和UE侧都可以装载多个天线，CSI中的“天线数”即BS和UE的天线数目的乘积，但这样会存在相邻天线信号之间的干扰问题，Massive MIMO技术就是用于降低干扰的技术，已经在5G中广泛使用。
通过该技术，同一个基站可以同时为多个UE服务，同一个UE也可以接受多个不同的基站的服务，由此可以在频谱资源固定的情况下，增大信道的容量。也就是说，在Massive MIMO系统中，一个基站通常会为多个不同的UE服务，而一个UE也常常会与多个基站进行通信。除此以外，即使同一个UE和同一个基站之间，由于信号可以在不同的频段上传输，所以也会存在多个频段(Multi-Band)同时服务的场景。需要说明的是，此处的频段是指“中心频点”，围绕一个中心频点，可以采集多个频点上的信道，这些频点组合称为CSI中的“频带”。
上面三种场景中，一个基站同时为多个UE服务，可以简称为多用户(Multi User，MU)；多个基站同时为一个UE服务可以简称为多站；在多个频段上服务，可以简称为多频(Multi-band)。
在这些场景下，同时传输的信道具有一定程度的相关性，例如，在MU场景下，由于基站是同一个，其周围的建筑物等环境信息是固定的，所以不同UE和该基站之间的环境具有一些共同的特点，此时不同UE和该基站的CSI之间具有较大的相关性，因此本申请实施例中，将多用户或多站或多频应用场景下的多信道，统称为“相关信道”或者“相关信道对”。
基于相关信道或者相关信道对的CSI联合压缩重构场景中，一般具有如下特性：
(1)UE与BS之间的环境复杂多变，导致相关信道之间的相关性也具有多样性，因此对压缩重构方案的通用性、泛化性具有较高的要求。(2)各个UE需要共享同样的编码器，例如编码网络的参数需要完全相同，但是希望各个UE提取出来的信息尽量不重复。(3)在压缩量化至同样长度的比特流下，重构CSI和原始CSI之间的NMSE(Normalized Mean Squared Error，归一化均方误差)越小越好。
起初，通过专家经验设计方案，进行CSI的压缩和重构；随着神经网络技术、人工智能技术的不断发展，一些方案将分布式信源编码技术和深度学习相结合进行相关信道对的联合压缩重构。
相关信道进行联合压缩重构的场景，与分布式信源编码的场景非常匹配，因此很多工作基于R16标准协议算法或者压缩感知算法，利用专家经验进行联合压缩重构。例如，一种基于专家经验的白盒算法中，基于多用户massive MIMO系统，由于共享共同的局部散射集合，多UE的CSI矩阵可能具有联合的相关性，因此，该基于专家经验的白盒算法，设计了一种分布式CS压缩重构框架，利用多UE的CSI矩阵之间的联合相关性进一步压缩CSI矩阵。
然而，基于CS、专家经验的方法，通常具有较强的假设条件，当满足这些假设条件时，这些方案可以表现得很好；但是，一旦假设条件不满足，这些方案的性能就会急剧下降，而实际应用环境复杂多变，经常出现无法满足假设条件的情形。可见，基于专家经验的白盒算法方案，方案的泛化能力较差，通用性有待提升。
此外，专家经验设计与神经网络难以结合。例如，在单UE对单基站或单频点等基础场景下，基于神经网络的方案相比专家白盒方案有着明显的优势，当从基础场景扩展到多UE等场景时，由于神经网络的不可解释性，白盒的方案很难与神经网络的基础场景方案相融合，使得基于神经网络的基础场景下的方案难以拓展到多用户或多站或多频应用场景中。
此外，在相关技术中，基于神经网络的压缩重构方案中，也存在一些问题需要解决，例如，一种基于深度学习的多用户下行CSI反馈方法，先是BS发送导频，UE进行信道估计得到下行CSI；之后，UE对CSI进行压缩与量化，并反馈给BS；BS收到所有UE反馈的下行CSI进行联合CSI重构；BS利用重构CSI实施预编码，同时与所有UE通信。
其中，该方案在UE侧对CSI进行压缩量化时，与单UE对单基站场景下的基础压缩重构方案是一样的，也就是在编码阶段、UE侧进行压缩时，并未考虑到多用户或者多站或多频场景下，相关信道之间的相关性。与单UE对单基站的基础压缩重构方案的区别，只在于在BS侧恢复信道时，使用了其他UE发送的CSI信息。
可见，此种基于深度学习的多用户下行CSI反馈方法，在UE侧编码时，没有考虑其他UE的存在，在编码时没有提取到更为全面的特征信息；完全寄希望于解码端可以指导编码端在压缩时去掉其他UE具有的信息，而解码端进行解码的对象为UE传输的比特流，在UE侧提取到的特征信息有限的情况下，解码端很难从中去掉其他UE具有的特征信息，增加了解码端的训练难度；并且，在解码端，也是直接寄希望于神经网络自动学习如何融合各个UE的信息，神经网络学习难度高，难以训练，训练结果难以达到预期，导致训练出的模型难以实际应用，不具备解决实际问题的能力。
此外，一些相关的基于神经网络的压缩重构方案中，让网络端到端学习哪些特征由哪个UE反馈，并在解码端将所有UE的CSI矩阵进行融合。此类方案，由于都是隐式地去除冗余特征，冗余特征与非冗余特征没有做显示区分或者区分的较为模糊，导致神经网络的学习难度大，神经网络难以学习到哪些是冗余特征，哪些是非冗余特征。如果能让神经网络显式地学习哪些特征可以是可以复用的冗余特征，哪些特征需要单独反馈，将更利于神经网络的学习和收敛，提升训练后的模型的性能。
鉴于上述方案涉及的问题或者技术需求，本申请实施例提出了一种信道状态信息压缩方法和/或重构方法，在一些实施例中可以是基于AI自编码器的相关信道(对)联合压缩重构方案，提升了自编码器压缩特征、重构CSI矩阵的能力，训练后的模型在实际应用中的泛化能力和通用性，并且模型的性能有所提升。
需要说明的是，模型性能提升的表现之一可以是，在压缩比相同的条件下可以获得更低的NMSE，即，在压缩比相同(原始CSI、压缩后的比特流之间的压缩比相同)下，重构CSI与原始CSI之间的NMSE下降。或者，在NMSE不变的情况下，实现了占用更少的通信资源完成了比特流的传输，即比特流的数据量下降，获得了更高的压缩比。
本申请实施例提供的方法，可以适用于MU、多站、多频等场景中的任一种场景多种场景。例如，可以应用于如图2所示的MU场景。图2所示的场景为j个UE对一个基站，1≤i≤j，i和j均为整数。第i个UEi对应的原始矩阵为Hi，压缩后的比特流为hi。以下说明将多以MU场景为例进行说明。
单UE信道压缩反馈的神经网络方案，在本申请实施例中定义为基础网络方案，本申请实施例提出的方法，可以适用于多UE场景，提出如何利用多个UE之间的共同特征，进一步提升压缩重构的性能，该种多UE联合压缩重构的方案，可以适配于不同的基础网络方案。例如，可以与各种基础网络方案兼容或者说结合。也就是说，本申请实施例提出的方案，可以在多UE场景下提升压缩重构性能的前提下，还可以充分发挥基础网络方案本身的优越性能。
具体地，本申请实施例提供的方法，通过显式地将CSI分为共同特征和个性化特征，其中共同特征，可以理解为多个UE对应的信道特征中的共同特征，而个性化特征可以理解为多个UE中的每个UE分别对应的信道特征中的专属特征。该方法降低了重构误差，在压缩至同样长度的bit流下，可以获得更小的NMSE。
在一些实施例中，在编码端，在显示地区分共同特征和个性化特征之后，可以通过Mask将冗余的特征置0，从而减少了需要传输的特征。在解码端，通过收集空域和/或频域邻近的信道特征，来补偿缺失的特征，最终在同样的反馈比特下，获得了更小的NMSE。
具体地，以一个基站同时为多个UE服务的场景为例，本申请实施例提供的信道状态信息压缩方法，可以基于如图3所示的系统架构执行。
在该系统架构中，在UE侧，增设Mask生成器，用于提取相关信道或者相关信道对之间的公共特征，将信道特征显示地区分为个性化特征和公共特征分别学习。
具体地，在UE侧部署的神经网络，可以简称为编码网络，编码网络至少包括编码器和Mask生成器。
为便于描述，以UE1(第一UE)为例展开说明。UE1为多个UE中的任一UE。UE1在获得原始CSI(第一CSI)矩阵之后，通过编码器对原始CSI矩阵进行编码，得到第一信道特征。第一信道特征可以理解为通过编码器提取到的信道个性化特征。
通过Mask生成器，输出掩码矩阵。需要说明的是，在一些实施例中，一个UE对应于一个专属的Mask生成器，将多个UE的信道状态信息(即原始CSI)分别输入到多个Mask生成器中，得到多个Mask矩阵U。例如，UE1对应的原始CSI，输入到UE1专属的Mask生成器中，输出UE1对应的掩码矩阵U1(第一掩码矩阵)。此种情况下，Mask生成器需要以相应的UE对应的原始CSI为输入。
或者，在其他实施例中，可以多个UE共同学习同一个掩码矩阵，直接将掩码矩阵U作为可学习的变量，多个UE只学习同一个U，此种情况下Mask生成器无需输入单个UE对应的原始CSI信息。
在得到第一信道特征和掩码矩阵U1之后，基于第一信道特征和掩码矩阵，获得第一UE对应的比特流。
掩码矩阵，用于指示掩码位置，可以利用掩码矩阵，对编码后的第一信道特征执行掩码处理，掩码处理的目的在于进一步去除信道特征中的冗余特征，减少需要传输的数据量。示例性地，利用掩码矩阵，对编码后的第一信道特征执行掩码处理，可以是计算掩码矩阵与第一信道特征矩阵的点乘，实现将第一信道特征矩阵中相应位置的元素的值置0的目的。
示例性地，基于第一信道特征和掩码矩阵，获得第一UE对应的比特流，可以是在执行掩码处理之后，继续对掩码后的第一信道特征矩阵进行量化、熵编码等操作，获得待传输的比特流。掩码操作可以去除冗余特征，冗余特征的去除，可以进一步降低待传输的数据量，获得压缩比更高的比特流。此处的压缩比，指的是原始CSI与比特流之间的压缩比。
具体地，在一些实施例中，编码器可以包括至少一个编码模块、第一处理层和第二处理层，以及与第二处理层连接的第一归一化层。例如，如图4所示，编码器包括编码模块100、编码模块101和编码模块102，以及处理层201和处理层202，以及与处理层202连接的归一化层301。其中，编码模块101至编码模块103依次串行连接，编码模块102的输出并行输入至处理层201和处理层202，处理层201的输出作为均值特征；处理层202后连接一个归一化层，经归一化处理后，得到尺度特征。
在一些实施例中，掩码生成器(Mask生成器)可以编码器共享至少一个编码模块，例如图4所示的架构中，编码器与Mask生成器共享编码模块。需要说明的是，图4所示的编码模块的数目为3仅为示例，在其他实施例中，可以适当增加或者减少编码模块的数量。
Mask生成器还包括第三处理层和与第三处理层连接的第二归一化层，至少一个编码模块的输出还并行输入至第三处理层。例如，Mask生成器包括处理层203和归一化层302，编码模块102的输出输入至处理层203，经过归一化层302之后，得到掩码矩阵。
在其他实施例中，如图5所示，Mask生成器包括专用的至少一个编码模块，例如，包括编码模块103、编码模块104和编码模块105，处理层203以及归一化层302。
其中，图4或图5中，处理层可以基于卷积层或者基于注意力(attention)机制的至少一层实现。例如，处理层201至203均为卷积层。归一化层301或者归一化层302，可以基于sigmoid函数实现。
示例性地，一个编码模块可以采用如图6所示的架构设计。例如，一个编码模块可以包括卷积层1001，归一化层1002和PReLU层1003。其中，归一化层1002，具体可以是批归一化层。在其他实施例中，激活函数PReLU(Parametric Rectified Linear Unit)，可以替换为Leaky ReLU、ELU(Exponential Linear Unit)或者Maxout。
基于上述网络架构，可以采用如下方式，确定UE1对应的第一信道特征：
通过编码器，输出均值特征和尺度特征，例如处理层201的输出即为均值特征，归一化层301的输出即为尺度特征。然后，根据均值特征、尺度特征和第一参数矩阵，得到UE1对应的第一信道特征。第一参数矩阵，为服从稀疏分布的矩阵，例如，第一参数矩阵可以服从多元拉普拉斯分布，从多元标准拉普拉斯分布中采样得到一个矩阵L，也就是说，矩阵L中的每个元素都服从标准拉普拉斯分布。
例如，其中M表示第一信道特征，表示均值特征(具体可以是均值特征矩阵)，为尺度特征(具体可以是尺度特征矩阵)，L为第一参数矩阵。
需要说明的是，第一参数矩阵不限于服从拉普拉斯分布，例如，在其他实施例中，还可以服从正态分布或β分布等稀疏分布。
得到第一信道特征之后，接下来，根据第一信道特征和掩码矩阵，得到掩码后的隐层特征。例如，计算第一信道特征和掩码矩阵的点乘，得到隐层特征，之后，对隐层特征做量化、熵编码等操作后，得到比特流。
掩码矩阵的作用在于将信道特征矩阵中相应位置的元素置0，通过点乘操作，可以实现掩码。
本申请实施例中，在对编码网络以及解码网络进行训练的阶段，在编码网络一侧，损失函数包括信道特征对应的部分损失和掩码矩阵对应的部分损失，为便于描述，两部分损失分别定义为第一损失和第二损失。第一损失，基于输出的第一信道特征获得，第二损失，基于掩码矩阵获得。训练阶段，可以根据损失函数的值，不断优化编码器和掩码生成器中的可训练参数。
示例性地，可以采用如下方式分别确定第一损失：
确定第一先验分布。第一先验分布可以是拉普拉斯分布、正态分布、β分布等稀疏分布中的一种；例如，第一先验分布为拉普拉斯分布。
接下来，基于KL散度损失函数，计算第一信道特征对应的第一分布与第一先验分布之间的距离，获得第一损失。
示例性地，可以采用如下方式确定第二损失：
确定第二先验分布。与第一先验分布类似地，第二先验分布可以是拉普拉斯分布、正态分布、β分布等稀疏分布中的一种；例如，第二先验分布为拉普拉斯分布。
接下来，基于KL散度损失函数，计算掩码矩阵对应的第二分布与第二先验分布之间的距离，获得第二损失。
得到第一损失和第二损失之后，可以得到编码网络对应的损失函数。
接下来，对BS侧展开说明。
在BS侧，部署有解码网络。解码网络包括解码器和共同特征提取器，还可以包括公共信道生成器。
在MU场景下，多个UE上传的比特流经解量化等操作之后，得到多个第二信道特征，将多个UE对应的多个第二信道特征，输入至公共特征提取器，输出公共特征；需要说明的是，本申请实施例提供的方案比较典型的适用场景为MU、多站或者多频场景，也可以应用于单UE场景，即在一些实施例中，公共特征提取器的输入可以是一个UE对应的信道特征。如此，基于至少一个UE对应的至少一个第二信道特征，获得公共特征。
得到公共特征之后，在一些实施例中，可以采用如下方式一或方式二，实现基于公共特征和至少一个UE分别对应的第二信道特征，获得至少一个UE分别对应的第二CSI：
方式一：将公共特征和至少一个UE对应的至少一个第二信道特征(未经解码)，输入至解码器，经过解码器解码，得到解码后的CSI(为防止混淆，定义为第二CSI)。
方式二：将第一UE对应的第二信道特征输入至解码器，输出解码后的第二信道特征，然后，将公共特征和解码后的第二信道特征，再次输入至第一UE对应的解码器，通过解码器，获得第一UE对应的第二CSI。
可选的，解码网络还包括公共信道生成器，公共特征输入至公共信道生成器，输出公共CSI，然后，根据公共信道状态信息和各UE分别对应的第二CSI，获得各UE分别对应的重构CSI。例如，将公共CSI添加到第二CSI中，得到重构CSI(第三CSI)。
如此，第二CSI是通过对第二信道特征和公共特征进行解码后得到，其中第二信道特征为各UE上传的比特流进行解量化等操作后得到，可以理解为个性化信道特征。公共特征是基于多个UE的信道特征提取出的多个信道的共有特征，在解码时，将特征显示地分为个性化特征和公共特征进行提取，然后再解码，使模型在训练阶段更易学习，提高训练出的模型的解码性能。
此外，结合公共信道生成器，进一步根据公共特征提取公共信道状态信息，即提取公共CSI，再将公共CSI添加到第二CSI中，得到的重构CSI既包含了个性化信道状态信息，还包含公共信道状态信息，还原度更高，相比于原始CSI，NMSE更小。
下面列举一个具体实施例。
本实施例中，整体流程可以基于数据采集模块、预处理模块、编码网络和解码网络等多个模块或网络架构实现，主体处理流程可以包括数据采集、预处理、编码和解码。
示例性地，本实施例采用无线信道模型COST2100产生的数据集，通过离散傅里叶变换(Discrete Fourier Transform，DFT)、矩阵截取等预处理步骤，得到稀疏的CSI矩阵，然后将稀疏的CSI矩阵送入到本申请实施例方案提出的核心网络架构中，得到重构的稀疏CSI矩阵，最后通过后处理步骤还原回原始的CSI矩阵。示例性地，核心网络架构如图7所示，本实施例中对CSI矩阵进行编码和解码等过程，可以基于图7所示的核心网络架构执行。
具体地，本实施例中，数据采集阶段，信道模型采用COST2100公开数据模型，中心频点取2.6GHz，基于非视域(Non-line-of-sight,NLOS)场景，一共有20个UE随机分布在5m*5m的范围内。设置子带数Nc＝1024，BS端的发射天线数Nt＝32，UE端的接受天线数为1。将生成的原始CSI矩阵记为H，H为维度为1024×32的复数矩阵。
在完成数据采集之后，后续流程可以包括如下步骤：
S801，将原始CSI矩阵H转为稀疏CSI矩阵X。
原始矩阵转换为稀疏矩阵的过程，可以通过预处理模块执行。
示例性地，具体转换方式可以如下：
将H输入到二维DFT中，得到角度-时延域的稀疏矩阵H′：
其中，Fc和均为DFT矩阵，中上标的H代表取共轭转置。H′为维度为1024×32的复数矩阵。
由于多径到达时间具有延迟，H′只有前32行有值，其余行均为0。因此，可以取H′的前32行，得到维度为32×32的复数矩阵，再将32×32的复数矩阵的实部和虚部单独作为一个维度，得到稀疏的CSI矩阵X，其维度为32×32×2。
在得到稀疏CSI矩阵X之后，将稀疏CSI矩阵输入至编码网络，通过编码网络，将稀疏CSI矩阵X编码为比特流。基于编码网络执行的操作具体可以包括步骤802至804。
S802，将多个UE对应的多个稀疏CSI矩阵X，分别输入到各UE的编码器中，得到多个压缩后的特征M。
如图7所示，编码网络包括编码器和Mask生成器。
以MU场景为例，经过预处理，得到了K个UE对应的k个稀疏CSI矩阵，分别为X1，X2…XK。将X1，X2…XK输入到K个编码器(Encoder)中，输出K个压缩后的特征分别为M1，M2…MK。
S803，将多个UE对应的多个稀疏CSI矩阵X，分别输入到Mask生成器，得到多个掩码矩阵U。
例如，在本实施例中，将各UE的稀疏CSI矩阵X作为Mask生成器的输入，例如，X1，X2…XK分别对应输入到K个Mask生成器中，分别输出U1，U2…UK。
需要说明的是，在其他实施例中，可以直接将U作为可学习的变量，多个UE共同学习同一个U，这种情况下Mask生成器无需输入稀疏CSI矩阵X。可以理解为，多个UE共享相同的mask生成器中的参数，多个UE中训练出的Mask生成器中的可训练参数是一致的。
示例性地，在本实施例中，编码器可以基于卷积神经网络(Convolutional Neural Network,CNN)实现，将X输入到基于全卷积网络实现的编码器，例如，编码器可以包括3个卷积模块，每个卷积模块包括一层CNN、一层批归一化层以及一层PReLU非线性层。
示例性地，在本实施例中，编码器Encoder和Mask生成器的具体架构设计可以如图9所示。其中，编码器包括3个编码模块(即卷积模块)，分别为编码模块100a，编码模块101a和编码模块102a，3个编码模块串行连接。其中，编码模块100a包括卷积层1001，批归一化层10021和PReLU层1003；编码模块101a包括卷积层1011，批归一化层1012和PReLU层1013；编码模块102a包括卷积层1021，批归一化层1022和PReLU层1023。卷积层1001、卷积层1011、卷积层1021的卷积核大小分别为9×9、5×5、5×5，对于卷积操作之后可能存在的输出特征的维度不同的问题，可以通过零填充的方式，使得卷积操作之后维度保持不变。
接着，将3个编码模块输出的特征并行送入到3个卷积层中，例如并行送入到卷积层2011、卷积层2021和卷积层2031中。示例性地，3个卷积层的卷积核大小均为5×5。
经过卷积层2011得到的输出，即为均值矩阵均值矩阵为均值特征的一种数据形式。卷积层2021之后连接一个Sigmoid层3011，卷积层2021的输出经过Sigmoid函数，得到尺度矩阵尺度矩阵为尺度特征的一种数据形式。尺度矩阵和均值矩阵这两个矩阵的维度均为32×32×2。卷积层2031之后连接一个Sigmoid层3011，卷积层2031的输出经过Sigmoid函数，得到Mask矩阵U，U的维度为32×32×1。
在S802中，在得到尺度矩阵和均值矩阵之后，采用如下方式确定M：
从多元标准拉普拉斯分布中采样得到一个矩阵L，也就是说，矩阵L中的每个元素都服从标准拉普拉斯分布，利用重参数化技巧，得到隐层特征矩阵M(维度同样为32×32×2)：
需要说明的是，上述S802与S803在执行时序上并无先后限定，可以并行执行，也可以先后执行，S803可以在S802之前或者之后执行。
S804，根据隐层特征矩阵M和掩码矩阵，得到掩码后的隐层特征矩阵。
在本实施例中，示例性地，执行Mask矩阵(即掩码矩阵)U与特征M的点乘运算，得到掩码后的隐层特征矩阵M′(图7未示出)：
M′＝M*U
将M′输入到量化器，得到特征再执行熵编码等操作之后，得到反馈比特流。
需要说明的是，训练阶段与测试阶段(应用阶段)，得到的量化方式会有不同。
示例性地，本实施例中，在训练阶段，可以从[-0.5,0.5]的均匀分布中采样得到矩阵ΔM，令
在测试阶段，对M′进行b维的均匀量化，得到量化后的例如，将[0,1]区间b等分，M′中的每个元素落在哪一个区间，就用落入的区间的中心点代替。
本申请实施例提出的方法中，在训练阶段，可以有两个优化目标：
优化目标一：
以基础模型(编码器)中的可训练参数为优化目标：通过优化神经网络的参数使得隐层特征矩阵M中的元素均服从均值为0，尺度为λ的拉普拉斯分布。
在本实施例中，选择拉普拉斯分布作为先验分布，用于控制稀疏程度。
在计算损失时，可以计算当前隐层特征服从的分布与先验分布这两个分布之间的KL散度损失函数(Kullback-Leibler Divergence，KLD)，作为encoder部分的Loss(损失值)：
具体可以令为隐层特征M的分布，p(M)表示M的先验分布，则有：


其中，λ为超参数，可以严格控制M的熵，从而控制比特流的长度。j代表隐层特征矩阵M中的元素序号。Lap表示拉普拉斯分布。
根据以上的公式分别计算出各个UE的Loss，最后相加。得到多个UE对应的总的损失函数的值，根据总的Loss，优化编码器中的可训练参数。
优化目标二：
第二个目标为多UE场景下Mask生成器中的可训练参数。
假设Mask矩阵U中的每个元素也服从一个稀疏分布，此处同样选择拉普拉斯分布,仍然表示需要优化的网络参数，令当前U的均值为μU，尺度为bU,用表示当前U的分布，p(U)表示U的先验分布，有：
则Loss为：
其中，l为U中的元素的序号，λU为超参数，可以控制U的稀疏程度。
综上，可以得到：
Lossencoder＝Lossencoder,U+Lossencoder,M
在训练阶段，根据损失函数的值，调节编码网络中的可训练参数，直至模型收敛。获得性能符合预期的，模型之后，在测试阶段，将输入到熵编码中得到比特流。比特流传输到BS端后，经解码得到得到重构CSI。
下面对BS侧的解码过程展开示例性描述。
S805，将多个UE对应的多个一起输入到公共特征提取器中，得到公共特征s。
如图7所示，为在BS侧执行解量化操作之后得到的特征。
公共特征提取的输入为K个UE对应的K个即一起输入至公共特征提取器，通过共同特征提取器输出的特征作为公共特征s。公共特征s的维度为32×32×4，其中4为超参数，可以设置。
S806，将各UE的与公共特征s一起输入到解码器(Decoder)，得到各自的信道
例如，s和一起输入Encoder，得到以此类推，得到
即解码后得到的个性化的信道状态信息，即第二CSI。
S807，将公共特征s输入到公共信道生成器中，得到公共信道Hcommon。
本实施例中，示例性地，公共信道Hcommon的维度为32×32×2。公共信道，即公共信道状态信息，即公共CSI。
S808，将公共信道Hcommon分别加到各自的信道中，得到作为最终的重构CSI。
如图7所示，公共CSI分别添加到各个UE对应的中，得到
需要说明的是，解码器Decoder，公共特征提取器，和/或，公共信道生成器，可以基于CNN或attention机制实现，或者基于CNN与attention组合实现。下面列举基于CNN实现的示例。
如图10所示，公共特征提取器，可以包括卷积层401、两个残差模块和卷积层404，两个残差模块分别为残差模块402和残差模块403。残差模块402可以包括卷积层4021和卷积层4022，残差模块403可以包括卷积层4031和卷积层4032。残差模块中的各个卷积层的卷积核的大小可以是5×5。
首先，将输入到卷积层401，卷积层401的卷积核大小可以是5×5，接着将输出拷贝两份，其中一份输入到残差模块402，之后，残差模块402的输出与残差模块402的输入相加，相加之后输入至残差模块403，残差模块403的输出与卷积层401的另一份输出相加。残差模块403的输出再经过一个5×5的卷积层404，即可得到输出，即得到公共特征s。
公共信道生成器可以采用与公共特征提取器相同或者不同的网络架构。
解码器(Decoder)与公共特征提取器的内部网络架构可以相同也可以不相同，解码器可以在公共提取器基础上进行适当调整。在本实施例中，解码器可以采用与公共特征提取器相同的架构，得到32×32×16的输出，其中16为超参数，然后把解码器的输出和公共s在通道维度连接(contact)起来，得到32×32×(16+4)的输出，再输入到Decoder中，得到个性化信道Hindividual，维度为32×32×2。
之后，将Hindividual与Hcommon逐元素相加，得到最终的重构稀疏CSI矩阵。然后，根据重构稀疏CSI矩阵计算重构LOSS，例如，在本实施例中，可以选择均方差损失函数(Mean Squared Error，MSE)或者KLD Loss作为解码一侧的损失函数。
在BS一侧还可以部署后处理模块，用于将重构稀疏CSI矩阵转为原始CSI矩阵。
例如，在本实施例中，后处理的具体步骤如下：
将重新组成32×32的复数矩阵，并补上全0行，得到维度为1024×32的复数矩阵，该矩阵只有32行的元素可以非0。
其中，为多个UE的重构组合得到。
将该矩阵输入到二维逆向DFT中，得到重构的空频域的CSI矩阵。
本申请上述实施例支持在单UE的AI信道压缩重构算法上添加多UE方案。如图11a和图11b所示的两个表格，图11a为不同压缩率下的2个UE场景下表现的性能，图11b为不同压缩率下的3个UE场景下表现的性能。其中，bitrate为比特率，compression rate为压缩率或压缩比，correlation为准确度。
再如图12所示，图12为MU场景下2个UE以及3个UE条件下实施本申请实施例提出的方案得到的NMSE与单UE对比图。其中，NMSE越小性能越好，在bit rate相同的情况下，2个UE和3个UE都获得了更好的性能。其中，MU 2UE，表示MU场景下有2个UE；MU 3UE，表示MU场景下有3个UE。
根据以上实施例的示例性说明，可以看出，本申请实施例提出的方案，可以应用于具有相关性(共同特征)的多个信道数据的压缩、量化、反馈和重构。具体可以首先建立Massive MIMO下行链路系统模型，通过该模型采集信道数据，然后将数据输入神经网络，每个UE的神经网络包括参数都是相同的，UE之间不进行交互。最后，每个UE各自将压缩量化后的比特流发送给基站，基站利用多个UE的CSI反馈信息重构出多个UE的信道。例如，首先建立Massive MIMO链路仿真系统，然后在UE侧获得CSI矩阵，通过编码模块，根据CSI矩阵计算比特流；之后，在BS侧，通过解码模块，收集多个UE的比特流，联合重构CSI矩阵。
综上，本申请实施例提出了一种针对相关信道对的、基于AI自编码器的联合压缩重构方案，可以进行冗余信息的剔除，包括空间中临近的多UE、多基站信道，以及频域上临近的多频信道，具体分析，本申请实施例至少在如下方面做出了改进，并取得了相应的技术效果：
1、利用Mask对编码器提取出的特征进行筛选，降低特征之间的冗余。
相关技术中的encoder网络，一般单UE的方案是一样的，本申请实施例提出直接显式地去除隐层特征中的冗余：在encoder处即会考虑多UE，通过Mask矩阵将不必要的特征置为0；提出了一种新的编码器结构，不仅仅输出特征，还输出掩码矩阵来控制哪些特征属于冗余信息，可以丢弃。
如此，在编码端显式进行冗余特征的去除，用Mask矩阵将冗余特征置0，这样获得的特征稀疏性更明显，具有更小的熵，因而能在同样的压缩比下，获得更小的重构损失。
为了防止平凡解，在编码阶段，还设计了Mask的先验分布，对其稀疏程度进行控制。
2、在解码时，对相关信道的特征进行联合处理，提取公共特征(s)和公共信道(Hcommon)，并把公共特征也输入解码器中，重构个性化信道：
具体提出将信道分解为个性化信道和公共信道两个部分，通过个性化特征和公共特征分别重构信道，最后融合得到最终的重构信道。相比于相关技术，本申请实施例提出的方案将信道分解成了两个部分进行重构。如此，在重构公共信道时，无需考虑个性化特征，而重构个性化信道时，免除了公共特征的干扰。相比于直接学习一个融合后的信道，网络学习的难度相对降低。
本申请实施例设计一种新的解码器网络，使其同时输入UE本身的特征和公共特征来计算个性化信道；即，将信道分解为公共信道和个性化信道两个部分：公共信道代表了环境中的共同信息，属于低频背景信息。
如此，在公共信道上叠加个性化信道，相当于补全细节、高频信息，可以同时兼顾两种信息，从而获得更小的重构损失。
并且，相比于基于专家经验的白盒算法，本申请实施例方案延续了神经网络的方案，还提升了泛化性和通用性：将单UE基础方案很好地扩展到了多UE方案，可以适配到任意的神经网络基础方案，并且不需要满足一系列假设条件，对于实际应用环境的适应性较强。
本申请实施例还提供一种电子设备，所述电子设备包括：处理器，所述处理器用于执行存储器中的计算机程序或指令以实现如上述任一实施例所述的方法。
示例性地，处理器可以包括一个或多个处理单元，例如：处理器包括神经网络处理器(neural-network processing unit，NPU)，还可以包括应用处理器(application processor，AP)，调制解调处理器，图形处理器(graphics processing unit，GPU)，图像信号处理器(image signal processor，ISP)，控制器，数字信号处理器(digital signal processor，DSP)，基带处理器等。其中，不同的处理单元可以是独立的器件，也可以集成在一个或多个处理器中。控制器可以根据指令操作码和时序信号，产生操作控制信号，完成取指令和执行指令的控制。
存储器可以用于存储计算机可执行程序代码，所述可执行程序代码包括指令。内部存储器可以包括存储程序区和存储数据区。其中，存储程序区可存储操作系统，至少一个功能所需的应用程序等。存储数据区可存储电子设备使用过程中所创建的数据(比如输入数据，输出数据)等。此外，内部存储器可以包括高速随机存取存储器，还可以包括非易失性存储器，例如至少一个磁盘存储器件，闪存器件，通用闪存存储器(universal flash storage，UFS)等。处理器通过运行存储在内部存储器的指令，和/或存储在设置于处理器中的存储器的指令，执行电子设备的各种功能应用以及数据处理。
可以理解的是，本发明实施例示意的结构仅为一个示例，并不构成对电子设备的限定。本申请实施例中的电子设备，可以包括比图示更多或更少的部件，或者组合某些部件，或者拆分某些部件，或者不同的部件布置。图示的部件可以以硬件，软件或软件和硬件的组合实现。
本申请实施例还提供一种计算机可读存储介质，所述计算机可读存储介质包括存储的程序，其中，所述程序被处理器执行时实现如上述任一实施例所述的方法。
本申请实施例还提供一种计算机程序产品，所述程序产品包括程序，当所述程序被电子设备运行时，使得所述电子设备实现如上述任一实施例所述的方法。
本申请实施例还提供一种芯片系统，包括：通信接口，用于输入和/或输出数据；处理器，用于执行计算机可执行程序，使得安装有所述芯片系统的设备执行如上述任一实施例所述的方法。
上述计算机可读存储介质可以采用一个或多个计算机可读的介质的任意组合。计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质。计算机可读存储介质例如可以是但不限于电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机存取存储器(RAM)、只读存储器(Read Only Memory；以下ROM)、可擦式可编程只读存储器(Erasable Programmable Read Only Memory；以下EPROM)或闪存、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本文件中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。
计算机可读的信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。
计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于无线、电线、光缆、RF等等，或者上述的任意合适的组合。
在本说明书的描述中，参考术语“一个实施例”、“一些实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本申请的至少一个实施例或示例中。在本说明书中，对上述术语的示意性表述不必须针对的是相同的实施例或示例。而且，描述的具体特征、结构、材料或者特点可以在任一个或多个实施例或示例中以合适的方式结合。此外，在不相互矛盾的情况下，本领域的技术人员可以将本说明书中描述的不同实施例或示例以及不同实施例或示例的特征进行结合和组合。
此外，术语“第一”、“第二”仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括至少一个该特征。在本申请的描述中，“多个”的含义是至少两个，例如两个，三个等，除非另有明确具体的限定。
流程图中或在此以其他方式描述的任何过程或方法描述可以被理解为，表示包括一个或更多个用于实现定制逻辑功能或过程的步骤的可执行指令的代码的模块、片段或部分，并且本申请的优选实施方式的范围包括另外的实现，其中可以不按所示出或讨论的顺序，包括根据所涉及的功能按基本同时的方式或按相反的顺序，来执行功能，这应被本申请的实施例所属技术领域的技术人员所理解。

Claims (17)

  1. 一种信道状态信息压缩方法，其特征在于，所述方法应用于第一用户设备UE，所述方法包括：
    确定所述第一UE对应的第一信道特征；所述第一信道特征，基于待编码的第一信道状态信息CSI获得；
    确定掩码矩阵；
    基于所述第一信道特征和所述掩码矩阵，获得所述第一UE对应的比特流。
  2. 如权利要求1所述的方法，其特征在于，
    基于所述第一信道特征和所述掩码矩阵，获得所述第一UE对应的比特流，包括：
    基于所述掩码矩阵，对所述第一信道特征执行掩码操作，获得所述第一UE对应的比特流。
  3. 如权利要求1或2所述的方法，其特征在于，
    确定掩码矩阵，包括：
    基于所述第一UE对应的所述第一CSI，确定所述第一UE对应的第一掩码矩阵；
    或者，
    确定多个UE对应的同一掩码矩阵，所述多个UE包括所述第一UE。
  4. 如权利要求1-3中任一项所述的方法，其特征在于，
    所述第一UE部署有编码网络，所述编码网络包括掩码生成器和编码器；
    确定所述第一UE对应的第一信道特征，包括：
    通过所述编码器，确定所述第一UE对应的第一信道特征；
    确定掩码矩阵，包括：
    通过所述掩码生成器，确定掩码矩阵。
  5. 如权利要求4所述的方法，其特征在于，
    通过所述编码器，确定所述第一UE对应的第一信道特征，包括：
    通过所述编码器，输出均值特征和尺度特征；
    基于所述均值特征和所述尺度特征以及第一参数矩阵，确定所述第一UE对应的第一信道特征；其中，所述第一参数矩阵服从稀疏分布。
  6. 如权利要求5所述的方法，其特征在于，
    所述编码器包括至少一个编码模块，以及第一处理层和第二处理层，以及与所述第二处理层连接的第一归一化层；其中，所述至少一个编码模块的输出并行输入至所述第一处理层和所述第二处理层；所述第一处理层或第二处理层，基于卷积层或注意力机制实现；
    通过所述编码器，输出均值特征和尺度特征，包括：
    通过所述第一处理层，输出所述均值特征；通过所述第二处理层和所述第一归一化层，输出所述尺度特征。
  7. 如权利要求6所述的方法，其特征在于，
    所述掩码生成器与所述编码器共享所述至少一个编码模块；所述掩码生成器还包括第三处理层和第二归一化层；所述至少一个编码模块的输出还并行输入至所述第三处理层；
    或者，所述掩码生成器包括专用的至少一个编码模块和所述第三处理层以及所述第二归一化层；
    通过所述掩码生成器，确定掩码矩阵，包括：
    通过所述第三处理层和所述第二归一化层，输出所述掩码矩阵。
  8. 如权利要求1-7中任一项所述的方法，其特征在于，
    基于所述第一信道特征和所述掩码矩阵，获得所述第一UE对应的比特流，包括：
    基于所述第一信道特征和所述掩码矩阵的点乘，得到隐层特征；
    基于所述隐层特征，得到所述第一UE对应的比特流。
  9. 如权利要求4-8中任一项所述的方法，其特征在于，
    确定所述第一UE对应的第一信道特征，以及确定掩码矩阵之前，所述方法还包括：
    基于所述第一信道特征，确定第一损失；
    基于所述掩码矩阵，确定第二损失；
    基于损失函数，优化所述编码器和所述掩码生成器中的可训练参数；其中，所述损失函数所述第一损失和所述第二损失。
  10. 如权利要求9所述的方法，其特征在于，
    基于所述第一信道特征，确定第一损失，包括：
    确定第一先验分布；所述第一先验分布为稀疏分布；
    基于KL散度损失函数，计算所述第一信道特征对应的第一分布与所述第一先验分布之间的距离，获得第一损失；
    基于所述掩码矩阵，确定第二损失，包括：
    确定第二先验分布；所述第二先验分布为稀疏分布；
    基于KL散度损失函数，计算所述掩码矩阵对应的第二分布与所述第二先验分布之间的距离，获得第二损失。
  11. 一种信道状态信息重构方法，其特征在于，所述方法应用于基站BS，所述方法包括：
    确定至少一个UE对应的至少一个第二信道特征；所述第二信道特征基于UE上传的比特流获得；
    基于至少一个所述第二信道特征，获得公共特征；
    基于所述公共特征和所述至少一个UE分别对应的第二信道特征，获得至少一个UE分别对应的第二CSI。
  12. 如权利要求11所述的方法，其特征在于，
    获得公共特征之后，所述方法还包括：
    基于所述公共特征，获得公共信道状态信息；
    获得至少一个UE分别对应的第二CSI之后，所述方法还包括：
    基于所述公共信道状态信息和至少一个UE分别对应的所述第二CSI，获得至少一个UE分别对应的第三CSI。
  13. 如权利要求11或12所述的方法，其特征在于，
    所述BS部署有解码网络，所述解码网络包括解码器和公共特征提取器；所述至少一个UE包括第一UE；
    基于至少一个所述第二信道特征，获得公共特征，包括：
    将至少一个所述第二信道特征输入至所述公共特征提取器，通过所述公共特征提取器，获得公共特征；
    基于所述公共特征和所述至少一个UE分别对应的第二信道特征，获得至少一个UE分别对应的第二CSI，包括：
    将所述公共特征和所述第一UE对应的第二信道特征，输入至所述第一UE对应的解码器，通过所述解码器，获得所述第一UE对应的第二CSI。
  14. 如权利要求12所述的方法，其特征在于，
    所述BS还包括公共信道生成器；
    基于所述公共特征，获得公共信道状态信息，包括：
    将所述公共特征输入至所述公共信道生成器，获得公共信道状态信息。
  15. 一种电子设备，其特征在于，所述电子设备包括：
    处理器，所述处理器用于执行存储器中的计算机程序或指令以实现如权利要求1-14中任一项所述的方法。
  16. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质包括存储的程序，其中，所述程序被处理器执行时实现如权利要求1-14中任一项所述的方法。
  17. 一种芯片系统，其特征在于，包括：
    通信接口，用于输入和/或输出数据；
    处理器，用于执行计算机可执行程序，使得安装有所述芯片系统的设备执行如权利要求1-14中任一项所述的方法。
PCT/CN2025/070654 2024-04-08 2025-01-06 信道状态信息压缩方法、重构方法和电子设备 Pending WO2025213896A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202410411951.1 2024-04-08
CN202410411951.1A CN120825255A (zh) 2024-04-08 2024-04-08 信道状态信息压缩方法、重构方法和电子设备

Publications (1)

Publication Number Publication Date
WO2025213896A1 true WO2025213896A1 (zh) 2025-10-16

Family

ID=97349316

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2025/070654 Pending WO2025213896A1 (zh) 2024-04-08 2025-01-06 信道状态信息压缩方法、重构方法和电子设备

Country Status (2)

Country Link
CN (1) CN120825255A (zh)
WO (1) WO2025213896A1 (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101166047A (zh) * 2006-10-17 2008-04-23 中国科学院上海微系统与信息技术研究所 利用信道几何均值分解的多天线通信系统的发送装置、接收装置、发送方法及接收方法
CN115642938A (zh) * 2021-07-20 2023-01-24 维沃移动通信有限公司 信息传输、信息接收方法、装置、终端及网络侧设备
WO2023211346A1 (en) * 2022-04-29 2023-11-02 Telefonaktiebolaget Lm Ericsson (Publ) A node and methods for training a neural network encoder for machine learning-based csi

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101166047A (zh) * 2006-10-17 2008-04-23 中国科学院上海微系统与信息技术研究所 利用信道几何均值分解的多天线通信系统的发送装置、接收装置、发送方法及接收方法
CN115642938A (zh) * 2021-07-20 2023-01-24 维沃移动通信有限公司 信息传输、信息接收方法、装置、终端及网络侧设备
WO2023211346A1 (en) * 2022-04-29 2023-11-02 Telefonaktiebolaget Lm Ericsson (Publ) A node and methods for training a neural network encoder for machine learning-based csi

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
MITSUBISHI ELECTRIC CORPORATION: "Comparison between Turbo and LDPC codes under the condition of parallel operation.", 3GPP DRAFT; R1-060526, 3RD GENERATION PARTNERSHIP PROJECT (3GPP), 9 February 2006 (2006-02-09), XP050101463 *

Also Published As

Publication number Publication date
CN120825255A (zh) 2025-10-21

Similar Documents

Publication Publication Date Title
CN108390706B (zh) 一种基于深度学习的大规模mimo信道状态信息反馈方法
US20140064479A1 (en) Method and systems for multilevel data security
Azghani et al. Multihypothesis compressed video sensing technique
CN115333900B (zh) 一种基于深度学习的无人机协同信道估计与csi反馈方法
Liu et al. Overcoming the channel estimation barrier in massive MIMO communication via deep learning
WO2023011472A1 (zh) 信道状态信息的反馈方法及接收方法、终端、基站、计算机可读存储介质
CN116569527A (zh) 信道信息反馈的方法、发送端设备和接收端设备
CN118355610A (zh) 处理数据的方法及装置
WO2018176340A1 (zh) 视频发送方法、接收方法、系统以及无人飞行器
KR20230163949A (ko) 통신 시스템에서 채널 추정 방법 및 장치
TWI856775B (zh) 通道狀態資訊處理方法及裝置
CN118120156A (zh) 信道信息反馈模型的训练方法、装置、设备及存储介质
CN114245117A (zh) 多采样率复用重建网络方法、装置、设备及存储介质
KR20250087554A (ko) 엔드-투-엔드 이미지/비디오 압축을 위한 잠재 코딩
WO2025213896A1 (zh) 信道状态信息压缩方法、重构方法和电子设备
CN120786067A (zh) 一种多点触觉智能编解码装置、系统及方法
WO2023108437A1 (zh) 一种信道状态信息csi压缩反馈的方法及其装置
WO2025214068A1 (zh) 一种量化参数确定方法及相关装置
CN117639873A (zh) 基于深度学习与注意力机制的mimo信道状态信息反馈方法
CN116015371A (zh) 一种基于神经网络的大规模mimo的csi压缩与反馈方法及系统
Marinkovic et al. Joint source-channel coding based on cosine-modulated filter banks for erasure-resilient signal transmission
Feng et al. mmWave RIS Phase Shift Feedback Based on Knowledge Base Autoencoder Framework
Samundiswary et al. An efficient pass parallel SPIHT based image compression using double density dual tree complex wavelet transform for WSN
CN112737728A (zh) 数据处理方法及装置
KR20240041231A (ko) 통신 시스템에서 신경망을 이용한 채널 정보 전달 방법 및 장치

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 25785378

Country of ref document: EP

Kind code of ref document: A1