Terminal device and information processing method

Abstract

The application discloses a terminal device and an information processing method. The information processing method comprises: jointly source channel encoding first channel state information (CSI) to obtain a first CSI feature vector, wherein the first CSI feature vector comprises a plurality of first feature units; determining a first codebook corresponding to the first feature unit according to a position index of the first feature unit and a plurality of codebooks, wherein the plurality of codebooks are obtained based on training of a plurality of second CSIs; determining a first bit sequence corresponding to the first CSI feature vector according to the first codebook corresponding to the first feature unit in the first CSI feature vector; determining a first orthogonal frequency division multiplexing (OFDM) feedback signal corresponding to the first CSI through modulation of the first bit sequence; and transmitting the first OFDM feedback signal to a network device.

Description

Technical Field

[0001] This application relates to the field of mobile communication technology, including but not limited to terminal equipment and information processing methods. Background Technology

[0002] In communication systems, base stations need to acquire downlink Channel State Information (CSI) to support functions such as precoding, beamforming, and radio resource scheduling. Since uplink and downlink channels in frequency division multiplexing (FDD) systems typically lack strict reciprocity, user equipment (UE) needs to estimate the downlink CSI based on the downlink reference signal and feed it back to the base station via the uplink feedback link. However, in wideband multi-antenna communication systems, CSI often has high dimensionality. Directly feeding back the complete CSI would consume significant uplink feedback resources, making it difficult to meet the requirements of practical systems for limited feedback overhead and low-latency transmission. Summary of the Invention

[0003] This application provides a terminal device and an information processing method.

[0004] The technical solution of this application embodiment is implemented as follows: This application provides a terminal device, which includes a transceiver and a processor coupled to the transceiver. The processor is configured to: perform joint source-channel coding on first channel state information (CSI) to obtain a first CSI feature vector, the first CSI feature vector including a plurality of first feature units; determine a first codebook corresponding to the first feature unit based on the position index of the first feature unit and a plurality of codebooks, the plurality of codebooks being trained based on a plurality of second CSIs; determine a first bit sequence corresponding to the first CSI feature vector based on the first codebook corresponding to the first feature unit in the first CSI feature vector; determine a first orthogonal frequency division multiplexing (OFDM) feedback signal corresponding to the first CSI by modulation of the first bit sequence; and send the first OFDM feedback signal to a network device via the transceiver.

[0005] This application provides an information processing method applied to a terminal device, comprising: performing joint source-channel coding on first channel state information (CSI) to obtain a first CSI feature vector, the first CSI feature vector including a plurality of first feature units; determining a first codebook corresponding to the first feature unit based on the position index of the first feature unit and a plurality of codebooks, the plurality of codebooks being trained based on a plurality of second CSIs; determining a first bit sequence corresponding to the first CSI feature vector based on the first codebook corresponding to the first feature unit in the first CSI feature vector; determining a first orthogonal frequency division multiplexing (OFDM) feedback signal corresponding to the first CSI through modulation of the first bit sequence; and sending the first OFDM feedback signal to a network device.

[0006] This application provides an information processing apparatus, the apparatus comprising: Memory is used to store executable instructions or computer programs. The processor, when executing computer-executable instructions or computer programs stored in the memory, implements the information processing method provided in the embodiments of this application.

[0007] This application provides a computer-readable storage medium storing a computer program or computer-executable instructions, which, when executed by a processor, implements the information processing method provided in this application.

[0008] This application provides a computer program product, including a computer program or computer executable instructions. When the computer program or computer executable instructions are executed by a processor, they implement the information processing method provided in this application.

[0009] In this embodiment, the terminal device can be deployed with multiple codebooks trained by multiple second CSIs, and select a first codebook for each feature unit from multiple codebooks by using the position index of each feature unit in the continuous CSI feature vector, thereby performing position-level codebook selection to generate a discrete feedback bit stream. The resulting feedback bit stream is a bit representation of the feedback link that is adaptively learned according to the CSI feature distribution, which can improve the insufficient feedback bit expression capability and improve the CSI reconstruction performance. Attached Figure Description

[0010] Figure 1 This is an example of an optional scenario for the communication system provided in this application. Figure 1 ; Figure 2 This is an optional flowchart illustrating the information processing method provided in the embodiments of this application. Figure 1 ; Figure 3This is an optional flowchart illustrating the information processing method provided in the embodiments of this application. Figure 2 ; Figure 4 This is an optional flowchart illustrating the information processing method provided in the embodiments of this application. Figure 3 ; Figure 5 This is an optional flowchart illustrating the information processing method provided in the embodiments of this application. Figure 4 ; Figure 6 This is an optional flowchart illustrating the information processing method provided in the embodiments of this application. Figure 5 ; Figure 7 This is a schematic diagram comparing the performance of different CSI feedback schemes under imperfect channel estimation conditions provided in the embodiments of this application; Figure 8 This is a schematic diagram comparing the SGCS performance of different CSI feedback schemes under perfect channel estimation conditions provided in the embodiments of this application; Figure 9 This is a schematic diagram comparing the normalized codebook entropy of different quantization schemes provided in the embodiments of this application; Figure 10 This is an optional structural diagram of the device provided in the embodiments of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0013] In the following description, the terms "first," "second," and "third" are used only to distinguish different objects and do not represent a specific order of objects, nor are they constituting a chronological order. It is understood that "first," "second," and "third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0014] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0015] The information processing method provided in this application embodiment can be applied to... Figure 1 The communication system shown, such as Figure 1 As shown, the communication system 100 may include a terminal device 110 and a network device 120. The network device 120 can communicate with the terminal device 110 via an air interface. Multi-service transmission is supported between the terminal device 110 and the network device 120. This application describes various embodiments in conjunction with the network device 120 and the terminal device 110, wherein the terminal device 110 may also be referred to as a user equipment (UE), access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent, or user device, etc.

[0016] The terminal device 110 in this application embodiment can be a device that provides voice / data / image connectivity to users, such as a handheld device, vehicle device, home device, home appliance, gaming device, etc., with wireless connection function or equipped with a wireless communication module. Examples include: mobile phones, tablets, laptops, PDAs, mobile internet devices (MIDs), virtual reality (VR) devices, augmented reality (AR) devices, wireless terminals in industrial control, wireless terminals in self-driving vehicles, drones or aerial photography equipment, wireless terminals in remote medical surgery, wireless terminals in smart grids, wireless terminals in transportation safety, wireless terminals in smart cities, wireless terminals in smart homes, cellular phones, cordless phones, session initiation protocol (SIP) phones, wireless local loop (WLL) stations, personal digital assistants (PDAs), handheld devices with wireless communication capabilities, computing devices or other processing devices connected to a wireless modem, terminal devices in 5G or 6G networks, or future public land mobile communication networks. Terminal devices in a network (PLMN), etc., are not limited to this in the embodiments of this application.

[0017] Network device 120 can be a device for communicating with mobile devices, and network device 120 may include access network equipment and / or core network equipment. That is, the wireless communication system also includes multiple core networks for communicating with the access network equipment. The access network equipment may be an evolved Node B (eNB or e-NodeB), macro base station, micro base station (also known as "small base station"), pico base station, transmission point (TP), or new generation Node B (gNodeB) in a long-term evolution (LTE) system, a next-generation (mobile communication system) (NR) system, a 6G communication system, or an authorized auxiliary access long-term evolution (LAA-LTE) system. In this embodiment, network device 120 can provide services to a cell. Terminal device 110 communicates with network device 120 through the transmission resources (e.g., frequency domain resources, or spectrum resources) used by the cell. The cell can be the cell corresponding to network device 120 (e.g., base station). The cell can belong to a macro base station or to a base station corresponding to a small cell. The small cell can include: metro cell, micro cell, pico cell, femto cell, etc. These small cells have the characteristics of small coverage area and low transmission power, and are suitable for providing high-speed data transmission services.

[0018] Figure 1 An exemplary illustration shows a base station and two terminal devices. Optionally, the wireless communication system 100 may include multiple base station devices and each base station may include other numbers of terminal devices within its coverage area. This application embodiment does not limit this.

[0019] Firstly, embodiments of this application can provide an information processing method, applied to a terminal device, such as... Figure 2 As shown, it includes S201 to S205.

[0020] S201. The terminal device performs joint source-channel coding on the first CSI to obtain a first CSI feature vector, wherein the first CSI feature vector includes multiple first feature units.

[0021] S202. The terminal device determines the first codebook corresponding to the first feature unit based on the location index of the first feature unit and multiple codebooks, wherein the multiple codebooks are obtained based on training of multiple second CSIs.

[0022] S203. The terminal device determines the first bit sequence corresponding to the first CSI feature vector based on the first codebook corresponding to the first feature unit in the first CSI feature vector.

[0023] S204. The terminal device determines the first orthogonal frequency division multiplexing (OFDM) feedback signal corresponding to the first CSI by modulating the first bit sequence.

[0024] S205. The terminal device sends the first OFDM feedback signal to the network device.

[0025] Secondly, embodiments of this application can provide an information processing method, applied to network devices, such as... Figure 3 As shown, it includes S301 to S304.

[0026] S301, The network device receives the first OFDM feedback signal.

[0027] S302. The network device performs channel estimation and equalization on the first OFDM feedback signal to obtain the modulation symbol.

[0028] S303. The network device demodulates the modulation symbol to obtain soft information.

[0029] S304. The network device performs joint source-channel decoding on the soft information to obtain the reconstructed CSI.

[0030] Below, on Figure 2 or Figure 3 The information processing method shown is described.

[0031] Terminal devices can receive downlink reference signals (RS) sent by network devices and perform channel estimation on the downlink reference signals to determine the downlink channel state (CSI). After determining the downlink CSI, the terminal device can feed it back to the network device. The downlink CSI can be understood as the CSI of the downlink channel. The CSI is a set of multi-dimensional information describing the channel state. The CSI can be represented as a CSI matrix. The CSI may include at least one of the following: Channel Quality Indicator (CQI), Precoding Matrix Indicator (PMI), Rank Indicator (RI), CSI-RS Resource Indicator (CRI), Layer Indicator (LI), and Layer 1 Reference Signal Received Power (L1-RSRP). In this embodiment, the content of the CSI is not limited.

[0032] After determining the first CSI, the terminal device performs joint source-channel coding on the first CSI to obtain the first CSI feature vector. The first CSI can be understood as any downlink CSI that needs to be fed back to the network device. The first CSI feature vector is the CSI feature vector obtained by performing joint source-channel coding on the first CSI; the CSI feature vector can also be called the CSI compressed feature vector.

[0033] In this embodiment, the first CSI can be feature extracted and compressed to obtain a first CSI feature vector. The first CSI feature vector is a real-valued vector used to represent the information extracted and compressed from the first CSI. The first CSI feature vector can be a continuous bit sequence, not a discrete bit sequence composed of 0s and 1s.

[0034] The terminal device can use a Joint Source-Channel Coding (JSCC) encoder to perform joint source-channel coding on the first CSI. In this case, the terminal device inputs the first CSI into the JSCC encoder, which outputs a first CSI feature vector. The JSCC encoder extracts and compresses features from the input first CSI to obtain the first CSI feature vector. The first CSI feature vector represents the information extracted and compressed by the joint source-channel encoder from the original CSI. For example, the first CSI feature vector retains information from the first CSI that is beneficial for CSI reconstruction.

[0035] The JSCC encoder, also known as a source-channel joint encoder, can employ a Transformer network structure, or a convolutional neural network, a multilayer perceptron, a lightweight attention network, or a combination thereof. For example, the source-channel joint encoder includes a linear embedding layer, a position coding layer, several Transformer coding layers, and an output projection layer. In this embodiment, the network structure used in the JSCC encoder is not limited.

[0036] The first CSI feature vector determined by the terminal device It can be represented as N represents the number of feature units included in the first CSI feature vector. Let N be the feature units at positions 1, 2, ..., N. The feature unit at position n out of N feature units can be represented as... This can also be referred to as the nth feature unit in the first CSI feature vector, where n ranges from 1 to N. In the embodiments of this application, This can also be understood as the feature value or feature representation at the nth position in the first CSI feature vector. In this embodiment, the position of the feature unit refers to the sequence index position in the first CSI feature vector, and is not limited to the actual physical space position, antenna position, or subcarrier position. For n, it can also be called the feature unit. The location index is different for different feature units in the first CSI feature vector.

[0037] The terminal device can determine the first codebook corresponding to each feature unit from multiple codebooks based on the position index of each feature unit in the first CSI feature vector. For a feature unit, the terminal device can determine the first codebook corresponding to that feature unit from multiple codewords based on the position index of that feature unit. In this embodiment, different codebooks in the multiple codebooks can be associated with different position indices. In this embodiment, the feature unit in the first CSI feature vector is referred to as the first feature unit.

[0038] A codebook can contain M codewords. Here, the codewords in the codebook can also be called candidate codewords. Each codeword is a soft bit vector of length B, where B represents the number of quantization bits corresponding to each feature position. For example, the first codebook of the nth feature unit in the first CSI feature vector can be represented as follows: , This represents the m-th codeword corresponding to the n-th feature unit, where m ranges from 1 to M.

[0039] In this embodiment of the application, feature units at different positions in the first CSI feature vector can correspond to codebooks at different positions, so that feature units at different positions can learn different codewords based on their own feature distribution.

[0040] In this embodiment, multiple codebooks are obtained based on multiple second CSIs (Code Sample Instructions for Training). These codebooks can be referred to as learnable codebooks. A learnable codebook is understood as a codebook that is not manually fixed, but rather one that can be trained using multiple second CSIs. In this embodiment, the training of multiple codebooks can be performed by a terminal device or by other training devices. When multiple codebooks are trained by other devices, the training device obtains multiple codebooks through training with the second CSIs and then sends these codebooks to the terminal device, thereby deploying the multiple codebooks on the terminal device. Here, the second CSI can be understood as a sample CSI used for codebook training.

[0041] The terminal device determines the first codebook corresponding to each of the multiple first feature units in the first CSI feature vector, thus obtaining multiple first codebooks. Based on these multiple first codebooks, the first bit sequence corresponding to the first CSI feature vector can be determined. Here, the first bit sequence is discrete; it can be understood as a bit sequence obtained by discretizing the first CSI feature vector based on multiple codebooks. The first bit sequence can also be called the discrete feedback bit stream obtained after discretizing the first CSI feature vector. The first bit sequence is a discrete bit sequence composed of 0s and 1s and can be directly used for digital modulation.

[0042] In this embodiment, the terminal device may be deployed with a learnable codebook quantization module, which is used to convert the first CSI feature vector into a first bit sequence. For example, it may be as follows: Figure 4 As shown, the CSI input is to the JSCC encoder, and the JSCC encoder outputs the CSI feature vector. The CSI feature vector is then converted into a bit sequence by the learnable codebook quantization module.

[0043] After determining the first bit sequence, the terminal device modulates and maps the first bit sequence to obtain the first OFDM feedback signal. And transmit the first OFDM feedback signal via a wireless channel The signal is sent to the network device. For example, the first OFDM feedback signal is sent to the network device via a wireless feedback link. Here, the first OFDM feedback signal sent by the terminal device is used by the network device to reconstruct the CSI to obtain the first CSI.

[0044] The terminal device sends the first OFDM feedback signal to the network device via a wireless channel. The network device receives the first OFDM feedback signal transmitted via the wireless channel. First feedback signal During wireless transmission, factors such as multipath fading, additive noise, and channel estimation errors may affect the first OFDM feedback signal received by the network device, thus affecting its performance. It may not be the first OFDM feedback signal sent by the terminal device. itself.

[0045] The network device performs channel estimation and linear equalization on the received first OFDM feedback signal to obtain equalized modulation symbols. Here, channel estimation and linear equalization on the received first OFDM feedback signal yield multiple modulation symbols, which constitute a modulation symbol vector. The network device demodulates the modulation symbols to obtain soft information. This soft information represents the soft reliability information of the values ​​of each feedback bit in the feedback bitstream, and can also be called the soft information corresponding to the feedback bitstream. For example, the soft information corresponding to the feedback bitstream represents the reliability of each feedback bit being 0 or 1. In this embodiment, the soft information may include a log-likelihood ratio (LLR) sequence. The network device can input the modulation symbols into a digital demodulation module to obtain the soft information output by the digital demodulation module.

[0046] Network devices can perform joint source-channel decoding on soft information to obtain the reconstructed CSI. In this embodiment, joint source-channel decoding of the soft information in the feedback bitstream by the network device can be understood as performing CSI reconstruction based on the soft information in the feedback bitstream to obtain the reconstructed CSI. Here, the reconstructed CSI can be understood as the CSI reconstructed by the network device based on the received first OFDM signal.

[0047] For example, a network device can input soft information into a Joint Source Channel (JSCC) decoder to obtain a matrix of reconstructed CSI output by the JSCC decoder. The joint source-channel decoder can also be called a source-channel joint decoder. The source-channel joint decoder can employ a Transformer decoding structure corresponding to the source-channel joint encoder, or it can employ a convolutional network, a multilayer perceptron, or a lightweight neural network structure. For example, the source-channel joint decoder includes an input projection layer, a position coding layer, several Transformer decoding layers, and an output projection layer, used to map the soft information recovered from the feedback link to the reconstructed CSI.

[0048] In this embodiment of the application, after receiving the first OFDM feedback signal, the network device does not need to determine the feedback bit stream, but instead determines the soft information corresponding to the feedback bit stream to reflect the information related to the feedback bit stream, and performs CSI reconstruction based on the soft information corresponding to the feedback bit stream.

[0049] Understandably, in this embodiment, no independent codebook dequantization step is provided. The source-channel joint decoder on the network device side can be a trained neural network decoder, which learns the mapping relationship from soft information to reconstructed CSI through training. Therefore, it can reconstruct CSI based on the soft information sequence of the feedback bitstream. For example, the source-channel joint decoder can be obtained through training multiple second CSIs.

[0050] In this embodiment, the terminal device can be deployed with multiple codebooks trained by multiple second CSIs, and select a first codebook for each feature unit from multiple codebooks by using the position index of each feature unit in the continuous CSI feature vector, thereby performing position-level codebook selection to generate a discrete feedback bit stream. The resulting feedback bit stream is a bit representation of the feedback link that is adaptively learned according to the CSI feature distribution, which can improve the insufficient feedback bit expression capability and improve the CSI reconstruction performance.

[0051] In some embodiments, the codebook is obtained by mapping codebook parameters to a first numerical interval, wherein the lower boundary and upper boundary of the first numerical interval are 0 and 1, respectively; the codebook parameters are updated as training parameters based on the training of the plurality of second CSIs.

[0052] A codebook can be obtained by mapping its codebook parameters to a first numerical interval [0,1]. The values ​​in the codebook parameters can be greater than or equal to 0, without the restriction of the [0,1] interval. A codebook is obtained by mapping its codebook parameters to the first numerical interval. For example, the codebook parameters are mapped to the [0,1] interval using the Sigmoid function. In this embodiment, the method of mapping the codebook parameters to the [0,1] interval is not limited.

[0053] For example, for codebook Its corresponding codebook parameter can be a size of That is, a matrix with M rows and B columns, where the elements of the m-th row, i.e., the parameters, are obtained after mapping to the first numerical interval. One parameter in the codebook is mapped to a bit in the codebook.

[0054] In this embodiment, during the codebook training phase, the codebook is represented by trainable codebook parameters, and the codebook parameters are mapped to the [0,1] interval using the Sigmoid function, so that each candidate codeword forms a learnable soft bit prototype. After training is completed, the codebook parameters are fixed and mapped to the codebook for use in generating the feedback bit sequence during the inference phase.

[0055] In some embodiments, S203 determines the first bit sequence corresponding to the first CSI feature vector based on the first codebook corresponding to the first feature unit in the first CSI feature vector, including the following steps: S231, determining the first target codeword corresponding to the first feature unit based on the first codebook corresponding to the first feature unit in the first CSI feature vector, obtaining a plurality of first target codewords, wherein the first target codeword is a soft bit vector; S232, converting the plurality of first target codewords into the first bit sequence.

[0056] In this embodiment of the application, for each first feature unit in the first CSI feature vector, the first target codeword corresponding to the first feature unit is determined according to the first codebook corresponding to the first feature unit, thereby obtaining multiple first target codewords. The first target codeword can be understood as the target codeword corresponding to the first feature vector.

[0057] Understandably, for the feature unit at the nth position... ,according to Determine the Corresponding target codeword Thus obtain , ... The terminal device will , ... The conversion yields the first bit sequence.

[0058] In this embodiment of the application, for the first CSI feature vector, the terminal device obtains the first bit sequence of the first CSI feature vector based on the target codeword of the feature unit at all positions, thereby ensuring the effectiveness and reliability of the first bit sequence.

[0059] In some embodiments, S231 determines the first target codeword corresponding to the first feature unit based on the first codebook corresponding to the first feature unit in the first CSI feature vector, including the following steps: S311, determining the first codeword selection scoring vector corresponding to the first feature unit based on the first CSI feature vector; the first codeword selection scoring vector includes multiple selection scores corresponding to different codewords; S312, determining the first target codeword corresponding to the first feature unit based on the first codeword selection scoring vector corresponding to the first feature unit and the first codebook corresponding to the first feature unit.

[0060] In this embodiment of the application, for each first feature unit in the first CSI feature vector, the following processing is performed to obtain the first target codeword corresponding to the first feature unit: determining the first codeword selection scoring vector corresponding to the first feature unit, and determining the first target codeword based on the first codeword selection scoring vector and the first codebook corresponding to the first feature unit. The first codeword selection scoring vector can be understood as the codeword selection scoring vector of the first feature unit.

[0061] For the feature unit at the nth position The first codeword selection scoring vector can be determined. Where M represents the number of codewords included in the codebook. The first codeword selection score represents the matching result or codeword selection score of the feature unit at position n selecting the candidate codeword at position m. In this embodiment, the first codeword selection score vector can be used to characterize the probability that a codeword at each position in the first codebook corresponding to the first feature unit is selected. It is understood that the codeword selection score is not the final discrete selection result, but rather the probability of each codeword being selected in subsequent calculations.

[0062] In some embodiments, S311, determining the first codeword selection scoring vector corresponding to the first feature unit based on the first CSI feature vector, may include: determining the first codeword selection scoring vector based on the first feature unit.

[0063] Here, for ,based on Sure The corresponding first codeword selection scoring vector; based on Sure The corresponding first codeword selection scoring vector, and so on, based on Sure The corresponding first codeword selection scoring vector; based on Sure The corresponding first codeword selection scoring vector, based on Sure The corresponding first codeword selection scoring vector is used to determine the first codeword selection scoring vector corresponding to each first feature unit in the first CSI feature vector.

[0064] For example, the terminal device can input the first feature unit to a selector to obtain the first codeword selection score vector corresponding to the first feature unit output by the selector. In this embodiment, for the first feature unit at each position in the first CSI feature vector, a selector can be set. For a first feature unit at a certain position, the first codeword selection score vector corresponding to the first feature unit is determined by the selector corresponding to that position. The terminal device can input the first feature unit to the selector corresponding to the position of the first feature unit to obtain the first codeword selection score vector corresponding to the first feature unit output by the selector.

[0065] The selector can be obtained based on training multiple second CSIs. In the embodiments of this application, when a selector for a position determines the codeword selection score vector of the feature unit at that position based on the feature unit at that position, the input of the selector for a position is the feature unit at that position during the training phase.

[0066] A selector, also known as a selector network, generates a selection score for a feature unit corresponding to multiple candidate codewords based on the feature value corresponding to the feature unit at the current position. For example, a selector network may include a one-dimensional convolutional layer, a hidden fully connected layer, and an output fully connected layer. The one-dimensional convolutional layer is used to capture local correlations between adjacent positions, and the fully connected layer is used to generate matching scores for candidate codewords. In this embodiment, the structure of the selector network can be other than the example described above; the model structure of the selector network is not limited in this embodiment.

[0067] In some embodiments, S311, determining the first codeword selection score vector corresponding to the first feature unit based on the first CSI feature vector includes: determining the first codeword selection score vector corresponding to the first feature unit based on the first feature unit and neighboring feature units adjacent to the first feature unit.

[0068] The terminal device can determine the first codeword selection scoring vector corresponding to the first feature unit based on the first feature unit and its neighboring feature units. Here, the terminal device can combine the feature value corresponding to the current feature unit with the feature values ​​of its neighboring feature units to generate a selection score for the feature unit corresponding to multiple candidate codewords. In this embodiment, the feature unit used to determine the first codeword selection scoring vector can be different for different first feature units. Here, for a feature unit, the feature unit at the corresponding position can be selected as a neighboring feature unit, where there may be a case where no feature unit exists at a certain position.

[0069] In this embodiment of the application, a first number of adjacent feature units can be selected as neighboring feature units. The value of the first number can be 1, 2, 3, etc. In this embodiment of the application, the value of the first number is not limited.

[0070] In one example, the first quantity is 1. For a feature cell, the first quantity of neighboring feature cells includes the next positional feature cell after that feature cell. , Neighboring feature units include Based on and Sure The corresponding first codeword selection scoring vector; The neighboring feature units include Based on and Sure The corresponding first codeword is selected as the scoring vector, and so on. The neighboring feature units include Based on and Sure The corresponding first codeword selection scoring vector; The neighboring feature units include Based on and Sure The corresponding first codeword selection scoring vector, If no neighboring feature units exist, then based on Sure The corresponding first codeword selection scoring vector is used to determine the first codeword selection scoring vector corresponding to each first feature unit in the first CSI feature vector.

[0071] In one example, the first quantity is 2. For a feature cell, the number of neighboring feature cells at the first quantity positions includes the feature cells on both sides of the feature cell. , The neighboring feature units include Based on and Sure The corresponding first codeword selection scoring vector; The neighboring feature units include and Based on , and Sure The corresponding first codeword is selected as the scoring vector, and so on. The neighboring feature units include and Based on , and Sure The corresponding first codeword selection scoring vector, The neighboring feature units include: Based on and Sure The corresponding first codeword selection scoring vector is used to determine the first codeword selection scoring vector corresponding to each first feature unit in the first CSI feature vector.

[0072] In this embodiment of the application, in the process of determining the first codeword selection scoring vector corresponding to a first feature unit, the neighboring feature units of the first feature unit are introduced, so that the neighboring feature units are used as the context of the first feature unit. The local correlation between the first feature unit and the neighboring feature units is used to determine the first codeword selection scoring vector corresponding to the first feature unit, thereby improving the codeword selection accuracy.

[0073] In some embodiments, determining a first codeword selection score vector corresponding to the first feature unit based on the first feature unit and neighboring feature units adjacent to the first feature unit includes: inputting the first feature unit and neighboring feature units adjacent to the first feature unit into a selector to obtain a first codeword selection score vector output by the selector; the selector is obtained based on training of the plurality of second CSIs.

[0074] In this embodiment, for each first feature unit in the first CSI feature vector, a selector can be set. For a first feature unit, the first codeword selection score vector corresponding to the first feature unit is determined by the selector corresponding to the position of the first feature unit. The terminal device can input the first feature unit and its neighboring feature units to the selector corresponding to the position of the first feature unit to obtain the first codeword selection score vector corresponding to the first feature unit output by the selector.

[0075] The selector can be obtained based on training multiple second CSIs. In the embodiments of this application, when a selector for a position determines the codeword selection score vector corresponding to the feature unit at that position based on the feature unit at that position and the neighboring feature units of adjacent positions, the input of the selector for a position during the training phase is the feature unit at that position and the neighboring feature units of adjacent positions.

[0076] A selector, also known as a selector network, generates a selection score for a feature unit corresponding to multiple candidate codewords based on the feature value of the feature unit at the current position and the feature values ​​of its neighboring feature units.

[0077] For example, the selector network may include a one-dimensional convolutional layer, a hidden fully connected layer, and an output fully connected layer, wherein the one-dimensional convolutional layer is used to capture local correlations between adjacent positions, and the fully connected layer is used to generate matching scores for candidate codewords. In this embodiment, the structure of the selector network may be other than the example described above, and the model structure of the selector network is not limited in this embodiment.

[0078] In this embodiment of the application, during the training process, the selector can learn the features of the codebook at the corresponding position of the selector, and then output the codeword selection score vector corresponding to the input feature unit based on the learned characteristics of the codeword and the input feature unit.

[0079] In some embodiments, S312 determines the first target codeword corresponding to the first feature unit based on the first codeword selection scoring vector corresponding to the first feature unit and the first codebook corresponding to the first feature unit, including: Based on the first codeword selection score vector corresponding to the first feature unit, the first target codeword corresponding to the first feature unit is determined from the multiple codewords included in the first codebook corresponding to the first feature unit. The first target codeword is the codeword with the highest selection score in the first codebook.

[0080] Here, for the first feature unit, the method of determining the first target codeword corresponding to the first feature unit based on the first codebook and the first codeword selection score vector includes method 1, which determines the codeword with the highest selection score in the first codebook as the first target codeword.

[0081] For the first feature unit Its corresponding first codebook is The corresponding first codeword selection scoring vector is ,in, for The corresponding selection rating, for The corresponding selection ratings, and so on, for The corresponding selection rating. Understandable. for The corresponding selection rating, then The codeword with the highest selection score is the codeword with the highest selection score, and the codeword with the highest selection score is determined as the first feature unit. The first target code Here, when the codeword is a soft bit vector of length B, the first target codeword is a soft bit vector of length B.

[0082] In this embodiment of the application, if the highest selection score in the first codeword selection score vector includes multiple selection scores with the same value, a selection score can be randomly selected from the multiple highest selection scores as the final highest selection score, or the selection score that appears first among the multiple highest selection scores can be selected as the final highest selection score, and the codeword corresponding to the final highest selection score can be selected as the first target codeword. Alternatively, the first target codeword can be obtained by averaging the multiple codewords corresponding to the multiple highest selection scores.

[0083] In this embodiment, the vector can be selected based on the codeword corresponding to the first feature unit, and the candidate codeword with the highest selection score in the first codeword can be directly selected as the first target codeword corresponding to the first feature unit, thereby reducing the implementation complexity of the scheme.

[0084] In some embodiments, S312 determines the first target codeword corresponding to the first feature unit based on the first codeword selection scoring vector corresponding to the first feature unit and the first codebook corresponding to the first feature unit, including: determining the first soft selection probability vector corresponding to the first feature unit based on the first codeword selection scoring vector corresponding to the first feature unit, wherein the first soft selection probability vector includes multiple soft selection probabilities; and performing weighted aggregation on multiple codewords included in the first codebook through the multiple soft selection probabilities to obtain the first target codeword corresponding to the first feature unit.

[0085] Here, the codewords in the first codebook can be considered as soft codewords. Therefore, weighted aggregation of multiple codewords included in the first codebook using multiple soft selection probabilities can also be called soft codeword aggregation.

[0086] For the first feature unit, the method of determining the first target codeword corresponding to the first feature unit based on the first codebook and the first codeword selection scoring vector includes method 2: weighted aggregation of the codewords in the first codebook based on the first selection scoring vector to obtain the first target codeword.

[0087] In this embodiment, the terminal device can first calculate a first soft selection probability vector based on a first selection scoring vector, and then perform weighted aggregation on the codewords in the first codebook based on the first soft selection probability vector. Here, calculating the first soft selection probability vector of the first codeword selection scoring vector can be understood as obtaining a soft selection probability with a value between 0 and 1 based on the mapping of each selection score in the first codeword selection scoring vector. The sum of all soft selection probabilities in the first soft selection probability vector can be 1.

[0088] For the first feature unit The corresponding first codebook is The corresponding first codeword selection scoring vector is The terminal device can select a scoring vector based on the first codeword. Calculate the first soft-selection probability vector ,in, for The corresponding soft selection probability represents the soft selection probability of the nth feature unit choosing the mth candidate codeword. Here, The value of is located in the interval [0,1], and The sum can be 1. In this embodiment, the first soft selection probability vector of the first codeword selection score vector can be calculated by an activation function (Softmax function), Gumbel-Softmax function or other differentiable discrete selection approximation methods. In this embodiment, the method of calculating the first soft selection probability vector corresponding to the first codeword selection score vector is not limited.

[0089] Terminal devices based on soft selection probability vectors We perform weighted aggregation on the candidate codewords in the codebook corresponding to the nth feature unit to obtain the soft bit representation of that position, i.e., the target codeword: .

[0090] because The soft selection probabilities are continuous, and the candidate codewords are... Since it is a soft bit vector with values ​​in the interval [0,1], the weighted aggregation is obtained. It is still represented by consecutive soft bits, rather than the final 0 / 1 binary bits.

[0091] In the embodiments of this application, If each codeword is a soft bit vector of length B, then... It is a soft bit vector of length B.

[0092] In this embodiment of the application, for a feature unit at a certain position, the codeword selection score vector at that position is converted into a soft selection probability vector, and the codewords in the codebook at that position are weighted and aggregated based on the soft selection probability vector. The target codeword of the feature unit at that position is determined by the aggregation of codewords, thereby improving the expressive power of the target codeword.

[0093] In some embodiments, S232 converts the plurality of first target codewords into the first bit sequence, including the following steps: S321. Hard quantize the plurality of first target codewords to obtain a plurality of first bit vectors; S322. Concatenate the plurality of first bit vectors to obtain the first bit sequence.

[0094] Hard quantization here can also be described as binarization. For each of the multiple first target codewords, the B values ​​included in the first target codeword are hard quantized to obtain a first bit vector of length B. For N first target codewords, N first bit vectors are obtained. The N first bit vectors are concatenated to obtain the first bit sequence. The first bit sequence can be understood as the final feedback bit stream.

[0095] If the first CSI feature vector includes N feature units, and the first bit vector corresponding to each feature unit includes B bits, then the final feedback bitstream length is... .

[0096] In some embodiments, in S232, multiple first target codewords can be concatenated to obtain concatenated codewords, and the concatenated codewords can be hard quantized to obtain a first bit sequence.

[0097] The first bit sequence is a discrete bit sequence consisting of 0s and 1s, which can be directly used for digital modulation.

[0098] In some embodiments, S204 determines the first OFDM feedback signal corresponding to the first CSI by modulating the first bit sequence, including: S241, digitally modulating the first bit sequence to obtain a first modulation symbol vector; S242, performing resource mapping on the first modulation symbol vector to generate the first OFDM feedback signal.

[0099] The terminal device digitally modulates the first bit sequence to obtain a first modulation symbol vector x, maps the first modulation symbol vector x onto an OFDM resource grid, and generates a first OFDM feedback signal. The first modulation symbol vector can be understood as the modulation symbol vector obtained by digitally modulating the first bit sequence.

[0100] In this embodiment of the application, the digital modulation method used to digitally modulate the first bit sequence may include Quadrature Amplitude Modulation (QAM), Quadrature Phase Shift Keying (QPSK), Phase Shift Keying (PSK), or other digital modulation methods. In this embodiment of the application, the digital modulation method used for the first bit sequence is not limited.

[0101] In this embodiment of the application, the terminal device can input the first bit sequence into the digital modulation module and perform digital modulation on the first bit sequence through the digital modulation module.

[0102] After generating the first OFDM feedback signal, the terminal device can perform power normalization processing on the first OFDM feedback signal before sending it to the network device through the wireless channel to ensure transmission stability.

[0103] In some embodiments, S201 performs joint source-channel coding on the first CSI to obtain a first CSI feature vector, including: separating the real and imaginary parts of the first CSI to obtain real data and imaginary data; concatenating the real data and the imaginary data to obtain a first real-valued CSI tensor; inputting the first real-valued CSI tensor into the joint source-channel encoder to obtain the first CSI feature vector output by the joint source-channel encoder; the joint source-channel encoder is obtained based on training of the plurality of second CSIs.

[0104] The terminal device can reassemble the first CSI before performing joint source channel coding on the first CSI. Here, the first CSI can be a complex CSI matrix. The terminal device separates the real and imaginary parts of the complex CSI matrix, obtaining real and imaginary data. The real data is a real matrix composed of the real parts of the CSI matrix, and the imaginary data is an imaginary matrix composed of the imaginary parts of the CSI matrix. The real and imaginary matrices are then concatenated along a new dimension to form a first CSI tensor. This first CSI tensor is the CSI tensor corresponding to the first CSI, and it is a real-valued CSI tensor, which can be called the first real-valued CSI tensor. The first real-valued CSI tensor can be understood as the real-valued CSI tensor after the first CSI is reassembled.

[0105] In one example, the first CSI is a matrix of size 32×13. Its separated real and imaginary parts are both 32×13. The real and imaginary parts are concatenated to obtain the first real-valued CSI tensor of size 2×32×13.

[0106] After obtaining the first real-valued CSI tensor, the terminal device inputs the first real-valued CSI tensor into the joint source-channel encoder to obtain the first CSI feature vector output by the joint source-channel encoder.

[0107] In this embodiment, the terminal device separates the real and imaginary parts of the complex CSI matrix and splices them along a new dimension to form a real-valued CSI tensor. The real-valued CSI tensor is used as the input of the source-channel joint encoder, thereby simplifying the processing of the source-channel joint encoder and improving its processing performance.

[0108] In some embodiments, the plurality of second CSIs are used to train the plurality of codebooks and at least one of the following: A joint source-channel encoder, wherein the joint source-channel encoder is used for joint source-channel coding of CSI to obtain CSI feature vectors; A selector is used to determine the codeword selection score vector corresponding to the feature units included in the CSI feature vector; A joint source-channel decoder, which is used for CSI reconstruction.

[0109] In this embodiment, in addition to training multiple codebooks, the multiple second CSIs can also train one or more of the following: a joint source-channel encoder, a selector, and a joint source-channel decoder. In this case, these training objects can learn CSI-related features during the CSI feedback process.

[0110] When multiple second CSI training objects include a joint source-channel encoder, the joint source-channel encoder can learn the feature representation of CSI, thereby improving the accuracy and compression performance of feature extraction and enhancing the semantic expressiveness of the CSI feature vector output by the joint source-channel encoder.

[0111] When multiple second CSI training objects include selectors, the selectors can effectively learn the characteristics of the feature units at the corresponding positions and evaluate the codeword selection score vector based on the distribution of CSI feature vectors.

[0112] When multiple training objects for the second CSI include the joint source-channel encoder, the joint source-channel decoder can learn the relationship between the soft information of the feedback bitstream and the CSI, thereby improving the CSI reconstruction performance of the joint source-channel decoder.

[0113] In this embodiment of the application, when the joint source-channel encoder, multiple codebooks, selectors, and joint source-channel decoders are trained by multiple second CSIs, the collaborative optimization of the joint source-channel encoder, learnable codebooks, selectors, and joint source-channel decoders can be achieved.

[0114] In some embodiments, the second CSI undergoes joint source-channel coding to obtain a second CSI feature vector, which includes multiple second feature units; a second codeword selection scoring vector corresponding to the second feature unit is used to determine a second soft-select probability vector corresponding to the second feature unit; a second codebook and the second soft-select probability vector corresponding to the second feature unit are used to determine a second target codeword corresponding to the second feature vector; the second target codeword corresponding to the second feature unit in the second CSI feature vector is used to determine a second bit sequence, which is digitally modulated to obtain a second OFDM feedback signal corresponding to the second CSI; the second OFDM feedback signal is reconstructed into a third CSI after being transmitted through a wireless channel.

[0115] During training, the source-channel joint encoder generates a continuous second CSI feature vector based on the input CSI, i.e., the second CSI. The learnable codebook quantization module generates a second bit sequence through differentiable soft selection, soft codeword aggregation, and hard binarization operations. After the second bit sequence is transmitted through the digital feedback link, the source-channel joint decoder on the base station side outputs the reconstructed CSI, i.e., the third CSI.

[0116] During training, the second CSI is encoded to obtain the second CSI feature vector, which includes multiple second CSI feature units. Based on the position index of the second CSI feature units, a second codebook corresponding to the second CSI feature vector is selected from multiple codebooks. Based on the second codebook corresponding to the second feature units in the second CSI feature vector, the second bit sequence corresponding to the second CSI feature vector is determined. By modulating the second bit sequence, the second OFDM feedback signal corresponding to the second CSI is determined. The second OFDM feedback signal is transmitted through a digital feedback link and then used to reconstruct the third CSI. The codewords during training can be represented by codeword parameters.

[0117] In this embodiment, training can be performed based on a wireless communication system including terminal devices and network devices, or it can be performed based on a simulated wireless communication environment of terminal devices and network devices. When training is performed based on a wireless communication system, the wireless channel for transmitting the second OFDM signal is the channel between the terminal device and the network device. When training is performed based on a simulated wireless communication environment, the wireless channel for transmitting the second OFDM signal can be a wireless channel simulated by a model or simulator.

[0118] In this embodiment, the process of reconstructing CSI based on the second CSI is the same as that of reconstructing CSI based on the first CSI, and will not be repeated here. It should be noted that, for the first CSI, the method of determining the first target codeword corresponding to the first feature unit based on the first codebook and the first codeword corresponding to the first feature unit includes method 1 or method 2; for the second CSI, the method of determining the second target codeword corresponding to the second feature unit based on the second codebook and the second codeword corresponding to the second feature unit includes method 2.

[0119] During training, the second CSI can be understood as the original CSI, and the third CSI as the reconstructed CSI. A reconstruction loss can be constructed based on the similarity between the second and third CSIs. This reconstruction loss can also be called the CSI reconstruction loss. For example, a squared generalized cosine similarity correlation loss can be used to improve the spatial orientation consistency between the reconstructed CSI and the original CSI. For the i-th subcarrier, the squared generalized cosine similarity between the original CSI and the reconstructed CSI can be used to measure their directional consistency, and further averaged over multiple subcarriers as an evaluation criterion for CSI reconstruction performance.

[0120] In this embodiment, end-to-end joint optimization of CSI feature extraction, feedback bit generation, wireless transmission, and base station-side CSI reconstruction is achieved through a location-level learnable codebook, selector network, soft aggregation, and hard binarization mechanism of pass-through estimator.

[0121] In this embodiment of the application, in order to solve the problem of non-differentiability of hard binarization operation, a pass-through estimator (STE) is used during training so that the forward propagation outputs hard binary bits, while the gradient is transmitted through the soft bit representation before hard binarization during back propagation, thereby maintaining effective gradient propagation between the encoder, selector network, learnable codebook and decoder.

[0122] In some embodiments, the reconstruction loss can be directly used as the training loss to determine whether to continue training and to update the parameters of the training object.

[0123] In some embodiments, the second CSI and the third CSI are used to determine the reconstruction loss; the second soft selection probability vector corresponding to each second feature unit in the second CSI feature vector is used to determine the diversity loss; the reconstruction loss and the diversity loss are used to determine the training loss, and the training loss is used to update the plurality of codebooks and one or more of the following: a joint source-channel encoder, a selector, and a joint source-channel decoder.

[0124] During training, the reconstruction loss can be determined based on the second CSI and the third CSI, and the diversity loss can be determined based on the second soft selection probability vector corresponding to each second feature unit in the second CSI feature vector. The training loss, i.e. the total training loss, is determined based on the reconstruction loss and the diversity loss.

[0125] For example, the overall training loss can be expressed as ,in, For total training losses, To rebuild the losses, For the loss of diversity, This represents the weight for the diversity loss. For example, The value of can be less than 1 and greater than 0. The diversity loss can be determined based on the second soft-selection probability vector corresponding to each second feature unit in the second CSI feature vector.

[0126] Diversity loss can also be called diversity regularization.

[0127] In this embodiment of the application, the final training loss can be composed of a weighted sum of reconstruction loss and codebook diversity regularization term.

[0128] In this embodiment, the following training parameters can be updated using the backpropagation algorithm: source-channel joint encoder parameters, selector parameters, learnable codebook parameters, and source-channel joint decoder parameters. After training, the source-channel joint encoder and learnable codebook quantization module are deployed on the user equipment side, and the corresponding receiver channel processing module, digital demodulation module, and source-channel joint decoder are deployed on the network equipment side for actual CSI feedback and CSI reconstruction.

[0129] In some embodiments, the diversity loss is determined based on a third soft-selection probability vector corresponding to a plurality of second feature units in the second CSI feature vector, the third soft-selection probability vector including an average selection probability determined based on the second soft-selection probability vector.

[0130] In this embodiment, for a second feature unit at a given location, the average selection probability can be obtained by averaging the soft selection probabilities of multiple second soft selection probability vectors for that location based on a training batch. This average selection probability then yields the third soft selection probability vector corresponding to the second feature unit at that location. Thus, the diversity loss is determined based on the third soft selection probability vectors corresponding to second feature units at multiple locations. For example, for a codeword, the average selection probability is obtained by averaging the soft selection probabilities of multiple codewords corresponding to that codeword based on a training batch.

[0131] In this embodiment of the application, when the second CSI feature vector includes N second feature units, there may be N third soft selection probability vectors, and the diversity loss is determined based on the average selection probability among the N third soft selection probability vectors.

[0132] In this embodiment of the application, the loss function used to determine the diversity loss based on the average selection probability among multiple third soft selection probability vectors is not limited.

[0133] In this embodiment, the diversity loss is calculated based on the average selection probability of different codewords in the training batch. This avoids insufficient codebook utilization or codeword collapse caused by long-term concentrated selection of a few codewords, thereby encouraging the full use of multiple codewords and improving the codebook's expressive power and the representation efficiency of the feedback bitstream.

[0134] The information processing method provided in the embodiments of this application will now be described in conjunction with specific scenarios.

[0135] The information processing method described in this application relates to the interdisciplinary fields of wireless communication and artificial intelligence, and particularly to a channel state information feedback method, apparatus, and communication system for frequency division duplexing-orthogonal frequency division multiplexing (FDD-OFDM) systems. Specifically, this invention introduces a source-channel joint encoder and a learnable codebook quantization module on the user equipment side to convert the continuous representation of downlink channel state information into a transmittable feedback bit stream. On the base station side, the channel state information is reconstructed through corresponding demodulation, equalization, and decoding modules, thereby reducing CSI feedback overhead and improving reconstruction performance under constrained feedback links.

[0136] In FDD-OFDM communication systems, base stations need to acquire downlink CSI to support functions such as precoding, beamforming, and radio resource scheduling. Since uplink and downlink channels in FDD systems typically lack strict reciprocity, user equipment (UE) needs to estimate the downlink CSI based on the downlink reference signal and then feed the CSI information back to the base station via the uplink feedback link. However, in wideband multi-antenna communication systems, CSI usually has high dimensionality. Directly feeding back the complete CSI would consume a significant amount of uplink feedback resources, making it difficult to meet the requirements of practical systems for limited feedback overhead and low-latency transmission.

[0137] To reduce CSI feedback overhead, various deep learning-based CSI compression feedback methods have been proposed. These methods typically utilize neural networks to extract and compress high-dimensional CSI features on the user equipment side, and then use a corresponding decoding network on the base station side for CSI reconstruction, thereby reducing the amount of feedback data and improving reconstruction performance to some extent. In recent years, joint source-channel coding methods have also been introduced into wireless feedback transmission scenarios. By jointly optimizing feature representation and channel transmission processes, these methods improve the system's robustness under noisy channels and limited feedback resources.

[0138] In practical digital feedback links, the compressed CSI representation typically needs to be further converted into a finite-length feedback bitstream, undergoing digital modulation, resource mapping, wireless channel transmission, demodulation, and decoding. To achieve the conversion from continuous features to discrete feedback information, related technologies generally employ fixed uniform quantization, traditional vector quantization, or codeword index lookup to convert the continuous feature representation output by the neural network into a discrete representation for subsequent feedback transmission.

[0139] Furthermore, while CSI feedback techniques can reduce feedback overhead to some extent, they still suffer from insufficient adaptation between continuous feature representations and actual digital feedback links. Specifically, some deep learning CSI feedback methods primarily focus on CSI compression and reconstruction, without fully considering how the compressed continuous CSI feature vectors can directly generate a transmittable feedback bitstream. Existing fixed uniform quantization, traditional vector quantization, or codeword index lookup methods typically treat quantization as an independent post-processing step, making it difficult to adaptively learn bit representations for the feedback link based on CSI feature distribution, and also difficult to unify and optimize with feature extraction, quantization, channel transmission, and CSI reconstruction processes. Simultaneously, the quantization operation itself is discrete, and direct hard decision or codeword selection can easily block gradient propagation, making end-to-end joint training of the encoder, quantizer, and decoder difficult. Therefore, in scenarios with low signal-to-noise ratios, imperfect channel estimation, or limited feedback resources, the methods in these techniques are prone to problems such as insufficient feedback bit representation capability, degraded reconstruction performance, and weak system robustness.

[0140] This application proposes a learnable codebook quantization CSI feedback method, device, and communication system for FDD-OFDM systems. It aims to solve the problems in CSI feedback methods, such as the difficulty in directly adapting continuous feature vectors to actual digital feedback links, the difficulty in jointly optimizing traditional quantization methods with neural network encoders and decoders end-to-end, and the insufficient performance and robustness of CSI reconstruction under limited feedback resources. The following solutions are provided.

[0141] In this embodiment of the application, for the downlink channel state information feedback scenario in an FDD-OFDM communication system, the system includes a user equipment, a base station, a source-channel joint encoder, a learnable codebook quantization module, a digital modulation module, an OFDM resource mapping module, a wireless feedback link, a receiver channel processing module, a digital demodulation module, and a source-channel joint decoder.

[0142] The user equipment (UE) first estimates the downlink channel state information (CSI) based on the downlink reference signal. Then, it extracts features from the CSI using a source-channel joint encoder to obtain a continuous CSI compressed feature vector. Subsequently, it uses a learnable codebook quantization module to convert the continuous feature vector into a transmittable discrete feedback bit stream, which is then transmitted to the base station after digital modulation and OFDM resource mapping. After receiving the feedback signal, the base station performs channel processing, equalization, demodulation, and decoding to finally reconstruct the downlink CSI.

[0143] It should be noted that, in this embodiment, a learnable codebook quantization module is set between the source-channel joint encoder on the user equipment side and the digital feedback link. This allows continuous CSI compression feature vectors to generate a feedback bitstream that can directly enter the digital modulation link through position-level codebook selection, soft codeword aggregation, and hard binarization. Digital modulation, OFDM resource mapping, wireless feedback link, equalization, and demodulation modules are used to construct the feedback transmission link, and their specific implementation can be adjusted according to the communication system configuration. In this embodiment, a new CSI feedback method is defined, along with the quantization codebook defined under this method, and joint training performed to optimize the system.

[0144] The following sections describe the end-to-end CSI feedback implementation process, the learnable codebook quantization module implementation process, and model training and optimization methods.

[0145] First, the end-to-end CSI feedback implementation process.

[0146] The end-to-end CSI feedback implementation process framework can be as follows: Figure 5 As shown, the specific process is as follows: Step 1: The user equipment estimates downlink channel state information based on the downlink reference signal transmitted by the base station. The downlink channel state information can be represented as a complex CSI matrix "H". Here, "H" represents the frequency domain channel information obtained by the user equipment across multiple subcarriers and multiple antenna dimensions.

[0147] Step 2: To facilitate neural network processing, the user equipment separates the real and imaginary parts of the complex CSI matrix "H" and concatenates them along a new dimension to form a real-valued CSI tensor, which serves as the input to the source-channel joint encoder. For example, if the original CSI matrix has a size of... Then both the real part and the imaginary part are After splicing, the resulting size is The real-valued CSI tensor.

[0148] Step 3: The user equipment inputs the real-valued CSI tensor into the source-channel joint encoder. The source-channel joint encoder can employ a Transformer network structure, or a convolutional neural network, a multilayer perceptron, a lightweight attention network, or a combination thereof. Preferably, the source-channel joint encoder includes a linear embedding layer, a positional coding layer, several Transformer coding layers, and an output projection layer, used to extract compressed features from the CSI tensor.

[0149] Step 4: The source-channel joint encoder extracts and compresses features from the input CSI, outputting a continuous CSI compressed feature vector z, which is the latent feature of the continuous CSI. This continuous CSI compressed feature vector is a real-valued vector, used to represent the information extracted and compressed by the encoder from the original CSI. Since this continuous CSI compressed feature vector is not a discrete bit sequence composed of 0s and 1s, it cannot be directly entered into the digital feedback link. Therefore, it needs to be further converted into a discrete feedback bit stream through a learnable codebook quantization module. It should be noted that the continuous CSI compressed feature vector is mainly used to retain information in the original CSI that is beneficial to subsequent CSI reconstruction. Therefore, in the context of task-oriented communication, it can also be understood as a continuous CSI semantic feature vector oriented towards the CSI reconstruction task. The CSI reconstruction goal refers to the base station reconstructing the CSI based on the feedback bit stream, making the reconstructed CSI as consistent as possible with the original CSI on the user equipment side.

[0150] Step 5: The user equipment inputs the continuous CSI compressed feature vector z into the learnable codebook quantization module. The learnable codebook quantization module performs codeword selection, codeword aggregation, and hard binarization operations based on the feature value or neighborhood features of each feature unit, generating a feedback bitstream b. The feedback bitstream is a discrete bit sequence composed of 0s and 1s, which can be directly used for subsequent digital modulation.

[0151] Step 6: The user equipment inputs the feedback bit stream b into the digital modulation module. The digital modulation module can use QAM modulation, or QPSK, PSK, or other digital modulation methods. After modulation, the feedback bit stream is mapped to a modulation symbol vector x.

[0152] Step 7: The user equipment maps the modulation symbol vector x to the OFDM resource grid to generate the OFDM feedback signal. The signal is then transmitted to the base station via a wireless feedback link. To ensure transmission stability, the OFDM signal can be power normalized before transmission.

[0153] Step 8: The base station (network) (NW) side receives the feedback signal y after transmission through the wireless channel. The feedback signal may be affected by factors such as multipath fading, additive noise, and channel estimation errors. The base station performs channel estimation, receiver-side channel processing, and linear equalization on the received signal to obtain the equalized modulation symbol.

[0154] Step 9: The base station inputs the equalized modulation symbols into the digital demodulation module to obtain the soft information corresponding to the feedback bits, i.e., the soft information sequence. Preferably, the digital demodulation module outputs a log-likelihood ratio (LLR) sequence (LLrs). Since the LLR sequence can preserve the soft reliability information of the feedback bits, it is used as the input to the source-channel joint decoder on the base station side. The LLR is used to indicate the reliability of whether the received bit is "0" or "1".

[0155] Step 10: The base station-side source-channel joint decoder reconstructs the CSI based on the LLR sequence and outputs the reconstructed CSI. The LLR sequence is used to represent the soft reliability information of each bit value in the feedback bitstream. The source-channel joint decoder can adopt a Transformer decoding structure corresponding to the encoder, or it can adopt a convolutional network, a multilayer perceptron, or a lightweight neural network structure. Preferably, the decoder includes an input projection layer, a position coding layer, several Transformer decoding layers, and an output projection layer, used to map the soft information obtained from the recovered feedback link to the reconstructed CSI. It should be noted that, in this embodiment, no independent codebook dequantization step is set. The base station-side source-channel joint decoder is a trained neural network decoder, which learns the mapping relationship from the LLR sequence to the reconstructed CSI through training, and therefore can output the reconstructed CSI based on the LLR sequence.

[0156] Second, the implementation process of the learnable codebook quantization module.

[0157] The learnable codebook quantization module is located between the source-channel joint encoder and the digital modulation module, and is used to convert continuous CSI compressed feature vectors into a transmittable feedback bit stream. Its specific process is as follows: Step 21: Let the continuous CSI compression feature vector output by the source-channel joint encoder be... Where N represents the number of feature units in the continuous CSI compressed feature vector. This represents the eigenvalue or eigenre representation corresponding to the nth eigenvalue location. It should be noted that the location of the eigenvalue refers to the sequence index position within the continuous CSI compressed eigenvector, and is not limited to the actual physical location, antenna location, or subcarrier location.

[0158] Step 22: For each feature unit in the continuous CSI compressed feature vector, a selector network is set up. The selector network generates a selection score for multiple candidate codewords corresponding to the current feature unit based on the feature value corresponding to the current feature unit, and may combine the feature values ​​of its neighboring feature units. Preferably, the selector network may include a one-dimensional convolutional layer, a hidden fully connected layer, and an output fully connected layer, wherein the one-dimensional convolutional layer is used to capture the local correlation between adjacent positions, and the fully connected layer is used to generate the matching score of the candidate codewords.

[0159] Step 23: For each feature unit in the continuous CSI compressed feature vector, set the corresponding learnable codebook. The codebook corresponding to the nth feature unit can be represented as: ,in, Let B represent the m-th candidate codeword corresponding to the n-th feature unit, and let B represent the number of quantization bits corresponding to each feature unit. Each candidate codeword is a soft bit vector of length B, where B represents the number of quantization bits corresponding to each feature position. Different feature units can correspond to different learnable codebooks, thus different positions can learn different candidate codewords based on their own feature distribution. The learnable codebook is not a manually fixed codebook, but is updated as trainable parameters during model training; after model training is completed, the codebook parameters are fixed and used for feedback bit generation during the inference phase.

[0160] A learnable codebook can be represented by trainable codebook parameters, and the codebook parameters can be mapped to the [0,1] interval using the Sigmoid function, so that each candidate codeword forms a learnable soft bit prototype. Here, the codebook parameters are numerical values ​​that are not limited to the range [0,1].

[0161] Step 24: For the nth feature unit, the selector network outputs a codeword selection score vector. Where M represents the number of candidate codewords, This represents the matching score when the nth feature unit selects the mth candidate codeword. The selection score is not the final discrete selection result, but is used to calculate the probability of each candidate codeword being selected in subsequent calculations.

[0162] Step 25: During the training phase, if a maximum value selection or hard decision method is directly used to select a codeword from multiple candidate codewords, the discrete selection process is usually non-differentiable, making it difficult to effectively update the selector network and learnable codebook parameters due to training loss. Therefore, this application adopts a differentiable soft selection method, calculating the soft selection probability of each candidate codeword based on the selection score. The soft selection probability is denoted as... denoted as , representing the soft selection probability of the nth feature unit choosing the mth candidate codeword. The soft selection probability can be obtained through Softmax, Gumbel-Softmax, or other differentiable discrete selection approximation methods.

[0163] Step 26: Based on the soft selection probability We perform weighted aggregation on the candidate codewords corresponding to the nth feature unit to obtain the soft bit representation at that position: ; because The soft selection probabilities are continuous, and the candidate codewords are... Since it is a soft bit vector with values ​​in the interval [0,1], the weighted aggregation still results in a continuous soft bit representation, rather than the final 0 / 1 binary bits. That is, it is a vector of length B.

[0164] Step 27: To meet the requirements of bit stream transmission in actual digital feedback links, the soft bit representation is hard-binarized, mapping continuous values ​​to 0 or 1 to obtain the binary bit representation corresponding to the feature unit. .

[0165] in, This indicates element-wise rounding (hard binarization). To address the issue of non-differentiability of hard binarization, this application employs a pass-through estimator (STE) during training. This ensures that the forward propagation outputs hard binary bits, while the gradient is propagated through the soft bit representation before hard binarization during backpropagation. This maintains effective gradient propagation between the encoder, selector network, learnable codebook, and decoder.

[0166] Step 28: After performing the codeword selection, soft aggregation, and hard binarization operations on all feature units, concatenate the binary bit representations corresponding to each position to obtain the final feedback bitstream b. If the number of feature units is N, and each unit corresponds to B bits, then the length of the final feedback bitstream is... .

[0167] Step 29: In the inference phase, instead of using a soft selection method with randomness, the candidate codeword with the highest score is directly selected based on the codeword selection score output by the selector network. Specifically, for each feature unit, the selector network outputs the selection score for multiple candidate codewords corresponding to that feature unit, and the user equipment determines a target candidate codeword from the multiple candidate codewords based on the selection score.

[0168] Based on the target candidate codeword, hard binarization is performed to generate the discrete binary bit vector corresponding to the feature unit. It should be noted that the output of each feature unit is not a single bit, but a binary bit vector of length B. After performing the above operation on all N feature units in the continuous CSI compressed feature vector, the binary bit vectors corresponding to each feature unit are concatenated to obtain the final feedback bitstream. If each feature unit corresponds to B bits, then the length of the final feedback bitstream is... This feedback bitstream consists of 0s and 1s and can be directly input into subsequent digital modulation modules.

[0169] In this embodiment of the application, the structure of the learnable codebook quantization module is as follows: Figure 6 As shown. CSI(z) is processed by the codeword selection module to obtain the soft selection probability. The soft selection probability and the learnable codebook are input into the codebook aggregation module, which aggregates them to obtain the soft bit representation. The soft bit representation is then hard quantized by the hard quantization module to obtain the feedback bit stream (b).

[0170] Third, model training and optimization methods To enable collaborative optimization of the source-channel joint encoder, learnable codebook quantization module, and source-channel joint decoder, this invention employs an end-to-end joint training approach. During the training phase, the user equipment-side source-channel joint encoder generates a continuous CSI compressed feature vector based on the input CSI. The learnable codebook quantization module generates a feedback bitstream through differentiable soft selection, soft codeword aggregation, and hard binarization operations based on a pass-through estimator. This feedback bitstream is transmitted via a digital feedback link and then used by the base station-side source-channel joint decoder to reconstruct the CSI.

[0171] During training, based on the original CSI matrix "H" and the reconstructed CSI matrix... The similarity between the original and reconstructed channel vectors is used to construct the CSI reconstruction loss. Preferably, the squared generalized cosine similarity correlation loss can be used to improve the spatial consistency between the reconstructed CSI and the original CSI. For the i-th subcarrier, the squared generalized cosine similarity between the original channel vector and the reconstructed channel vector can be used to measure their directional consistency, and further averaged over multiple subcarriers as the evaluation criterion for CSI reconstruction performance.

[0172] Furthermore, to avoid the selector network consistently selecting only a few codewords, leading to insufficient codebook utilization or codeword collapse, this invention introduces a codebook diversity regularization term. This term is calculated based on the average selection probability of different candidate codewords in the training batch, and is used to encourage the full utilization of multiple candidate codewords, thereby improving codebook expressive power and the representation efficiency of the feedback bitstream.

[0173] The final training loss is a weighted sum of the CSI reconstruction loss and the codebook diversity regularization term. The parameters of the source-channel joint encoder, selector network, learnable codebook, and source-channel joint decoder are updated using the backpropagation algorithm. After training, the source-channel joint encoder and learnable codebook quantization module are deployed on the user equipment side, while the corresponding receiver channel processing module, digital demodulation module, and source-channel joint decoder are deployed on the base station side for actual CSI feedback and CSI reconstruction.

[0174] Through the above technical solution, the present invention can convert the continuous CSI compressed feature vector output by the neural network on the user equipment side into a discrete feedback bit stream that can be directly entered into the digital modulation link. Through the position-level learnable codebook, selector network, soft aggregation and hard binarization mechanism of pass-through estimator, the end-to-end joint optimization between CSI feature extraction, feedback bit generation, wireless transmission and base station-side CSI reconstruction can be achieved.

[0175] To verify the effectiveness of this invention under limited feedback resources and different channel conditions, this paper conducts experimental analysis on the proposed JSCC+ learnable codebook quantization scheme from two aspects: CSI reconstruction performance and codebook utilization. The results are as follows: Figures 7 to 9 As shown.

[0176] from Figure 7 It can be seen that, under conditions of imperfect channel estimation, the JSCC+learnable codebook quantization scheme proposed in this invention achieves good CSI reconstruction performance across different signal-to-noise ratio (SNR) ranges. Compared to JSCC+uniform quantization, JSCC+vector quantization, and traditional discrete feedback schemes, this invention shows a more significant advantage in the low SNR region. This indicates that this invention, through joint source-channel coding and learnable codebook quantization, can improve the robustness of the feedback link to noise, maintaining high CSI reconstruction accuracy even under conditions of limited feedback resources and poor channel conditions.

[0177] from Figure 8 As can be seen, under ideal channel estimation conditions, the overall performance of various CSI feedback schemes is improved, and the JSCC+ learnable codebook quantization scheme proposed in this invention still maintains superior performance over most signal-to-noise ratio (SNR) ranges. Compared with fixed uniform quantization and traditional vector quantization, this invention can adaptively generate feedback bitstreams based on CSI feature distribution, enabling finite feedback bits to more effectively represent the original CSI information. Therefore, this invention is not only suitable for low SNR or non-ideal channel conditions, but also has a relatively stable CSI reconstruction capability under more ideal feedback link conditions.

[0178] from Figure 9 As can be seen, the normalized codebook entropy of this invention remains at a high level throughout the training process, indicating that multiple candidate codewords in the learnable codebook can be effectively utilized without significant codeword collapse. Compared to a fixed uniform quantization scheme, this invention can adaptively adjust codeword usage according to the training process; compared to traditional vector quantization schemes, the codebook usage of this invention is more aligned with the feedback bitstream generation and CSI reconstruction goals. This demonstrates that this invention can improve codebook utilization and feedback bit representation capabilities, thereby enhancing model training stability and CSI feedback performance.

[0179] The information processing method provided in this application can be applied to CSI feedback scenarios of FDD-OFDM multi-antenna base stations, CSI feedback scenarios with limited feedback overhead and learnable codebook quantization.

[0180] Example 1: CSI Feedback Scenario for FDD-OFDM Multi-Antenna Base Station In FDD-OFDM multi-antenna communication systems, the base station needs to acquire downlink channel state information from user equipment (UE) for precoding, beamforming, and resource scheduling. Since uplink and downlink channels lack strict reciprocity, UE typically needs to feed back the estimated downlink CSI to the base station. However, in wideband multi-antenna systems, the CSI dimension is high; directly feeding back the complete CSI would consume significant uplink feedback resources. Deploying this invention in this scenario allows the estimated CSI from the UE side to be compressed and converted into a transmittable discrete feedback bitstream, thereby reducing feedback overhead and improving the CSI reconstruction performance on the base station side.

[0181] During the deployment phase: The user equipment (UE) side deploys a source-channel joint encoder and a learnable codebook quantization module. The source-channel joint encoder is used to extract and compress features from the downlink CSI estimated by the UE, outputting continuous features; the learnable codebook quantization module is used to convert the continuous feature vector into a feedback bit stream composed of 0s and 1s.

[0182] The base station deploys a receiver channel processing module, a digital demodulation module, and a source-channel joint decoder. The receiver channel processing module performs channel estimation and equalization on the received feedback signal, the digital demodulation module obtains the soft information corresponding to the feedback bits, and the source-channel joint decoder reconstructs the downlink CSI based on the soft information. During the debugging and operation phase: Debugging and running phase: 1) CSI Acquisition: The user equipment estimates the downlink CSI based on the downlink reference signal sent by the base station, and separates the real and imaginary parts of the complex CSI matrix to form a real-valued CSI tensor.

[0183] 2) CSI feature extraction and compression: The user equipment inputs the real-valued CSI tensor into the source-channel joint encoder to obtain a continuous CSI compressed feature vector.

[0184] 3) Codebook Quantization and Bit Generation: The user equipment inputs the continuous CSI compressed feature vector into the learnable codebook quantization module. During the training phase, differentiable training can be achieved through soft selection, weighted aggregation of candidate codewords, and hard binarization. During the inference phase, the target candidate codeword can be directly selected based on the selection score output by the selector network, and then hard binarized to generate a discrete binary bit vector. The binary bit vectors corresponding to all feature units are concatenated to obtain the final feedback bitstream. If the number of feature units is N, and each feature unit corresponds to B bits, then the length of the final feedback bitstream is... Therefore, the feedback overhead is fixed and configurable.

[0185] 4) Feedback transmission: The user equipment inputs the feedback bit stream into the digital modulation module to generate modulation symbols, and after mapping the modulation symbols to the OFDM resource grid, it sends them to the base station through the uplink feedback link.

[0186] 5) CSI Reconstruction: After receiving the feedback signal, the base station performs channel processing, equalization, and digital demodulation to obtain the LLR soft information corresponding to the feedback bits. This LLR soft information is then input into the source-channel joint decoder to reconstruct the downlink CSI. The reconstructed CSI can be used for subsequent precoding, beamforming, and resource scheduling by the base station.

[0187] Example 2: Learnable Codebook Quantization CSI Feedback Scenario with Limited Feedback Overhead In practical FDD-OFDM communication systems, uplink feedback resources are often limited, and user equipment cannot directly transmit complete CSI. If fixed uniform quantization or traditional vector quantization is used, the quantization results are difficult to adaptively adjust according to the CSI feature distribution, easily leading to insufficient utilization of limited feedback bits. Deploying this invention in scenarios with limited feedback overhead allows the use of a learnable codebook quantization module to convert continuous CSI compressed features into a fixed-length discrete feedback bit stream, thereby improving the expressive power of the feedback bits and the CSI reconstruction performance on the base station side.

[0188] During the deployment phase: 1. Deployment Phase: The user equipment side deploys a source-channel joint encoder, a selector network, a position-level learnable codebook, and a hard binarization module. The source-channel joint encoder is used to compress the input CSI into a continuous CSI compressed feature vector; the selector network is used to generate candidate codeword selection scores for each feature unit; the position-level learnable codebook is used to provide candidate soft bit representations for different units; and the hard binarization module is used to convert the soft bit representations into feedback bits composed of 0s and 1s.

[0189] The base station side deploys a receiver channel processing module, a digital demodulation module, and a source-channel joint decoder. The digital demodulation module outputs LLR soft information corresponding to the feedback bits, and the source-channel joint decoder reconstructs the CSI based on this LLR soft information.

[0190] During the debugging and operation phase: 1) CSI compression: The user equipment uses the estimated downlink CSI input source channel to jointly encode and obtain continuous feature units.

[0191] 2) Codeword Selection: For each feature unit in the continuous CSI compressed feature vector, the selector network outputs the selection score for the feature unit corresponding to multiple candidate codewords. During the training phase, the soft selection probability of the candidate codeword is obtained based on the selection score; during the inference phase, the target candidate codeword is determined based on the selection score.

[0192] 3) Codebook Quantization: Each feature unit corresponds to a learnable codebook, and each candidate codeword in the codebook represents a soft bit vector. During the training phase, candidate codewords can be weighted and aggregated according to their soft selection probabilities to obtain continuous soft bit representations and maintain gradient propagation; during the inference phase, the target candidate codeword can be directly determined based on the selection score output by the selector network.

[0193] 4) Bitstream generation: Hard binarize the soft bit representation to obtain the binary bits corresponding to the feature unit. Concatenate the binary bits of all feature units sequentially to form a fixed-length feedback bitstream.

[0194] 5) Feedback and Reconstruction: The feedback bitstream is digitally modulated and OFDM resource mapped before being sent to the base station. Upon receiving the feedback bitstream, the base station performs channel processing, equalization, and digital demodulation to obtain LLR soft information, which is then input into the source-channel joint decoder to reconstruct the CSI. The reconstructed CSI can be used for subsequent precoding, beamforming, and resource scheduling at the base station.

[0195] Thirdly, to implement the information processing method, embodiments of this application provide a terminal device, including: The coding unit is used to perform joint source-channel coding on the first channel state information (CSI) to obtain a first CSI feature vector, which includes multiple first feature units. The selection unit is used to determine the first codebook corresponding to the first feature unit based on the position index of the first feature unit and multiple codebooks, wherein the multiple codebooks are obtained based on the training of multiple second CSIs; The first determining unit is used to determine the first bit sequence corresponding to the first CSI feature vector based on the first codebook corresponding to the first feature unit in the first CSI feature vector. The second determining unit is used to determine the first OFDM feedback signal corresponding to the first CSI by modulation of the first bit sequence; The transmitting unit is used to send the first OFDM feedback signal to the network device.

[0196] In some embodiments, the codebook is obtained by mapping codebook parameters to a first numerical interval, the lower boundary of the first numerical interval being 0 and the upper boundary being 1, respectively; the codebook parameters are updated as training parameters based on the training of multiple second CSIs.

[0197] In some embodiments, the first determining unit is further configured to: determine the first target codeword corresponding to the first feature unit according to the first codebook corresponding to the first feature unit in the first CSI feature vector, thereby obtaining a plurality of first target codewords, wherein the first target codeword is a soft bit vector; and convert the plurality of first target codewords into a first bit sequence.

[0198] In some embodiments, the first determining unit is further configured to: determine a first codeword selection scoring vector corresponding to the first feature unit based on the first CSI feature vector; the first codeword selection scoring vector includes multiple selection scores corresponding to different codewords; and determine a first target codeword corresponding to the first feature unit based on the first codeword selection scoring vector corresponding to the first feature unit and the first codebook corresponding to the first feature unit.

[0199] In some embodiments, the first determining unit is further configured to: determine a first codeword selection scoring vector corresponding to the first feature unit based on the first feature unit and neighboring feature units adjacent to the first feature unit.

[0200] In some embodiments, the first determining unit is further configured to: input the first feature unit and the neighboring feature units adjacent to the first feature unit into the selector to obtain the first codeword selection score vector output by the selector; the selector is obtained based on training of multiple second CSIs.

[0201] In some embodiments, the first determining unit is further configured to: determine a first target codeword corresponding to the first feature unit from among a plurality of codewords included in the first codebook corresponding to the first feature unit according to the first codeword selection scoring vector corresponding to the first feature unit, wherein the first target codeword is the codeword with the highest selection score in the first codebook.

[0202] In some embodiments, the first determining unit is further configured to: determine a first soft selection probability vector corresponding to the first feature unit based on the first codeword selection scoring vector corresponding to the first feature unit, wherein the first soft selection probability vector includes multiple soft selection probabilities; and perform weighted aggregation on multiple codewords included in the first codebook using the multiple soft selection probabilities to obtain a first target codeword corresponding to the first feature unit.

[0203] In some embodiments, the first determining unit is further configured to: perform hard quantization on a plurality of first target codewords to obtain a plurality of first bit vectors; and concatenate the plurality of first bit vectors to obtain a first bit sequence.

[0204] In some embodiments, the second determining unit is further configured to: digitally modulate the first bit sequence to obtain a first modulation symbol vector; and perform resource mapping on the first modulation symbol vector to generate a first OFDM feedback signal.

[0205] In some embodiments, the encoding unit is further configured to: separate the real and imaginary parts of the first CSI to obtain real data and imaginary data; concatenate the real data and imaginary data to obtain a first real-valued CSI tensor; input the first real-valued CSI tensor into the joint source-channel encoder to obtain the first CSI feature vector output by the joint source-channel encoder; the joint source-channel encoder is obtained based on training of multiple second CSIs.

[0206] In some embodiments, multiple second CSIs are used to train multiple codebooks and at least one of the following: The joint source-channel encoder is used for joint source-channel coding of CSI to obtain CSI feature vectors; The selector is used to determine the codeword selection score vector corresponding to the feature units included in the CSI feature vector. Joint source-channel decoder, used for CSI reconstruction.

[0207] In some embodiments, the second CSI is jointly source-channel coded to obtain a second CSI feature vector, which includes multiple second feature units; the second codeword selection scoring vector corresponding to the second feature unit is used to determine the second soft-select probability vector corresponding to the second feature unit; the second codebook and the second soft-select probability vector corresponding to the second feature unit are used to determine the second target codeword corresponding to the second feature vector; the second target codeword corresponding to the second feature unit in the second CSI feature vector is used to determine the second bit sequence, and the second bit sequence is digitally modulated to obtain the second OFDM feedback signal corresponding to the second CSI; the second OFDM feedback signal is reconstructed into the third CSI after being transmitted through the wireless channel.

[0208] In some embodiments, the second CSI and the third CSI are used to determine the reconstruction loss; the second soft selection probability vector corresponding to each second feature unit in the second CSI feature vector is used to determine the diversity loss; the reconstruction loss and the diversity loss are used to determine the training loss, which is used to update multiple codebooks and one or more of the following: a joint source-channel encoder, a selector, and a joint source-channel decoder.

[0209] In some embodiments, the diversity loss is determined based on a third soft-selection probability vector corresponding to a plurality of second feature units in the second CSI feature vector, the third soft-selection probability vector including the average selection probability determined based on the second soft-selection probability vector.

[0210] Fourthly, to implement the above information processing method, an embodiment of this application provides a device 1000 (such as a terminal device), such as... Figure 10 As shown, it may include at least one processor 1001 and at least one transceiver 1002 coupled to at least one processor 1001. The transceiver 1002 may include at least one separate receiving circuitry and a transmitting circuitry, or at least one integrated receiving circuitry and a transmitting circuitry. The at least one processor 1001 may be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field-programmable gate array (FPGA), etc.

[0211] According to some embodiments of this application, the processor in the terminal device is configured to: perform joint source-channel coding on the first channel state information (CSI) to obtain a first CSI feature vector, the first CSI feature vector including multiple first feature units; determine a first codebook corresponding to the first feature unit based on the position index of the first feature unit and multiple codebooks, the multiple codebooks being trained based on multiple second CSIs; determine a first bit sequence corresponding to the first CSI feature vector based on the first codebook corresponding to the first feature unit in the first CSI feature vector; determine a first orthogonal frequency division multiplexing (OFDM) feedback signal corresponding to the first CSI by modulating the first bit sequence; and send the first OFDM feedback signal to the network device via a transceiver.

[0212] In some embodiments, the codebook is obtained by mapping codebook parameters to a first numerical interval, the lower boundary of the first numerical interval being 0 and the upper boundary being 1, respectively; the codebook parameters are updated as training parameters based on the training of multiple second CSIs.

[0213] In some embodiments, the processor determines a first bit sequence corresponding to the first CSI feature vector based on a first codebook corresponding to a first feature unit in the first CSI feature vector, and is configured to: determine a first target codeword corresponding to the first feature unit based on the first codebook corresponding to the first feature unit in the first CSI feature vector, thereby obtaining a plurality of first target codewords, wherein the first target codewords are soft bit vectors; and convert the plurality of first target codewords into a first bit sequence.

[0214] In some embodiments, the processor determines the first target codeword corresponding to the first feature unit based on the first codebook corresponding to the first feature unit in the first CSI feature vector, and is configured to: determine the first codeword selection score vector corresponding to the first feature unit based on the first CSI feature vector; the first codeword selection score vector includes multiple selection scores corresponding to different codewords; and determine the first target codeword corresponding to the first feature unit based on the first codeword selection score vector corresponding to the first feature unit and the first codebook corresponding to the first feature unit.

[0215] In some embodiments, the processor determines a first codeword selection score vector corresponding to a first feature unit based on a first CSI feature vector, and is configured to: determine the first codeword selection score vector corresponding to the first feature unit based on the first feature unit and neighboring feature units adjacent to the first feature unit.

[0216] In some embodiments, the processor determines a first codeword selection score vector corresponding to a first feature unit based on a first feature unit and neighboring feature units adjacent to the first feature unit, and is configured to: input the first feature unit and neighboring feature units adjacent to the first feature unit to a selector to obtain a first codeword selection score vector output by the selector; the selector is obtained based on training of multiple second CSIs.

[0217] In some embodiments, the processor selects a scoring vector based on the first codeword corresponding to the first feature unit and a first codebook corresponding to the first feature unit to determine a first target codeword corresponding to the first feature unit. The processor is configured to: select a scoring vector based on the first codeword corresponding to the first feature unit, and determine a first target codeword corresponding to the first feature unit from multiple codewords included in the first codebook corresponding to the first feature unit. The first target codeword is the codeword with the highest selection score in the first codebook.

[0218] In some embodiments, the processor selects a scoring vector based on the first codeword corresponding to the first feature unit and a first codebook corresponding to the first feature unit to determine the first target codeword corresponding to the first feature unit. The processor is configured to: select a scoring vector based on the first codeword corresponding to the first feature unit to determine a first soft selection probability vector corresponding to the first feature unit, the first soft selection probability vector including multiple soft selection probabilities; and perform weighted aggregation on multiple codewords included in the first codebook using multiple soft selection probabilities to obtain the first target codeword corresponding to the first feature unit.

[0219] In some embodiments, the processor converts multiple first target codewords into a first bit sequence, and is configured to: hard quantize the multiple first target codewords to obtain multiple first bit vectors; and concatenate the multiple first bit vectors to obtain a first bit sequence.

[0220] In some embodiments, the processor determines the first OFDM feedback signal corresponding to the first CSI by modulating the first bit sequence, and is configured to: digitally modulate the first bit sequence to obtain a first modulation symbol vector; and perform resource mapping on the first modulation symbol vector to generate the first OFDM feedback signal.

[0221] In some embodiments, the processor performs joint source-channel coding on the first CSI to obtain a first CSI feature vector, which is then configured to: separate the real and imaginary parts of the first CSI to obtain real data and imaginary data; concatenate the real data and imaginary data to obtain a first real-valued CSI tensor; input the first real-valued CSI tensor into the joint source-channel encoder to obtain the first CSI feature vector output by the joint source-channel encoder; the joint source-channel encoder is obtained based on training of multiple second CSIs.

[0222] In some embodiments, a plurality of second CSIs are used to train a plurality of codebooks and at least one of the following: a joint source-channel encoder, which is used to obtain a CSI feature vector by joint source-channel coding of the CSI; a selector, which is used to determine the codeword selection score vector corresponding to the feature units included in the CSI feature vector; and a joint source-channel decoder, which is used to reconstruct the CSI.

[0223] In some embodiments, the second CSI is jointly source-channel coded to obtain a second CSI feature vector, which includes multiple second feature units; the second codeword selection scoring vector corresponding to the second feature unit is used to determine the second soft-select probability vector corresponding to the second feature unit; the second codebook and the second soft-select probability vector corresponding to the second feature unit are used to determine the second target codeword corresponding to the second feature vector; the second target codeword corresponding to the second feature unit in the second CSI feature vector is used to determine the second bit sequence, and the second bit sequence is digitally modulated to obtain the second OFDM feedback signal corresponding to the second CSI; the second OFDM feedback signal is reconstructed into the third CSI after being transmitted through the wireless channel.

[0224] In some embodiments, the second CSI and the third CSI are used to determine the reconstruction loss; the second soft selection probability vector corresponding to each second feature unit in the second CSI feature vector is used to determine the diversity loss; the reconstruction loss and the diversity loss are used to determine the training loss, which is used to update multiple codebooks and one or more of the following: a joint source-channel encoder, a selector, and a joint source-channel decoder.

[0225] In some embodiments, the diversity loss is determined based on a third soft-selection probability vector corresponding to a plurality of second feature units in the second CSI feature vector, the third soft-selection probability vector including the average selection probability determined based on the second soft-selection probability vector.

[0226] The descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0227] It should be noted that, in the embodiments of this application, if the above-described information processing method is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0228] Fifthly, to implement the above information processing method, embodiments of this application provide an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor, and when the processor executes the program, it implements the steps in the information processing method provided in the above embodiments.

[0229] Sixthly, embodiments of this application provide a storage medium, namely a computer-readable storage medium, on which a computer program is stored, which, when executed by a processor, implements the steps in the information processing method provided in the above embodiments.

[0230] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0231] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0232] 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.

[0233] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0234] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0235] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0236] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0237] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0238] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A terminal device, the terminal device comprising a transceiver; and A processor coupled to the transceiver; the processor is configured to: Joint source-channel coding is performed on the first channel state information (CSI) to obtain a first CSI feature vector, which includes multiple first feature units. Based on the position index of the first feature unit and multiple codebooks, a first codebook corresponding to the first feature unit is determined, wherein the multiple codebooks are obtained based on training of multiple second CSIs; Based on the first codebook corresponding to the first feature unit in the first CSI feature vector, determine the first bit sequence corresponding to the first CSI feature vector; The first orthogonal frequency division multiplexing (OFDM) feedback signal corresponding to the first CSI is determined by the modulation of the first bit sequence. The first OFDM feedback signal is sent to the network device via the transceiver.

2. The device according to claim 1, wherein the codebook is obtained by mapping codebook parameters to a first numerical interval, wherein the lower boundary and the upper boundary of the first numerical interval are 0 and 1, respectively; The codebook parameters are updated as training parameters based on the training of the multiple second CSIs.

3. The device according to claim 1, wherein the processor determines the first bit sequence corresponding to the first CSI feature vector based on the first codebook corresponding to the first feature unit in the first CSI feature vector, and configures it to: Based on the first codebook corresponding to the first feature unit in the first CSI feature vector, the first target codeword corresponding to the first feature unit is determined, and multiple first target codewords are obtained. The first target codeword is a soft bit vector. The plurality of first target codewords are converted into the first bit sequence.

4. The device according to claim 3, wherein the processor determines the first target codeword corresponding to the first feature unit based on the first codebook corresponding to the first feature unit in the first CSI feature vector, and configures it to: Based on the first CSI feature vector, a first codeword selection score vector corresponding to the first feature unit is determined; the first codeword selection score vector includes multiple selection scores corresponding to different codewords. The first target codeword corresponding to the first feature unit is determined by selecting the scoring vector based on the first codeword corresponding to the first feature unit and the first codebook corresponding to the first feature unit.

5. The device according to claim 3, wherein the processor, when converting the plurality of first target codewords into the first bit sequence, is configured to: Hard quantize the plurality of first target codewords to obtain a plurality of first bit vectors; The plurality of first bit vectors are concatenated to obtain the first bit sequence.

6. The device according to claim 1, wherein the processor, after modulating the first bit sequence to determine the first OFDM feedback signal corresponding to the first CSI, is configured to: The first bit sequence is digitally modulated to obtain a first modulation symbol vector; Resource mapping is performed on the first modulation symbol vector to generate the first OFDM feedback signal.

7. The device according to claim 1, wherein the processor performs joint source-channel coding on the first CSI to obtain a first CSI feature vector, and configures it to: Separate the real and imaginary parts of the first CSI to obtain real data and imaginary data; The real part data and the imaginary part data are concatenated to obtain the first real-valued CSI tensor; The first real-valued CSI tensor is input into the joint source-channel encoder to obtain the first CSI feature vector output by the joint source-channel encoder. The joint source channel encoder is obtained based on the training of the multiple second CSIs.

8. The device according to any one of claims 1 to 7, wherein the plurality of second CSIs are used to train the plurality of codebooks and at least one of the following: A joint source-channel encoder, wherein the joint source-channel encoder is used for joint source-channel coding of CSI to obtain CSI feature vectors; A selector is used to determine the codeword selection score vector corresponding to the feature units included in the CSI feature vector; A joint source-channel decoder, which is used for CSI reconstruction.

9. The device according to claim 8, The second CSI is jointly source-channel coded to obtain a second CSI feature vector, which includes multiple second feature units; The second codeword selection scoring vector corresponding to the second feature unit is used to determine the second soft selection probability vector corresponding to the second feature unit; the second codebook and the second soft selection probability vector corresponding to the second feature unit are used to determine the second target codeword corresponding to the second feature vector. The second target codeword corresponding to the second feature unit in the second CSI feature vector is used to determine the second bit sequence, and the second bit sequence is modulated to obtain the second OFDM feedback signal corresponding to the second CSI; The second OFDM feedback signal is reconstructed into the third CSI after being transmitted through the wireless channel.

10. An information processing method, the method comprising: The terminal device performs joint source-channel coding on the first channel state information (CSI) to obtain a first CSI feature vector, which includes multiple first feature units. The terminal device determines the first codebook corresponding to the first feature unit based on the position index of the first feature unit and multiple codebooks, wherein the multiple codebooks are obtained based on training of multiple second CSIs; The terminal device determines the first bit sequence corresponding to the first CSI feature vector based on the first codebook corresponding to the first feature unit in the first CSI feature vector; The terminal device determines the first orthogonal frequency division multiplexing (OFDM) feedback signal corresponding to the first CSI by modulating the first bit sequence; Send the first OFDM feedback signal to the network device.

Images