WO2023077274A1

WO2023077274A1 - Csi feedback method and apparatus, and device and storage medium

Info

Publication number: WO2023077274A1
Application number: PCT/CN2021/128219
Authority: WO
Inventors: 刘文东; 田文强; 肖寒
Original assignee: Oppo广东移动通信有限公司
Priority date: 2021-11-02
Filing date: 2021-11-02
Publication date: 2023-05-11
Also published as: CN117751559A

Abstract

A CSI feedback method and apparatus, and a device and a storage medium, which relate to the field of mobile communications. The method comprises: encoding CSI by using an encoder (12), so as to obtain CSI feedback information, wherein the encoder (12) is obtained by training a real training set and a first supplementary training set, the first supplementary training set is generated by means of a generator in a generative adversarial network, and the generative adversarial network is obtained by performing training on the basis of the real training set; and sending the CSI feedback information to an access network device (20).

Description

CSI feedback method, device, equipment and storage medium

technical field

The present application relates to the field of mobile communications, in particular to a channel state information (Channel State Information, CSI) feedback method, device, device and storage medium.

Background technique

In the current New Radio (NR) system, for the CSI feedback scheme, the terminal usually adopts codebook-based eigenvector feedback, so that the base station can obtain the CSI of the downlink channel. Specifically, the base station sends downlink CSI reference signals (CSI Reference Signals, CSI-RS) to the user, and the terminal uses the CSI-RS to estimate the CSI of the downlink channel, and performs eigenvalue decomposition on the estimated downlink channel to obtain the corresponding eigenvectors of . Furthermore, NR provides Type 1 and Type 2 codebook design schemes, of which Type 1 codebook is used for conventional precision CSI feedback and single-user MIMO (Single-User MIMO, SU-MIMO) and multi-user MIMO (Multi- User MIMO, MU-MIMO) transmission, the Type 2 codebook is used to improve the transmission performance of MU-MIMO.

Contents of the invention

Embodiments of the present application provide a CSI feedback method, device, device, and storage medium, and propose a CSI feedback scheme based on an adversarial generative network. The technical scheme is as follows.

According to one aspect of the present application, a CSI feedback method is provided, which is applied to a terminal, and the method includes:

Using an encoder to encode the CSI to obtain CSI feedback information; the encoder is trained by the real training set and the first supplementary training set, the first supplementary training set is generated by the generator in the confrontation generation network, The confrontation generation network is trained based on the real training set;

Send the CSI feedback information to the access network device.

According to one aspect of the present application, a CSI feedback method is provided, which is applied to an access network device, and the method includes:

receiving CSI feedback information sent by the terminal, where the CSI feedback information is obtained by encoding the CSI by the terminal through an encoder;

Use a decoder to decode the CSI feedback information to obtain the CSI measured by the terminal; the encoder and the decoder are trained by the real training set and the first supplementary training set, and the first supplementary training set is generated by a generator in an adversarial generative network trained based on the real training set;

Send the CSI feedback information to the access network device.

According to one aspect of the present application, a CSI feedback device is provided, the device comprising:

The encoding module is used to encode the CSI using an encoder to obtain CSI feedback information; the encoder is obtained by training the real training set and the first supplementary training set, and the first supplementary training set is an adversarial generation network Generated by a generator, the confrontation generation network is trained based on the real training set;

A sending module, configured to send the CSI feedback information to the access network device.

A receiving module, configured to receive CSI feedback information sent by the terminal, where the CSI feedback information is obtained by encoding the CSI by the terminal through an encoder;

A decoding module, configured to use a decoder to decode the CSI feedback information to obtain the CSI measured by the terminal; the encoder and the decoder are obtained by training the real training set and the first supplementary training set, so The first supplementary training set is generated by a generator in the adversarial generation network, and the adversarial generation network is trained based on the real training set.

According to one aspect of the present application, a terminal is provided, and the terminal includes: a processor; a transceiver connected to the processor; a memory for storing executable instructions of the processor; wherein, the processing The device is configured to load and execute the executable instructions to implement the CSI feedback method as described in the above aspect.

According to one aspect of the present application, a network device is provided, and the network device includes: a processor; a transceiver connected to the processor; a memory for storing executable instructions of the processor; wherein, the The processor is configured to load and execute the executable instructions to implement the CSI feedback method as described in the above aspects.

According to one aspect of the present application, a computer-readable storage medium is provided, wherein executable instructions are stored in the computer-readable storage medium, and the executable instructions are loaded and executed by the processor to implement the above aspects. The CSI feedback method described above.

According to one aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program comprising computer instructions stored in a computer-readable storage medium readable by a processor of a computer device from a computer The storage medium reads the computer instruction, and the processor executes the computer instruction, so that the computer device executes the CSI feedback method described in the above aspect.

According to one aspect of the present application, a chip is provided, the chip includes a programmable logic circuit or a program, and the chip is used to implement the CSI feedback method as described in the above aspect.

The technical solutions provided by the embodiments of the present application at least include the following beneficial effects:

When there are few training samples in the real training set, use the adversarial generative network to generate a supplementary training set, so as to train an encoder and decoder with excellent performance. Using the encoder and decoder to complete the CSI feedback can improve the terminal performance. CSI feedback accuracy between network equipment and network equipment, and use less feedback data to represent more complete and detailed channel information.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.

FIG. 1 is an architecture diagram of a CSI feedback method provided by an exemplary embodiment of the present application;

FIG. 2 is a flowchart of a CSI feedback method provided by an exemplary embodiment of the present application;

Fig. 3 is a flowchart of the confrontation generating network provided by an exemplary embodiment of the present application;

Fig. 4 is a schematic diagram of the training of the confrontation generation network provided by an exemplary embodiment of the present application;

Fig. 5 is the flowchart of the training method of encoder and decoder that an exemplary embodiment of the present application provides;

Fig. 6 is a schematic diagram of training of an encoder and a decoder provided by an exemplary embodiment of the present application;

FIG. 7 is a flowchart of a CSI feedback method provided by an exemplary embodiment of the present application;

Fig. 8 is a flowchart of a model updating method provided by an exemplary embodiment of the present application;

Fig. 9 is a flowchart of a model updating method provided by an exemplary embodiment of the present application;

Fig. 10 is a flowchart of a model updating method provided by an exemplary embodiment of the present application;

Fig. 11 is a flowchart of a model updating method provided by an exemplary embodiment of the present application;

Fig. 12 is a flowchart of a model updating method provided by an exemplary embodiment of the present application;

Fig. 13 is a structural block diagram of a CSI feedback device provided by an exemplary embodiment of the present application;

Fig. 14 is a structural block diagram of a CSI feedback device provided by an exemplary embodiment of the present application;

Fig. 15 is a schematic structural diagram of a communication device provided by an exemplary embodiment of the present application.

Detailed ways

In order to make the purpose, technical solution and advantages of the present application clearer, the implementation manners of the present application will be further described in detail below in conjunction with the accompanying drawings.

Fig. 1 shows a schematic diagram of a mobile communication system provided by an embodiment of the present application. The mobile communication system may include: a terminal 10 and an access network device 20 .

The number of terminals 10 is generally multiple, and one or more terminals 10 may be distributed in a cell managed by each access network device 20 . The terminal 10 may include various handheld devices with mobile communication functions, vehicle-mounted devices, wearable devices, computing devices or other processing devices connected to a wireless modem, as well as various forms of user equipment (User Equipment, UE), mobile station ( Mobile Station, MS) and so on. For convenience of description, in the embodiment of the present application, the above-mentioned devices are collectively referred to as terminals.

The access network device 20 is a device deployed in an access network for providing mobile communication functions for the terminal 10 . The access network device 20 may include various forms of macro base stations, micro base stations, relay stations, access points, location management function entities (Location Management Function, LMF) and so on. In systems using different wireless access technologies, the names of devices with access network device functions may be different. For example, in 5G NR systems, they are called gNodeB or gNB. With the evolution of communication technology, the name "access network equipment" may change. For the convenience of description, in the embodiment of the present application, the above-mentioned devices that provide mobile communication functions for the terminal 10 are collectively referred to as access network devices. A connection may be established between the access network device 20 and the terminal 10 through an air interface, so as to perform communication through the connection, including signaling and data interaction. The number of access network devices 20 may be multiple, and two adjacent access network devices 20 may also communicate in a wired or wireless manner. The terminal 10 can switch between different access network devices 20 , that is, establish connections with different access network devices 20 .

The "5G NR system" in the embodiments of the present disclosure may also be called a 5G system or an NR system, but those skilled in the art can understand its meaning. The technical solution described in the embodiments of the present disclosure can be applied to the 5G NR system, and can also be applied to the subsequent evolution system of the 5G NR system.

In the embodiment of the present application, the terminal 10 is provided with an encoder 12 , and the access network device 20 is provided with a decoder 22 . The access network device 120 sends the CSI-RS to the terminal 10 on a downlink channel. Based on the CSI-RS, the terminal 10 measures and obtains the CSI of the downlink channel. The terminal 10 encodes the CSI through the encoder 12 to obtain CSI feedback information. The terminal 10 reports the CSI feedback information to the access network device 20 . The access network device 20 decodes through the decoder 22 to obtain the CSI of the terminal 10 .

Since the encoder 12 and the decoder 22 are models based on artificial intelligence, they need to be pre-trained using real training samples. However, since the training samples in the real training set are relatively small, this application also proposes a sample supplement scheme based on Generative Adversarial Networks (GAN), which can supplement a sufficient number of training samples, and the supplemented training samples are consistent with real The similarity of the training samples is very high.

Fig. 2 shows a flowchart of a CSI feedback method provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to the terminal 10 and the network device 20 shown in FIG. 1 as an example. The method includes:

Step 202: the terminal uses an encoder to encode the CSI to obtain CSI feedback information;

The encoder is an AI encoding model for encoding CSI into CSI feedback information. The access network device sends the CSI-RS to the terminal on the downlink channel, and the CSI is obtained by the terminal after measuring the CSI-RS.

The CSI feedback information is a feedback bit sequence or a feedback codebook obtained after the encoder encodes the CSI. The terminal uses an encoder to encode or compress the CSI to obtain CSI feedback information. That is, the AI coding model has a nonlinear fitting capability, and the CSI is compressed and fed back using the nonlinear fitting capability. Encoders are also called channel encoders.

Schematically, the CSI feedback information is at least one of a feedback codebook, a feature vector, a matrix, and a bit sequence.

Step 204: the terminal sends CSI feedback information to the access network device;

The terminal sends CSI feedback information to the access network device through the uplink feedback channel. The uplink feedback channel may be a physical uplink control channel (Physical Uplink Control Channel, PUCCH), and the uplink feedback channel may also be a physical uplink shared channel (Physical Uplink Shared Channel, PUSCH).

Step 206: The access network device receives the CSI feedback information sent by the terminal, and the CSI feedback information is obtained by the terminal encoding the CSI through an encoder;

The access network device receives the CSI feedback information sent by the terminal through the uplink feedback channel.

Step 208: the access network device uses a decoder to decode the CSI feedback information to obtain the CSI measured by the terminal.

Wherein, the encoder and the decoder are trained by the real training set and the first supplementary training set, the first supplementary training set is generated by the generator in the adversarial generative network, and the adversarial generative network is trained based on the real training set. Decoders are also called channel decoders. The encoder and decoder can be collectively referred to as a CSI autoencoder.

The access network device uses the decoder to decode or reconstruct the CSI feedback information to obtain the CSI of the downlink channel measured by the terminal.

Schematically, the neural network structure of the generator and discriminator can use deep neural network (Deep Neural Networks, DNN), convolutional neural network (Convolutional Neural Network, CNN), long short-term memory (Long short-term memory, LSTM ), Gated Recurrent Unit (Gate Recurrent Unit, GRU), Recurrent Neural Network (Recurrent Neural Network, RNN) and at least one of other possible neural network architectures, the specific network of generator and discriminator in this embodiment Architecture is not limited.

To sum up, the method provided in this embodiment uses the adversarial generative network to generate a supplementary training set when there are few training samples in the real training set, so as to train an encoder and decoder with excellent performance. The CSI feedback is completed by the decoder and the decoder, which can improve the accuracy of the CSI feedback between the terminal and the network device, and use less feedback data to represent more complete and detailed channel information.

It should be noted that step 202 and step 204 in the embodiment of FIG. 2 can be independently implemented as a CSI feedback method on the terminal side, and steps 206 and 208 in the embodiment of FIG. 2 can be independently implemented as a CSI feedback method on the access network device side. method. Similarly, the steps performed by the terminal in other embodiments may be individually implemented as corresponding methods on the terminal side, and the steps performed by the access network device in other embodiments may be independently implemented as corresponding methods on the access network device side.

The training process of the adversarial generative network:

Fig. 3 shows a flow chart of a method for training an adversarial generative network provided by an exemplary embodiment of the present application. The method may be performed by an access network device or a terminal or other devices, and the method includes:

The confrontation generation network includes: Generator Neural Network and Discriminator Neural Network. Adversarial Generative Networks are also known as Generative Adversarial Networks. The generator neural network is referred to as the generator, and the discriminator neural network is referred to as the discriminator.

Inspired by the zero-sum game in game theory, GAN regards the generation problem as a confrontation and game between the two networks of the discriminator and the generator: the generator generates a composite from a given noise (generally refers to a uniform distribution or a normal distribution). data, the discriminator distinguishes between the output of the generator and the real data. The former tries to produce data that is closer to the real, and the latter, in turn, tries to more perfectly distinguish between real and generated data. As a result, the two networks progress in the confrontation, and continue to confront after progress, and the data obtained by the generative network will become more and more perfect, approaching the real data, so that the desired data can be generated.

Schematically, the neural network structure of the generator and the discriminator can use at least one of DNN, CNN, LSTM, GRU, RNN and any other possible neural network architectures. The specific details of the generator and the discriminator in this embodiment The network architecture is not limited.

Step 302: Input the training samples in the real training set into the discriminator to obtain the first discriminant result;

The training samples in the real training set include at least one of the following:

·CSI

· CSI and CSI feedback information appearing in pairs (for example, a feedback codebook obtained by high-precision quantization).

The present application does not limit the form of the training samples, and this embodiment is described by taking the training samples in the real training set as CSI.

Input the training samples in the real training set into the discriminator to obtain the first discriminant result. Schematically, the first discrimination result may be 0 or 1, 0 representing false, and 1 representing true. Alternatively, the first discrimination result may be a probability value in the form of a percentage, which is used to indicate the probability that the discrimination result is true. For example, 80% means that the probability of the discrimination result being true is 80%, and it can be regarded as true if it exceeds the threshold of 50%. 50% is considered false.

In this embodiment, the discriminator may also be called a channel discriminator.

Step 304: Input the noise signal into the generator to obtain supplementary training samples;

The noise signal may be random noise, such as a noise signal conforming to a Gaussian distribution, or a noise signal conforming to a uniform distribution, a noise signal conforming to a Bernoulli two-dimensional distribution, or a noise signal conforming to other distributions. The noise signal may also be a noise signal containing known information.

The noise signal is input into the generator, and the generator will generate supplementary training samples based on the noise signal. The goal of the supplementary training samples is to be as similar or identical as possible to the real training samples in the real training set. A generator may also be called a channel generator.

Step 306: Input the supplementary training samples into the discriminator to obtain the second discriminant result;

The supplementary training samples are input into the discriminator to obtain the second discriminant result.

It should be noted that step 302 and steps 304-306 may be executed alternately or simultaneously, and the present application does not limit the execution sequence of step 302 and steps 304-306. For example, step 302 can be performed first, and then steps 304-306 can be performed; or steps 304-306 can be performed first, and then step 302 can be performed. For another example, based on different training samples, step 302 and steps 304-306 are executed alternately; or, step 302 is executed once, and steps 304-306 are executed multiple times; or, steps 304-306 are executed once, and step 302 is executed multiple times.

With reference to FIG. 4 , taking the access network device as the training device as an example, the access network device may collect CSI of multiple terminals as the real training set 31 , for example, the CSI of multiple terminals within the same geographic range as the real training set 31 . On the one hand, the real training samples 32 in the real training set 31 are input to the discriminator D, and the discriminator D will output a corresponding discriminant result 35 , that is, the first discriminant result. On the other hand, the random noise 33 is input to the generator G, the generator G will output a false (supplementary) training sample 34, and the false training sample 34 is input into the discriminator D, and the discriminator D will output a corresponding discriminant result 35, also That is, the second discrimination result. Then, the generator and the discriminator are trained based on the loss functions of the first discriminant result and the second discriminative result.

Step 308: Based on the first discrimination result and the second discrimination result, train a generator and a discriminator.

Among them, the loss function of the generator is set with the target that both the first discriminant result and the second discriminant result are true, and the loss function of the discriminator is set with the target that the first discriminant result is true and the second discriminant result is false .

Schematically, in the training phase of the discriminator, the neural model parameters of the generator are fixed, and the discriminator is trained based on the loss function of the discriminator; in the training phase of the generator, the neural model parameters of the fixed discriminator are not The generator is trained based on the generator's loss function. The above two training processes are alternately executed until the training end condition is satisfied.

Schematically, the training end condition includes: the number of training times reaches a number threshold, or the loss function converges.

This embodiment does not limit the training method of the confrontation generation network, for example, it can also be implemented in an improved form such as Wasserstein GAN (WGAN-GP) with gradient penalty.

To sum up, the method provided in this embodiment can train an adversarial generation network based on real training samples in a real training set. The adversarial generation network can generate fake training samples that are as identical or similar to the real training samples as possible. In the case of limited real training samples in the real training set, the adversarial generation network can generate enough fake training samples as supplementary training samples. .

The training process of encoder and decoder:

Fig. 5 shows a flowchart of a training method for an encoder and a decoder provided by an exemplary embodiment of the present application. The method may be performed by an access network device or a terminal or other devices, and the method includes:

Step 402: using the generator in the confrontational generative network to generate a first supplementary training set;

Assume that the number of training samples needed to support the training of the encoder and decoder is M, and the number of real training samples in the real training set (or called the original data set) is m. If the real training set has acquisition limitations, time constraints, or cost constraints, and the size m of the real training set is much smaller than the size M of the required training set, that is, m<<M, then directly use the real training set to train the encoder , an encoding model with better performance cannot be obtained.

After the adversarial generative network is trained, the generator has a good ability to generate samples. Based on the generator in the adversarial generative network, a first supplemental training set can be generated. The size of the first supplementary training set is not less than (M-m).

Step 404: Mixing the real training set and the first supplementary training set to obtain a joint training set;

A joint training set can be obtained by mixing the real training samples in the real training set and the supplementary training samples in the first supplementary training set. The size of the joint training set is equal to or larger than M.

That is, the size of the joint training set is sufficient to support the number of training samples required by the encoder and decoder.

Step 406: Use the joint training set to train the encoder and/or decoder to obtain a trained encoder and/or decoder.

The encoder and/or decoder are trained by using the joint training set to obtain the trained encoder and/or decoder.

In one example, each training sample in the joint training set is CSI, and the encoder and decoder are trained in an end-to-end training manner. In another example, each training sample in the joint training set includes a set of CSI and CSI feedback codebooks. Based on each training sample, the encoder can be trained separately, the decoder can also be trained separately, and the encoder and The decoder is jointly trained end-to-end.

This embodiment does not limit the training manner of the encoder and/or decoder. Referring schematically to FIG. 6 , the training device mixes the real training set 41 and the supplementary training set 42 into a joint training set 43 . The joint training set 43 includes a plurality of training samples, for example, each training sample is a CSI, and the encoder and the decoder are trained using the CSI in the joint training set 43 . That is, the training samples are input to the encoder, and the encoder performs training on the training samples to obtain CSI feedback information 44 . The decoder decodes the CSI feedback information 44 and outputs the restored channel 45 . The error between the restored channel 45 and the training samples is calculated based on the loss function 46, and the encoder and decoder are trained end-to-end by using the error back propagation method.

To sum up, the method provided in this embodiment uses the adversarial generative network to generate the first supplementary training set, and then uses the joint training set based on the mixture of the real training set and the first supplementary training set, which can be trained using sufficient training samples to obtain Encoder and decoder with better performance to improve compression efficiency and feedback accuracy during CSI feedback.

The CSI feedback process of the training device based on the access network device as the model:

Fig. 7 shows a flowchart of a CSI feedback method provided by an exemplary embodiment of the present application. The method can be performed by an access network device or a terminal, and the method includes:

Step 502: The access network device is trained to obtain an adversarial generation network based on the real training set;

For the training process of the adversarial generative network, reference may be made to the training process shown in the embodiment shown in FIG. 3 above, and details are not repeated here.

Step 504: The access network device trains an encoder and a decoder based on the joint training set;

For the training process of the encoder and the decoder, reference may be made to the training process shown in the embodiment shown in FIG. 5 above, and details are not repeated here.

Step 506: the access network device issues an encoder to the terminal;

The access network device sends the encoder or the model parameters of the encoder to the terminal through downlink signaling, and the terminal builds the encoder by itself according to the model parameters of the encoder.

Schematically, the downlink signaling includes at least one of downlink control information (Downlink ControlInformation, DCI), radio resource control (Radio Resource Control, RRC), and medium access control (Medium Access Control Control Element, MAC CE). The downlink signaling may also be dedicated signaling and channel resources dedicated to model delivery, which is not limited in this embodiment of the present application.

Schematically, the model parameters of the encoder and/or decoder include: the type of neural network, the number of layers of the neural network, the type of the layer of the neural network, the type of neurons in the neural network, the number of neurons in the neural network, the number of neurons in the neural network, At least one of the matrix weights of neurons in .

Step 508: The terminal uses an encoder to encode the CSI to obtain CSI feedback information;

Step 510: the terminal sends CSI feedback information to the access network device;

The terminal uses the uplink feedback channel to send CSI feedback information to the access network device.

Step 512: the access network device receives the CSI feedback information sent by the terminal, and the CSI feedback information is obtained by the terminal encoding the CSI through an encoder;

Step 514: The access network device uses a decoder to decode the CSI feedback information to obtain the CSI measured by the terminal.

In some other embodiments, the above training device may also be executed by a terminal. The terminal is trained based on the real training set to obtain the confrontation generation network; the terminal is trained based on the joint training set to obtain the encoder and decoder; the terminal reports the model parameters of the decoder or decoder to the access network device. The terminal uses an encoder to encode the CSI to obtain CSI feedback information. The terminal sends CSI feedback information to the access network device. The access network device receives the CSI feedback information sent by the terminal, and the CSI feedback information is obtained by the terminal encoding the CSI through an encoder. The access network device uses a decoder to decode the CSI feedback information to obtain the CSI measured by the terminal.

In some other embodiments, the adversarial generation network may also be trained by the first device, and the adversarial generation network or model parameters of the adversarial generation network may be sent to the second device. The second device generates a supplementary training set based on the confrontation generation network, and obtains an encoder and a decoder based on joint training set training. The first device is an access network device, and the second device is a terminal; or, the first device is a terminal, and the second device is an access network device.

To sum up, the method provided in this embodiment can quickly train an encoder and a decoder with excellent performance by utilizing the strong computing capability of the access network device. Then, the access network device sends the encoder to the terminal to complete the AI-based CSI feedback process and improve the compression efficiency and feedback accuracy during CSI feedback.

An updated training process for encoders and/or decoders based on periodic feedback:

It is limited by the generalization of the encoder and the complexity and variability of the channel environment. When the channel environment changes and the encoder does not adapt to the channel environment, resulting in performance degradation, the model parameters need to be updated. However, since the encoder of the CSI autoencoder is deployed on the UE side and the decoder is deployed on the network side, when model mismatch occurs, the network side cannot obtain a large number of changed high-quality effective channel datasets through CSI feedback. Continuously adopting a high-precision codebook to perform compressed feedback on changing channels will introduce high feedback overhead. Therefore, this embodiment adopts the method of confrontation generation network to reduce the feedback density of the high-precision codebook, and generates the second supplementary data set based on a small amount of channel feedback to complete the model update of the encoder.

Fig. 8 shows a flowchart of a model updating method provided by an exemplary embodiment of the present application. The method can be performed by an access network device and a terminal, and the method includes:

Step 602: The terminal periodically sends the first CSI feedback codebook based on codebook quantization to the access network device;

The first CSI feedback codebook is a CSI feedback codebook obtained based on a codebook quantization manner, rather than a CSI feedback codebook obtained based on an AI model or coder compression. Schematically, the first CSI feedback codebook is a high-quality CSI feedback codebook that can effectively express a changed channel environment.

Since there may be multiple first CSI feedback codebooks, the first CSI feedback codebook may also be called a second original training set, a second real training set, an updated training set, and the like.

Schematically, the first CSI feedback codebook includes CSI feedback information at multiple sampling time points, such as multiple CSI feedback information within a time window W, or multiple CSI feedback information sampled according to a specified period within the time window W.

Schematically, the first CSI feedback codebook includes CSI feedback information on multiple widebands, subbands or frequency points, such as multiple CSI feedback information in frequency window B, or multiple widebands and/or subbands of frequency window B. Multiple CSI feedback information on the belt.

Schematically, the feedback period of the first CSI feedback codebook may be determined by a configuration parameter T. The configuration parameter T of the first CSI feedback codebook may also carry other configuration information required in the codebook feedback process.

In an example, before step 602, the access network device sends the first report configuration to the terminal, for example, sends the first report configuration to the terminal through downlink signaling. The terminal receives the first report configuration sent by the access network device, where the first report configuration is used to indicate the first report parameter of the first CSI feedback codebook. Schematically, the first reporting parameter includes at least one of the following: a time window W; a frequency window B; and a configuration parameter T of the first CSI feedback codebook.

In another example, before step 602, the terminal determines the first reporting configuration by itself, and the first reporting configuration is used to indicate the first reporting parameter of the first CSI feedback codebook. The terminal sends the first reported configuration to the access network device. For example, the terminal sends the first reported configuration to the access network device through uplink signaling.

Step 604: The access network device receives the first CSI feedback codebook sent by the terminal;

Step 606: The access network device constructs a second supplementary training set based on the first CSI feedback codebook through an adversarial generation network;

The access network device constructs a second supplementary training set based on the first CSI feedback codebook through the confrontation generation network; or, the access network device determines the CSI corresponding to the first CSI feedback codebook, and uses the confrontation generation network based on the first CSI feedback codebook The corresponding CSI constructs a second supplementary training set; or, the access network device determines the CSI corresponding to the first CSI feedback codebook, and constructs the second supplementary training set based on the first CSI feedback codebook and the CSI corresponding to the first CSI feedback codebook through the confrontation generation network. Two supplementary training sets.

The manner of training the adversarial generation network based on the first CSI feedback codebook is similar to the embodiment shown in FIG. 3 , and will not be repeated in this embodiment.

In an example, the adversarial generation network is the adversarial generation network shown in the embodiment shown in FIG. 3 .

In an embodiment, the adversarial generation network is an adversarial generation network additionally trained based on the first CSI feedback codebook, and the adversarial generation network is different from the adversarial generation network shown in the embodiment shown in FIG. 3 .

Step 608: The access network device updates and trains at least one of the encoder and the decoder through the first CSI feedback codebook and the second supplementary training set;

Schematically, the access network device performs joint update training on the encoder and the decoder through the first joint update training set composed of the first CSI feedback codebook and the second supplementary training set. The updated encoder can be well adapted to the changed channel information.

The access network device may only perform update training on the encoder; or, perform update training only on the decoder; or, perform update training on both the encoder and the decoder.

Step 610: The access network device delivers the updated encoder to the terminal.

In the case of updating the encoder, the access network device delivers the updated encoder to the terminal, and the whole process can be referred to as shown in FIG. 9 . Among them, the quantization accuracy of the high-precision codebook in step 602 can be further enhanced based on the Type2 codebook to improve the recovery accuracy of CSI and ensure the accuracy of the first CSI feedback codebook; meanwhile, the feedback of the first CSI feedback codebook Relevant configurations: cycle parameter T, time window W, frequency window B, etc. can be configured by the network side and notified to the UE through DCI. The process of generating the second supplementary data set based on the first CSI feedback codebook in step 606 is the same as the production method of the first supplementary data set in the embodiment in FIG. 3 . However, since the model is updated, the sample size of the generated second supplementary data set is generally less than the sample size M of the required data set in Example 1 shown in Figure 3, so as to ensure that the model against the generative network can be quickly completed Convergence and fast updating of CSI autoencoder models.

In another embodiment, the device for updating the encoder and/or decoder may also be a terminal. The terminal periodically measures the first CSI to form an updated training set; the terminal constructs a second supplementary training set based on the updated training set through the confrontation generation network; the terminal feeds back the codebook and the second supplementary training set through the first CSI, and the encoder and decoder At least one of them is updated for training. The terminal reports the updated decoder to the access network device. The specific process is similar to the embodiment shown in FIG. 8 and will not be repeated here.

To sum up, the method provided in this embodiment is based on the periodic update method during the online deployment of the CSI autoencoder, and configures a lower-density high-precision codebook for feedback update data sets, and cooperates with The adversarial generation network can realize the online update of the CSI autoencoder model and ensure the accuracy of CSI feedback and recovery.

An updated training process for encoders and/or decoders based on aperiodic feedback:

Fig. 10 shows a flowchart of a model updating method provided by an exemplary embodiment of the present application. The method can be performed by an access network device and a terminal, and the method includes:

Step 702: When the trigger condition is met, the terminal sends the second CSI feedback codebook based on codebook quantization to the access network device;

The second CSI feedback codebook is a CSI feedback codebook obtained based on a codebook quantization manner, rather than a CSI feedback codebook obtained based on an AI model or coder compression. Schematically, the second CSI feedback codebook is a high-quality CSI feedback codebook that can effectively express a changed channel environment.

Since there may be multiple second CSI feedback codebooks, the second CSI feedback codebook may also be called a second original training set, a second real training set, an updated training set, and the like.

The trigger condition includes: the change of channel state or channel information or channel parameter is greater than a preset threshold. Channel status or channel information or channel parameters include: Reference Signal Received Power (Reference Signal Received Power, RSRP), Reference Signal Received Quality (Reference Signal Received Quality, RSRQ), Received Signal Strength Indicator (Reference Signal Strength Indicator, RSSI), At least one of CSI.

The preset threshold may be predefined by the communication protocol; or, the preset threshold may be preconfigured; or, the preset threshold may be configured by the access network device to the terminal.

Schematically, the second CSI feedback codebook includes CSI feedback information at multiple sampling time points, such as multiple CSI feedback information within the time window W, or multiple CSI feedback information sampled at a specified period within the time window W.

Schematically, the second CSI feedback codebook includes CSI feedback information on multiple widebands, subbands or frequency points, such as multiple CSI feedback information in frequency window B, or multiple widebands and/or subbands of frequency window B Multiple CSI feedback information on the belt.

Schematically, the feedback period of the second CSI feedback codebook may be determined by a configuration parameter T. The trigger condition of the second CSI feedback codebook may also carry other configuration information required in the codebook feedback process.

In an example, before step 702, the access network device sends the second report configuration to the terminal, for example, sends the second report configuration to the terminal through downlink signaling. The terminal receives the second report configuration sent by the access network device, where the second report configuration is used to indicate the second report parameter of the second CSI feedback codebook. Schematically, the second reporting parameter includes at least one of the following: a time window W; a frequency window B; a trigger condition of the second CSI feedback codebook, as shown in FIG. 11 .

In another example, before step 702, the terminal determines the second reporting configuration by itself, and the second reporting configuration is used to indicate the second reporting parameter of the second CSI feedback codebook. The terminal sends the second reported configuration to the access network device. For example, the terminal sends the second reported configuration to the access network device through uplink signaling, as shown in FIG. 12 .

Step 704: The access network device receives the second CSI feedback codebook sent by the terminal;

Step 706: The access network device constructs a third supplementary training set based on the second CSI feedback codebook through the confrontation generation network;

The access network device constructs a third supplementary training set based on the second CSI feedback codebook through the confrontation generation network; or, the access network device determines the CSI corresponding to the second CSI feedback codebook, and uses the confrontation generation network based on the second CSI feedback codebook The corresponding CSI constructs a third supplementary training set; or, the access network device determines the CSI corresponding to the second CSI feedback codebook, and constructs the second CSI feedback codebook based on the second CSI feedback codebook and the CSI corresponding to the second CSI feedback codebook through the confrontation generation network. Three supplementary training sets.

The manner of training the adversarial generation network based on the second CSI feedback codebook is similar to the embodiment shown in FIG. 3 , and will not be repeated in this embodiment.

In an embodiment, the adversarial generation network is an adversarial generation network additionally trained based on the second CSI feedback codebook, and the adversarial generation network is different from the adversarial generation network shown in the embodiment shown in FIG. 3 .

Step 708: The access network device updates and trains at least one of the encoder and the decoder through the second CSI feedback codebook and the third supplementary training set;

Schematically, the access network device performs joint update training on the encoder and the decoder through the second joint update training set composed of the second CSI feedback codebook and the third supplementary training set. The updated encoder can be well adapted to the changed channel information.

Step 710: The access network device delivers the updated encoder to the terminal.

In the case of updating the encoder, the access network device delivers the updated encoder to the terminal. Among them, the quantization accuracy of the high-precision codebook in step 702 can be further enhanced based on the Type2 codebook to improve the recovery accuracy of CSI and ensure the accuracy of the second CSI feedback codebook; meanwhile, the feedback of the second CSI feedback codebook Relevant configurations: cycle parameter T, time window W, frequency window B, etc. can be configured by the network side and notified to the UE through DCI. The process of generating the second supplementary data set based on the second CSI feedback codebook in step 706 is the same as the production method of the first supplementary data set in the embodiment in FIG. 3 . However, since the model is updated, the sample size of the generated second supplementary data set is generally less than the sample size M of the required data set in Example 1 shown in Figure 3, so as to ensure that the model against the generative network can be quickly completed Convergence and fast updating of CSI autoencoder models.

In another embodiment, the device for updating the encoder and/or decoder may also be a terminal. The terminal periodically measures the second CSI to form an updated training set; the terminal constructs a third supplementary training set based on the updated training set through the confrontation generation network; the terminal feeds back the codebook and the third supplementary training set through the second CSI, and the encoder and decoder At least one of them is updated for training. The terminal reports the updated decoder to the access network device. The specific process is similar to the embodiment shown in FIG. 8 and will not be repeated here.

To sum up, the method provided in this embodiment, by updating the CSI autoencoder model online based on trigger conditions during the online deployment of the CSI autoencoder, compared with the previous embodiment, can reduce the number of terminals and servers The amount of communication data when communicating with each other.

FIG. 13 shows a block diagram of a CSI feedback device provided by an exemplary embodiment of the present application. The device can be implemented as a terminal or a functional module in the terminal, and the device includes:

The encoding module 1320 is configured to use an encoder to encode the CSI to obtain CSI feedback information; the encoder is trained from the real training set and the first supplementary training set, and the first supplementary training set is obtained from the adversarial generation network Generated by the generator, the confrontation generation network is trained based on the real training set;

The sending module 1340 is configured to send the CSI feedback information to the access network device.

In an optional embodiment, the encoder is trained as follows:

generating the first supplemental training set using a generator in the adversarial generative network;

mixing the real training set and the first supplementary training set to obtain a joint training set;

The encoder is trained by using the joint training set to obtain the trained encoder.

In an optional embodiment, the adversarial generation network includes a generator and a discriminator, and the adversarial generation network is trained in the following manner:

Inputting the training samples in the real training set into the discriminator to obtain a first discriminant result;

inputting the noise signal into the generator to obtain a supplementary training sample; inputting the supplementary training sample into the discriminator to obtain a second discrimination result;

training the generator and the discriminator based on the first discrimination result and the second discrimination result;

Wherein, the loss function of the generator is set with the target that both the first discrimination result and the second discrimination result are true, and the loss function of the discriminator is set with the first discrimination result being true and The second discrimination result is set for a false target.

In an optional embodiment, the device also includes:

The receiving module 1360 is configured to receive the encoder issued by the access network device, the encoder is obtained through training by the access network device.

In an optional embodiment, the sending module 1340 is further configured to periodically send the first CSI feedback codebook based on codebook quantization to the access network device; the receiving module 1360 is configured to receive the An updated coder issued by the access network device, where the updated coder is the updated coder performed by the access network device based on the first CSI feedback codebook and the second supplementary training set obtained later, the second supplementary training set is constructed by the adversarial generation network based on the first CSI feedback codebook.

In an optional embodiment, the receiving module 1360 is further configured to receive a first report configuration sent by the access network device, where the first report configuration is used to indicate the first CSI feedback codebook The first reported parameter;

or,

The sending module 1340 is further configured to send a first report configuration to the access network device, where the first report configuration is used to indicate a first report parameter of the first CSI feedback codebook.

In an optional embodiment, the first reporting parameter includes at least one of the following: a time window W; a frequency window B; and a configuration parameter T of the first CSI feedback codebook.

In an optional embodiment, the sending module 1340 is further configured to send a second CSI feedback codebook based on codebook quantization to the access network device when a trigger condition is met; the receiving module 1360, further receiving an updated encoder issued by the access network device, where the updated encoder is based on the second CSI feedback codebook and the third supplementary training set by the access network device obtained after updating the encoder, and the third supplementary training set is constructed by the adversarial generation network based on the second CSI feedback codebook.

In an optional embodiment, the trigger conditions include:

The change value of the channel parameter is greater than a preset threshold;

Wherein, the channel state or channel information or channel parameters include: at least one of RSRP, RSRQ, RSSI, and CSI.

In an optional embodiment, the device also includes:

The receiving module 1360 is further configured to receive a second report configuration sent by the access network device, where the second report configuration is used to indicate a second report parameter of the second CSI feedback codebook; or, the The sending module 1340 is further configured to send a second reporting configuration to the access network device, where the second reporting configuration is used to indicate a second reporting parameter of the second CSI feedback codebook.

In an optional embodiment, the second reporting parameter includes at least one of the following: a time window W; a frequency window B; and a trigger condition of the second CSI feedback codebook.

Fig. 14 shows a block diagram of a CSI feedback device provided by an exemplary embodiment of the present application. The device can be realized as an access network device or a functional module in the access network device, and the device includes:

The receiving module 1420 is configured to receive CSI feedback information sent by the terminal, where the CSI feedback information is obtained by the terminal encoding CSI through an encoder;

The decoding module 1440 is configured to use a decoder to decode the CSI feedback information to obtain the CSI measured by the terminal; the encoder and the decoder are obtained by training the real training set and the first supplementary training set, The first supplementary training set is generated by a generator in the adversarial generation network, and the adversarial generation network is trained based on the real training set.

In an optional embodiment, the encoder and the decoder are trained in the following manner:

The encoder and the decoder are trained by using the joint training set to obtain the trained encoder and the decoder.

In an optional embodiment, the sending module 1460 is configured to send the encoder to the terminal.

In an optional embodiment, the receiving module 1420 is configured to periodically receive the first CSI feedback codebook based on codebook quantization sent by the terminal;

The training module 1480 is used to construct a second supplementary training set based on the first CSI feedback codebook through the confrontation generation network; through the first CSI feedback codebook and the second supplementary training set, the At least one of the encoder and the decoder undergoes update training.

In an optional embodiment, the receiving module 1420 is configured to receive a first reporting configuration sent by the terminal, where the first reporting configuration is used to indicate a first reporting parameter of the first CSI feedback codebook or, a sending module 1460, configured to send a first reporting configuration to the terminal, where the first reporting configuration is used to indicate a first reporting parameter of the first CSI feedback codebook.

In an optional embodiment, the first reporting parameter includes at least one of the following:

time window W;

frequency window B;

The configuration parameter T of the first CSI feedback codebook.

In an optional embodiment, the receiving module 1420 is configured to receive the second CSI feedback codebook based on codebook quantization sent by the terminal when the trigger condition is met;

In an optional embodiment, the sending module 1460 is configured to send the updated coder to the terminal when the coder is updated and trained.

In an optional embodiment, the trigger condition includes: a change value of the channel parameter is greater than a preset threshold. The channel status or channel information or channel parameters include: at least one of RSRP, RSRQ, RSSI, and CSI.

In an optional embodiment, the receiving module 1420 is configured to receive a second reporting configuration sent by the terminal, where the second reporting configuration is used to indicate a second reporting parameter of the second CSI feedback codebook or, a sending module 1460, configured to send a second reporting configuration to the terminal, where the second reporting configuration is used to indicate a second reporting parameter of the second CSI feedback codebook.

Figure 15 shows a schematic structural diagram of a communication device (terminal or access network device) provided by an exemplary embodiment of the present application, the communication device includes: a processor 101, a receiver 102, a transmitter 103, a memory 104 and a bus 105 .

The processor 101 includes one or more processing cores, and the processor 101 executes various functional applications and information processing by running software programs and modules.

The receiver 102 and the transmitter 103 can be implemented as a communication component, which can be a communication chip.

The memory 104 is connected to the processor 101 through the bus 105 .

The memory 104 may be used to store at least one instruction, and the processor 101 is used to execute the at least one instruction, so as to implement various steps in the foregoing method embodiments.

In addition, the memory 104 can be implemented by any type of volatile or non-volatile storage device or their combination, volatile or non-volatile storage devices include but not limited to: magnetic disk or optical disk, electrically erasable and programmable Electrically-Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Static Random Access Memory (SRAM), Read-Only Memory (Read-Only Memory, ROM), magnetic memory, flash memory, programmable read-only memory (Programmable Read-Only Memory, PROM).

In an exemplary embodiment, a computer-readable storage medium is also provided, the computer-readable storage medium stores at least one instruction, at least one program, a code set or an instruction set, the at least one instruction, the At least one program, the code set or the instruction set is loaded and executed by the processor to implement the CSI feedback method performed by the first terminal or the second terminal or the network device provided in the above method embodiments.

In an exemplary embodiment, there is also provided a computer program product or computer program, the computer program product or computer program comprising computer instructions, the computer instructions are stored in a computer-readable storage medium, the processor of the communication device can read from the computer The computer instruction is read by reading the storage medium, and the processor executes the computer instruction, so that the communication device executes the CSI feedback method performed by the first terminal or the second terminal or the network device described in the above aspect.

The above are only optional embodiments of the application, and are not intended to limit the application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the application shall be included in the protection of the application. within range.

Claims

A channel state information CSI feedback method, characterized in that it is applied to a terminal, and the method includes:

Use an encoder to encode the CSI to obtain CSI feedback information; the encoder is trained by the real training set and the first supplementary training set, and the first supplementary training set is generated by the generator in the confrontation generation network , the confrontation generation network is obtained based on the real training set training;

Send the CSI feedback information to the access network device.
The method according to claim 1, wherein the encoder is trained as follows:

generating the first supplemental training set using a generator in the adversarial generative network;

mixing the real training set and the first supplementary training set to obtain a joint training set;

The encoder is trained by using the joint training set to obtain the trained encoder.
The method according to claim 1, wherein the confrontation generation network includes a generator and a discriminator, and the confrontation generation network is trained based on the following method:

Inputting the training samples in the real training set into the discriminator to obtain a first discriminant result;

inputting the noise signal into the generator to obtain a supplementary training sample; inputting the supplementary training sample into the discriminator to obtain a second discrimination result;

training the generator and the discriminator based on the first discrimination result and the second discrimination result;

Wherein, the loss function of the generator is set with the target that both the first discrimination result and the second discrimination result are true, and the loss function of the discriminator is set with the first discrimination result being true and The second discrimination result is set for a false target.
The method according to any one of claims 1 to 3, wherein the method further comprises:

receiving the encoder sent by the access network device, where the encoder is trained by the access network device.
The method according to any one of claims 1 to 3, wherein the method further comprises:

periodically sending the first CSI feedback codebook based on codebook quantization to the access network device;

Receive an updated encoder sent by the access network device, where the updated encoder is the codebook that the access network device performs on the encoder based on the first CSI feedback codebook and the second supplementary training set Obtained after updating training, the second supplementary training set is constructed by the adversarial generation network based on the first CSI feedback codebook.
The method according to claim 5, wherein the method further comprises:

Receive a first report configuration sent by the access network device, where the first report configuration is used to indicate a first report parameter of the first CSI feedback codebook;

or,

Sending a first reporting configuration to the access network device, where the first reporting configuration is used to indicate a first reporting parameter of the first CSI feedback codebook.
The method according to claim 6, wherein the first reporting parameter includes at least one of the following:

time window W;

frequency window B;

The configuration parameter T of the first CSI feedback codebook.
The method according to any one of claims 1 to 3, wherein the method further comprises:

When the trigger condition is met, send the second CSI feedback codebook based on codebook quantization to the access network device;

Receive an updated encoder delivered by the access network device, where the updated encoder is the code that the access network device performs on the encoder based on the second CSI feedback codebook and the third supplementary training set Obtained after updating training, the third supplementary training set is constructed by the adversarial generation network based on the second CSI feedback codebook.
The method according to claim 8, wherein the trigger conditions include:

The change value of the channel parameter is greater than a preset threshold.
The method according to claim 8, characterized in that the method further comprises:

Receive a second report configuration sent by the access network device, where the second report configuration is used to indicate a second report parameter of the second CSI feedback codebook;

or,

Sending a second reporting configuration to the access network device, where the second reporting configuration is used to indicate a second reporting parameter of the second CSI feedback codebook.
The method according to claim 10, wherein the second reporting parameters include at least one of the following:

time window W;

frequency window B;

The configuration parameter T of the second CSI feedback codebook.
A channel state information CSI feedback method, characterized in that it is applied to an access network device, and the method includes:

receiving CSI feedback information sent by the terminal, where the CSI feedback information is obtained by the terminal encoding the CSI through an encoder;

Use a decoder to decode the CSI feedback information to obtain the CSI measured by the terminal; the encoder and the decoder are trained by the real training set and the first supplementary training set, and the first supplementary training The set is generated by the generator in the adversarial generative network, which is trained based on the real training set.
The method according to claim 12, wherein the encoder and the decoder are trained as follows:

generating the first supplemental training set using a generator in the adversarial generative network;

mixing the real training set and the first supplementary training set to obtain a joint training set;

The encoder and the decoder are trained by using the joint training set to obtain the trained encoder and the decoder.
The method according to claim 12, wherein the confrontation generation network includes a generator and a discriminator, and the confrontation generation network is trained based on the following method:

Inputting the training samples in the real training set into the discriminator to obtain a first discriminant result;

inputting the noise signal into the generator to obtain a supplementary training sample; inputting the supplementary training sample into the discriminator to obtain a second discrimination result;

training the generator and the discriminator based on the first discrimination result and the second discrimination result;

Wherein, the loss function of the generator is set with the target that both the first discrimination result and the second discrimination result are true, and the loss function of the discriminator is set with the first discrimination result being true and The second discrimination result is set for a false target.
The method according to any one of claims 12 to 14, further comprising:

delivering the encoder to the terminal.
The method according to any one of claims 12 to 14, further comprising:

periodically receiving the first CSI feedback codebook based on codebook quantization sent by the terminal;

A second supplementary training set constructed by the confrontation generation network based on the first CSI feedback codebook;

At least one of the encoder and the decoder is updated and trained by using the first CSI feedback codebook and the second supplementary training set.
The method according to claim 16, further comprising:

Receive a first report configuration sent by the terminal, where the first report configuration is used to indicate a first report parameter of the first CSI feedback codebook;

or,

Sending a first reporting configuration to the terminal, where the first reporting configuration is used to indicate a first reporting parameter of the first CSI feedback codebook.
The method according to claim 17, wherein the first reporting parameter includes at least one of the following:

time window W;

frequency window B;

The configuration parameter T of the first CSI feedback codebook.
The method according to any one of claims 12 to 14, further comprising:

receiving the second CSI feedback codebook based on codebook quantization sent by the terminal when the trigger condition is met;

A second supplementary training set constructed by the confrontation generation network based on the first CSI feedback codebook;

At least one of the encoder and the decoder is updated and trained by using the first CSI feedback codebook and the second supplementary training set.
The method according to claim 16 or 19, further comprising:

In a case where the encoder is updated and trained, the updated encoder is delivered to the terminal.
The method according to claim 19, wherein the trigger conditions include:

The change value of the channel parameter is greater than a preset threshold.
The method according to claim 19, further comprising:

Receive a second reporting configuration sent by the terminal, where the second reporting configuration is used to indicate a second reporting parameter of the second CSI feedback codebook;

or,

Sending a second reporting configuration to the terminal, where the second reporting configuration is used to indicate a second reporting parameter of the second CSI feedback codebook.
The method according to claim 22, wherein the second reporting parameters include at least one of the following:

time window W;

frequency window B;

A trigger condition of the second CSI feedback codebook.
A channel state information CSI feedback device, characterized in that the device comprises:

An encoding module, configured to use an encoder to encode the CSI to obtain CSI feedback information; the encoder is trained from a real training set and a first supplementary training set, and the first supplementary training set is an adversarial generation network Generated by the generator in , the confrontation generation network is trained based on the real training set;

A sending module, configured to send the CSI feedback information to the access network device.
The device according to claim 24, wherein the encoder is trained in the following manner:

generating the first supplemental training set using a generator in the adversarial generative network;

mixing the real training set and the first supplementary training set to obtain a joint training set;

The encoder is trained by using the joint training set to obtain the trained encoder.
The device according to claim 24, wherein the confrontation generation network includes a generator and a discriminator, and the confrontation generation network is trained in the following manner:

Inputting the training samples in the real training set into the discriminator to obtain a first discriminant result;

inputting the noise signal into the generator to obtain a supplementary training sample; inputting the supplementary training sample into the discriminator to obtain a second discrimination result;

training the generator and the discriminator based on the first discrimination result and the second discrimination result;

Wherein, the loss function of the generator is set with the target that both the first discrimination result and the second discrimination result are true, and the loss function of the discriminator is set with the first discrimination result being true and The second discrimination result is set for a false target.
The device according to any one of claims 24 to 26, wherein the device further comprises:

The receiving module is configured to receive the coder issued by the access network device, the coder is obtained through training of the access network device.
The device according to any one of claims 24 to 26, wherein the device further comprises:

The sending module is further configured to periodically send the first CSI feedback codebook based on codebook quantization to the access network device;

A receiving module, configured to receive an updated encoder issued by the access network device, where the updated encoder is based on the first CSI feedback codebook and the second supplementary training set by the access network device obtained after updating the encoder, and the second supplementary training set is constructed by the adversarial generation network based on the first CSI feedback codebook.
The device according to claim 28, further comprising:

The receiving module is further configured to receive a first report configuration sent by the access network device, where the first report configuration is used to indicate a first report parameter of the first CSI feedback codebook;

or,

The sending module is further configured to send a first report configuration to the access network device, where the first report configuration is used to indicate a first report parameter of the first CSI feedback codebook.
The device according to claim 29, wherein the first reporting parameter includes at least one of the following:

time window W;

frequency window B;

The configuration parameter T of the first CSI feedback codebook.
The device according to any one of claims 24 to 27, wherein the device further comprises:

The sending module is further configured to send a second CSI feedback codebook based on codebook quantization to the access network device when a trigger condition is met;

The receiving module is further configured to receive an updated coder delivered by the access network device, where the updated coder is obtained by the access network device based on the second CSI feedback codebook and the third supplementary training The third supplementary training set is obtained by updating and training the encoder on the set, and the third supplementary training set is constructed by the adversarial generation network based on the second CSI feedback codebook.
The device according to claim 31, wherein the trigger conditions include:

The change value of the channel parameter is greater than a preset threshold.
The device according to claim 31, further comprising:

The receiving module is further configured to receive a second report configuration sent by the access network device, where the second report configuration is used to indicate a second report parameter of the second CSI feedback codebook;

or,

The sending module is further configured to send a second reporting configuration to the access network device, where the second reporting configuration is used to indicate a second reporting parameter of the second CSI feedback codebook.
The device according to claim 33, wherein the second reporting parameter includes at least one of the following:

time window W;

frequency window B;

A trigger condition of the second CSI feedback codebook.
A channel state information CSI feedback device, characterized in that the device comprises:

A receiving module, configured to receive CSI feedback information sent by the terminal, where the CSI feedback information is obtained by the terminal encoding the CSI through an encoder;

A decoding module, configured to use a decoder to decode the CSI feedback information to obtain the CSI measured by the terminal; the encoder and the decoder are obtained by training the real training set and the first supplementary training set, so The first supplementary training set is generated by a generator in the adversarial generation network, and the adversarial generation network is trained based on the real training set.
The device according to claim 35, wherein the encoder and the decoder are trained in the following manner:

generating the first supplemental training set using a generator in the adversarial generative network;

mixing the real training set and the first supplementary training set to obtain a joint training set;

The encoder and the decoder are trained by using the joint training set to obtain the trained encoder and the decoder.
The device according to claim 35, wherein the adversarial generation network includes a generator and a discriminator, and the adversarial generation network is trained in the following manner:

Inputting the training samples in the real training set into the discriminator to obtain a first discriminant result;

inputting the noise signal into the generator to obtain a supplementary training sample; inputting the supplementary training sample into the discriminator to obtain a second discrimination result;

training the generator and the discriminator based on the first discrimination result and the second discrimination result;

Wherein, the loss function of the generator is set with the target that both the first discrimination result and the second discrimination result are true, and the loss function of the discriminator is set with the first discrimination result being true and The second discrimination result is set for a false target.
The device according to any one of claims 35 to 37, wherein the device further comprises:

A sending module, configured to send the encoder to the terminal.
The device according to any one of claims 35 to 37, wherein the device further comprises:

The receiving module is configured to periodically receive the first CSI feedback codebook based on codebook quantization sent by the terminal;

The training module is used to construct a second supplementary training set based on the first CSI feedback codebook through the confrontation generation network; through the first CSI feedback codebook and the second supplementary training set, the encoding At least one of the decoder and the decoder is updated for training.
The device according to claim 39, further comprising:

The receiving module is configured to receive a first report configuration sent by the terminal, where the first report configuration is used to indicate a first report parameter of the first CSI feedback codebook;

or,

A sending module, configured to send a first reporting configuration to the terminal, where the first reporting configuration is used to indicate a first reporting parameter of the first CSI feedback codebook.
The device according to claim 40, wherein the first reporting parameter includes at least one of the following:

time window W;

frequency window B;

The configuration parameter T of the first CSI feedback codebook.
The device according to any one of claims 35 to 37, wherein the device further comprises:

The receiving module is configured to receive the second CSI feedback codebook based on codebook quantization sent by the terminal when the trigger condition is met;

The training module is used to construct a second supplementary training set based on the first CSI feedback codebook through the confrontation generation network; through the first CSI feedback codebook and the second supplementary training set, the encoding At least one of the decoder and the decoder is updated for training.
The device according to claim 39 or 42, further comprising:

A sending module, configured to deliver the updated encoder to the terminal when the encoder is updated and trained.
The device according to claim 42, wherein the trigger conditions include:

The change value of the channel parameter is greater than a preset threshold.
The device according to claim 42, further comprising:

The receiving module is configured to receive a second report configuration sent by the terminal, where the second report configuration is used to indicate a second report parameter of the second CSI feedback codebook;

or,

A sending module, configured to send a second reporting configuration to the terminal, where the second reporting configuration is used to indicate a second reporting parameter of the second CSI feedback codebook.
The device according to claim 45, wherein the second reporting parameter includes at least one of the following:

time window W;

frequency window B;

A trigger condition of the second CSI feedback codebook.
A terminal, characterized in that the terminal includes:

processor;

a transceiver connected to the processor;

memory for storing executable instructions of the processor;

Wherein, the processor is configured to load and execute the executable instructions to implement the CSI feedback method according to any one of claims 1-11.
A network device, characterized in that the network device includes:

processor;

a transceiver connected to the processor;

memory for storing executable instructions of the processor;

Wherein, the processor is configured to load and execute the executable instructions to implement the CSI feedback method according to any one of claims 12-23.
A computer-readable storage medium, characterized in that executable instructions are stored in the readable storage medium, and the executable instructions are loaded and executed by the processor so as to implement any one of claims 1 to 23. The CSI feedback method described above.
A computer program product, characterized in that executable instructions are stored in the computer program product, and the executable instructions are loaded and executed by the processor to implement the CSI according to any one of claims 1 to 23 Feedback method.
A chip, characterized in that the chip is used to implement the CSI feedback method according to any one of claims 1-23.