WO2023115254A1 - Data processing method and device - Google Patents

Abstract

The present application provides a data processing method and device. The data processing method comprises: generating a plurality of pieces of training data by using a codebook, the plurality of pieces of training data being used for training a neural network model, and the neural network model being used for performing channel information feedback. According to the present application, the plurality of pieces of training data are generated by using the codebook, and a channel feedback model based on a neural network is trained by using the training data. Therefore, according to the present application, the requirement and overhead of obtaining actual scene channel data by the neural network can be reduced, and the generalization of the neural network is enhanced by using the generalization ability of the codebook.

Description

Method and device for processing data technical field

The present application relates to the field of communication technology, and more specifically, to a method and device for processing data.

Background technique

In a communication system, a scheme based on a codebook can be used to realize channel information feedback. Due to the exquisite design of the codebook, the channel information feedback scheme based on the codebook can adapt to the channel information feedback in different scenarios, that is, the scheme has high generalization. The channel information needs to be mapped to the precoding matrix in the codebook, and the feedback of the channel information is realized by feeding back the precoding matrix. Since the precoding matrix is discrete, the mapping process is quantized and lossy, which leads to low accuracy of the codebook-based precoding scheme.

With the development of artificial intelligence technology, related technologies have proposed a channel information feedback scheme based on neural networks. This scheme can directly compress and feed back channel information. Therefore, the accuracy of the channel information feedback scheme based on the neural network is relatively high.

In order to improve the generalization of the neural network-based channel feedback model, related technologies use multi-scenario training data to train the neural network model. In the actual operation process, it is difficult to obtain large-scale and high-diversity training data. Therefore, obtaining training data will consume a lot of manpower, material resources, financial resources and time.

Contents of the invention

The present application provides a data processing method and device to solve the problem of difficulty in obtaining training data for a neural network-based channel feedback model.

In the first aspect, a method for processing data is provided, the method includes: using a codebook to generate a plurality of training data, the plurality of training data is used to train a neural network model, and the neural network model is used to perform channel information feedback.

In a second aspect, a device for processing data is provided, and the device includes: a first generation unit, configured to generate a plurality of training data using a codebook, the plurality of training data are used to train a neural network model, and the neural network The network model is used for channel information feedback.

In a third aspect, a device for processing data is provided, including a processor, a memory, and a communication interface, the memory is used to store one or more computer programs, and the processor is used to call the computer programs in the memory to make the The terminal device executes the method described in the first aspect.

In the fourth aspect, the embodiment of the present application provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and the computer program enables the terminal device to perform some or all of the steps in the method of the first aspect above .

In a fifth aspect, an embodiment of the present application provides a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to enable the terminal to execute the above-mentioned first Some or all of the steps in the method of one aspect. In some implementations, the computer program product can be a software installation package.

In a sixth aspect, an embodiment of the present application provides a chip, the chip includes a memory and a processor, and the processor can call and run a computer program from the memory to implement some or all of the steps described in the method of the first aspect above .

In a seventh aspect, a computer program product is provided, including a program, the program causes a computer to execute the method described in the first aspect.

In an eighth aspect, a computer program is provided, the computer program causes a computer to execute the method described in the first aspect.

The present application uses a codebook to generate a plurality of training data, and uses the training data to train a neural network-based channel feedback model. Therefore, the present application can reduce the requirement and overhead of the neural network for acquiring actual scene channel data, and enhance the generalization of the neural network by using the generalization capability of the codebook.

Description of drawings

Fig. 1 is a wireless communication system applied in the embodiment of the present application.

FIG. 2 is a structural diagram of a neural network applicable to an embodiment of the present application.

FIG. 3 is a structural diagram of a convolutional neural network applicable to an embodiment of the present application.

Fig. 4 is an example diagram of an image compression process based on an autoencoder.

Fig. 5 is an example diagram of an image generation process based on a variational autoencoder.

Fig. 6 is a schematic diagram of a neural network-based channel feedback process.

Fig. 7 is a schematic flowchart of a method for processing data provided by an embodiment of the present application.

FIG. 8 is a schematic diagram of a method for grouping precoding matrices of a codebook provided by an embodiment of the present application.

FIG. 9 is a channel information feedback model based on a variational autoencoder provided in an embodiment of the present application.

FIG. 10 is a schematic structural diagram of a variational autoencoder provided by an embodiment of the present application.

Fig. 11 is a schematic structural diagram of an apparatus for data processing provided by an embodiment of the present application.

Fig. 12 is a schematic structural diagram of a device provided by an embodiment of the present application.

Detailed ways

The technical solution in this application will be described below with reference to the accompanying drawings.

Communication Systems

FIG. 1 is a wireless communication system 100 applied in an embodiment of the present application. The wireless communication system 100 may include a network device 110 and a terminal device 120 . The network device 110 may be a device that communicates with the terminal device 120 . The network device 110 can provide communication coverage for a specific geographical area, and can communicate with the terminal device 120 located in the coverage area.

Figure 1 exemplarily shows one network device and two terminals. Optionally, the wireless communication system 100 may include multiple network devices and each network device may include other numbers of terminal devices within the coverage area. The embodiment does not limit this.

Optionally, the wireless communication system 100 may further include other network entities such as a network controller and a mobility management entity, which is not limited in this embodiment of the present application.

It should be understood that the technical solutions of the embodiments of the present application can be applied to various communication systems, for example: the fifth generation (5th generation, 5G) system or new radio (new radio, NR), long term evolution (long term evolution, LTE) system , LTE frequency division duplex (frequency division duplex, FDD) system, LTE time division duplex (time division duplex, TDD), etc. The technical solutions provided in this application can also be applied to future communication systems, such as the sixth generation mobile communication system, and satellite communication systems, and so on.

The terminal equipment in the embodiment of the present application may also be called user equipment (user equipment, UE), access terminal, subscriber unit, subscriber station, mobile station, mobile station (mobile station, MS), mobile terminal (mobile terminal, MT) ), remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent, or user device. The terminal device in the embodiment of the present application may be a device that provides voice and/or data connectivity to users, and can be used to connect people, objects and machines, such as handheld devices with wireless connection functions, vehicle-mounted devices, and the like. The terminal device in the embodiment of the present application can be mobile phone (mobile phone), tablet computer (Pad), notebook computer, palmtop computer, mobile internet device (mobile internet device, MID), wearable device, virtual reality (virtual reality, VR) equipment, augmented reality (augmented reality, AR) equipment, wireless terminals in industrial control, wireless terminals in self driving, wireless terminals in remote medical surgery, smart Wireless terminals in smart grid, wireless terminals in transportation safety, wireless terminals in smart city, wireless terminals in smart home, etc. Optionally, UE can be used to act as a base station. For example, a UE may act as a scheduling entity that provides sidelink signals between UEs in V2X or D2D, etc. For example, a cell phone and an automobile communicate with each other using sidelink signals. Communication between cellular phones and smart home devices without relaying communication signals through base stations.

The network device in this embodiment of the present application may be a device for communicating with a terminal device, and the network device may also be called an access network device or a wireless access network device, for example, the network device may be a base station. The network device in this embodiment of the present application may refer to a radio access network (radio access network, RAN) node (or device) that connects a terminal device to a wireless network. The base station can broadly cover various names in the following, or replace with the following names, such as: Node B (NodeB), evolved base station (evolved NodeB, eNB), next generation base station (next generation NodeB, gNB), relay station, Access point, transmission point (transmitting and receiving point, TRP), transmission point (transmitting point, TP), primary station MeNB, secondary station SeNB, multi-standard radio (MSR) node, home base station, network controller, access node , wireless node, access point (access point, AP), transmission node, transceiver node, base band unit (base band unit, BBU), remote radio unit (Remote Radio Unit, RRU), active antenna unit (active antenna unit) , AAU), radio head (remote radio head, RRH), central unit (central unit, CU), distributed unit (distributed unit, DU), positioning nodes, etc. A base station may be a macro base station, a micro base station, a relay node, a donor node, or the like, or a combination thereof. A base station may also refer to a communication module, modem or chip used to be set in the aforementioned equipment or device. The base station can also be a mobile switching center, a device that undertakes the function of a base station in D2D, vehicle-to-everything (V2X), machine-to-machine (M2M) communication, and a device in a 6G network. Network-side equipment, equipment that assumes base station functions in future communication systems, etc. Base stations can support networks of the same or different access technologies. The embodiment of the present application does not limit the specific technology and specific device form adopted by the network device.

Base stations can be fixed or mobile. For example, a helicopter or drone can be configured to act as a mobile base station, and one or more cells can move according to the location of the mobile base station. In other examples, a helicopter or drone may be configured to serve as a device in communication with another base station.

In some deployments, the network device in this embodiment of the present application may refer to a CU or a DU, or, the network device includes a CU and a DU. A gNB may also include an AAU.

Network equipment and terminal equipment can be deployed on land, including indoors or outdoors, hand-held or vehicle-mounted; they can also be deployed on water; they can also be deployed on aircraft, balloons and satellites in the air. In the embodiment of the present application, the scenarios where the network device and the terminal device are located are not limited.

It should be understood that all or part of the functions of the communication device in this application may also be realized by software functions running on hardware, or by virtualization functions instantiated on a platform (such as a cloud platform).

channel feedback

In a wireless communication system, a codebook-based scheme can be used to realize channel feature extraction and feedback. That is, after the receiver performs channel estimation, according to the result of channel estimation, according to a certain optimization criterion, select the precoding matrix that best matches the current channel from the pre-set precoding codebook, and pass the feedback link of the air interface to the Precoding matrix index (precoding matrix index, PMI) information is fed back to the transmitter for the transmitter to implement precoding. In some implementation manners, the receiver may also feed back the measured channel quality indication (CQI) to the transmitter for the transmitter to implement adaptive modulation and coding. Channel feedback may also be called channel state information (channel state information-reference signal, CSI) feedback.

Codebook-Based Channel Feedback Scheme Due to the sophisticated design of the codebook, multiple precoding matrices in the codebook can represent channel information in different scenarios. Therefore, the codebook-based channel feedback scheme can adapt to channel information feedback tasks in different scenarios.

The precoding matrix in the codebook is discrete. Therefore, mapping channel information to the precoding matrix in the codebook is a process of discretely quantizing continuous channel information, which makes the mapping process lossy in quantization, resulting in a decrease in the accuracy of the feedback channel information, thereby reducing the accuracy of the precoding matrix. Encoding performance.

Neural Networks

In recent years, artificial intelligence (AI) research has achieved great results in many fields such as computer vision and natural language processing, and it will also play an important role in people's production and life for a long time in the future . The field of communication has also begun to try to use AI technology to seek new technical ideas to solve technical problems that are limited by traditional methods.

Neural networks are commonly used architectures in AI. Common neural networks include convolutional neural network (CNN), recurrent neural network (RNN), deep neural network (DNN), etc.

The neural network applicable to the embodiment of the present application is introduced below with reference to FIG. 2 . The neural network shown in FIG. 2 can be divided into three types according to the position of different layers: input layer 210 , hidden layer 220 and output layer 230 . Generally speaking, the first layer is the input layer 210 , the last layer is the output layer 230 , and the middle layer between the first layer and the last layer is the hidden layer 220 .

The input layer 210 is used to input data. Taking a communication system as an example, the input data may be, for example, a received signal received by a receiver. The hidden layer 220 is used to process the input data, for example, to decompress the received signal. The output layer 230 is used for outputting processed output data, for example, outputting a decompressed signal.

As shown in Figure 2, the neural network includes multiple layers, each layer includes multiple neurons, and the neurons between layers can be fully connected or partially connected. For connected neurons, the output of neurons in the previous layer can be used as the input of neurons in the next layer.

With the continuous development of neural network research, neural network deep learning algorithms have been proposed in recent years. The neural network deep learning algorithm introduces more hidden layers in the neural network. Through layer-by-layer training of multi-hidden neural network for feature learning, the learning and processing capabilities of the neural network are greatly improved. This neural network model is widely used in pattern recognition, signal processing, optimization combination, anomaly detection and so on.

CNN is a deep neural network with a convolutional structure, and its structure can be shown in Figure 3. The neural network shown in FIG. 3 may include an input layer 310 , a convolutional layer 320 , a pooling layer 330 , a fully connected layer 340 , and an output layer 350 .

Each convolutional layer 320 can include many convolution kernels. The convolution kernel is also called an operator. Its function can be regarded as a filter for extracting specific information from the input signal. The convolution kernel can be a weight in essence. matrix, this weight matrix is usually predefined.

The weight values in these weight matrices need to be obtained through a lot of training in practical applications, and each weight matrix formed by the weight values obtained through training can extract information from the input signal, thereby helping CNN to make correct predictions.

When CNN has multiple convolutional layers, the initial convolutional layer often extracts more general features, which can also be called low-level features; as the depth of CNN deepens, the later convolution The features extracted by the layers are getting more and more complex.

Pooling layer 330, because it is often necessary to reduce the number of training parameters, it is often necessary to periodically introduce a pooling layer after the convolutional layer, for example, it can be a layer of convolutional layer followed by a layer of pooling layer as shown in Figure 3 , can also be a multi-layer convolutional layer followed by one or more pooling layers. In signal processing, the sole purpose of pooling layers is to reduce the spatial size of the extracted information.

The introduction of the convolutional layer 320 and the pooling layer 330 effectively controls the sharp increase of network parameters, limits the number of parameters and taps the characteristics of the local structure, improving the robustness of the algorithm.

The fully connected layer 340, after being processed by the convolutional layer 320 and the pooling layer 330, CNN is not enough to output the required output information. Because as mentioned above, the convolutional layer 320 and the pooling layer 330 only extract features and reduce the parameters brought by the input data. However, in order to generate the final output information (eg, the bitstream of the original information transmitted by the transmitter), the CNN also needs to utilize the fully connected layer 340 . Generally, the fully connected layer 340 may include a plurality of hidden layers, and the parameters contained in the multi-layer hidden layers may be pre-trained according to relevant training data of a specific task type, for example, the task type may include receiving For another example, the task type may also include performing channel estimation based on the pilot signal received by the receiver.

After the multi-layer hidden layers in the fully connected layer 340, that is, the last layer of the entire CNN is the output layer 350 for outputting results. Usually, the output layer 350 is provided with a loss function (for example, a loss function similar to classification cross entropy), which is used to calculate the prediction error, or to evaluate the result (also called predicted value) output by the CNN model and the ideal result (also called The degree of difference between the true value).

In order to minimize the loss function, the CNN model needs to be trained. In some implementations, the CNN model may be trained using a backpropagation algorithm (BP). The training process of BP consists of forward propagation process and back propagation process. In the process of forward propagation (the propagation from 310 to 350 in Fig. 3 is forward propagation), the input data is input into the above layers of the CNN model, processed layer by layer and transmitted to the output layer. If the result output at the output layer is quite different from the ideal result, the above loss function is minimized as the optimization goal, and transferred to backpropagation (as shown in Figure 3, the propagation from 350 to 310 is backpropagation), and the calculation is obtained layer by layer The partial derivative of the optimization target to the weight of each neuron constitutes the gradient of the optimization target to the weight vector, which is used as the basis for modifying the model weight. The training process of CNN is completed in the weight modification process. When the above error reaches the expected value, the training process of CNN ends.

It should be noted that the CNN shown in Figure 3 is only an example of a convolutional neural network. In a specific application, the convolutional neural network can also exist in the form of other network models, and this embodiment of the present application does not make any reference to this. limited.

autoencoder

Autoencoders are a class of artificial neural networks used in semi-supervised and unsupervised learning. An autoencoder is a neural network that takes an input signal as the training target. An autoencoder can include an encoder (encoder) and a decoder (decoder).

Taking the image compression shown in Figure 4 as an example, the self-encoder is described.

The input of the encoder may be an image to be compressed. In the embodiment shown in FIG. 4 , the image to be compressed occupies 28×28=784 bits. After the image to be compressed is compressed by the encoder, a code stream (code) is output. The number of bits occupied by the code stream output by the encoder is generally smaller than the number of bits occupied by the image to be compressed. For example, the number of bits occupied by the code stream output by the encoder shown in FIG. 4 may be less than 784 bits. From this, it can be seen that the encoder can achieve a compressed representation of the entity input to the encoder.

The input of the decoder can be code stream. The code stream may be a code stream output by an encoder. The output of the decoder is the decompressed image. It can be seen from Fig. 4 that the decompressed image is consistent with the image to be compressed input to the encoder. Therefore, the decoder can realize the reconstruction of the original entity.

In the process of training the self-encoder, the data to be compressed (such as the picture to be compressed in Figure 4) can be used as the input of the self-encoder (ie, the input of the encoder) and the label (ie, the output of the decoder), and the encoder and the decoder for end-to-end joint training.

Variational Autoencoder

Variational autoencoders introduce probability distributions in the encoder. The probability distribution can make the entity input to the encoder be mapped to the code stream conforming to the probability distribution. For the same input entity, different values collected in the probability distribution can be combined, so that the input of the decoder of the variational autoencoder can be multiple different code streams. The decoder can decode different code streams output by the encoder, so as to output multiple entities similar to the input entity.

The principle of the variational autoencoder will be described below in combination with FIG. 5 . FIG. 5 is an example diagram of an image generation process based on a variational autoencoder. The input of the encoder can be the original image, and the output of the encoder can be the mean and variance. The code stream can be calculated by the mean and variance of the encoder output and the collected values collected from the probability distribution. The code stream can be input into the decoder, and the decoder decodes the code stream and outputs an image similar to the original image. Wherein, the probability distribution may be a normal distribution.

Channel Feedback Based on Neural Networks

Channel feedback can be realized based on AI, such as neural network-based channel feedback. The network device side can restore the channel information fed back by the terminal device side as much as possible through the neural network. This neural network-based channel feedback can restore channel information, and also provides the possibility of reducing channel feedback overhead on the terminal device side.

As an example, a deep learning autoencoder can be used to implement channel feedback. The input of the AI-based channel feedback model can be channel information, that is, the channel information can be regarded as the compressed image input to the self-encoder. The AI-based channel feedback model can perform compressed feedback on channel information. At the sending end, the AI-based channel feedback model can reconstruct the compressed channel information, thereby retaining the channel information to a greater extent.

FIG. 6 is a schematic diagram of an AI-based channel feedback process. The channel feedback model shown in Fig. 6 includes an encoder and a decoder. The encoder and decoder are respectively deployed at the receiving end (receive, Rx) and the sending end (transmit, Tx). The receiving end can obtain the channel information matrix through channel estimation. The channel information matrix can be compressed and encoded by the neural network of the encoder to form a compressed bit stream (codeword). The compressed bit stream can be fed back to the receiving end through an air interface feedback link. The sending end can decode or restore the channel information according to the feedback bit stream through the decoder, so as to obtain complete feedback channel information.

The AI-based channel feedback model may have the structure shown in FIG. 6 . For example, the encoder may include several fully connected layers, and the decoder may include a residual network. It can be understood that FIG. 6 is only an example, and the present application does not limit the structure of the network model inside the encoder and decoder, and the structure of the network model can be flexibly designed.

The channel feedback based on the neural network can directly compress the channel information. Therefore, the accuracy of the channel information fed back by the neural network-based channel feedback scheme is relatively high.

In the case where the training data is matched with the test data, the channel feedback scheme based on neural network has better performance. The performance of neural network-based channel feedback schemes is poor when the training data does not match the test data. For example, neural network-based channel feedback schemes perform poorly when the scenarios of training data and test data are inconsistent. Therefore, channel feedback schemes based on neural networks have the problem of low generalization. Related technologies use multi-scenario training data to train a neural network model to improve generalization. In the actual operation process, it is more difficult to obtain a large-scale and high-diversity training data set.

In view of the above problems, the present application proposes a method for processing data. FIG. 7 is a schematic flowchart of a method for processing data provided by an embodiment of the present application. The method shown in FIG. 7 may include step S710.

Step S710, using the codebook to generate a plurality of training data. Multiple training data can be used to train the neural network model. The neural network model can be used for channel information feedback.

Training data may include channel information. The present application does not limit the representation form of the channel information. For example, the channel information may be a tensor composed of channel feature vectors of multiple subbands and multiple layers. As an embodiment, the channel information may be a tensor with a dimension of N×M×S×I. N may be the number of transmitting antennas, and the value of N may be greater than or equal to 2. M may be the number of subbands, and the value of M may be greater than or equal to 1. S may be the number of transmission layers, and the value of S may be greater than or equal to 1. I may represent a real part and an imaginary part, and a value of I may be I=2. Alternatively, I may represent amplitude and phase, and the value of I may be I=2. Alternatively, I may be real part, imaginary part, amplitude and phase, and the value of I may be I=4.

The codebook may be the codebook used in codebook-based channel feedback in the related art. A codebook may include multiple precoding matrices. The present application does not limit the form of the precoding matrix. For example, the precoding matrix may be in the form of a vector, that is, the codebook may include multiple precoding vectors.

The present application does not limit the type of the codebook. For example, the codebook may be implemented based on a discrete Fourier transform (discrete fourier transform, DFT) vector, that is, the precoding matrix may include a DFT vector. As an embodiment, the length of the DFT vector may be N, and the number of DFT vectors in the codebook may be N ₁ N ₂ . Wherein, N ₁ and N ₂ may be the number of antenna ports. Taking the two-dimensional planar array antenna at the transmitting end as an example, N ₁ may be the number of antenna ports in the first dimension, and N ₂ may be the number of antenna ports in the second dimension. N may be the total number of antenna ports. For example, the total number of antenna ports may be N=N ₁ N ₂ .

It can be seen from the above that the precoding matrix in the codebook can represent channel information in different scenarios, that is, the codebook naturally has generalization ability. It can be understood that the multiple pieces of training data generated by using the codebook may also include channel information of multiple scenarios. Therefore, using the training data generated by the codebook to train the neural network model can enhance the generalization of the neural network model. Moreover, the process of using codebooks to generate training data is much easier than obtaining diverse training data in actual scenarios. Therefore, using the technical solution of the present application, a neural network model with high generalization can be trained more simply, so as to realize high-precision and high-performance channel information feedback.

As an embodiment, the method shown in FIG. 7 may further include step S720.

Step S720, perform oversampling processing on multiple precoding matrices in the codebook.

Oversampling can be to insert more precoding matrices between adjacent precoding matrices. It can be understood that each precoding matrix may correspond to a spatial angle orientation, and there is a certain angular interval between adjacent precoding matrices. Through oversampling processing, the angle interval between adjacent precoding matrices can be made smaller, so as to achieve the purpose of fine quantization of angle.

As an embodiment, the oversampling factors of the first dimension and the second dimension of the two-dimensional array antenna may be O ₁ and O ₂ respectively. There may be N ₁ O ₁ N ₂ O ₂ oversampled precoding matrices. Taking the precoding matrix as a two-dimensional DFT vector as an example, the oversampled two-dimensional DFT vector a _m,n can be expressed as:

Wherein, m=0,1,2,...N ₁ O ₁ -1, n=0,1,2,...N ₂ O ₂ -1. x _m and u _n are oversampled one-dimensional DFT vectors respectively, and x _m and u _n can be expressed as:

x _m =[1,...,exp(j2π(N ₁ -1)m)/N ₁ O ₁ ] ^T ,

u _n =[1,...,exp(j2π(N ₂ −1)m)/N ₂ O ₂ ] ^T .

It can be understood that, from a physical point of view, a precoding matrix can represent a beam. The actual channel information can consist of information from one or more beams. Therefore, one or more precoding matrices can be selected in the codebook to form a training data to simulate real channel information. The present application does not limit the method of selecting the precoding matrix in the codebook to generate the training data.

As an embodiment, the precoding matrix may be selected in a grouping manner to generate training data. Multiple precoding matrices can belong to multiple groups. The plurality of groups may include a first group, and the plurality of training data may include the first training data. The first training data can be generated using the precoding matrix in the first packet. For example, all the precoding matrices contained in the first training data can be selected from the first group. It should be noted that, the present application does not limit the number of precoding matrices in the first group, for example, it may be one or more.

The grouping can be determined according to the beam situation corresponding to the precoding matrix. For example, real channel information often includes information about multiple similar beams. Therefore, precoding matrices corresponding to adjacent or similar beams may be grouped into one group. It can be seen that selecting training data from a group can make the training data closer to real channel information.

FIG. 8 is a schematic diagram of a grouping method provided by an embodiment of the present application. In FIG. 8, a grid point (circular point in FIG. 8) on the two-dimensional grid represents a precoding matrix. A grid block (rectangular box in FIG. 8 ) on a two-dimensional grid can represent a group. The first precoding matrix may be any grid point on the two-dimensional grid. For example, in FIG. 8 , grid point 811 may represent the first precoding matrix. The first group can be any rectangular box on the two-dimensional grid. For example, in FIG. 8, block 821 may represent the first grouping. By adjusting the first beam of the grid block, ie the first beam (the shaded circular point in Fig. 8), the starting position of the packet can be adjusted. By adjusting the size of the grid block, the number of precoding matrices in the group can be adjusted. It should be noted that the precoding matrices in one group may be continuous or spaced, which is not limited in this application.

As an example, grouping can be controlled by parameters L ₁ , L ₂ , s ₁ , s ₂ , p ₁ , p ₂ . L ₁ represents the size of the grid in the first dimension, L ₂ represents the size of the grid in the second dimension, and L ₁ ×L ₂ represents the size of the grid. s ₁ represents the distance between the first beams of adjacent grid blocks in the first dimension. s ₂ represents the distance between the first beams of adjacent grid blocks in the second dimension. p ₁ represents the distance between adjacent beams in a grid block in the first dimension. p ₂ represents the distance between adjacent beams in a grid block in the second dimension. In the grouping method shown in FIG. 8 , the values of each parameter are: L ₁ =4, L ₂ =2, s ₁ =s ₂ =2, p ₁ =p ₂ =1.

The training data can be selected from the precoding matrix in the packet. For example, MS precoding matrices can be selected in the first group, and further processed (for example, perform dimension conversion) to generate a tensor with a dimension of N×M×S×I, and use this tensor as the first training data. The present application does not limit the selection manner of the precoding matrix. For example, the first training data may randomly select a precoding matrix in the first group.

It can be understood that each group can generate at least one training data, and multiple groups can generate multiple training data. For example, the first group can generate the first training data and the second training data, and the second group can generate the third training data.

When enough training data is obtained, a data set composed of multiple training data can cover all precoding matrices in the codebook. It can be understood that all precoding matrices in the codebook can cover various scenarios. Therefore, the neural network model trained with this dataset can integrate the high generalization performance of the codebook.

Optionally, precoding matrices in multiple groups may overlap, for example, multiple groups may include a first group and a second group, multiple precoding matrices may include a first precoding matrix, and the first precoding matrix may be both Belonging to the first group also belongs to the second group. Continuing to take FIG. 8 as an example, the second grouping may be a grid block 822 . The first precoding matrix 811 may belong to both the first group 821 and the second group 822 . It can be seen from the example shown in FIG. 8 that the first group 821 and the second group 822 may overlap in the first dimension. It can be understood that the first group and the second group may overlap in the second dimension. Alternatively, the first grouping and the second grouping overlap in both the first dimension and the second dimension.

It can be understood that the overlap of precoding matrices of multiple groups can make the codebook be divided into more groups, thereby obtaining more training data.

The application does not limit the type of neural network model. As an example, the neural network model may include a variational autoencoder.

It can be understood that during the training process, the variational autoencoder can map the input training data to a probability distribution, so that multiple code streams conforming to the probability distribution can be obtained from one training data. Multiple codestreams provide more training data for the decoder, which can improve the generalization of the decoder and thus the generalization of the variational autoencoder. In the actual channel information feedback process (also called the testing phase), the probability distribution may not be introduced, and the variational autoencoder can become an autoencoder, so that the decoder can accurately restore the channel information input to the encoder.

FIG. 9 is a channel information feedback model based on a variational autoencoder provided in an embodiment of the present application.

Both the input data of the encoder 910 and the output data of the decoder 920 may be channel information. The channel information may be training data generated by a codebook, or actually obtained channel information. In the training phase, the input data of the encoder 910 or the output data of the decoder 920 may include training data generated based on a codebook, or may include training data collected in an actual scene. In the test phase, the input data of the encoder 910 or the output data of the decoder 920 may be channel information obtained after actual channel estimation.

As an embodiment, the channel information may be reshaped to change the dimension of the channel information. For example, in a training phase or a testing phase, training data or channel information obtained from channel estimation may be reshaped. For the input data of the encoder 910, reshaping can make the input data of the encoder 910 adapt to the model structure of the encoder 910, thereby simplifying the operation of the neural network model. For the output data of the decoder 920, reshaping can make the output data suitable for subsequent processing, thereby simplifying subsequent operations.

Take the four-dimensional tensor whose channel information is N×M×S×I as an example. This tensor can be reshaped into other dimensional formats as input data to encoder 910 or output data to decoder 920 . For example, the shaped channel information can be a three-dimensional tensor, and the dimension format of the three-dimensional tensor can be NM×S×I, NS×M×I, NI×M×S, N×MS×I, N×MI×S Or N×M×SI. Alternatively, the shaped channel information may be a two-dimensional matrix, and the dimension format of the two-dimensional matrix may be NMS×I, NMI×S, NSI×M or MSI×N. Alternatively, the shaped channel information may be a vector, and the length of the vector may be NMSI.

The output of the encoder 910 may include two vectors m=[m_1, . . . , m_p] and v=[v_1, . . . , v_p]. Among them, m can be related to the mean, and v can be related to the variance. m and v can be of length p. The output length p can be determined according to the quantization scheme and the size of the feedback resource. For example, under the condition that the quantization scheme is 3-bit uniform quantization and the feedback resource is 48 bits, the value of p is 48/3=16 bits.

The input to the decoder 920 may be a vector of length p. In the training phase and the testing phase, the input to the decoder 920 can be different. For example, in the training phase, the input of the decoder 920 may be a quantized and dequantized vector of the vector s×exp(v)+m. The vectors v and m are obtained from the output of the encoder 910, and the vector s can be collected from a standard normal distribution. The vector s can also be of length p. In the test phase, the input of the decoder 920 may be a quantized and dequantized vector m, and m may be obtained from the output of the encoder 910 .

This application does not limit the specific structure of the encoder or decoder in variational self-encoding. For example, it can be constructed by using one or more network structures of fully connected network, convolutional neural network, residual network, and self-attention mechanism network.

FIG. 10 is a schematic structural diagram of a variational autoencoder provided by an embodiment of the present application. The variational autoencoder shown in FIG. 10 includes an encoder 1010 and a decoder 1020 . The encoder 1010 may include a fully connected layer 1012 , a fully connected layer 1013 , a fully connected layer 1014 , a fully connected layer 1015 and a fully connected layer 1016 . The decoder 1020 may include a fully connected layer 1022 , a fully connected layer 1023 and a fully connected layer 1024 . The quantization scheme of the variational autoencoder shown in FIG. 10 may be 3-bit uniform quantization, and the feedback resource may be 48 bits.

The input to the encoder 1010 may be a vector of dimension 768 (eg NMSI=768). The fully connected layer 1012 can output a vector with a dimension of 1024. The fully connected layer 1013 can output a vector with a dimension of 256. The fully connected layer 1014 can output a vector with a dimension of 128. Both the input of the fully connected layer 1015 and the fully connected layer 1016 may be the output of the fully connected layer 1014 . The fully connected layer 1015 can output a vector m with a dimension of 16. The fully connected layer 1016 can output a vector v with dimension 16.

The input of decoder 1020 may be a vector with dimension 16. The fully connected layer 1022 can output a vector with a dimension of 2048. The fully connected layer 1023 can output a vector with a dimension of 1024. The fully connected layer 1024 can output a vector with a dimension of 768.

The multiple training data generated by using the codebook can be used to train the neural network-based channel information feedback model. This application does not limit the training method of the model. As an implementation, W can be used as the input of the encoder, and the output of the decoder can be W`. W may include training data generated using a codebook, or may include training data collected in actual scenarios.

This application does not limit the type of loss function. For example, the loss function can be:

loss＝||WW`|| _F +(exp(v)-(z+v)+m⊙m)

Where ||·|| _F is the Frobenius norm, z=[1,1,...,1] ^T ∈{1} ^N , ⊙ is the Hadamard product.

It should be noted that, in some embodiments, training data is also referred to as training samples or samples. This application is not limited to this.

The method embodiment of the present application is described in detail above with reference to FIG. 7 to FIG. 10 , and the device embodiment of the present application is described in detail below in conjunction with FIG. 11 . It should be understood that the descriptions of the method embodiments correspond to the descriptions of the device embodiments, therefore, for parts not described in detail, reference may be made to the foregoing method embodiments.

FIG. 11 is a schematic structural diagram of an apparatus 1100 for processing data provided by an embodiment of the present application. The apparatus 1100 may include a generating unit 1110 .

The first generating unit 1110 may be configured to generate a plurality of training data by using a codebook, the plurality of training data are used to train a neural network model, and the neural network model is used for channel information feedback.

Optionally, the neural network model includes a variational autoencoder.

Optionally, the codebook includes multiple precoding matrices.

Optionally, the apparatus 1100 may further include an oversampling unit 1120 . The oversampling unit 1120 may be configured to perform oversampling processing on the multiple precoding matrices.

Optionally, the multiple precoding matrices belong to multiple groups, the multiple groups include a first group, the multiple training data include first training data, and the first generating unit 1110 may include: a second A generating unit, configured to generate the first training data by using the precoding matrix in the first group.

Optionally, the multiple precoding matrices include a first precoding matrix, the multiple groups include a second group, and optionally, the precoding matrix includes a discrete Fourier transform (DFT) vector.

Optionally, the device further includes: a reshaping unit, configured to reshape the training data, so as to change the dimensions of the training data.

FIG. 12 is a schematic structural diagram of an apparatus for processing data according to an embodiment of the present application. The dashed line in Figure 12 indicates that the unit or module is optional. The apparatus 1200 may be used to implement the methods described in the foregoing method embodiments. Apparatus 1200 may be a chip, a terminal device or a network device.

Apparatus 1200 may include one or more processors 1210 . The processor 1210 can support the device 1200 to implement the methods described in the foregoing method embodiments. The processor 1210 may be a general purpose processor or a special purpose processor. For example, the processor may be a central processing unit (central processing unit, CPU). Alternatively, the processor can also be other general-purpose processors, digital signal processors (digital signal processors, DSPs), application specific integrated circuits (application specific integrated circuits, ASICs), off-the-shelf programmable gate arrays (field programmable gate arrays, FPGAs) Or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.

Apparatus 1200 may also include one or more memories 1220 . A program is stored in the memory 1220, and the program can be executed by the processor 1210, so that the processor 1210 executes the methods described in the foregoing method embodiments. The memory 1220 may be independent from the processor 1210 or may be integrated in the processor 1210 .

The apparatus 1200 may also include a transceiver 1230 . The processor 1210 can communicate with other devices or chips through the transceiver 1230 . For example, the processor 1210 may send and receive data with other devices or chips through the transceiver 1230 .

The embodiment of the present application also provides a computer-readable storage medium for storing programs. The computer-readable storage medium can be applied to the terminal or the network device provided in the embodiments of the present application, and the program enables the computer to execute the methods performed by the terminal or the network device in the various embodiments of the present application.

The embodiment of the present application also provides a computer program product. The computer program product includes programs. The computer program product can be applied to the terminal or the network device provided in the embodiments of the present application, and the program enables the computer to execute the methods performed by the terminal or the network device in the various embodiments of the present application.

The embodiment of the present application also provides a computer program. The computer program can be applied to the terminal or the network device provided in the embodiments of the present application, and the computer program enables the computer to execute the methods performed by the terminal or the network device in the various embodiments of the present application.

It should be understood that the terms "system" and "network" may be used interchangeably in this application. In addition, the terms used in the application are only used to explain the specific embodiments of the application, and are not intended to limit the application. The terms "first", "second", "third" and "fourth" in the specification and claims of the present application and the drawings are used to distinguish different objects, rather than to describe a specific order . Furthermore, the terms "include" and "have", as well as any variations thereof, are intended to cover a non-exclusive inclusion.

In the embodiments of the present application, the "indication" mentioned may be a direct indication, may also be an indirect indication, and may also mean that there is an association relationship. For example, A indicates B, which can mean that A directly indicates B, for example, B can be obtained through A; it can also indicate that A indirectly indicates B, for example, A indicates C, and B can be obtained through C; it can also indicate that there is an association between A and B relation.

In this embodiment of the application, "B corresponding to A" means that B is associated with A, and B can be determined according to A. However, it should also be understood that determining B according to A does not mean determining B only according to A, and B may also be determined according to A and/or other information.

In this embodiment of the application, the term "corresponding" may indicate that there is a direct or indirect correspondence between the two, or that there is an association between the two, or that it indicates and is instructed, configures and is configured, etc. relation.

In this embodiment of the application, "predefined" or "preconfigured" can be realized by pre-saving corresponding codes, tables or other methods that can be used to indicate relevant information in devices (for example, including terminal devices and network devices). The application does not limit its specific implementation. For example, pre-defined may refer to defined in the protocol.

In the embodiment of the present application, the "protocol" may refer to a standard protocol in the communication field, for example, may include the LTE protocol, the NR protocol, and related protocols applied to future communication systems, which is not limited in the present application.

The term "and/or" in the embodiment of the present application is only an association relationship describing associated objects, which means that there may be three relationships, for example, A and/or B, which can mean: A exists alone, and A and B exist at the same time , there are three cases of B alone. In addition, the character "/" in this article generally indicates that the contextual objects are an "or" relationship.

In various embodiments of the present application, the sequence numbers of the above-mentioned processes do not mean the order of execution, and the execution order of each process should be determined by its functions and internal logic, rather than the implementation process of the embodiments of the present application. constitute any limitation.

In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods may be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or May be integrated into another system, or some features may be ignored, or not implemented. In another point, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit.

In the above embodiments, all or part of them may be implemented by software, hardware, firmware or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application will be generated in whole or in part. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website, computer, server or data center Transmission to another website site, computer, server or data center by wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be read by a computer, or a data storage device such as a server or a data center integrated with one or more available media. The available medium may be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a digital versatile disc (digital video disc, DVD)) or a semiconductor medium (for example, a solid state disk (solid state disk, SSD) )wait.

The above is only a specific implementation of the application, but the scope of protection of the application is not limited thereto. Anyone familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the application. Should be covered within the protection scope of this application. Therefore, the protection scope of the present application should be determined by the protection scope of the claims.

Claims (22)

1. A method for processing data, comprising:

   A codebook is used to generate a plurality of training data, the plurality of training data are used to train a neural network model, and the neural network model is used to perform channel information feedback.
2. The method according to claim 1, wherein the neural network model comprises a variational autoencoder.
3. The method according to claim 1 or 2, wherein the codebook comprises a plurality of precoding matrices.
4. The method according to claim 3, characterized in that the method further comprises:

   Perform oversampling processing on the multiple precoding matrices.
5. The method according to claim 3 or 4, wherein the plurality of precoding matrices belong to a plurality of groups, the plurality of groups include a first group, and the plurality of training data includes first training data, so The use of codebooks to generate multiple training data includes:

   The first training data is generated by using the precoding matrix in the first group.
6. The method according to claim 5, wherein the plurality of precoding matrices include a first precoding matrix, the plurality of groups include a second group, and the first precoding matrix belongs to the first Groups also belong to said second group.
7. The method according to any one of claims 1-6, wherein the precoding matrix comprises a discrete Fourier transform (DFT) vector.
8. The method according to any one of claims 1 to 7, wherein the method further comprises:

   Reshaping the training data to change the dimensions of the training data.
9. A device for processing data, characterized in that it comprises:

   The first generation unit is configured to generate a plurality of training data using a codebook, the plurality of training data are used to train a neural network model, and the neural network model is used for channel information feedback.
10. The apparatus according to claim 9, wherein the neural network model comprises a variational autoencoder.
11. The device according to claim 9 or 10, wherein the codebook comprises a plurality of precoding matrices.
12. The device according to claim 11, further comprising:

    An oversampling unit, configured to perform oversampling processing on the plurality of precoding matrices.
13. The device according to claim 11 or 12, wherein the plurality of precoding matrices belong to a plurality of groups, the plurality of groups include a first group, and the plurality of training data includes first training data, so The first generating unit includes:

    A second generating unit, configured to generate the first training data by using the precoding matrix in the first group.
14. The device according to claim 13, wherein the multiple precoding matrices include a first precoding matrix, the multiple groups include a second grouping, and the first precoding matrix belongs to the first Groups also belong to said second group.
15. The device according to any one of claims 9-14, wherein the precoding matrix comprises a discrete Fourier transform (DFT) vector.
16. The device according to any one of claims 9-15, characterized in that the device further comprises:

    A reshaping unit, configured to reshape the training data to change the dimensions of the training data.
17. A device for processing data, characterized by comprising a memory and a processor, the memory is used to store a program, and the processor is used to call the program in the memory to execute any one of claims 1-8 the method described.
18. A device for processing data, characterized by comprising a processor for invoking a program from a memory to execute the method according to any one of claims 1-8.
19. A chip, characterized by comprising a processor, configured to call a program from a memory, so that a device installed with the chip executes the method according to any one of claims 1-8.
20. A computer-readable storage medium, characterized in that a program is stored thereon, and the program causes a computer to execute the method according to any one of claims 1-8.
21. A computer program product, characterized by comprising a program, the program causes a computer to execute the method according to any one of claims 1-8.
22. A computer program, characterized in that the computer program causes a computer to execute the method according to any one of claims 1-8.

Images